Symposium EQ11—Neuromorphic Computing and Biohybrid Systems—Materials and Devices for Brain-Inspired Computing, Adaptive Biointerfacing and Smart Sensing

Memory and memristive devices have shown strong ability to support brain-inspired computing primitives, thanks to their tunable volatile/nonvolatile characteristics, their good scaling in both 2D and 3D, their ability to run in-memory compute algorithms, and their unique physical properties that can mimic the individual building blocks in the brain, such as neurons, synapses and dendrites. Recently, the domain of resistive switching materials has enlarged from conventional memory devices, such as crosspoint structures, to alternative technologies, such as solid/liquid electrolytes, nanowires and 2D semiconductors. Exploring alternative materials/devices may disclose novel physical phenomena and different scales of voltage, current and time, thus providing new opportunities for neuromorphic engineering technologies.
This talk will present the recent progress on resistive switching devices for brain-inspired spike-based sensing and computing. First, volatile switching devices based on Ag migration will be presented, illustrating their physical switching mechanism and their application for spatiotemporal pattern sensing and recognition. Then I will illustrate novel volatile memtransistor devices based on 2D semiconductors such as MoS2, based on resistive switching or interface charge storage. Their potential applications in 3D high-density memory or spike-based computing will be described. Finally, I will summarize recent results on memristive nanowire structures, capable of reservoir computing in combination with a fully-connected network of nonvolatile resistive switching memories. The prospects for hybrid memristive-CMOS circuits for neuromorphic computing will be discussed in terms of scaling and energy efficiency.

We demonstrate experimentally a spintronic neural network that communicates with radio-frequency (RF) signals between neurons and synapses, and that natively processes RF signals as inputs. Radio-frequency signals are ubiquitous and are widely employed to convey information. These signals require quick and accurate analysis for diverse applications ranging from medical diagnosis, assessing and issuing warnings of disruptive solar events as well as identification of drones from their RF-ID tags [1,2]. To this end, artificial neural networks are an invaluable tool but suffer from the time and energy consuming digitisation of the analogue inputs. Due to their sensitivity to RF signals, magnetic tunnel junctions are ideal candidates for the implementation of hardware neural networks that intrinsically process radio-frequency waveforms. Spin-torque nano-oscillators implement neurons efficiently. They take DC signals as inputs and broadcast their output in RF form [3]. Meanwhile, chains of spin-diodes achieve synaptic multiply and accumulate functions directly on RF inputs, and produce a DC output [4]. In this work we demonstrate experimentally the principle of a multilayer neural network that performs non-linear operations on RF inputs. We will show that the hardware neural network, composed of an input layer with two neurons, a hidden layer with two neurons and a single neuron as output, can classify complex shapes in a 2D plane with high accuracy. This work opens the path to embedded spintronic neural networks that classify complex RF inputs with high accuracy, high speed and low energy.
[1] N. Chrysaphi et al. Astrophys. J. 893, 115 (2020).
[2] M. F. Al-Sa’d et al, Future Gener. Comput. Syst. 100, 86–97 (2019).
[3] J. Torrejon et al., Nature 547, 428–431 (2017).
[4] N. Leroux et al. Neuromorphic Comput. Eng. 1, 011001 (2021).

Materials that show multiple responses depending of the history of the applied stimuli, are able to emulate the plasticity and memory functionality of synapses in the brain, appearing as the ideal basic elements for neuromorphic or in-memory computing.
On the one hand, thin metal oxides that experience ionic migration or electrochemical reactions under the application of an electric field display resistive switching between multiple resistive states that allows storage of information in a non-digital manner at the material's level. On the other hand, dense networks of these devices are needed in order to achieve the necessary richness that will allow information processing and computing. Nanoscale metal oxides that self-assemble into complex networks have, thus, great potential as materials that can process information and learn. In this work, we use polymer templating and polymer imprinting to achieve the synthesis of interconnected networks of metal oxides. We will present a variety of different oxides prepared in this manner as well as the characterization of their properties, with specific focus on the characterization of the electrical properties of these networks, that are relevant for their application as self-learning materials. This approach could lead to the optimization of the connectivity of sub-micrometer memristive channels, as well as the number of accessible memory states.

Artificial Intelligence (AI) technology based on artificial neural networks is making a huge business and socioeconomic impact, changing the way of our daily life. With an expectation of achieving a brain-like efficiency and large performance gain, there has been increasing interest to implement artificial neural networks using novel non-volatile memory devices. Novel cross-point array architectures based on resistive memory devices, which can be operated in a massively parallel manner, have shown potential to achieve a large acceleration in AI computation. For realization of such architectures in hardware, the specifications for synaptic devices are revealed to be different in many aspects with those for traditional memory devices, necessitating new material systems and device designs [1, 2] as well as novel algorithms and architectures for efficient AI computation. In this talk, I will overview the recent progress and effort to achieve ideal synaptic device characteristics for novel neuromorphic architectures, as exemplified in our recent experimental results on capacitor-based approach [1, 3, 4], and 3-terminal ionic switching devices [5, 6]. In addition, I will discuss novel algorithmic and architecture-level remedies to overcome the device non-idealities for neural network training [7].

[1] S. Kim et al., 2017 IEEE 60th MWSCAS, 2017. [2] T. Gokmen et al., Frontiers in Neuroscience 10, 2016. [3] Y. Li et al., Symposia on VLSI Technology and Circuits, 2018. [4] Y. Kohda et al., 2020 IEEE International Electron Devices Meeting, 2020. [5] J. Tang et al., 2018 IEEE International Electron Devices Meeting, 2018. [6] S. Kim et al., 2019 IEEE International Electron Devices Meeting, 2019. [7] C. Lee et al., Frontiers in Neuroscience (Accepted)

Continued miniaturization in digital electronics over the past several decades has led to modern ubiquitous technologies such as the Internet of Things, edge computing, artificial intelligence (AI), and machine learning (ML) that are impacting nearly all aspects of society. However, current AI/ML algorithms incur undesirable energy costs on conventional von Neumann hardware platforms, motivating exploration of more efficient neuromorphic architectures exemplified by the human brain. Compared to previous two-terminal neuromorphic architectures such as memristors and phase change memory, three-terminal memtransistors have emerged as a tunable biorealistic neuromorphic system due to co-location of a field-effect transistor and non-volatile memory within one device. By integrating atomically thin two-dimensional materials that enable enhanced electrostatic tunability, monolayer polycrystalline MoS₂ memtransistors achieve gate-tunable memristive switching, linearity, and reconfigurability [1]. Similarly, four-terminal dual-gated MoS₂ memtransistors minimize crosstalk and sneak currents in scalable crossbar architectures, simplifying integration challenges that have hindered memristive architectures based on bulk materials [2]. Despite the unique attributes of memtransistors, their implementation in neuromorphic architectures has been limited to conventional artificial neural networks [3] suggesting that their full potential for AI/ML has not yet been realized.

This presentation will introduce MoS₂ memtransistors with a wide range of learning behaviors achieved through a combination of enhanced electrostatic control and tailored gate bias pulsing profiles. Using monolayer MoS₂ grown on sapphire, long-term potentiation and depression behaviors are singularly modulated by the gate electrode polarity, an observation that parallels the synaptic weight update and neuroplasticity in biological systems [4]. By enhancing the relative importance of the vertical field effect from the gate voltage compared to the lateral field from the drain voltage, these devices show a greater reconfigurability of synaptic behavior compared to previously reported MoS₂ memtransistors grown on SiO₂. Different gate bias pulsing strategies further diversify the library of learning curves. The resulting gate-tunable learning behavior is then modeled in a simplified spike-timing-dependent plasticity (STDP) scheme to perform unsupervised continuous learning in a simulated spiking neural network (SNN). We show that continuous learning, a previously underexplored cognitive concept in hardware neuromorphic computing, circumvents traditional trade-offs between image recognition accuracy and resource allocation. Overall, this work demonstrates that reconfigurable MoS₂ memtransistors provide unique hardware accelerator opportunities for energy-efficient artificial intelligence.
References
[1] V. K. Sangwan, H.-S. Lee, H. Bergeron, I. Balla, M. E. Beck, K.-S. Chen, and M. C. Hersam, Nature, 554, 500-504 (2018)
[2] H.-S. Lee, V. K. Sangwan, H. Bergeron, H. Y. Jeong, K. Su, and M. C. Hersam, Adv. Funct. Mater., 30, 2003683 (2020).
[3] X. Feng; S. Li; S. L. Wong.; S. Tong; L. Chen; P. Zhang; L. Wang; X. Fong; D. Chi; K.-W. Ang, ACS Nano, 15, 1764-1774 (2021).
[4] J. Yuan, S. E. Liu, A. Shylendra, W. A. Gaviria Rojas, S. Guo, H. Bergeron, S. Li, H.-S. Lee, S. Nasrin, V. K. Sangwan, A. R. Trivedi, and M. C. Hersam, Nano Letters, 21, 6432-6440 (2021).

A hardware implementation inspired by brain functionalities - Neuromorphic Computing, allows non-deterministic polynomial-time (NP) hard combinatorial optimization problems to be efficiently solved. A representative implementation of a Boltzmann Machine (BM), which is a recurrent artificial neural network with stochastic features analogous to the thermodynamics of real-world physical systems, requires the dynamically tunable decision-making process in the stochastic neurons. However, a common challenge to emulate neurons with previous stochastic devices lacks the means to precisely control and modulate the probability distribution associated with intrinsic randomness. Here, we report a three-terminal tin oxide (SnO_x)/molybdenum disulfide (MoS₂) heterogeneous memristive device (hetero-memristor) to effectively realize tunable stochastic dynamics through gate modulation in its output sampling characteristics. The device can sample exponential-class sigmoidal distributions with quantitatively defined tunable “temperature” effect, analogous to the Fermi-Dirac distribution of physical systems. A BM composed of these tunable stochastic neuron devices, together with simulated annealing of designed “cooling” strategies is conducted to solve the MAX-SAT, a representative in NP-hard combinatorial optimization problems. A quantitative discussion of Russell-Rao (RR) similarity profiles is also provided to understand the temperature dependence associated with neuron selectivity and specificity in the BM optimization process. These stochastic neurons based on tin oxide/MoS₂ hetero-memristors with reconfigurable statistical behavior pave the way for implementing selected “cooling” strategies in BM to reach optimal convergence efficiency and can find broad applications in energy-efficient computing for learning, clustering, and classification.

Flexible Electrocorticography (ECoG) electrode arrays that conform to and record surface field potentials from multiple brain regions provide unique insights into how computations occurring in distributed brain regions mediate behavior. Typically, current flexible ECoG devices require highly specialized microfabrication methods precluding the ability to easily fabricate customizable and inexpensive flexible ECoG devices for basic and clinical neuroscience applications. Here, we bring together multiple innovations to create fully desktop fabricated flexible graphene electrodes that enable chronic in vivo neural recordings in mice. First, we synthesized a highly stable, conductive ink via liquid exfoliation of graphene in Cyrene solvent. Next, we have established a fabrication process for stencil-printing this ink via laser-cut stencils on flexible polyimide substrate to realize multi-channel graphene ECoG arrays. Bench-top tests indicate that the graphene electrode has good conductivity of ~ 1.1 × 10³ S●cm^-1, flexibility to maintain its electrical connection under static bending, and material stability in a 15-day accelerated corrosion test. Further, chronically implanted graphene ECoG devices remained fully functional for long durations (> 180 days), with a low average in vivo impedances of 24.72 ± 95.23 kΩ measured at 1 kHz. The ECoG device can measure spontaneous surface field potentials from mice under both awake and anesthetized states as well as localize evoked responses triggered by sensory stimuli. The stencil-printing fabrication process can be used to create graphene ECoG devices with customized electrode layouts within 24 hours using commonly available laboratory equipment.

As the world becomes more interconnected and data-driven, effective deployment of data-intensive computation methods become more critical, driving the need for neuromorphic networks to overcome the memory wall bottleneck in von Neumann architectures. Previously proposed multi-weight synaptic devices often have drawbacks like nonlinear and asymmetric update response and low energy efficiency, which is detrimental to online supervised learning. Furthermore, most proposed neuromorphic devices, along with previously proposed graphene synapses, are stiff, rigid, and generally incompatible with biological tissue. In this work, we propose a graphene-based artificial synapse featuring highly linear and symmetric update response, superior energy efficiency, mechanical flexibility, and bio-compatibility as a platform for neuromorphics in biological interfaces.
We fabricated mesoscale (few mm²) and microscale (few µm²) synaptic transistor devices by interfacing graphene with a Nafion membrane. The conductance of the device is controlled by applying a current pulse through the gate (Nafion) and the synaptic weight is read by applying voltage across the channel (graphene). Multiple different graphene structures are characterized and tested for long term potentiation, evaluating optimal write pulse duration and amplitude, write-read delay, switching energy, endurance, linearity, number of states, and temperature dependence. The devices were found to have high linearity and symmetry, important for implementation of backpropagation. In addition, they were found to have extremely low power updates, opening the possibility for online learning. CrossSim was used to evaluate neural network performance on the Fashion-MNIST clothing article dataset using a multilayer perceptron. Due to the unique synaptic update response, training performance was found to exceed that of an ideal linear software synapse. These results propose an artificial synapse that is bio-compatible, flexible, energy efficient, and tailored for learning applications, paving the way for a platform for neuromorphic interfacing with biological systems.

Tuning electronic conductance through solid state electrochemical ion insertion has emerged as a promising technology to enable next-generation ultra low energy computing architectures. Unlike two-terminal non-volatile memory elements, the three-terminal redox transistor, also known as electrochemical random access memory (ECRAM), decouples the ‘write’ and ‘read’ operations using a ‘gate’ electrode to tune the conductance state through charge transfer reactions involving ion injection into the channel electrode through a solid-state electrolyte. The insertion of ions into the bulk of the channel acts to dope the material through a gradual composition modulation that leads up to thousands of finely spaced conductance levels (synaptic weights) with near-ideal analog behavior. These properties enable low-energy operation without compromising analog performance. However, the strong coupling of ionic and electronic processes sharply challenges our current understanding of solid-state electrochemical systems, particularly at decreasing dimensions and timescales relevant to computing technology. In my presentation, I will summarize the rich portfolio of challenging, exciting fundamental science questions about ion tunable electronic materials systems and how we can harness these to realize a new paradigm for low power neuromorphic computing.

Despite the excellent performance of Von Neumann architecture-based computing systems, their large power consumption and the memory bottleneck make them unsuitable of efficiently handling modern intelligent tasks such as visual and speech recognition and real time analysis of large data sets. On the other hand, the human brain excels at such complicated tasks with excellent energy efficiency, since the storage and processing of information occurs at the same locations in massively parallel and fault-tolerant neural networks. With this in mind, neuromorphic computing inspired by the brain have attracted a great deal of attention from researchers as a new computing paradigm. In particular, spiking neural networks (SNNs), with their event driven nature and inherently good energy efficiency and bandwidth, offer promising alternatives to complement the deep neural networks (DNNs). However, the challenge to implement temporal dynamics with CMOS devices has hindered the development of SNNs.
In this work, we present a dynamic, electrochemical synapses suitable for event-based computing with tunable temporal dynamics and excellent energy efficiency. We choose the molybdenum disulfide (MoS₂) as the channel material of our devices because its two-dimensional (2D) nature makes it a good candidate for both ionic intercalation and electric-double-layer (EDL) gating. We adopt a novel dual-gate design to implement both short-term and long-term synaptic plasticity that have been difficult to achieve in CMOS devices. When a programming pulse (+1 V, 900 µs) is sent to the inert Au gate, we observe a 64 % (or 0.85 µS) increase in synaptic weight (represented by the channel conductance) due to induced charge carriers in MoS₂ via ionic gating. This response mimics the short-term plasticity in biological synapses since the induced carriers dissipates once the programming gate voltage is removed. On the other hand, when we apply a similar programming pulse (+1 V, 900 µs) through the electrochemically reactive gate (LiFePO₃ or LFP), we can induce long-term change (31%) in the synaptic weight via electrochemical intercalation since charge transfer doping via intercalation is non-volatile.
We can achieve both potentiation and depression by changing the polarity of the pulses and also modulate the change in synaptic response by varying the pulse amplitude and/or width. These short-term and long-term temporal dynamics are instrumental for implementing bio-realistic spike-based computations while the dual-gate structure allows us to elucidate the doping mechanisms from ionic gating and intercalation processes. We demonstrate basic neuronal functions such as paired pulse facilitation (PPF), spike-rating dependent plasticity (SRDP), and temporal filtering. Simulation results suggest that our synapses can lead to high accuracy in pattern recognition applications.
In summary, we develop a dual-gate MoS₂ synapse with tunable short-term and long-term temporal dynamics. The 2D synapse exhibits low energy consumption (<700 aJ/switching), good programming speed (>10 MHz), high precision level (>5000 states), as well as high linearity and symmetry. Capable of implementing basic neuronal functions, this dynamic synapse can lead to the hardware implementation of SNNs for event-based computations.

State-of-the-art Flash is currently the dominating non-volatile memory technology. Its operation principle is based on storing charge on a floating-gate. Advanced scaling of the device geometry required to decrease the floating-gate charge down to some tens of electrons. Thus, to ensure that information is stored for years, the leakage current of the gate-dielectric needs to be in the order of 10e-26 A. This ultra-low leakage current requirement is opposed to a high tunneling rate required for device programming. Thus, a tradeoff between state retention, programming speed, and endurance is made, which limits the device performance. Here we demonstrate a novel concept to overcome the limitations of Flash memories. Our concept is based on a memristive switch integrated within the gate-terminal of a transistor. In a sense, the operation principle is similar to floating-gate transistors, but the advantage is that ions instead of electrons are used for programming. Since the mass of ions is larger than of electrons, gate-dielectrics with higher leakage current levels can be used. Our results demonstrate the practicability of the concept in an experimental proof-of-concept study and are further supported by LT-Spice simulations. This work offers a guideline for development of scalable non-volatile memories providing high endurance and retention times, and fast programming speeds.

Deep neural networks are well suited for edge computing tasks, such as real time sensor processing and data fusion. However, they are computationally expensive and hard to implement in the SWaP (size, weight, and power) constrained environments encountered by most edge applications. Analog in-memory computing can potentially solve this challenge and process deep neural inference with millisecond latencies and milliwatt power draw, due to the two or more orders of magnitude energy efficiency improvement over state-of-the-art digital systems. Efficiency improvements are a result of holding neural network weights stationary in arrays of nonvolatile memory, and analog input data from sensors is formatted and applied directly to these weights. Each memory’s physical state (often conductance) represents a range of numerical values with precision limited by the memory device physics and peripheral circuitry. Unlike traditional CMOS processors, the accuracy of the neural network is directly dependent on the collective precision of these physical weights, creating new memory device requirements which differ significantly from traditional binary and multi-level cell memory schemes. To assess the neural network accuracy of candidate devices, we have created an electrical characterization and modeling framework for analog in-memory computing. Determining deep neural net inference accuracy requires devices to be programmed to specific resistances with up to 128 levels, but significantly more overlap in spread is allowed than with a standard multi-level memory scheme. Accuracy is also strongly affected by how the precision varies across the 128 conductance levels, which directly interacts with the algorithm’s distribution of weight values. This analog characterization dataset can be used to statistically model the accuracy of an in-memory analog accelerator. We have characterized and modeled the potential accuracy of a commercial 40nm semiconductor-oxide-nitride-oxide-semiconductor (SONOS) technology, Sandia’s prototypical tantalum oxide resistive memory (ReRAM), and the emerging battery-inspired electrochemical memory (ECRAM). In the near term, embedded SONOS is an attractive option capable of high accuracy even on full deep neural networks, such as ImageNet. In the longer term, TaOx ReRAM promises a lower switching voltage, excellent scalability, ns switching, and high radiation resilience. However, the stochastic behavior associated with filamentary switching has created difficulty achieving high accuracy on highly complex deep network datasets such as ImageNet. Novel three terminal electrochemical memory (ECRAM) devices combine low voltage, low energy programming, and excellent analog tunability due to their bulk switching mechanism. Furthermore, fab-compatible metal oxide-based ECRAM devices use similar materials to ReRAM. In addition to precise initial weight values, edge computing typically requires the stable retention of weight values over time and across temperature ranges. We have explored the impact of weight drift on neural network accuracy, which has allowed us to understand how often refresh cycles might be needed. Finally, we have explored the effect of ionizing radiation which can be encountered in edge computing applications in space. The eventual degradation of neural network accuracy has been quantified as a function of total ionizing dose, and the effect of physical weight mapping on this degradation will be discussed. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

With growing interest in the feasibility of conductive bridge random access memory (CBRAM) as a synaptic application, the ability to control the size of the nanosized conductive filament (CF) and the resistance remains a challenge. Precise control of the nanofilament is critical for the switching reliability and the multiple resistance states’ formation, which are required characteristics for being adapted as neuromorphic computing applications. To improve the switching reliability of CBRAMs for its further application, various methods were suggested for the past years, including extreme scaling down, limiting the cation injection, and integration into the 1T1R (1-transistor-1-resistor) structure.
This study suggested the Cu-cone inserted CBRAM device (Pt/TiN/TiO₂/Cu-cone/TiN/Pt, Cu-cone device) to improve the resistance controllability originated from the well-controlled single CF, which could serve as a feasible artificial synaptic device. The problem of random and stochastic dynamics of weak bundles of Cu CFs in Cu-based CBRAM was largely resolved by inserting the Cu cone into the memory cell. In the authors’ previous work, it was reported that including a Cu conical structure (≈300 nm in base-diameter and ≈400 nm in height) into the CBRAM as a cation source can largely enhance the reliability issue. The strong electric field concentrated on the inserted Cu cone induced a single CF near the tip of the Cu cone, making the device feature distinctive from the conventional planar structured Cu CBRAM (Pt/ TiN/TiO₂/Cu/TiN/Pt, planar Cu device).
The Cu-cone device exhibited improved control of the resistance states attributed to the well-controlled CF, which improved switching reliability and the stable formation of multiple resistance states. The analog resistive switching and enhanced controllability of the multiple resistance states of the Cu-cone inserted CBRAM were demonstrated in direct current – voltage sweep, closed-loop pulse switching (CLPS) test, and synaptic device characterization protocols. By exploiting the enhanced controllability of the resistance, highly improved endurance could be achieved, and the potentiation/depression performance could be emulated in the Cu-cone device. This is ascribed to the well-controlled formation/dissolution of Cu CF only near the tip region of the Cu-cone, where the strength of the Cu CF can be well manipulated as intended. Single filament formation could be induced attributed to the concentrated field on the conical structure, and the abrupt and stochastic resistive switching could be prevented. As a result, the ability to precisely control the size of the nanometer-sized CF was hugely improved, which is providing the high potential for the applicability of the Cu-cone device. The strong and single CF in the Cu-cone device offered a well-controlled formation/ dissolution depending on the potentiation and depression pulses, making them appropriate as an artificial synaptic device. The planar Cu device was also evaluated for comparison but failed to show these performances because of the stochastic and abrupt formation and rupture of bundles of weak Cu CFs.
The gradually degrading linearity during the early stage of the potentiation/depression still needs to be enhanced, and other synaptic functions, such as the spike-timing-dependent-plasticity, also need to be proved for the Cu-cone device to be utilized as a synaptic device. Even though further research is still required to realize fully-functioning artificial synapses, the feasibility of the Cu-cone device based on the improved resistance controllability was suggested in this work.

Oxide-based non-volatile memory devices such as Resistive Random Access Memory (RRAM) have broad applications, particularly in the field of analog neuromorphic computing_.¹ Analog computing requires a linear change in resistance as a function of applied pulses. Resistance change involves the motion of oxygen ions across a conductive filament in filamentary oxide-based memory. However, if the oxygen ion motion is too abrupt with an applied pulse, the resistance change becomes non-linear, degrading the neuromorphic performance.² A diffusion barrier layer with higher activation energy for oxygen ion motion (E_A,O) can be used at the interface (where the filament is expected to reform during the set) to improve the linearity. A previous study showed that HfO_x RRAM with an AlO_x barrier exhibited analog set behavior with an on/off ratio of ~3.³ However, SiO_x has a higher diffusion barrier for oxygen ion motion as compared to AlO_x.
In this study, the effect of two different barrier layers, AlO_x (E_A,O= 1.3 eV) and SiO_x (E_A,O= 2.7 eV), on the RRAM switching characteristics and analog temporal responses (e.g., long-term-potentiation and depression, spike-time-dependent-plasticity) of HfO_x based filamentary memory are explored. A thin ~0.8 nm barrier layer was deposited at the HfO_x/BE interface via atomic layer deposition (ALD), followed by a ~4.2 nm HfO_x active layer. A detailed X-ray photoelectron spectroscopy (XPS) depth profile analysis was performed to characterize the composition of the stack. Electrical characterization shows that devices with an AlO_x barrier layer have a comparable forming voltage (~3.1 V) and set-reset behavior than the HfO_x device without a barrier layer. However, the devices with a SiO_x barrier layer show a higher forming voltage (~4.5 V) and an on/off ratio of ~5. The SiO_x barrier layer devices demonstrate a more gradual set, improved linearity in the potentiation, and significantly less spike-to-spike variation during the analog pulsing compared to HfO_x/AlO_x and HfO_x devices. A COMSOL Multiphysics^® model is used to develop a fundamental understanding of the observed results. Overall, the experimental results demonstrate that reducing the oxygen ion motion using a SiO_x diffusion barrier layer at the bottom electrode interface improves the linearity in the analog potentiation of HfO_x-based memory for neuromorphic applications.
1. J. J. Yang, et al., Nature nanotechnology 8 (1), 13-24 (2013).
2. K. Moon, et al., Faraday discussions 213, 421-451 (2019).
3. J. Woo, et al., IEEE Electron Device Lett. 37 (8), 994-997 (2016).
Acknowledgments: This work was supported by the AFOSR MURI under Award No. FA9550-18-1-0024. In part at the Georgia Tech Institute for Electronics and Nanotechnology, a member of the NNCI, which is supported by the NSF (ECCS-2025462). This material is based upon the work supported by the NSF GRFP under Grant No. DGE-1650044.

Bioinspired neuromorphic system is an emerging technology that could revolutionize the present computer architecture. Chalcogenide materials show unique material system with properties for memory and switching devices, thus having a potential to be a crucial element for a reliable neuromorphic computing. To mimic synaptic functions with achieving the highest integration density, one-selector one-resistor (1S-1R) crossbar arrays are becoming next generation memory technology. 1S1R crossbar arrays using non-volatile memory devices and volatile selector devices are capable of desired behavior such as efficient spike-event due to its advantages of low power consumption, large endurance, and great switching characteristics. However, the complexity of the lithographic process for multi-functional layers and the poor stability due to the integration of heterogeneous thin film materials in the 1S1R structure remain challenging to develop technology further.
At present, the chalcogenide materials are attracting research interest for their extensive applications ranging from phase-change materials in nonvolatile memories to volatile threshold switching materials in selector devices or optical, thermoelectric devices. Among these materials, some chalcogenide materials show intriguing properties with the conduction characteristic, which has led to the applications for nonvolatile memory and selector devices. While some of them allow rapid and reversible switching between the amorphous and crystalline state for phase change memory (PCM) application, others exhibit an uncommon conductive behavior under high electric field, called the Ovonic threshold switching (OTS) effect.
In this study, we demonstrate that memory switching and threshold switching can be modulated in a single Ge_xTe_1-x material system by controlling the atomic ratio of Ge-Te system. In the memory switching mode both low resistive state (LRS) and high resistive state (HRS) can be maintained after removing the external voltage, while the LRS in the threshold switching mode will return to the HRS once the applied voltage becomes smaller than a critical value. With Te deficient Ge-Te system close to the atomic ratio of 1:1, the device based on this material shows the characteristics of memory switching with high switching speed (<10 ns) and superior endurance (>10⁸ cycle). When the Ge-Te material system is Te rich with the atomic ratio close to 1:6, it exhibits threshold switching characteristics as an excellent bi-directional selector with great selectivity (>10⁵ ON/OFF ratio) and rapid switching speed (<10 ns), suitable for high-density memory application. Monolithic fabrication of 1S1R structure is demonstrated using our GeTe memory and GeTe₆ selector with good selectivity (>10² ON/OFF ratio) and enough voltage margin. This work can significantly reduce the complexity in the fabrication of stackable devices and improve the understanding of the switching mechanisms in chalcogenide materials, thus opening up wider applications from memory to selector devices for neuromorphic computing.
This research was supported by Korea Semiconductor Research Consortium (KSRC) through a project for developing source technologies for future semiconductor devices and also supported by Samsung Electronics Co. Ltd. (Project number IO2102021-08356-01)

Resistive switching (RS) devices have attracted increasing attention for their application for non-volatile memories and neuromorphic computing systems. In my talk, I will focus on the design of novel memristive heterostructures using a perovskite-related oxide with high oxygen mobility and p-type semiconducting electronic transport, namely La₂NiO_4+δ (L2NO4). Dense L2NO4 thin films were grown using pulsed-injection metal-organic chemical vapor deposition (PI-MOCVD) under optimized conditions. While our first studies were dedicated to epitaxial films on single-crystal substrates, we then moved towards industrial-relevant substrates and thus deposited polycrystalline L2NO4 films on Si-based CMOS-compatible wafers. The epitaxial films allowed to construct planar memristive devices, while the polycrystalline ones enabled for both planar and vertical configurations. The electrical response of these devices could be tuned both by changing the electrode materials and the amount of point defects (oxygen interstitials) of the sandwiched film. The chemical, structural, microstructural of the films will be presented, showing the high crystal quality, together with the change in Ni oxidation state and the concomitant change in resistivity after annealing. Furthermore, by several electrical characterization measurements, the key role played by the oxygen interstitials on the initial resistance state, the initialization step, and the memristive characteristics of the devices will be presented and discussed.^[1,2] An analog bipolar counter-eightwise RS behavior with multiple resistance states has been proven both for the planar and vertical devices, using continuous sweeps and voltage pulses. Selected devices were further analyzed by X-ray absorption spectroscopy (XAS) using synchrotron radiation. Changes in the Ni K-edge were measured for different annealings, in different regions of the device and under operando conditions, allowing for a better understanding of the RS mechanisms taking place in the device. Furthermore, gradual changes in conductance after the application of repetitive DC sweeps, which can be regarded as the evolution of the synaptic weight between neurons, were obtained for these novel L2NO4-based devices. These promising results open the door to the use of L2NO4-based memristors as artificial synapses for neuromorphic computing.
[1] K. Maas, E. Villepreux, D. Cooper, C. Jiménez, H. Roussel, L. Rapenne, X. Mescot, Q. Rafhay, M. Boudard, M. Burriel, J. Mater. Chem. C 2020, 8, 464.
[2] K. Maas, E. Villepreux, D. Cooper, E. Salas Colera, J. Rubio Zuazo, G. R. Castro, O. Renault, C. Jimenez, H. Roussel, X. Mescot, Q. Rafhay, M. Boudard, M. Burriel, Adv. Funct. Mater. 2020, 30, 1909942.

Resistive switching devices are integral for realising physical computing systems. A large number of switching mechanisms have been observed and depend strongly on the material system. They can loosely be distinguished by the switching location as either occurring through the material bulk between two electrodes, or interface-type where switching takes place in a region localised to the area underneath the electrodes. This first kind of is widely studied and has been identified in a wide range of materials. The scalability of memristors is important for the development of large networks. For example, in valence change memristors multilevel analogue switching is observed in larger area memristors, but when downscaled this behaviour changes, and most devices show an abrupt transition.
We have demonstrated that interface-based memristors on Schottky contacts on semiconductor complex-oxide platforms have interesting properties for computing purposes, such as multilevel continuous switching, energy landscapes that can be tuned by electric fields [1], and responses that can be used in learning algorithms [2]. The scalability of memristors is important for the development of large networks, but this area is so far largely unexplored for interface memristors. By performing transmission electron microscopy measurements on biased and unbiased samples we show that interfacial defects, particularly oxygen vacancies, play a key role in governing resistive switching. This raises the important question of whether memristive action is sustained when devices are downscaled and the number of defects is reduced. We address this by investigating Nb-doped SrTiO₃ electronic devices of areas spanning several orders of magnitude. Unexpectedly, we find that smaller devices exhibit larger resistance windows, which is desirable for network implementation. We explain these surprising results by considering the kinetics of trapping and de-trapping in dielectric materials from first principles to relate the decay constants from retention measurements to the trapping density. Larger values are found for smaller junctions, suggesting an increase in the density of traps is responsible for enhancing the resistance ratio.
Typically, models of the Schottky barrier and interfacial electric field consider material-specific parameters and assume spatial homogeneity. In finite devices, however, this does not describe the system completely. It is expected that field lines will accumulate near the device edges, resulting in local enhancement of the field. This is especially important to consider for Nb-doped SrTiO₃ as the dielectric constant, and consequently, the Schottky barrier profile strongly depends on electric fields. We extend existing models of the electric field to include the device edge. Our work indicates an increased contribution of the device edge compared to the inner area when downscaled. This is an important finding for network scaling as it allows for a controlled way of increasing the memory window while limiting the electric field.

[1] A. S. Goossens, A. Das, and T. Banerjee. Journal of Applied Physics 124, 15 (2018)
[2] T. F. Tiotto, A. S. Goossens, T. Banerjee, J. P. Borst, N. A. Taatgen. Frontiers in Neuroscience 19, 1456 (2021)

Recent developments in resistive switching technologies have been focused on the fabrication and characterization of new memristor configurations in an attempt to combat sneak path problems and enhance the already attractive low consumption properties of memristors, while ensuring high number of states and configurability. We have developed a three terminal, double-stack memristor comprising of two fully working memristive half-cells. This device was fabricated using conventional photolithography techniques for patterning, followed by evaporation of electrodes or sputtering of metal oxide active material. The final device has a Metal-Insulator-Metal-Insulator-Metal material stack.

This configuration provides access to both half-cells, through the shared electrode in the middle of the stack, enabling fine tuning capabilities of the total device ,which can also be read as a single unit cell, by using the top and bottom electrodes to assess its state. Preliminary characterization results indicate an increase in tunability and memory capacity. The ability to control either one of the half-cells independently, as well as switch them simultaneously, improves controllability of the states screened when reading the entire cell. Moreover, simultaneous reset of both half-cells, due to application of voltage pulsing between the top and bottom electrodes , can lead to greatly reduced memory-erase power consumption. The additional memristor also functions as a tunable serial resistance for overvoltage protection, when the device is probed or switched as a single cell. Finally, this configuration enables in-depth testing of complementary and memristor-fuse theories by allowing separate forming of each half cell and the possibility to test multiple configurations of device materials, device types and device polarities.

The dramatic rise of data-intensive workloads has revived special-purpose hardware and architectures for continuing improvements in computational speed and energy efficiency. Several promising special-purpose approaches take inspiration from the brain, which outperforms digital computing in power and performance in key tasks such as pattern matching. One of these computing approaches inspired by the brain’s dataflow architecture is called “in-memory computing” which stores and computes stable computational kernels within specialized circuit geometries. While traditional CMOS ASICs implementing in-memory computing deliver some performance gains, such approaches still suffer from low power efficiency. Therefore, new proposals leveraging non-volatile memristive devices for in-memory computation are highly attractive in a variety of application domains. While originally developed for as digital (binary) high density non-volatile memories, metal oxide memristive devices have demonstrated a wide range of behaviors and properties – such as a wide range of tunable analog resistance and non-linear dynamics – which motivate their use in novel functions and new computational models. Many recent in-memory compute studies have focused on crossbar circuit architectures, demonstrating their application for neural networks, scientific computing and signal processing. However, other circuit primitives – such as content addressable memories (CAMs) and combined systems such as crossbar arrays and non-linear elements– have shown further promise for mapping a diverse range of complimentary computational models such as finite state machines, pattern matching, hashing algorithms and Hopfield networks for tackling optimization problems. In this talk, I will review the varied opportunities by co-designing from algorithms and architectures to circuits and devices for enabling low power, high-throughput computation in a variety of application domains.

Stochastic neural networks have become the state-of-the-art approach for solving problems in machine learning, information theory, and statistics. The critical operation in such networks is a stochastic dot-product. While there have been many demonstrations of dot-product circuits and, separately, stochastic neurons, the efficient hardware implementation combining both functionalities is still missing. In my talk, I will discuss our recent work on addressing this need. I will start with the discussion of compact, fast, energy-efficient, and scalable stochastic dot-product circuits based on passively integrated metal-oxide memristors. The high performance of such circuits is due to mixed-signal implementation enabling energy-saving in-memory computing. The stochastic functionality is achieved by operating the circuit in a low signal-to-noise ratio regime by utilizing the circuit’s noise, intrinsic and/or extrinsic to the memory cell array. I will then discuss several application demonstrations based on passively integrated TiO2-x memristors, including Hopfield networks for solving optimization problems and a Boltzmann machine. In the proposed circuits the neuron outputs can be selectively scaled which allows adjusting coupling between the neurons and/or controlling the signal-to-noise ratio at runtime, without the need to rewrite the memory weights. This feature enables, for example, efficient implementation of different annealing approaches that improve the quality of the solution for the solved optimization problems. I will conclude my talk with a comparison of the proposed approach to other solutions based on the conventional and emerging technologies, and a discussion of the future important work.

The paradigm of neuromorphic computing in hardware is becoming increasingly urgent due to the ubiquity of portable electronics and the Internet of Things (IoT) that acquire and process vast amounts of data. Neuromorphic computing also promises to dramatically reduce power consumption compared to conventional digital computing by providing non-volatile memory and threshold switches that can store and process data locally to overcome the von Neumann bottleneck. While first-order linear memory devices have been developed that can store weight matrices in a dot product machine, higher-order memristive switches are desired for in-situ data processing. In this context, two-dimensional (2D) materials are being explored for memristive applications due to tunable electrostatic coupling, novel neuromorphic functions, and ultrathin geometries that facilitate scaling [1-3]. Solution-processed 2D materials provide additional advantages for large-area, flexible, and printed neuromorphic circuits. However, the memristive switching mechanisms in solution-processed 2D materials are poorly understood. In particular, understanding charge transport in composite films of solution-processed 2D materials is complicated by heterogeneous morphologies, the interplay of multiple physical processes, and the inaccessibility of the active area in two-terminal vertical geometries.

This presentation will discuss thermally activated memristive switching mechanisms in percolating networks of diverse solution-processed 2D semiconductors including MoS₂, ReS₂, WS₂, and InSe [5]. This threshold switching behavior is distinct from the widely reported non-volatile resistive behavior arising from electrochemically active metal filaments or related conductive filaments bridging the device channel. The mechanism underlying thermally activated memristive switching is elucidated by engineering large-area lateral memristors that allow the direct observation of filament formation using spatially resolved optical and chemical analyses and in situ thermal analysis. These devices show high switching ratios (up to 10³) at low global electric fields (≈4 kV cm⁻¹) that is explained by a thermally assisted electrical discharge that preferentially occurs at the sharp edges of 2D nanosheets. X-ray photoelectron spectroscopy and infrared imaging show that the filaments consist of oxygen-deficient regions that get heated (>115 °C) during switching. Using the insight gained from the thickness dependence of switching in lateral devices, vertical devices have been designed and demonstrated with operating voltages as low as 2 V. A novel electroforming scheme is employed to tune the operating voltage in voltage-controlled threshold switches and current-controlled negative differential resistance devices. Overall, this work establishes percolating networks of solution-processed 2D semiconductor nanoflakes as a platform for high-order non-dynamical devices for neuromorphic circuits and systems.
References

[1] V. K. Sangwan and M. C. Hersam, Nature Nanotechnology, 15, 517 (2020).
[2] V. K. Sangwan, H.-S. Lee, H. Bergeron, I. Balla, M. E. Beck, K.-S. Chen, and M. C. Hersam, Nature, 554, 500-504 (2018).
[3] H.-S. Lee, V. K. Sangwan, H. Bergeron, H. Y. Jeong, K. Su, and M. C. Hersam, Advanced Functional Materials 30, 2003683 (2020).
[4] M. E. Beck and M. C. Hersam, ACS Nano, 14, 6498 (2020).
[5] V. K. Sangwan, S. V. Rangnekar, J. Kang, J. Shen, H.-S. Lee, D. Lam, J. Shen, X. Liu, A. C. M. de Moraes, L. Kuo, J. Gu, H. Wang, and M. C. Hersam, Advanced Functional Materials DOI:10.1002/adfm.202107385 (2021).

Metal oxide-based resistive random-access memory (RRAM) offers several excellent performances such as lower power consumption, fast switching speed, simple structure, and compatibility to the complementary metal oxide semiconductor (CMOS) technologies which can potentially replace the traditional memory. The functionality of a metal oxide-based RRAM stems from the conductive filament (CF) of high concentration of oxygen vacancy, which initially forms, and later grows or dissolves inside the oxide switching layer during the resistive switching process. To date, various metal oxide-based RRAMs have been reported in literatures. However, the multi-physics processes and their complicated interplays during the formation, growth and rupture of the CFs are not fully understood, which severely limits the application of RRAMs in semiconductor industry. In this study, we develop a comprehensive physical model based on defect chemistry and charge transport theory to investigate the electroforming process and subsequent resistive switching behavior, using HfO_2-x as a prototypical model system. Our simulation results indicate that the CF formation is assisted by the supply of oxygen vacancies at the anode/oxide interface and the oxygen vacancy transport in the bulk. The effects of the electrode properties and the intrinsic metal oxide properties on the CF growth behavior during the electroforming process are further explored. Next, the roles of the electrical bias, heat transport and Vegard strain effect, as well as the electrical and thermal properties of the switching layers in the resistive switching processes are systematically investigated. Finally, high-throughput simulations and machine learning method are performed to derive interpretable analytical models for device performance metrics in terms of key material parameters. These analytical models reveal that optimal resistive performance can be achieved in materials with a low Lorenz number. This work provides a fundamental understanding of the CF formation and growth behaviors in the electroforming and resistive switching processes, and demonstrates a computational data-driven methodology of materials selection for improved RRAM performance.

Synapses play a key role in learning and perception and avail as a model for efficient brain-inspired computing. Devices that emulate bio-synaptic plasticity present an auspicious route to energy-efficient and agile neuromorphic computing. Synaptic plasticity refers to the ability of synapses to alter interneural strength in an event-based manner, according to the history and timing characteristics of preceding trigger impulses[1]. Synaptic plasticity entails activity-dependent transitions from short-term potentiation (STP) to long-term potentiation (LTP) (memorization effect) and vice versa (forgetting effect). The trigger firing conditions (sequence, timing, intensity) under which these effects occur govern memorization and adaptive learning[2], thus constitute the foundation of information processing and in-memory computing in neuromorphic electronics. The most promising candidates for computational nodes are the two-terminal resistive switching (RS) memristive devices due to their intrinsic property to alter their resistance states in a step wise manner when voltage trigger applied across their terminal, enabling analog RS with both volatile and non-volatile properties[3]. The dynamic nature of their transitional properties is shown to have commonalities with bio-synaptic plasticity, and this is a rich field far from fully explored[4].
The present work demonstrates an alternative approach to the realization of electronic synapses for in-memory computing. The fabrication was realized through low-cost and self-sufficient processes. A solely additive approach, the emerging inkjet printing technique [5–8] was used as a universal fabrication method. The developed fully printed devices were based on Ag/a-TiO₂/Ag structures and developed by implementing an a-TiO₂ custom-made ink, functionalised for inkjet printing compatibility, while also retaining its printing and electrical properties for an optimal period (>five months). The ink produced crack-free nanolayers with adjustable functional thickness in the range of 80 to 350 nm.
The printed nano-electronic synapses were characterised electrically through voltage pulses in order to investigate their dynamic properties in time. Due to the amorphous phase of the active layer, the RS was able to be induced under low voltage pulses, thus reducing the overall power requirement. The devices demonstrated transitions from STP to LTP that occurred in a step wise and event-based manner, emulating bio-synaptic plasticity. This memorisation effect along with the switching rate and reset process were all effectively controlled by changing solely the trigger’s duty cycle (α = pulse duration / period (%)). Notably, there is an indication of homeostatic regulation in RS. Similar to biology, the synaptic strength (here conductance) did not increase perpetually but was bounded within a range thus maintaining an equilibrium.
Consequently, these minimally structured, low-cost and eco-friendly neuromorphic nodes constitute a fruitful avenue toward the development of flexible neural networks. The optimized ink formulation and the thorough investigation of the electrical framework under which these devices respond synaptically represent a path to unconventional computing with applications in biomimetic sensing and prosthetics.
[1] W.A.Catterall, A.P. Few, Neuron 2008, 59, 882
[2] J.Fiser, P.Berkes, G.Orbán, M.Lengyel,Trends Cogn. Sci. 2010, 4,119
[3] C.Wang, W.He, Y.Tong, R.Zhao, Sci. Rep. 2016, 6, 1
[4] S.Choi, J.Yang, G.Wang, Adv. Mater. 2020,32,1
[5] W.Lee, T.Someya, Chem. Mater. 2019, acs. chemmater
[6] W.Banerjee, S.H.Kim, S.Lee, D.Lee, H.Hwang, 2021
[7] T.Someya, Z. Bao, G. G. Malliaras, Nature 2016, 540, 379
[8] B.Salonikidou, Y.Takeda, B.Le Borgne, J.England, T.Shizuo, R.A.Sporea, ACS Appl. Electron. Mater. 2019, 1

Filamentary adaptive oxide neuromorphic devices are emerging as a candidate for on-chip, bio-inspired circuits. These devices consist of a metal-insulator-metal structure with typical insulators being dioxides, such as, HfO₂, ZrO₂, SiO₂, etc. When electrically biased, oxygen vacancies are created within the oxide forming a nanoscale conductive bridge, known as the filament. Once formed, the resistance of the filament can be increased or decreased by applying voltages of either positive or negative polarity. For “in-memory” computing based filamentary adaptive oxides, ideally the device’s resistance should linearly increase with the number of applied voltage pulses and decrease linearly with a voltage of opposite polarity. Many oxides have shown promise in achieving these resistance changes, however, comparing their performance in neuromorphic applications is not straightforward. Many studies do not report the specifics of the forming step or any pre-stabilization/ conditioning of the filament before the pulsing experiment. This work demonstrates how the reset condition during conditioning leading up to analog pulsing influences the resistance response with identical pulsing schemes.

Neuromorphic devices consisting of a titanium oxygen reservoir, HfO_x active layer, and gold electrodes are fabricated with photolithography. After forming all the devices with a current limit of 0.1 mA, devices are initially reset with either -1 V, -1.3 V, or -1.5 V. Following the initial reset, 10 consecutive set/reset cycles are conducted. All devices are set to the low resistance state with a voltage sweep from 0 to 1.2 V and a current limit of 0.5 mA, however, the reset sweep is varied between the three reset conditions. Devices from each of the reset conditions are subject to -0.7 V, -0.85 V, and -1 V pulses with 1 μs width. Even though the starting resistance, the device structure, and the analog voltages pulses are identical for all three conditions, the rate of change of the resistance and the overall resistance change is lower for the devices that are pre-conditioned with a lower magnitude voltage. For example, after 30 pulses with -0.7 V amplitude and 1 μs width, the devices pre-conditioned at -1 V have an 83% change in resistance compared to devices pre-conditioned at -1.5 V having a 193% change. One possibility for the varied analog response is that the devices reset with a higher magnitude voltage have a larger filament, suggesting that the geometry of the filament is not solely determined from the forming step. Temperature-calibrated, scanning thermal microscopy (SThM) is used to measure the temperature rise at the top electrode during the pre-conditioning voltage sweeps and during analog pulsing. This temperature rise is then correlated to the filament temperature using a COMSOL Multiphysics® model. It is believed that devices with a larger filament would reach a higher overall temperature and thus exhibit the largest resistance change during pulsing. This work suggests that the current-voltage history of the filament affects the resistance vs. pulse number trend, increasing the difficulty to predict the pulsing scheme needed to program a device’s resistance during neuromorphic circuit training.
This work was supported by the AFOSR MURI under Award No. FA9550-18-1-0024. In part at the Georgia Tech Institute for Electronics and Nanotechnology, a member of the NNCI, which is supported by the NSF (ECCS-2025462). This material is based upon the work supported by the NSF GRFP under Grant No. DGE-1650044.

Memristor, which simply consists of a switching layer inserted between two electrodes, is one of the most strong candidates to become a device platform for imitating the principal characteristics of the biological neural network due to its nonlinear and dynamic electrical characteristics depending on the history of applied electrical programming [1-3]. In this study, we fabricated a silicon oxide (SiO_x) nanorod memristor using E-beam evaporator with glancing angle deposition at the wafer-scale and utilized the device as an artificial neuron for probabilistic computing applications [4]. Notably, the SiO_x nanorod structure provides the random distribution of multiple nanopores all across the active area, capable of forming a multitude of Si filaments at many SiO_x nanorod edges after the electromigration process. This can facilitate probabilistic switching that can mimic integrate-and-fire signaling and the stochastic dynamics of biological neurons with a very high dynamic range (~5.15 × 10¹⁰) and low energy (~4.06 pJ). Different probabilistic activation (ProbAct) functions in a sigmoid form are implemented, showing its controllability with low variation by manufacturing and electrical programming schemes. Then, as a proof of concept, based on the suggested memristive neuron, we demonstrated the self-resting neural operation with local circuit configuration and revealed the probabilistic Bayesian inferences for genetic regulatory networks with low errors (< ~2.41 × 10^-2) with its robustness to the ProbAct variation. Taken all together, our study shows the availability of the designed SiO_x nanorod memristive neuron for probabilistic neural networks that can be efficiently used for a variety of uncertainty quantification problems.

References [1] S. Choi, J.-W. Choi, J. C. Kim, H. Y. Jeong, J. Shin, S. Jang, S. Ham, N.-D. Kim, and G. Wang, Nano Energy, 84, 105947 (2021)
[2] S. Choi, J. Yang, G. Wang, Adv. Mater., 32, 2004659 (2020)
[3] S. Choi, S. Jang, J.-H. Moon, J. C. Kim, H. Y. Jeong, P. Jang, K.-J. Lee, and G. Wang. NPG Asia Mater. 10, 1097–1106 (2018)
[4] S. Choi G. S. Kim, J. Yang, H. Cho, C. –Y. Kang and G. Wang, Adv. Mater., Accepted (2021)

Artificial synaptic devices are up-and-coming for futuristic semiconductor computing devices. However, up-to-date optoelectronic approaches remain challenging, due to their unstable charge trap states and limited integration. In this presentation, we demonstrate wide-bandgap (WBG) III−V materials for photoelectronic neural networks system. The experimental analysis shows that the enhanced crystallinity of WBG synapses develop better synaptic characteristics, such as effective multilevel states, a wider dynamic range, and linearity, allowing the better power consumption, training, and recognition accuracy of artificial neural networks. Additionally, light-frequency-dependent memory characteristics suggest that artificial optoelectronic synapses with advanced crystallinity support the transition from short-term potentiation to long-term potentiation, implying a clear emulation of the psychological multistorage model. This is attributed to the charge trapping in deep-level states and suppresses fast decay and nonradiative recombination in shallow traps. We believe that the fingerprints of these WBG synaptic characteristics provide an effective strategy for establishing an artificial optoelectronic synaptic architecture for innovative neuromorphic computing.

Abstract:
Neuromorphic computing systems have been required recently due to the nature of the von Neumann structure, which causes high power consumption and more time to require the process of big data and deep learning tasks [1].
In new hardware architectures, two terminal-based neuromorphic devices with synaptic crossbar arrays have been studied widely but the regarding leakage current, reduction of Snake path has still been challenged [2]. On the other hand, synapse transistors have been attracting attentions in the substitute technology to overcome the leakage current and control the weight of the synapse independently [1].
Here, we conducted to create synapse plasticity with transistors with potentiation and depression, which were consisted of IGZO transistors with Ga2O3 charge trap layer using a solution process and an ALD process. The linearity of synaptic devices plays an important role in evaluating synaptic plasticity.
The magnitude of charge trapping in transistors highly depend on inducing pulse numbers and frequencies of applied voltages. However, detail explorations of the assessment in this factor have not been revealed widely. In order to support the guidance of the assessment, we studied this aspect of trap control and guide how to approach optimal condition of own devices as function of pulse time, input voltage, and initialization.
As a result, we obtained high linearity of synaptic plasticity by several steps inducing different voltages and pulse numbers.
It was also confirmed that the changes in weight were significantly changed according to the difference in roughness of the Ga2O3 surface and the defect-free tunneling layer through the ALD process.
Acknowledgements
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2020R1A2C1013693), Korea Institute for Advancement of Technology (KIAT) grant funded by By the Ministry of Trade, Industry & Energy (MOTIE, Korea) (P0012451, The Competency Development Program for Industry Specialist) and the Technology Innovation Program - (20016102, Development of 1.2kV Gallium oxide power semiconductor devices technology) funded by MOTIE.
References
[1] IGZO/Al2O3 based depressed synaptic transistor”, Yu Liu1, Xiawa Wang1, Wenjie Chen, Linyuan Zhao, Wei Zhang, Weijun Cheng, Ziting Zhuo, Jing Wang, Tian-Ling Ren, Jun Xu, Superlattices and Microstructures 128 (2019)
[2] Neuromorphic computing using non-volatile memory”, Geoffrey W. Burr, Robert M. Shelby, Abu Sebastian, Sangbum Kim, Seyoung Kim, Severin Sidler, Kumar Virwani, Masatoshi Ishii, Pritish Narayanan, Alessandro Fumarola, Lucas L. Sanches, Irem Boybat, Manuel Le Gallo,Kibong Moon, Jiyoo Woo, Hyunsang Hwang & Yusuf Leblebici, Taylor & Francis 2 (2017)

Gallium (Ga)-based Liquid metals (LMs) have attracted much attention for its application in bioelectronics due to their unique properties of non-toxicity, high electrical/thermal conductivity, deformability, and unique surface chemistry. Herein, we present an artificially intelligent synaptic organic field effect transistor (OFET) based on magneto-responsive LM (m-LM) as the top gate electrode. Presence of magnetic field induces morphological change of m-LM, which modifies the conductance of the transistor device in accordance with the field magnitude. By integrating the ferroelectric property of Poly(vinylidene fluoride-co-trifluoro ethylene) P(VDF-TrFE) with morphology and locomotion control of m-LM, an artificially intelligent synaptic device is shown. An array of 4 x 4 synaptic devices is developed to exhibit 3D pattern encryption learning and recognition, enabled by spatial specific magnetic field. The device further demonstrates possible application in the field of augmented reality by utilizing the touchless characteristic of magnetic field.

Recently, brain-inspired artificial neuromorphic computing system has widely been researched to replace conventional von Neumann-based computation because of their advantages such as low power consumption and efficient space usage. The artificial neuromorphic computing system is typically divided into transistor and mersister type according to the type of unit element, and optical and electrical type according to the type of input signal. Among them, optical synaptic transistor type is favorable option with high bandwidth, ultrafast operation speed, and low crosstalk. Oxide semiconductor is powerful candidate for active materials of optical synaptic transistor with low off-current, capability of large-are deposition, and high photosensitivity. Additionally, oxygen vacancies in oxide semiconductor materials are ionized with light absorption resulting in slow recovery of photo induced currents and persist photoconductance (PPC). The optical synaptic behavior can be originated from these oxygen vacancy and PPC. However, oxide semiconductor materials have limitation of detection wavelength range because of their wide bandgap (> 3eV) to be driven with visible light region. Furthermore, there is tradeoff between synaptic behavior and switching characteristic when oxygen vacancies in channel layer increase.
In this paper, we suggested Indium-Gallium-Zinc-Oxide (IGZO) optical synaptic transistor with solution processed IGZO layer with quasi 2-dimensional thickness between sputtered IGZO layer and gate insulator and absorption layer above channel layer. Inserted defective interface caused interfacial trapping & de-trapping resulting in increase of PPC and accumulated current. Furthermore, additional solution processed absorption layer (SAL) could increase photo generated carrier. IGZO optical synaptic transistor with defective interface shows the peak of photo induced postsynaptic current (PSC) of 1171 pA and maximum gain of 24.67 under red the light illumination of 635 nm wavelength.
We compared parameter (on/off current ratio and S.S) derived from transfer characteristics of IGZO TFT with each condition (concentration of IGZO precursor: 0.2, 0.1, and 0.05 / spin coating rotation speed: 6000 and 3000 rpm) for defective interface and pristine IGZO TFT and optimized concentration of precursor solution to 0.1 M. The gain was improved from 2.53 to 9.61 in the red light region with the addition of defective interface compared to pristine IGZO TFT and increased from 9.61 to 24.67 under red light illumination (635 nm) with additional absorption layer. Short term plasticity (STP)-to-long term plasticity (LTP) transition is important factor which showed learning and decision-making functions for a biological neural system, which can be well simulated by varying the number of pulses, pulse frequency, pulse width, and even light power density. A series of light pulses with different frequencies are applied with wavelength of 635 nm. We showed that the peak PSC increased from 58 to 1171 pA with the increase of the pulse frequency from 0.02 Hz to 1 Hz. We compared the gain (A25/A1) of the IGZO optical synaptic transistor with defective interface with various the pulse frequency. As the pulse frequency increases from 0.2 to 1 Hz, the gain increased from 2.62 to 27.46, which means the STP-to-LTP transition can be caused by increase of frequency. Additionally, we suggest Pavlov’s associative learning mimicked by using photonic spike and electric spikes and their computation of IGZO optical synaptic transistor with defective interface.

Along with the advent of memory industry and technology, different forms of memory array have been studied and commercialized. However, those memory architectures suffered critical trade-off, cost and efficiency. 3D X-point memory array gathered attention for future memory architecture because of its cost efficiency and high density. Since chalcogenide based phase change random access memory (PCRAM) was commercialized for various advantages, coupling chalcogenide based OTS selector is the most suitable candidate along with simple process to enable 3D stacking array. Conventional binary OTS selectors suffered from thermal instability, which disturb keeping amorphous state and cause low on current. Therefore, different materials are being doped to overcome those limitations, which make process much complex. Here, we introduce Ge_xS_1-x films for future OTS selector application. Ge_xS_1-x films have higher bandgap compare to other binary selectors, which enable large on current. Along with bandgap, Ge_xS_1-x films are thermally stable because of size of covalent radius. Various compositions were extracted from Auger Electron Spectroscopy (AES), then tested electrical characteristics using Ru (Top electrode) - GexS1-x (Selector) - TiN (Bottom electrode) shaped device. Also, optical parameters were calculated from results came from reflectance and transmittance. Bandgap, defect, etcs were extracted using F(R) formula.

A first principle multi-scale model is used to study the charge dynamics of a neuromorphic device based on PEDOT:PSS in an NaCl electrolyte system. We compare the results with and without the application of an orthogonal electric field to study the device operation properties. The first step of the model is the classical molecular dynamics (MD) simulation of the device whose atomistic morphology is then partitioned into hopping sites, for which energies are calculated by a QM/MM method.
The coordinates of the sites are used to construct a list of molecule pairs, for which reorganization energies, electronic coupling elements and eventually the Marcus hopping rate are calculated.
A directed graph is defined from the neighbor list and hopping rates where kinetic Monte Carlo (KMC) and graph theory are used to analyze the charge dynamics.
Our results reveal a different energy landscape when an external electric field is applied orthogonal to the device. Also, both the distribution of coupling elements and site energy correlation functions are altered by the electric field. Accordingly, KMC results show that when the electric field drives the Na+ ions out of the device channel, charge dynamics are faster and have a large flow order parameter.
Random walk theory is used for studying the resistance/conductance for this neuromorphic charge network, and percolation is simulated numerically to reveal its connectivity. Electric field-sensitive regions are identified by the distribution of site voltages and Ricci curvature in the graph, which is related to the fast/slow dynamics from the KMC simulations. Our study reveals the linkage between the elementary process on atomistic/electronic level and the device performance.

Redox-based memristive devices are one of the most attractive candidates as a synapse for neuromorphic computing system due to its non-volatility and high operating speed. It is necessary to analyze the behavior of oxygen ions to better understand the resistive switching mechanism. Previously, we discovered the resistive switching behavior through topotactic phase transition of SrFeO_x through Au/SrFeO_x/SrRuO₃ device. Here, we studied Au/SrFeO_x/Nb: SrTiO₃ device grown epitaxially in (111) direction to understand the migration of oxygen ions during resistive switching compared to Au/SrFeO_x/SrRuO₃ device. This Au/SrFeO_x/Nb: SrTiO₃ device has a Schottky-like I-V curve, and shows a difference in resistance of about 100 to 1000 times. In order to precisely study the mechanism of resistance change, in-situ I-V transmission electron microscopy (TEM) was used. In-situ TEM enables to observe changes in atomic structure under different resistance states by applying external field to the specimen in a TEM. Moreover, oxygen ionic migration as well as oxidation states of transition metals can be measured simultaneously by electron energy loss spectroscopy.
In the pristine state, the SrFeO_x layer mainly exists as a perovskite structure. However, after the forming process, a phase change occurred to the brownmillerite structure near the interface with bottom electrode Nb: SrTiO₃. The structure during SET and RESET cycles did not change significantly, and only the area of perovskite in HRS was partially reduced. Unlike the previous work, the main factor of the resistance change comes from the Nb: SrTiO₃ interface close to SrFeO_x layer not from SrFeO_x layer itself. When a positive bias is applied to the top electrode, oxygen vacancies are accumulated near the Nb: SrTiO₃ interface under the brownmillerite SrFeO_2.5 and the resistance value is lowered. Conversely, as a negative bias is applied to the top electrode, oxygen ions piles up at the interface and the device’s resistance increases. In this system, the SrFeO_x layer acts as an oxygen reservoir and transport path, and it can alter the oxygen ions concentration on the Nb: SrTiO₃ surface easily. This study provides a better understanding and applicability of perovskite-based memristive devices.

The ion-gel gated synaptic transistor has been widely researched to achieve low voltage operation, mimicking synaptic functions for neuro-inspired electronics. Although organic materials have been widely studied for ion-gel gated synaptic transistors, the reliability and retention time of long-term potentiation is still short for neuromorphic computing. To overcome these issues, exploration for new materials in an ion-gel gated synaptic transistor is required. Graphene has been widely studied with various advantages such as zero bandgap property, mechanical flexibility, and biocompatibility.
Here, we fabricated ion-gel gated monolayer graphene synaptic transistors (IG-MGST). Due to ion trapping between graphene monolayer/substrate and ion adsorption on a graphene substrate, various synaptic plasticity long-term plasticity has been realized. Because of the ion-gel gated device, and zero bandgap characteristic of IG-MGST, the device can operate with low voltage consumption with the ambipolar operation. To tune the synaptic plasticity of IG-MGSTs, the ionic size effect of ion-gel has been investigated. Also, Dirac point and resulted operation voltage have been controlled by introduction of various kinds of self-assembled layers. By temperature-dependent plasticity measurement and Raman spectroscopy, the origin of synaptic plasticity has been explored which has not been discovered before. Finally, we simulated the recognition of handwritten digits in both LiTFSI-based ion-gel and [EMIM][TFSI]-based ion-gel, thus improved linearity has been obtained in IG-MGST using LiTFSI-based ion-gel. Our study provides not only a graphene-based synaptic transistor but also provides fundamental information for 2D material-based synaptic transistors for future neuromorphic electronics.

Emulation of synaptic functions is an emerging method for enhancing the performance and efficiency of computational systems. Artificial synaptic devices should exhibit a gradual change in conductivity to emulate synaptic functions. For artificial synaptic devices, ions are suitable materials to emulate the analog characteristics of synapses because they can move gradually. However, to develop efficient neuromorphic devices, the power required for controlling the ions in the devices should be minimized. In this study, we report liquid-type artificial synaptic devices with low power consumption and high ionic mobility. These devices consist of top and bottom electrodes that sandwich liquid materials. Because the liquid materials exhibit high ionic movement, the ions require less energy to migrate. Therefore, the developed liquid-type artificial synaptic devices can achieve low operating power consumption. The devices exhibit neuromorphic characteristics, including potentiation, depression, excitatory postsynaptic current, and paired-pulse facilitation. In addition, they can be employed as flexible neuromorphic devices because their main component is liquid. This study can provide a platform for the development of flexible and highly energy-efficient brain-mimicking devices. In this presentation, the characteristics of the proposed liquid-based neuromorphic devices characteristics will be presented in detail.

Recently, topotactic phase transformation of brownmillerite (BM) oxides has attracted great attention due to their potential physical and electronical properties for resistive random access memory (RRAM) devices. Compared with other non-volatile memory (NVM), RRAM exhibits low power consumption, high stability, and simple structure properties. However, the development of switching mechanism and the oxygen vacancy transition for multilevel operation in nanometer-scale memory devices is still incompletely understanding .In this work, epitaxial BM-LaCoO_2.5 thin films were grew on conductive Nb-doped SrTiO₃ as the resistive switching layer, and deposited the Au/Ti metal as the top electrode. In order to reveal the resistive switching characteristics, we use the high-resolution transmission electron microscope (TEM) and Atomic-scale scanning transmission electron microscopy (STEM) to observe the topotactic phase change in LaCoO_x. Under positive bias, the device switches from the pristine high resistance state (HRS) to the LRS, and the structure transformed from brownmillerite LaCoO_2.5 to perovskite LaCoO₃ filament. As we continuously modulate the SET/RESET voltage and compliance current (CC), the device exhibited multilevel resistance switching. This continuous change in the resistance could also influence the retention characteristics. Additionally, we use X-ray photoelectron spectroscopy (XPS) to demonstrate the Co valence changed for further studying micro-structural evolution and the resistive switching mechanism. This study indicated that topotactic phase transition in the LaCoO_x thin film layer, provides a new perspective for RRAM multilevel resistive switching application.

In this talk I will describe an organic electrochemical neuromorphic device that switches at record-low energy (<0.1 fJ projected, <10 pJ measured) and voltage (< 1mV, measured), displays >500 distinct, non-volatile conductance states within a ~1 V operating range. Furthermore, it achieves record classification accuracy when implemented in neural network simulations. Our organic neuromorphic device works by combining ionic (protonic) and electronic conduction and is essentially similar to a concentration battery. Our synapses display outstanding speed (<20 ns) and endurance achieving over 10⁹ switching events with very little degradation all the way to high temperature (up to 120°C). Having developed these devices for a few years, we have recently started working with collaborators on demonstrating what they can do. As a result, I will describe two demonstrations of learning using our devices. In the first one, a crossbar array of synapses “learns” to perform basic logic operations. In the second one, a simple electronic circuit that incorporates the synapse “learns” when to turn right and when to turn left in a predetermined maze. In the last demonstration, we take advantage of the inherent compatibility of our polymers with living matter. We show that a chemical signal, such as dopamine secreted from cells can be used to generate electronic updates to the device, as a first step towards integrated brain-machine interfaces.

The field of neuromorphic electronics is promising for local training and learning in energy restricted environments as well as in bioelectronics. Various neuromorphic devices will be presented that are based on organic-mixed conductors (ionic-electronic), materials that are traditionally used in organic bioelectronics. A prominent example of a device in bioelectronics that exploits mixed conductivity phenomena is the organic electrochemical transistor (OECT). Devices based on OECTs show volatile, non-volatile and tunable dynamics suitable for the emulation of synaptic plasticity and neuronal functions for the mapping of artificial neural networks in physical circuits. Finally, small-scale organic neuromorphic circuits enable the local sensorimotor control and learning in robotics.

Evolvable organic electrochemical transistors (EOECTs) rely on the polymerization of a semiconducting monomer to grow the transistor channel. In contrast to traditional OECTs, the channel conductance at a given combination of drain and gate voltages can be modulated reversibly in situ, which allows for in-memory computing at the level of a single device and renders EOECTs uniquely well-suited as a basis for neuromorphic hardware. Herein, we demonstrate how several commonly used machine learning algorithms can be implemented on an array of EOECTs by using the conductance of the transistor channel as an analog representation of weight. We implement the Widrow-Hoff learning scheme, an early version of iterative least-mean-square optimization, for binary pattern classification. Additionally, we implement two modes of multi-class classification—two-factor classification and a decision tree—on a random weight matrix to distinguish between digits written to a 3x5 grid.

Neurodegenerative diseases and brain damage require the constant development of new technologies able to replicate synaptic functionalities and replace damaged connections. As widely reported in literature, synaptic communication is based on electrochemical mechanisms: certain stimuli might induce the release of neurotransmitters from the pre-synaptic site and the consequent recognition of these molecules at the post-synaptic receptors, thus causing the propagation of an action potential from one cell to another.¹ In this scenario, neuromorphic devices represent a powerful platform to recreate artificial neural networks, which will offer the chance to improve implantable devices towards the replacement and repair of injured neuronal network areas. In particular, organic neuromorphic devices, such as PEDOT:PSS organic electrochemical transistor (OECT), offer a unique approach: in addition to the well-known biocompatibility of PEDOT:PSS, owing to its ionic-to-electronic current transduction, the conductance of these devices can be modulated through external stimuli (i.e. ionic flow) and the altered conductance state is retained over long time, replicating to some extent the physiological synaptic plasticity.² Recently we have successfully established a direct coupling between PC12 neuron-like cells, able to secrete dopamine (DA)³, and an organic neuromorphic device, thereby recreating an in vitro model of the synaptic system which simulate both long and short-term plasticity through the oxidation of DA.⁴ However, such biohybrid synapses still lack biomimetic features, which could promote their seamless integration within the neuronal network.
Here, we aim to recapitulate not only the same functionalities of neurons (i.e. synaptic plasticity) using an organic neuromorphic device, but also to have a biomimetic platform replicating physiological cell membranes by means of supported lipid bilayers (SLBs). We investigated how the bilayer fluidity affects the ions flow towards the organic neuromorphic channel, thus modulating the de-doping level of PEDOT:PSS and the short-term memory of the neuromorphic OECT: For this reason, biomembranes with two different compositions were investigated: a fluid homogeneous bilayer formed by a single-type lipid molecule, and a more complex SLB which displays a higher rigidity and the same composition of biological neuronal membranes. We expect that according to lipids diffusion, such artificial bilayers could modulate differently the passage of ions through the membrane, with rigid SLB acting like a barrier and fluid membranes displaying a more sponge-like behavior. Furthermore, such platforms thanks to their learning capabilities and biomimetic properties, could really interact with biological cells, monitoring their activities and providing feedback accordingly, thus creating an artificial biohybrid functional network. This biohybrid platform might be further exploited to engineer future implantable devices able to replace damaged connections either in pathological networks in neurodegenerative diseases, or in case of amputations to create adaptive prosthetics able to reduce the mismatch between these artificial parts and the human body.
References.
1. Burns, M. E. & Augustine, G. J. Synaptic structure and function: Dynamic organization yields architectural precision. Cell 83, 187–194 (1995).
2. Gkoupidenis, P., Schaefer, N., Garlan, B. & Malliaras, G. G. Neuromorphic Functions in PEDOT:PSS Organic Electrochemical Transistors. Adv. Mater. 27, 7176–7180 (2015).
3. Westerink, R. H. S. & Ewing, A. G. The PC12 cell as model for neurosecretion. Acta Physiol. 192, 273–285 (2008).
4. Keene, S. T., Lubrano, C., Kazemzadeh, S., Melianas, A., Tuchman, Y., Polino, G., Scognamiglio, P., Cinà, L., Salleo, A., van de Burgt, Y., Santoro, F. A biohybrid synapse with neurotransmitter-mediated plasticity. Nat. Mater. (2020).

Organic electrochemical transistors (OECTs) provide a device structure that emulates key aspects of biological neuronal function, including utilization of ion migration to induce changes in conductivity state, inherent nonlinearity in the transfer function, history dependent behavior, and the ability to accept multiple, simultaneous inputs from neighboring units [1,2]. The last trait, implemented in the form of multi-gate devices, provides design flexibility in network topology that is well-suited toward the development of reservoir computing architectures. These have the potential to operate with very low energy-per-computation values, into the femtojoule range [2], while taking advantage of dynamical noise that can enhance the accuracy of the computation [3]. Here we present studies of multi-gate OECT reservoir computing circuits via numerical simulation and subsequent fabrication by direct-write methods. The simulations are based on a robust, scalable method to calculate drain-source currents as a function of multiple gate inputs and to then cascade the output to multiple devices elsewhere in the reservoir. We show that doing so involves careful balance of the calculation timescales of the internal dynamics of the OECTs versus the dynamics of the network as a whole. We benchmark the simulations with classification tasks of simple time series data, e.g., square wave versus triangle wave input and find that near unity efficiency can be obtained with small reservoirs of approximately ten OECTs with randomized connections. We then present experimental studies of inkjet printing of OECT circuits in reservoir architectures, aimed at fabrication of the complete reservoir with scalable, additive manufacturing methods optimized for soft and viscoelastic materials.
[1] Y. Tuchman, et al., “Organic neuromorphic devices: Past, present, and future challenges”, MRS Bulletin 45 (8), 619-630 (2020).
[2] J. C. Perez and S. E. Shaheen, “Neuromorphic-based Boolean and reversible logic circuits from organic electrochemical transistors”, MRS Bulletin 45 (8), 649-654 (2020).
[3] A. Banerjee, J. Pathak, R. Roy, J. G. Restrepo, and E. Ott, “Using machine learning to assess short term causal dependence and infer network links”, Chaos 29 (12), 121104 (2019).

Drug discovery has become slow and increasingly expensive. One of the primary causes being the 'better than the Beatles' problem, whereby each new drug discovered needs to be a significant improvement to the back catalogue of approved medicines^[1]. As an alternative, researchers have been developing bioelectronic therapeutics which leverage electrical signals originating from the patients to modulate biological activity in a closed feedback loop; examples of these bioelectronic systems extend from treatment of diabetics to neuromodulation devices for mediation of epileptic seizures^[2],[3]. Implementation of these bioelectronics requires precise control and personalisation, leading to increasing interest in smart bioelectronics^[4] which are specifically designed to fit the patient in contrast to a one-size-fits-all approach. Smart bioelectronics use specific brain-like neuromorphic circuits which emulate characteristics of biological synapses through the co-location of information storage and processing to directly modulate biological environments^[5]. However, to achieve the necessary personalisation to tailor these smart devices to individual patients, a modular and adaptable manufacturing technique should be implemented. Herein, we introduce the use of additive manufacturing as a tool to fabricate and deliver personalised neuromorphic devices. Additive manufacturing has moved from rapid prototyping to being used as a direct digital manufacturing tool for end-use electronic components and devices^[6–8]. By using additive manufacturing to fabricate neuromorphic devices, we build components layer-by-layer from mutable CAD models, enabling rapid design change and fabrication of more geometrically complex components ^[9,10]. This rapid feedback in the development of neuromorphic devices allows for smarter, more complex, and personalised bioelectronics. In this presentation, we demonstrate how accessible, low-cost fuse deposition modelling and inkjet printing additive manufacturing techniques can be used to deliver a wide range of neuromorphic devices and circuit architectures. We go on to show how additive manufacturing design geometries and print parameters can be used to control the electronic properties of the devices as well as the functionality of integrated systems. Finally we demonstrate how additive manufacturing can be used to fabricate complex neuromorphic circuits such as spiking neural networks. We hope the work presented shows that cheap and accessible additive manufacturing techniques can be used to deliver smart and personalised bioelectronics into healthcare.
References
[1] J. W. Scannell, A. Blanckley, H. Boldon, B. Warrington, Nat Rev Drug Discov 2012, 11, 191.
[2] C. Diacci, J. W. Lee, P. Janson, G. Dufil, G. Méhes, M. Berggren, D. T. Simon, E. Stavrinidou, Advanced Materials Technologies 2020, 5, 1900262.
[3] C. M. Proctor, A. Slézia, A. Kaszas, A. Ghestem, I. del Agua, A.-M. Pappa, C. Bernard, A. Williamson, G. G. Malliaras, Science Advances 2018, 4, eaau1291.
[4] E. R. W. van Doremaele, P. Gkoupidenis, Y. van de Burgt, Journal of Materials Chemistry C 2019, 7, 12754.
[5] Y. Tuchman, T. N. Mangoma, P. Gkoupidenis, Y. van de Burgt, R. A. John, N. Mathews, S. E. Shaheen, R. Daly, G. G. Malliaras, A. Salleo, MRS Bulletin 2020, 45, 619.
[6] C. Kim, D. Espalin, M. Liang, H. Xin, A. Cuaron, I. Varela, E. Macdonald, R. B. Wicker, IEEE Access 2017, 5, 25286.
[7] T. Tilford, S. Stoyanov, J. Braun, J. C. Janhsen, M. Burgard, R. Birch, C. Bailey, Microelectronics Reliability 2018, 85, 109.
[8] D. Espalin, D. W. Muse, E. MacDonald, R. B. Wicker, Int J Adv Manuf Technol 2014, 72, 963.
[9] J. Gardan, International Journal of Production Research 2016, 54, 3118.
[10] T. N. Mangoma, S. Yamamoto, G. G. Malliaras, R. Daly, Advanced Materials Technologies, 2000798.

Biological neurons mainly communicate with each other through sparse-in-time trains of spikes. However, in nature homogeneous networks of neurons do not exist: in the human retina a complex circuitry comprising spiking neurons and non-spiking interneurons encode and process the light signal coming from the photoreceptors. In the field of neuromorphic computing,exploiting conventional Si-based electronics, spiking neural networks (SNNs) emulate the biological spiking elements since they are specifically based on spikes communication among neurons whose connections are weighted by spike timing dependent plasticity (STDP)¹. On the other hand, non-spiking interneurons functions have been replicated (still with inorganic circuits) by “sub-threshold encoding”². Although the software design in neuromorphic computing has been based on inorganic materials hardware, particularly on silicon spiking circuits³, organic materials are emerging.
Beside displaying a set of advantageous circuit design properties ( reduced number of devices and pads for circuitry connections⁴, flexibility of architectures⁵, decreased fabrication cost⁶) organic devices promise to replicate the intricate cooperation mechanisms of biological circuit elements.
Here an organic spiking/non-spiking circuit is developed to emulate biological heterogeneous neural networks, particularly focusing on the human retina circuitry, to execute SNNs-based algorithms. Indeed, organic devices such as OECTs and ENODEs⁷operate by design through graded potential signals so that they directly mimic non-spiking interneurons, while offering alternative strategies (conductance and capacitance modulations ) to control the weight of synaptic connections⁸, avoiding to recur to complex methods. Spiking neurons operations are then recreated by an organic spiking unit where depletion-mode PEDOT:PSS and enhancement-mode PEDOT:PSS:DEMTA OECTs are integrated in a zero-V_GS inverter configuration which represents the best solution in terms of gain, noise margin and transistor count topology. The inverter characteristics (gain, frequency of operation and voltage range) can be tuned by matching OECTs with different channel lengths. Indeed, this integrate-and-fire spiking neuron allows to use and adapt the same algorithms so far developed for inorganic hardware to perform retina-like pre-processing functions such as spatio-temporal pattern recognition. As such, a model neuronal network pattern relying on the cooperation of spiking and non-spiking elements is recreated with neuromorphic computing functions. The use of organic materials, enabling tissue-integration and the direct transduction at the interface of chemical, physical and biochemical signals via organic biosensors⁹, match the “in-sensor computing” paradigms¹⁰. Finally, this neuromorphic spiking/non-spiking circuit promises to perform the fundamental tasks of data generation, collection and computation within the sensory device opening to future prosthetic applications.
1 Nguyen, D.-A. et al. JLPEA 11, 23 (2021)
2 Strohmer, B. et al. Front. Neurosci. 15, 633945 (2021)
3 Serb, A. et al. Scientific Reports 10, 1–7 (2020)
4 Pecqueur, S. et al. Adv. Electron. Mater. 4, 1800166 (2018)
5 Spyropoulos, G. D. et al. Science Advances 5, eaau7378 (2019)
6 Rashid, R. B. et al. Biosensors and Bioelectronics 190, 113461 (2021)
7 Burgt, Y. van de et al. Nat Electron 1, 386–397 (2018)
8 Mirshojaeian Hosseini, M. J. et al. J. Phys. D: Appl. Phys. 54, 104004 (2021)
9 Shen, H. et al. ACS Appl. Electron. Mater. 3, 2434–2448 (2021)
10 Zhou, F. et al. Nat Electron 3, 664–671 (2020)

Electrochemical resistive switching is among one of the various mechanisms found in different kinds of resistive switching devices that have been reported. One of the major interests in these devices is due to their potential for application in next-generation memory devices such as biological neuron-inspired neuromorphic computing. There have been various materials used for the development of resistive switching devices ranging from inorganic materials to polymers to biologically sourced materials. Here, we report the facile construction of resistive switching devices based on pristine leaves of a household plant - Golden Pothos. We were able to switch between the high and low resistance states in these devices for hundreds of cycles with an average ON/OFF ratio of about 73, demonstrating a high level of endurance and their potential for random access memory devices. These leaf-based devices demonstrate a non-pinched current-voltage (IV) hysteresis loop observed among some electrochemical memristors. We will discuss about the capacitive origin of the non-pinched IV behavior based on our experiments and make a case for the capacitive-coupled memristor model that has been suggested by others[1].
[1] Sun, B. et. al. Nano Lett. 19, 6461-6465 (2019).

Organic electrochemical transistors (OECTs) are in a stage of rapid development as novel applications that use these versatile devices continue to emerge. OECTs are characterized by the coupling of both ionic and electronic inputs to modulate transistor channel conductance. This attribute renders OECTs ideal for interfacing electronics with biological systems, which use ionic and biochemical currents and gradients for signaling. Here, we will make use of OECTs to develop organic electrochemical neurons and synapses with ion-modulated spiking. We will discuss their facile bio-integration, using Venus Flytrap as a bio-signaling model system, and demonstrate a simple neurosynaptic circuit with Hebbian learning capabilities. These soft and flexible organic electrochemical neurons and synapses operate at low voltage and respond to multiple stimuli, defining a new era for bio-integrable artificial neuronal systems.

With the advent of next-generation technologies such as the internet of things (IoT) and artificial intelligence (AI), it became important to efficiently process a massive amount of information. However, it is difficult to catch up with such an overflow of information using conventional von Neumann architecture where data bottleneck occurs due to serial processing. As an alternative, a new system emulating the human nervous system called ‘neuromorphic computing’ has recently attracted significant attention. Similar to parallelly connected synapses of the human nervous system, neuromorphic computing can simultaneously perform functions including memory, reasoning, and learning with low power consumption.
Among various neuromorphic devices, interest in bio-implantable electronics is especially growing along with the increase of interest in healthcare. Such devices require bio-implantability, a property of not affecting cells or organs once implanted into the human body, as well as flexibility for mechanical compatibility with adjacent organs. However, conventional inorganic materials for neuromorphic devices cannot meet the requirements mentioned above.
Herein, we introduce a promising biomaterial called L-DOPA (3,4-Dihydroxy-L-phenylalanine), a type of mussel protein. It has been used for a natural suture in vascular surgery due to no antigen-antibody reaction with surrounding tissues and cells. Also, L-DOPA is used as an additive to improve the stretchability and flexibility of carbon-based fibers, thereby implying bio-implantability and flexibility and thus enabling its use for implantable neuromorphic devices. L-DOPA solution was prepared by dissolving the L-DOPA powder into 2-methoxyethanol with 5 wt. %. After aging 24 hours, L-DOPA thin film was deposited by spin-coating (3000 rpm, 30 sec) on p⁺ Si substrate. The samples were then annealed at 85^oC for 1 hour on a hot plate in an ambient air condition. Finally, L-DOPA based neuromorphic device was fabricated into a form of metal-insulator-metal configuration with deposition of Mg top electrode by e-beam evaporation. In such a structure, the top electrode and bottom electrode respectively correspond to pre-synapse and post-synapse of neurotransmission, and synaptic weight is expressed as a form of resistance of the active layer.
To confirm non-linear transmission properties, five positive (0 to 2 V) voltage sweeps and negative (0 to –2 V) voltage sweeps were applied to the device. The current level was changed depending on the direction of voltage sweep, resulting in a gradual increase of positive voltage and a gradual decrease in negative voltage for each case, demonstrating that the device can modulate synaptic weights like biological synapses.
The excitatory postsynaptic current (EPSC) was measured by applying a voltage of 2 V to the top electrode. Once the spike pulse with a duration of 10 ms was applied, the current level increased to have the peak value of 400 μA and then returned to the initial state. When the spike pulse width was increased from 10 ms to 100 ms, both the EPSC level and the time required for returning to the initial value increased. Finally, while fixing the pulse duration into 10 ms, the EPSC level increased as the pulse amplitude increased from 1.8 to 2.3 V at an interval of 0.1 V.
To confirm the short-term plasticity in biological behavior, paired-pulse facilitation (PPF) characteristic was demonstrated by applying two consecutive pre-synaptic spike pulses (2 V, 10 ms) and varying the pulse width from 10 to 180 ms. As the pulse interval became smaller, the EPSC level did not return to the initial value and rather increased after applying the second pulse. From bio-implantability and flexibility of the material and demonstration of various neuromorphic characteristics, this research implies the future applicability of L-DOPA-based neuromorphic devices for bio-implantable electronics.

The rise of artificial intelligence and big data applications in biological domains requires a new branch of biocompatible, flexible, untethered, and therefore, energy efficient or self-sustainable, computational devices. Such devices must be able to preprocess bulk collected data and only output the relevant, classified data to an interface. Generally, conventional electronics do not meet these requirements, thus, finding equivalents in soft materials or existing biological materials is necessary. Further, traditional methods of data processing using von Neumann computers would not translate well into biological systems as energy efficiency is relatively low, even in optimized silicon chips. Other energy efficient methods of data processing, such as reservoir computing neural networks, may prove more viable, as they are modeled on the brain’s architecture. Recently, volatile memristors have been shown to be an effective way to implement a reservoir layer into a neural network. An example of such devices that are suitable for biological applications are lipid-membrane-based memristors. Specifically, lipid membranes doped with voltage-dependent, pore-forming peptides, such as alamethicin, demonstrate highly nonlinear volatile memory in the current-voltage plane unmatched in their silicon-based counterparts. However, the main current limitation to this technology lies in the fact that they have never been scaled to the point where bulk calculations could be carried out to perform computational tasks. Here, we develop a novel technique of scaling up biological memristors based on the crossbar architecture that is ubiquitous in semiconductor-based memristor arrays using only soft materials. We fabricate a three-layered microfluidic chip, where lipid-encased aqueous droplets generated off chip move through a continuous oil phase. These droplets are then captured in pairs by geometric constraints in precise locations that match with small openings from the top and bottom layers. These openings will allow contact between the droplets and hydrogel crossbar electrodes, granting two-terminal access to each membrane while minimizing footprint. When two droplets are captured adjacent to one another, they form a lipid bilayer at their interface. Each bilayer that is doped with alamethicin exhibits memristive behavior. Finally, we demonstrate these biological memristors’ ability to perform as a reservoir layer in neural network classification tasks using the MNIST dataset of handwritten digits. Our work paves the way toward a new generation of computational devices that are completely organic and modeled on the brain architecture to perform efficient computations in biological domains.

Hardware implementation of neural networks is vital to realize power-efficient and fast artificial intelligence (AI) systems for a plethora of applications, e.g., bio-signal classification or real-time monitoring of industrial processes. Following the traditional paradigms of machine learning, efficient hardware implementation of AI systems requires versatile artificial synapses with well-adjustable synaptic strength, also referred to as synaptic plasticity.
However, besides these traditional approaches, it is worth taking a look at the function of biological neural networks as they possess superior properties in terms of power efficiency, fault tolerance, and flexibility. In particular, biological neural networks attain their excellent efficiency from employing specific non-linear properties of the synapses and neurons as well as of the entire system.
In this contribution, I discuss the non-linear properties of organic electrochemical transistors (OECTs), which are in the spotlight of research as they enable the hardware implementation of versatile, low-voltage artificial synapses. The non-linear properties are based on the complex interaction between polarons and ions, which can be manipulated, e.g., by chemical treatment of the semiconducting polymer. I will show how the non-linearity can be quantified and I will demonstrate three approaches to employ it for efficient computation. More specifically, I will discuss how the non-linear response of OECTs can be used to mimic the functions of biological neuron models (spiking neuron), help in decision making using Bayesian inference, and classify complex bio-signals using reservoir computing.

In bioelectronics, the rise of electrolyte-gated transistors, based on conducting polymers (CPs), enabled the development of efficient platforms interfacing with biological tissues¹. Recently, thanks to the development of inexpensive, tuneable and energy efficient devices, organic neuromorphic engineering emerged as promising approach to overcome conventional silicon technology limitations. In particular, PEDOT:PSS-based organic electrochemical transistors (OECTs), have come to play an important role for their natural ionic-to-electronic signal transduction and biocompatibility. In addition, the intrinsic bulk capacitance changes, resulting from the doping/de-doping processes occurring upon the application of electrical signals, endows such devices with increased sensitivity and possibility to operate at low voltages².
To date, the operation mechanism of numerous OECTs has been widely studied: top and planar gating configurations were investigated, along with polarizable and non-polarizable gate materials³. On the other hand, their combination with biological systems makes it difficult to obtain a proper model able to incorporate complex phenomena like diffusion and surface reactions.
Importantly, supported lipid bilayers (SLBs), widely employed to study membrane functions and transmembrane proteins’ activities⁴, are often coupled with OECTs and have been studied and modelled through passive electrical models⁵. This modelling is crucial in biosensing applications, however there is a lack of studies of dynamical modelling of OECTs coupled to such membranes.
In this regard, the present work aims to study and model the physics and the operation regimes of OECTs coupled with SLBs, investigating how the presence and the position of such biomembranes affect the behaviour of transistors fabricated on different inorganic conductors. In fact, PEDOT:PSS is patterned on both gold and indium tin oxide (ITO) substrates. The two obtained architectures are characterized and differences in terms of response time and ion diffusion inside the bulk of the CP channel are highlighted. An SLB formed by POPC and a brain cell-like SLB composition are presented and characterized through impedance measurements and pulsed transistor operations, particularly relevant in neuromorphic operation regime of the above-mentioned devices. As a result, the two examined SLBs compositions lead to different modulations of the ionic circuit of the devices, in terms of response time and amplitude of the transistor channel current. This ultimately leads to the computation of a model of the SLB-OECT system, able to depict complex ionic interactions between CPs and phospholipid membranes. In contrast to passive electrical modelling of the bilayers, this approach allows to characterize and exploit the biological tissue itself as a tool to tune the electrical properties of an organic electrochemical transistors, at the same time enhancing the biomimetic potential of the CP-based electronics.
1. Mariano, A., Lubrano, C., Bruno, U., Ausilio, C., Dinger, N. B. & Santoro, F. Advances in Cell-Conductive Polymer Biointerfaces and Role of the Plasma Membrane. Chem. Rev. (2021).
2. Friedlein, J. T., McLeod, R. R. & Rivnay, J. Device physics of organic electrochemical transistors. Org. Electron. 63, 398–414 (2018).
3. Tan, S. T. M., Giovannitti, A., Melianas, A., Moser, M., Cotts, B. L., Singh, D., McCulloch, I. & Salleo, A. High-Gain Chemically Gated Organic Electrochemical Transistor. Adv. Funct. Mater. 31, 2010868 (2021).
4. Pappa, A.-M., Liu, H.-Y., Traberg-Christensen, W., Thiburce, Q., Savva, A., Pavia, A., Salleo, A., Daniel, S. & Owens, R. M. Optical and Electronic Ion Channel Monitoring from Native Human Membranes. ACS Nano 14, 12538–12545 (2020).
5. Zhang, Y., Inal, S., Hsia, C.-Y., Ferro, M., Ferro, M., Daniel, S. & Owens, R. M. Supported Lipid Bilayer Assembly on PEDOT:PSS Films and Transistors. Adv. Funct. Mater. 26, 7304–7313 (2016).

In the era of artificial intelligence of things (AIoT), flexible piezoelectric acoustic sensors (f-PAS) have been spotlighted as a promising candidate for voice user interfaces (VUI) by mimicking the human cochlea (trapezoidal membrane and ~10,000 hair cell channels). However, biomimetic f-PAS can induce the signal distortion in real-life applications, due to the fundamental difference and high sensitivity compared with the conventional microphones.
Herein, we demonstrate a deep learning-based noise-robust flexible piezoelectric acoustic sensor (NPAS) for speech processing. The noise-robust response was achieved via three methods: i) the frequency coverage of multi-channel NPAS, ii) the optimized seven-signal processing by convolutional neural network (CNN), iii) the newly designed deep U-net model for speech enhancement. The NPAS achieved noise-robust response and 0.1 – 8 kHz coverage by designing the multi-resonant bands outside the noise dominant spectrum, and using Nb-doped PZT (PNZT) membrane, respectively. Compared to the condenser microphones, the highly sensitive NPAS showed the clear sound detection with 35 times higher sensitivity and SNR. Adopting the newly optimized CNN model with the channel attention method, the NPAS exhibited a 62% reduction in error rate compared to the MEMS microphone. Deep U-net based speech enhancement of the NPAS was also achieved via the selective processing of multi-channel signals. Last, the AI-based NPAS demonstrated the separation of multi-speaker’s voices from a crowd.

McCulloch and Pitts proposed that the brain is composed of logic-gates. The idea of Boolean neurons had a huge effect on artificial intelligence. To investigate whether Boolean logic gates can be implemented using biological neurons, neural networks with a crossbar array structure were constructred on multi-electrodes arrays using PDMS microfluidic channels. The fluorescence optical images revealed that the crossbar array consisted of axons. The activation curves measured using the stimulus-evoked spike probability exhibited the spike timing dependent plasticity learning effects. Furthermore, reconfigurable logic gates of AND and OR were successfully implemented through learning. Most neuromorphic devices have the crossbar array structure ; thus, these studies may contribute to the development of neuromorphic studies.

Despite the exponential rise in the field of brain–computer interfaces (BCI), achieving long-term stability of microelectrodes in the nervous system remains a challenge. This is due to the inflammatory encapsulation of implanted neural probes, leading to a gradual decrease of device functionality after implantation, as a result of the foreign body reaction (FBR) triggered by the foreign material. The device morphology, biocompatible coating, and insertion method are key factors to consider in mitigating the FBR. Recently, biohybrid implants have gained popularity as a design for housing cultured cells by acting as a biological intermediate layer at the tissue-electrode interface and potentially improving biocompatibility and in vivo performance. In vitro, native supported lipid bilayers (SLBs) formed on thin film arrays have demonstrated to be powerful multimodal sensors used for a wide range of applications such as ion channel studies and understanding protein-lipid interactions. Herein, we introduce the first use of human-derived SLBs, integrated into biomimetic electronic devices, as a meditator between the implanted electrodes and neural tissue for enhanced tissue integration. This biohybrid modal offers a unique advantage to its counterparts by preserving the native environment and compositional complexity of the cell membrane, and implanting it to reduce inflammation rate and improve the device performance functionally and anatomically.

As the information technology industry continues to develop, conventional von Neumann architecture is facing the limit for managing the increasing amount of big data. It is difficult to improve data throughput in von Neumann architecture due to the bottleneck between the computing and memory units. Especially, in the implantable bioelectronic devices, this bottleneck is becoming more severe since the devices should deal with various signals for monitoring blood glucose, blood pressure, heart rate, and electrocardiogram signal, etc.
To overcome this bottleneck issue, an alternative architecture called neuromorphic computing architecture is being studied. Neuromorphic computing architecture is inspired by the human brain so that it can emulate the biological synapses to perform cognition, learning, and memory concurrently. Neuromorphic systems for implantable bioelectronics should be consist of biocompatible materials to prevent inflammatory reactions when applied to the human body. Moreover, it should have biodegradable properties so as not to cause an issue that additional operation is required to remove the device implanted in the body at the end of its lifespan or after fulfilling its duties.
In this research, we propose a silibinin (SB) extracted from milk thistle seeds as the active layer for implantable neuromorphic devices. SB is biocompatible and biodegradable enough to be used as a pharmaceutical drug for fatty liver and cancer treatment. In addition, since the top and bottom electrodes of the neuromorphic device consist of Mg and W, which have biocompatible and biodegradable properties, the devices were fabricated only with materials suitable for implantable bioelectronics.
The biological nervous system of the human body is composed of numerous neurons and synapses, and neural signals are transmitted through synapses between neurons. To emulate this biological nervous system, we fabricated an artificial synapse with a two-terminal metal-insulator-metal structure neuromorphic device. To measure the nonlinear transmission characteristics of the SB-based neuromorphic device, 10 consecutive positive and negative voltage sweeps were applied, respectively. During 10 consecutive positive voltage sweeps, the current level was gradually increased for each sweep which means a partially set process. Conversely, during 10 consecutive negative voltage sweeps, the current level was gradually decreased (partially reset process) was proceeding. Through the gradual change in the current, the SB-based neuromorphic device control synaptic weights.
The electrical characteristics of the SB-based neuromorphic device could be modulated even when a continuous programmed pulse was applied similar to the operating environment of the biological synapse. Excitatory post-synaptic current (EPSC) was measured according to the pre-synaptic spike pulse width. A 2.5 V of the pre-synaptic spike was applied to the Mg top electrode and the spike pulse width was gradually increased from 5 to 100 ms to read the current changes after stimulation. As the spike pulse was applied, EPSC increased abruptly and returned to the initial value almost immediately. Another characteristic of biological synaptic systems is paired-pulse facilitation (PPF). When paired pre-synaptic spikes with a pulse width of 10 ms apply with an interval time of 10 ms, the second EPSC was triggered a larger peak value than the first EPSC. In the case of the SB-based neuromorphic device, it was confirmed that the PPF index was over 126%. Furthermore, the potentiation-depression property of SB-based neuromorphic device according to consecutive 50 positive pulses and followed by 50 negative pulses. The EPSC of the SB-based neuromorphic device can be precisely controlled by cyclical multiple positive and negative spike pulses. Through these synaptic characteristics, it was confirmed that the SB-based neuromorphic device would open a new era in future implantable bioelectronics.

Bio-inspired artificial perception system has been attracted interest to realize cognition, learning, and memory functions for conferring visual, auditory, and tactile perception system to human and robots. In particular, since the visual perception system is directly related to object recognition, various efforts are being conducted to mimic this system. Visual perception in the human body is achieved by photoreceptors, neurons, and the visual cortex. Photoreceptors in the retina convert external light inputs into electrical spikes. After generating electrical spikes, neurons transfer the electrical spikes to the visual cortex. Therefore, to emulate the visual perception system, an optical neuromorphic transistor (ONT) which can detect light and show neuromorphic functions is required.
Recently, indium-gallium-zinc oxide (IGZO) has been introduced as a channel of the ONT due to its inherent photoresponse characteristic called persistent photoconductivity (PPC). It shows long decay time in drain current after light irradiation having high energy enough to generate photoexcitation. Accordingly, drain current shows an increasing tendency during continuous light pulse similar to potentiation characteristic. It is believed that the ionized oxygen vacancies (V_o⁺ or V_o²⁺) are the origin of the PPC characteristic in IGZO. However, IGZO ONT has a limited range of light detection due to wide bandgap of IGZO (> 3 eV). To expand the light detection range, the light absorbing material with narrow bandgap and high optical sensitivity is required.
In this study, Bi₂Se₃, one of the topological insulators, is applied to the light absorbing layer of IGZO ONT for realizing the artificial visual perception system. Bi₂Se₃ has narrow bandgap (~300 meV) enough to expand the light detection range from UV to mid-infrared (IR) range, fast charge transport capability, and high carrier mobility due to its helical band structure. To fabricate the IGZO ONT, the IGZO channel is deposited on a p⁺-Si substrate with thermally grown SiO₂ by radio frequency (RF) magnetron sputtering and annealed at 300^oC in the air atmosphere for 1 h. For source and drain, Al electrodes were deposited by RF magnetron sputtering with a shadow mask. The width and length of channel were 1,000 and 150 μm, respectively. For the light absorbing layer, Bi₂Se₃ layer was deposited by drop-casting on the IGZO channel.
To confirm optoelectronic characteristics, transfer curves were measured under red, green, and blue light irradiation with the wavelength of 630, 532, and 405 nm, respectively, at intensity of 5 mW. The IGZO ONT without Bi₂Se₃ layer exhibited poor optoelectronic characteristics such as photoresponsivity of 24.82, 225.05, and 665.83 A/W, photosensitivity of 1.17 × 10², 7.70 × 10⁴, and 1.88 × 10⁸, and detectivity of 1.19 × 10⁷, 2.06 × 10⁹, and 3.65 × 10¹² Jones under red, green, and blue light irradiation, respectively. The optoelectronic characteristics were improved by applying Bi₂Se₃ layer, showing photoresponsivity of 176.39, 735.88, and 879.92 A/W, photosensitivity of 1.65 × 10⁶, 5.85 × 10⁸, and 2.63 × 10⁹, and detectivity of 2.52 × 10¹⁰, 8.91 × 10¹², and 2.27 × 10¹³ Jones under red, green, and blue light irradiation, respectively. Furthermore, neuromorphic functions such as paired-pulse facilitation (PPF), short-term plasticity, and long-term plasticity were measured under continuous green light pulses. Frequency and duty ratio of green light pulses were 1 Hz and 50 %, respectively. By applying 2 consecutive green light pulses, the IGZO ONT with Bi₂Se₃ layer showed the PPF index of 202.44 %. With consecutive 5 and 50 green light pulses, the IGZO ONT with Bi₂Se₃ layer showed short-term plasticity and long-term plasticity, respectively. As a result, it was confirmed that the IGZO ONT can mimic photoreceptors and neurons which convert light input to electrical spikes and show neuromorphic functions by applying Bi₂Se₃ layer.

Artificial photonic synapses with morphologically controlled photoreception, allowing for area-dependent tunable light reception as well as information storage and learning, have the potential in emerging photo-interactive neuro-computing technologies. Herein, we present an artificially intelligent photonic synapse with structurally tunable perovskite nano-cone arrays templated in a self-assembled block copolymer (BCP), which is based on a field-effect transistor with a floating gate of photoreceptive perovskite crystal arrays preferentially synthesized in a micro-phase-segregated BCP film. These arrays are capable of electric charge (de)trapping and photo-excited charge generation, and they exhibit versatile synaptic functions of the nervous system. The area-density variable perovskite floating gate developed by the off-centered spin coating process allows for emulating the human retina with a position-dependent spatial distribution of cones. 60 × 12 arrays of the developed synapse devices exhibit position-dependent dual functions of receptor and synapse. They are artificially intelligent and exhibit a pattern recognition accuracy of up to approximately 90% when examined using the Modified National Institute of Standards and Technology (MNIST) handwritten digit pattern recognition test.

The electronic and magnetic properties of the prototypical Mott insulator V₂O₃ are very sensitive to stress, which strongly modulate the stability of its various structural, magnetic and electronic phases. In this work we use this sensitivity to manipulate these phases by substrate-induced strain. By growing thin films of V₂O₃ on sapphire substrates of different orientations, we can induce both compressive and tensile stress allowing access to phases which are inaccessible in bulk samples. This kind of manipulation can be useful for developing novel functionalities and tailoring electronic properties for applications such as memory devices, optical switches and neuromorphic hardware.
In this poster I will focus on results obtained on thin films of V₂O₃ grown on sapphire substrates cut along the (100) plane. We find large compressive strain along the c-axis and an expansion perpendicular to it. This anisotropic strain drives V₂O₃ into a regime which has only been shown to be induced by Cr doping, and exhibits multiple transitions as a function of temperature: antiferromagnetic insulating (AFI) → paramagnetic metal (PM) → paramagnetic insulator (PI) → critical point. To track the structural phase transition and the changes in global and local electrical conductivity as a function of temperature, we use a combination of reciprocal space mapping using X-ray diffraction, electrical transport and atomic force microscopy.

In the last few years, organic optoelectronic devices, notably organic photodetectors (OPD) and organic light emitting diodes (OLED) have demonstrated exceptional performances in recording and stimulating neurons. Intrinsic properties of these polymers, electronic materials and devices, as well as their soft nature and mechanical conformability, make them excellent candidates for interacting with biological systems. Despite these important advancements, neuroscience applications of these devices are still limited due to important constraints of scalability and miniaturization technique. These challenges prevent these devices from efficiently interfacing with the brain for single cell recording and stimulation, due to low spatial resolution. Micro patterning and optimization of the optoelectronic properties of µ-scale OPD pixels will provide the advantage of optical recording from the brain at both cellular and population levels. Furthermore, a reliable mechanically-flexible encapsulation technology is essential to prevent the degradation of these devices while interfacing the brain. We have successfully patterned the active layer of OPDs and achieved miniaturized devices. Then, subsequently we studied the influence of micro-patterning on the fundamental physical properties of these devices. These OPDs are based on fullerene as acceptor and P3HT as donor. Different sizes of the active layer were patterned and encapsulated with a novel encapsulation approach. These miniaturization and encapsulation methods are also expandable to OLEDs.

Threshold switching selector is an essential component within the 3D cross-point (X-Point) array that enables stable synaptic operation in neuromorphic devices. The necessity of high-performance selector is especially emphasized in the 3D X-Point due to requirements of low-power operation and drastic switching characteristic for neuromorphic operation, as well as low off-state leakage and device stability for data reliability of integrated circuit. In this respect, research on selector devices with outstanding electrical performances and stability have been conducted. Among various candidates for high-performance and 3D-stackable selector device, Ovonic threshold switching (OTS) device showed excellent properties that can potentially satisfy all the requirements that can be successfully adopted into the 3D X-Point memory.

OTS phenomenon is characterized by a drastic on/off switching operation within the amorphous chalcogenide film. Thanks to its potential for meeting various requirements, OTS selectors are considered strong candidates for neuromorphic and high-density memory applications. However, the demand for improvements in OTS selector has increased due to its poor device characteristics such as on/off ratio, delay time, endurance, and thermal stability. Preceding research had initially endeavored to satisfy the requirements mainly with binary OTS systems but failed to satisfy all the parameters due to limited composition controllability when fabricating the device. Such limitation motivated researchers to proceed to ternary and quaternary material systems where positive results were obtained with arsenic-based quaternary systems, but the toxicity of arsenic remains a major concern. In this research, we first demonstrate arsenic-free quaternary OTS device based on N-Si-Ge-Te chalcogenide system that shows excellent electrical performance such as high on/off ratio with an intrinsic selectivity of 10⁶, delay time of less than 5 nanoseconds while providing thermal stability over 400°C and an endurance of 10⁸ cycles. All of the parameters can be simultaneously achieved by understanding the role of each element in the material system that contribute to device’s overall characteristics, and thereby concurrently optimizing the device performances through precise modification of material composition. We hope that this result will help the development of high performance As-free OTS devices for neuromorphic applications as well as next-generation electronics.

This research was supported by the Ministry of Trade, Industry & Energy (MOITIE)/Korean Evaluation Institute Industrial Technology (KEIT) (Project No. 10080625) and Korea Semiconductor Research Consortium (KSRC) program for the development of the future semiconductor devices, and by Samsung Electronics Co. Ltd. (Project number IO2102021-08356-01).

Deep Learning algorithms have revolutionized many aspects of machine learning, especially in the fields of image recognition, speech detection and translation. Such improvement has been made possible by the increased computational capabilities of modern CPUs and GPUs, which have enabled the training of huge networks based on extremely large datasets.

While neural networks do not seem to slow down the need for increasingly higher number of trainable parameters, the long processing times are requesting novel and more specialized hardware capable of reducing both computational time and power consumption. Many examples are already available, reporting for specialized ASICs and computing clusters targeting well-defined networks.

Here, we describe the recent advances in reconfigurable hardware based on analog Phase-Change Memory (PCM) in 14-nm technology, able to perform inference of a variety of neural networks. Input signals are encoded using constant-amplitude and variable pulse-width pulses which are routed using a massively parallel 2D mesh. Multiply and Accumulate (MAC) operations are then performed on analog tiles based on 512x512 unique weights, each of them encoded using four PCM devices. Analog peripheral circuitry is then able to collect the integrated charge and convert it in pulses, enabling signal cascading between different tiles. Results showing end-to-end 2-layer MNIST network and full-MAC on-chip LSTMs reveal high accuracy.

While such implementation shows promising perspectives, additional insight is taken into future larger networks. The special case of Transformers is then addressed: such networks, which have a very large number of weights, can potentially be accelerated by leveraging analog hardware. Finally, impact of noise and weight quantization is described, revealing advantages for analog computing.

Phase change memory (PCM) is a promising candidate for brain-inspired and in-memory computing, which could enable data-intensive artificial intelligence applications [1,2]. Such systems rely on key attributes of PCM including gradual change of resistance states, low power switching, and multilevel operation with low resistance drift [2]. However, achieving all attributes simultaneously in the same device remains a challenge, especially with conventional phase change materials such as Ge₂Sb₂Te₅ (GST225) [2].

Here we show, for the first time, energy-efficient and neuro-inspired PCM devices using the recently discovered phase change nanocomposite Ge₄Sb₆Te₇ (GST467) [3] as well as Sb₂Te₃/GeTe superlattices (SLs) [4]. We also unveil the operation of such novel PCM devices through detailed materials characterization, electrical and thermal transport measurements, as well as simulations.

We demonstrate bi-directional and gradual resistance change (10x resistance window) in GST467 mushroom cell PCM devices with 110 nm bottom electrode diameter. For depression (low to high resistance transition), we apply consecutive reset pulses of the same magnitude and shape (1.45 V; 45 ns pulse width) while for potentiation (high to low resistance transition) we use 0.72 V; 10 ns set pulses. The minimum extracted energy-per-pulse is 1.53 pJ for depression and 86 fJ for potentiation, among the lowest for neuromorphic PCM [5,6]. These devices also show 13 resistance states achieved with several reset pulses (20 ns) of increasing amplitude (1.35 V to 2.1 V), resistance on/off ratio of ~1000, and low cycle-cycle variation suitable for high-density analog memory.

Extensive characterization reveals the unique properties of GST467 which enable the observed PCM behavior. Temperature (T) dependent X-ray diffraction spectra reveal a higher crystallization T but a lower melting T in GST467 compared to GST225. This enables high thermal stability yet lower energy operation in GST467 PCM. High-resolution transmission electron microscopy (HRTEM) shows the presence of SbTe nanophase grown coherently within a GST matrix. The SbTe nanophase has a lower local melting T compared to the overall nanocomposite [7], promoting the gradual change of resistance in GST467. While promising for gradual change in resistance, the resistance drift and switching current of GST467 need improvement.

To address the resistance drift and relatively high switching current in PCM, we also fabricated superlattice PCM devices with alternating layers of Sb₂Te₃ (4 nm) and GeTe (1 nm), both on rigid and flexible substrates. HRTEM confirms SL layers with sharp interfaces revealing van der Waals (vdW) gaps. Our transport measurements [8] uncover low cross-plane thermal conductivity and highly anisotropic electrical resistivity of the SL stack, leading to electrothermal confinement within the SL.

Our SL-PCM devices show multibit operation with a resistance drift coefficient of ≈0.01 and >10⁴ s retention. Such low resistance drift is an attribute of SL structures with vdW gaps. We also demonstrate record-low switching current density (from ~0.1 MA/cm² on a low thermal conductivity flexible substrate [4] to ~2.5 MA/cm² on rigid Si substrates) in these devices. This is enabled by heat confinement within the SL-PCM device, further confirmed through our electrothermal simulations.

In summary, we addressed the fundamental challenges of PCM for neuromorphic applications by examining novel epitaxial nanocomposites and superlattice materials. These results pave the way for energy-efficient neuromorphic computing using PCM.

1. A. Sebastian et al., Nat. Nanotech. 15, 529 (2020)
2. D. Kuzum et al., Nano Lett. 12, 2179 (2012)
3. A. Kusne, I. Takeuchi et al., Nat. Commun. 11, 5966 (2020)
4. A.I. Khan, E. Pop et al., Science 373, 1243 (2021)
5. S. Go et al., APL Mater. 9, 91103 (2021)
6. K. Ding et al., Science 366, 210 (2019)
7. T. Zhang et al., E\PCOS, 105 (2009)
8. H. Kwon, E. Pop et al., Nano Lett. 21, 5984 (2021)

Integrated nanophotonics is an emerging research direction that has attracted great interests for technologies ranging from classical to quantum computing. One of the key-components in the development of nanophotonic circuits is the phase-change unit that undergoes solid-state phase transformation upon thermal excitation. The quaternary alloy, Ge₂Sb₂Se₄Te (GSST), is one of the most promising material candidates for application in photonic circuits due to its broadband transparency and large optical contrast in the infrared spectrum. Here, we investigate the thermal properties of GSST and show that upon substituting Te with Se, the thermal transport transitions from an electron dominated to a phonon dominated regime. By implementing an ultrafast mid-infrared pump-probe spectroscopy technique that allows for direct monitoring of electronic and vibrational energy carrier lifetimes in these materials, we find that this reduction in thermal conductivity is a result of a drastic change in electronic lifetimes of GSST, leading to a transition from an electron-dominated to a phonon-dominated thermal transport mechanism upon Se substitution. In addition to thermal conductivity measurements, we provide an extensive study on the thermophysical properties of GSST thin films such as thermal boundary conductance, specific heat, and sound speed from room temperature to 400 °C across varying thicknesses.

Phase change chalcogenide materials, are good candidates for neuromorphic hardware intended to be embedded in microprocessors for the automotive industry, where devices must be able to operate reliably at high temperatures (> 160 °C) for at least ten years. Alloys like Ge₂Sb₂Te₅ (GST225) are already a standard to realize memory devices. However, one of the major disadvantages of GST225 single layers is the low crystallization temperature (T_c), which results in low thermal stability and limited data retention. An interesting option to increase T_c is to grow Ge-rich GST alloys. Here, several Ge_xSb₂Te₅ and Ge_xSb₂Te₃ layers were grown by RF-sputtering and their thermal stability was studied by X-ray diffraction (XRD) and Raman spectroscopy as a function of temperature. The compositional analysis was performed by X-ray Fluorescence (XRF). The results showed that the T_c amount grows linearly with the Ge content in the alloys; the amount of Ge segregation upon crystallization was also studied.
This project has received funding from the European Union's Horizon 2020 research and innovation program under Grant Agreement No. 824957 (“BeforeHand:” Boosting Performance of Phase Change Devices by Hetero- and Nanostructure Material Design).

The strong increase in digital computing power in combination with the availability of large amounts of data has led to a revolution in machine learning. Computers now exhibit superhuman performance in activities such as pattern recognition and board games. However, the implementation of machine learning in digital computers is intrinsically wasteful, with energy consumption becoming prohibitively high for many applications. For that reason, people have started looking at natural information processing systems, in particular the brain, that operate much more efficiently. Whereas the brain utilizes wet, soft tissue for information processing, one could in principle exploit any material and its physical properties to solve a problem. Here we give examples of how nanomaterial networks, in particular dopant network processing units (DNPUs), can be trained using the principle of material learning to take full advantage of the computational power of matter¹.
Previously, we have shown that a disordered (or ‘designless’) network of gold nanoparticles can be electronically configured into arbitrary Boolean logic gates using artificial evolution². We demonstrated that this principle is generic and can be transferred to other material systems. By exploiting the nonlinearity of variable-range hopping transport in a nanoscale network of boron dopants in silicon, we can significantly facilitate classification. Using a convolutional neural network approach, it becomes possible to use our device for handwritten digit recognition³. An alternative material-learning approach is followed by first mapping our Si:B network on a deep neural network model, which allows for applying standard machine-learning techniques in finding functionality⁴. Finally, we show that the widely applied machine learning technique of gradient descent can be directly applied in materio, using the principle of homodyne gradient extraction, opening up the pathway for autonomously learning hardware systems⁵.
References
[1] C. Kaspar, B. J. Ravoo, W. G. van der Wiel, S. V. Wegner and W. H. P. Pernice, The rise of intelligent matter, Nature 594, 345 (2021).
[2] S. K. Bose, C. P. Lawrence, Z. Liu, K. S. Makarenko, R. M. J. van Damme, H. J. Broersma and W. G. van der Wiel, Evolution of a Designless Nanoparticle Network into Reconfigurable Boolean Logic, Nature Nanotechnol. 10, 1048 (2015).
[3] T. Chen, J. van Gelder, B. van de Ven, S. Amitonov, B. de Wilde, H.-C. Ruiz Euler, H. J. Broersma, P. A. Bobbert, F. A. Zwanenburg, W. G. van der Wiel , Classification with a disordered dopant-atom network in silicon, Nature 577, 341 (2020).
[4] H.-C. Ruiz Euler, M. N. Boon, J. T. Wildeboer, B. van de Ven, T. Chen, H.J. Broersma, P. A. Bobbert and W. G. van der Wiel, A deep-learning approach to realizing functionality in nanoelectronic devices, Nature Nanotechnol. 15, 992 (2020).
[5] M. N. Boon, H.-C. Ruiz Euler, T. Chen, B. van de Ven, U. Alegre Ibarra, P. A. Bobbert and W. G. van der Wiel, Gradient Descent in Materio, arxiv.org/abs/2105.11233 (2021).

Traditional computing systems are unable to capture the capability of the brain in real world information processing as evidenced by the anticipated end to Moore’s law. Neuromorphic computing aims to mimic the efficiency of the brain in complex classification tasks by applying learning directly in hardware and is promising for local edge computing applications such as smart point-of-care devices (1). However, straightforward tuneability, low power consumption and high accuracy are necessary for successful implementation of such applications (2). Organic materials have emerged as building blocks for neuromorphic systems due to their low operation voltage, biocompatibility and excellent tunability properties (3). Nevertheless, fully autonomous bioelectronic applications demand not only the acquisition of biological signals, but also local data processing, storage and the extraction of specific features of merit (4). Here we show a smart biosensor based on organic neuromorphic devices that can be fully trained in hardware to classify a model disease (e.g. cystic fibrosis). We demonstrate the flexibility and versatility of the neuromorphic system by on-chip retraining with different input signals as well as by classification of logic gates. We further show adaptive sensing and local tuning using a single active material, an ambipolar mixed ionic-electronic conductor P-3O (5). These results pave the way for adaptive biosensors, on-site processing, and smart point-of-care devices.
1. E. J. Fuller et al., Science, 2019
2. van de Burgt et al., Nature Materials, 2017
3. van de Burgt et al., Nature Electronics, 2018
4. van Doremaele et al., J. Mater. Chem. C, 2019
5. van Doremaele et al., ChemRxiv, 2021

Organic electrochemical transistors (OECTs) emerged as promising building blocks for brain-inspired hardware thanks to the interplay of electronic and ionic species in their operation. Several studies demonstrated OECTs with synaptic properties, e.g., short- and long-term plasticity and spike-time-dependent plasticity. However, to go beyond academic demonstrations, the design and fabrication of neuromorphic circuits require high-density integration of OECTs, hardly possible due to the conductive nature of the electrolyte.
Here, we present a photopatternable solid electrolyte that we use to produce high-density OECT integrated circuits with a resolution of 10 µm. We use the patternability of the electrolyte to decouple adjacent devices and minimize the gate current. At the single-device level, the electrolyte offers strong gate-channel coupling, resulting in outstanding performance: the OECT exhibit an on-off ratio of 10⁶ and a subthreshold swing of 61 mv/dec close to the thermodynamic minimum.
This allows us to integrate OECTs into traditional and neuromorphic circuits without crosstalk between devices as well as adjust the heterosynaptic plasticity in OECT circuits. Therefore, we discuss the influence of the channel material and device dimensions on the device properties. We then tune the threshold voltage, switching speed, hysteresis, and retention time to our needs. As such, we control the device synaptic properties such as short- and long-term plasticity of single devices in a circuit. Finally, we use the device as building blocks in a Morris-Lecar spiking neuron circuit. We discuss the influence of tunable device properties on the behavior of the neuron’s subcircuits.

Organic transistor architectures that are able to interact with ions such as electrolyte-gated field effect transistors or electrochemical transistors are highly attractive devices for acquisition, transmission and processing of physiological signals. These devices are able to convert ions to electrons effectively and provide mechanical flexibility owing to their inherently soft channel materials. However, the temporal response of these devices, which directly translates into their operation speed, is defined by ion drift and mobility. As a result, they are limited to a few KHz speed regime. In addition, the lack of physically independent gate limits their ability to form integrated-circuits (ICs). Here we use internal ion-gated organic electrochemical transistors (IGTs) that are able to operate beyond the speeds dictated by ion mobility by creating local ion reservoirs inside the channel bulk. We demonstrate IGT-based amplifiers with uniform gain across their MHz range operation frequency. We characterize the electrical performance of these transistors with respect to various transistor sizes, contact area, and channel material composition, and determined the relationship between device geometry and gain, off current, leakage, and inter-transistor cross talk. Using these data and a high- resolution vertical fabrication approach (capable of creating <1 µm transistors), we then created high-density IGT-based conformable circuits that performed time-division multiplexing at speeds comparable to silicon-based ICs for physiological signal acquisition and multiplexing. The unique features of IGT-based devices facilitate potential applicability to a broad range of flexible circuits that require safe, chronic implantation in humans, including brain-machine interfaces, wearable electronics, and therapeutic responsive stimulation devices.

Emerging nanoelectronics devices offer an attractive solution to minimize off-chip data transfer and parallelize on-chip computations for hardware implementation of neural networks. While substantial progress has been made towards development of synaptic devices, compact nanodevices implementing non-linear activation functions have been yet to be demonstrated. In this talk, I will present our work on energy-efficient and compact activation neurons and their successful integration with a conductive bridge random access memory (CBRAM) crossbar arrays in hardware. The activation neuron implements the rectified linear unit function in the analogue domain. I will discuss system level performance gains enabled by activation devices and conclude my talk by presenting large-scale image edge detection using the Mott activation neurons integrated with a CBRAM crossbar array in hardware.

Superconducting, metal-to-insulator, and paraelectric-to-ferroelectric phase transitions often occur in condensed matter systems with electronic and lattice instabilities. These phase change materials and their transitions have emerged as a promising platform for next-generation computing technologies such as quantum computing, neuromorphic computing, and energy-efficient memory and storage^1-2. BaTiS₃ belongs to a ternary transition metal chalcogenide family with hexagonal symmetry³. For many years this material has been considered a trivial small bandgap semiconductor mainly due to the nominally unoccupied Ti 3d orbitals⁴. Here, we report the observation of a series of electronic phase transitions in single crystals of BaTiS₃ from electrical transport measurements. Two different phase transitions are identified from abrupt hysteric jumps in electrical resistance at 150-190 K (Transition I) and 245-255 K (Transition II) respectively, complemented by other characterizations such as synchrotron X-ray diffraction, optical spectroscopies, and DFT calculations. We also demonstrate reversible resistive switching with negative differential resistance (NDR) for Transition II, and a memristive switching between multiple metastable states within Transition I, both of which are key device ingredients for emerging neuromorphic computing circuits. Our study identifies a novel electron-lattice instability that leads to a series of electronic phase transitions in quasi-1D hexagonal chalcogenide with a strong potential for next-generation electronic applications.

1. Liu, M. et al. Terahertz-field-induced insulator-to-metal transition in vanadium dioxide metamaterial. Nature 487, 345-348, doi:10.1038/nature11231 (2012).
2. Chanthbouala, A. et al. A ferroelectric memristor. Nat Mater 11, 860-864, doi:10.1038/nmat3415 (2012).
3. Niu, S. et al. Giant optical anisotropy in a quasi-one-dimensional crystal. Nature Photonics 12, 392-396 (2018).
4. Massenet, O., et al. Magnetic and electrical properties of BaVS₃ and BaV_xTi_1-xS₃. Journal of Physics and Chemistry of Solid 40, 573-577 (1979).

The unique electrical characteristics and crystal structure of NbO₂ has attracted much interest in recent years as this material undergoes an insulator metal transition (IMT) between 750 and 800°C accompanied by a change from a distorted rutile to rutile structure [1]. This transition can be utilized in nanoscale devices such as selectors for high density memory arrays, in conjunction with novel two terminal non-volatile devices, and as temporal synapses in neuromorphic circuit architectures. For those future applications, fundamental understanding of the IMT must be better understood to tailor material switching properties. One way to gain additional insight into the IMT, is to dope our material and observe electrical and structural changes. For NbO₂, Ti is a prime candidate, it has a stable dioxide rutile structure and lacks a +5 oxidation state. In conjunction, these differences should lower the transition temperature as the combined lattice structure should be less favorable to the distorted rutile structure. Secondly, the Ti dopant should stabilize the metastable NbO₂ phase due to the lack of the +5 oxidation state. Nb_xTi_1-xO₂ samples were prepared via reactive sputtering in an oxygen deficient atmosphere resulting in phase pure films after a 750 °C annealing step in an inert N₂ environment. Ti content was adjusted from 0 to 30% to observe gradual changes in the crystal structure and electrical performance. Electrical results obtained from NbO₂ and Nb_0.9Ti_0.1O₂ via nanoscale TiN contacts (50x50 nm²) and an Ir top electrode show bulk resistive switching without the requirement for an electroforming step. This approach yielded reliable devices switching more than 10 million times in an operational range between -1.5 and 1.5 V. A reduction in the switching threshold voltage and threshold variation was confirmed with these two concentrations and further substantiated by Nath et. al [1] in an unrelated effort. Bulk switching was confirmed during a thermal anneal with in-situ XRD at the Canadian Light Source (CLS) showing a crystallographic change at around 750 °C accompanied by a sheet resistance change. Furthermore, structural characterization was performed on the full range of Ti doped NbO₂ samples. XRD results show a gradual shift from the low symmetry space group I41/a:2 of NbO₂ to the higher symmetry space group P42/mnm related to NbTiO₄. Most recently, XAFS data of the K-Nb and K-Ti edge were collected and the NbO₂ structure was confirmed via EXAFS modelling with Artemis and FEFF. The remaining data will be used to extract the Nb-Nb (Nb-Ti) dimer distance which is anticipated to be an indicator for the electrically obtained threshold voltage. Furthermore, additional electrical samples will be prepared matching the film properties of the XAFS samples to yield a more detailed picture of the threshold voltage and dimer distance correlation.

[1] Rana, R., Klopf, J. M., Grenzer, J., Schneider, H., Helm, M., & Pashkin, A. (2019). Nonthermal nature of photoinduced insulator-to-metal transition in NbO2. Physical Review B, 99(4). https://doi.org/10.1103/PhysRevB.99.041102
[2] Nath, S. K., Nandi, S. K., Ratcliff, T., & Elliman, R. G. (2021). Engineering the Threshold Switching Response of Nb2O5-Based Memristors by Ti Doping. ACS Applied Materials & Interfaces, 13(2), 2845–2852. https://doi.org/10.1021/ACSAMI.0C19544

Certain correlated oxides feature an insulator-to-metal transition which can be triggered by applying an external voltage: the material becomes conducting if a threshold electric field is exceeded [1]. This phenomenon is known as voltage-driven IMT, and it has very promising applications in emerging technologies such as optoelectronics and neuromorphic computing. While it is known that this process takes place in a filamentary way, it is not yet known what sets the characteristic lengthscales, or the dynamic path through which these filaments nucleate, grow and relax. We combine reflectivity and transport measurements to image metallization with spatial and temporal resolution [2]. Five systems featuring an IMT from two different families are analyzed: VO2, V2O3, V3O5, NdNiO3 and SmNiO3 [3]. By comparing these systems and with further insight from numerical simulations, we identify the key parameters that govern the dynamics of the voltage-driven IMT, presenting a unified and simple interpretation of the transition dynamics.
References:
[1] J. del Valle et al. Nature 569, 388 (2019).
[2] J. del Valle et al. Science 373, 907 (2021).
[3] J. del Valle et al. Phys. Rev. B 104, 165141 (2021).

Vanadium dioxide is an excellent candidate for a material for neuromorphic devices, already having successfully imitated many neurological functions such as tonic spiking, phasic spiking, and tonic bursting in simple circuits consisting of two VO₂ devices .^[1] This characteristic behavior is possible due to an abrupt negative differential resistance (NDR) associated with the material’s metal-insulator transition (MIT), which is associated with a symmetry-breaking first-order phase transition from a high-temperature rutile phase to a low temperature monoclinic phase. Much prior research has focused on understanding the nature of the temperature- or current-driven MIT, including modulating the onset temperature, or the hysteresis of the transition through different processing vectors including chemical doping, external strain, or epitaxial strain. However, in addition to the first-order phase transformation, NDR can also be derived from other internal instabilities within a material, which result in transitions from homogeneous current density throughout a film to localized domains of high current density.^[2]
Recent work has found that during a quasi-dc current-driven transition, a stable electro-thermal localization, which is a region within the device material that is stable at a higher temperature and with higher current density than the surrounding material, is present under conditions of forced current prior to the onset of the MIT. Under an applied dc current, this is evidenced both by an NDR region in the device’s I-V trace and an obvious thermal localization in in situ temperature maps of the device under conditions of applied current. However, the driving factors behind this electro-thermal localization are unknown. The objective of this study is to investigate the origin of stable electro-thermal localizations in VO₂ devices under conditions of forced voltage and current. Here we demonstrate that electrothermal localization can be caused by an initial small temperature perturbation, or local variation in material properties (e.g., electrical resistance), and is only possible when a specific combination of the materials properties exist that cause additional heating of a local volume to result in further intensification of current flow and local heat generation. This condition can be further clarified as a competition between heat lost through the device’s substrate, the magnitude of which is dictated by a net thermal boundary resistance , and the heat generated within the localized region according to Joule heating, which is a function of the net applied current. A 3D multiphysics model was used to carry out a parametric analysis of the localization of charge and heat in a VO₂ device. It was found that critical parameters for electro-thermal localization included an initial temperature perturbation of at least 0.1 K that occupied a volume of at least 100³ nm³ when the thermal conductivity of the substrate was set between 0.5 Wm^-1K^-1 and 10 Wm^-1K^-1. For the variety of conditions tested, the critical current for localization was between 0.5 – 2 mA. This research provides information on the mechanism of electro-thermal localization, which at least in some cases, precedes, and can serve to nucleate the MIT and its associated structural phase transition. Furthermore, understanding of the origins and control vectors for different NDR phenomena could potentially form the basis for future devices.
[1] W. Yi, K. K. Tsang, S. K. Lam, X. Bai, J. A. Crowell, and E. A. Flores, "Biological plausibility and stochasticity in scalable VO2 active memristor neurons," Nature Communications, vol. 9, no. 1, p. 4661, 2018/11/07 2018.
[2] B. K. Ridley, "Specific Negative Resistance in Solids," Proceedings of the Physical Society, vol. 82, no. 6, pp. 954-966, 1963/12 1963.

Memristors, resistive memory devices, have gained popularity in numerous applications, from nonvolatile memory to artificial neural networks. They present an opportunity for massive energy savings, as well as improvements in physical density. That density is commonly achieved with crossbar arrays of two-terminal memristors. An abundance of research has demonstrated synaptic behavior from transistors and memristors, but neuronal behavior (i.e. spiking) remains underexplored. Recent work has demonstrated artificial neurons based on locally active memristors (operating in the negative differential resistance region) using VO₂, which has a Mott insulator-metal transition (IMT) at 340 K. The IMT of a memristor material is ideally above room temperature so that the circuit does not require cooling and unselected devices in an array are insensitive to noise, but alternative materials such as NbO₂ possess too high of an IMT, 1040 K, resulting in slower switching speeds and higher energy use.
Rare-earth nickelates (RNOs) are perovskite oxides in which a rare-earth ion sits on the A site. Apart from LaNiO₃, RNOs exhibit a charge-transfer IMT at temperatures spanning 100-600 K. The transition temperature varies with rare-earth cation and strain, among other things, making RNOs an intriguing class of materials for spiking neuronal behavior. In our work, we use epitaxial thin-film growth to stabilize RNO films and, along with rare-earth cation alloying, to control the IMT transition temperature.
In addition, we aim to demonstrate on-chip spiking, without external RC elements, using the capacitance of the insulating active layer that surrounds the selected device, rather than an external capacitor. We use a physics-based model that relates the Joule heating of a device to the thermodynamics of its IMT to determine the amount of metallic material in an active device at a given moment.The size of that metallic channel controls the device’s conductivity. Using LT-SPICE as a non-linear solver for these governing equations, we model a single memristoras part of an oscillator circuit. As an initial model, we focus on a Pearson-Anson oscillator circuit topology because it is commonly used to describe a biological neuron ion channel.
Our modeled devices consist of LaNiO₃ grown on an insulating crystalline substrate and etched into read lines, then an active layer of RNO alloy, and metallic write lines. The experimental RNO layers are grown as single-orientation thin films by physical vapor deposition, and the properties of preliminary growths are used in the SPICE modeling. For example, we measure an average sheet resistance of our LaNiO₃ films of 4e-4 Ω-cm and can use this value to design the resistive component of the RC oscillator. In general, we find that the design space is limited by the higher line resistance of LaNiO₃ contacts and lower capacitance of the RNO active layer, compared to previous studies using metallic contacts and external components in the oscillator circuit. However, with careful geometric design all on-chip spiking is achievable in this system.
In conclusion, we describe the first round co-design of active memristors for biomimetic neurons fabricated from RNOs. First, we take measured parameters of test films to model the proposed oscillator circuit. Using physical governing equations, we construct a SPICE model of a single oscillator representing a single on-chip spiking device. Future rounds of design will consider the capabilities of nanoimprint lithography, which is very sensitive to feature size and fill factor, when determining geometric parameters such as crossbar widths, to ultimately produce an array of devices displaying biological spiking behavior.

Vanadium oxides (VO_x) are of great interest for electrical, optical, and energy applications due to their wide spectrum of physical and chemical properties. We separately stabilize various VO_x phases by the epitaxial growth technique and investigate the electrical properties of each phase. By carefully selecting the crystal orientation of Y-stabilised ZrO₂(YSZ) substrates and precisely controlling oxygen partial pressure below 500°C, we grow pure phases of VO₂(M1), VO₂(A), VO₂(B), V₆O₁₃, and V₂O₅ epitaxial films [1, 2]. We find that VO₂(M1) and V₆O₁₃ display insulator-to-metal transition upon heating at 68°C and -123°C, respectively, while VO₂(A) and V₂O₅ are insulating and VO₂(B) is semiconducting. In addition, we observe nanoisland formation above 500°C, and propose that this degradation is due to rapid diffusion of V and Y [3]. Our newest growth phase diagram and better understanding of the degradation mechanismwill promote VO_x applications for next-generation electronic and energy devices.
This work was supported by the Defense Acquisition Program Administration and the Agency for Defense Development of Korea (project no.911223001).
[1] Songhee Choi et al., Adv. Electron. Mater. 4, 1700620 (2018)
[2] Songhee Choi et al., Phys. Rev. Mater. 3, 063401 (2019)
[3] Songhee Choi et al., Adv. Eng. Mater. 21, 1900918 (2019)

Silicon compatible ferroelectric Hf_0.5Zr_0.5O₂ (HZO) [1] opens new opportunities for single-layer ferroelectric tunnel junction (FTJ)[2] non volatile memories (NVM) integrated with CMOS. Semiconductors are good candidates for the bottom electrode [3] since they allow for partial compentation of the polarization charge, consequently the efficient modulation of the potential barrier necessary for high tunneling electroresistance (TER).
Using Ge [4] and ~ 0.7% Nb-doped SrTiO₃ (NSTO) [3] semiconductor substrates, amorphous HZO is prepared with plasma assisted molecular beam deposition [4] at 120 ⁰C, followed by in-situ deposition of TiN or W top electrodes. Crystallization annealing is performed at 400 ⁰C - 550 ⁰C for 45 -400 sec.
Robust ferroelectricity is obtained after wake-up on capacitors with 5 nm HZO showing coercive voltage V_c below 1V and remanent polarization 10-20 μC/cm². Ferroelectricity persists in HZO as thin as 4 nm albeit with reduced P_r ~ 5 μC/cm². Best FTJ results are obtained for W/HZO/NSTO FTJs showing TER > 4 at a read voltage of 0.4 V. The memory window (current vs programming voltage) is from -0.7 V to +1.3V which makes these devices suitable for low votage/low power applications. Despite the relatively low ON current of ~5μA/cm² and conductance ~ 0.1nS, the devices show more than 32 stable non-volatile intermediate states between conductance extrema corresponding to 5-bit memory cell. Synaptic plasticity in the form of long term potentiation (LTP) and depression (LTD) of the conductance synaptic weights is obtained by applying consecutive pulses of varying amplitude (1 to 1.5V for potentiation, -0.4 to -1.2 V for depression) or varying width (1μs-10ms). The LTP and LTD curves for varying amplitude are nearly symmetric and linear as required for analog in-memory accelerators with good inference accuracy. The ON and OFF states have an endurance of >10⁵ cycles with cycling field of 1.6 V. The ON, OFF and all intermediate states show no retention loss for >10⁴ sec. The low programming voltage and the good retention present advantages compared to the double-layer HZO/AlOx gate stacks [5] more often reported in the literature.
For a deeper understanding of the mechanism giving rise to TER, temperature dependent I-V measurements up to 100 ⁰C are performed. A weak temperature dependence in both ON and OFF state current is observed indicating a thermal activation process with a low activation barrier of the order of 0.1 V. This is an indication of thermal injection over the Schottky barrier formed at the HZO/semiconductor interface which is modulated by the ferroelectric polarization. Appropriate Schottky barrier engineering by changing the Nb doping and the top metal workfunction could enhance the TER and the ON current.
Given the high quality epitaxial STO on Si(100) [6], an STO/NSTO multilayer can be epitaxially grown on Si, then by etching and ALD conformal deposition of HZO/metal, it is possible to define a number of vertically stacked side wall HZO/NSTO FTJs. This can lead to a large and dense 3D cross bar array integrated on Si. Our FTJs operating at low read and write voltage (<1 V) and low output current are ideally suited for the implementation of dense analog in-memory accelerators for ultralow power massively parallel data processing.
We acknowledge financial support from the European Union through the H2020 project BeFerroSynaptic -contract No. 871737.
References
1. J. Müller et al, Appl. Phys. Lett. 99, 112901 (2011)
2. Wen, Z., Wu, Adv. Mater. 32, 1904123 (2020)
3. Xi, Z., Ruan, J., Li, C. et al, Nat. Commun. 8, 15217 (2017)
4. C. Zacharaki, Appl. Phys. Lett. 117, 212905 (2020)
5. B. Max et al, IEEE J. Electron Device Soc. 7, 1175-1181 (2019)
6. McKee R. A. et al, Phys. Rev. Lett. 81 3014 (1998)

We report the growth of smooth AlN and AlScN thin films on Si(111) using reactive sputtering. X-ray diffraction shows that both AlN and AlScN films are epitaxial single crystals with a [0001] out-of-plane orientation. We report a suite of characterization results, including X-ray reflectivity, atomic force microscopy, Raman spectroscopy, and instrumented nanoindentation. Our study of conductivity switching in these materials is motivated by recent interests in AlScN for ferroelectric random access memory applications. We found that applying electric fields in excess of 2×10⁶ V/cm resulted in more than one thousand-fold increase in conductivity. However, applying an electric field with a similar magnitude but the reverse polarity caused the conductivity to revert to the initial state. In AlN, this effect was observed stably in excess of 100 cycles, and in AlScN, we report up to 10 manual switching cycles. Epitaxial AlScN/Si offers the opportunity for epitaxial integration of resistive and ferroelectric random access memory on Si(111) with the possibility of additional epitaxial layers.

Recently, neuromorphic hardware accelerators for image classification or speech recognition have attracted attention owing to their energy-efficient big-data processing capability compared to conventional computing hardware based on von-Neumann architecture. Emerging nonvolatile memory devices such as resistive switching memory, phase change memory, and ionic transistors are potential candidates for these hardware accelerators because of their analog conductance modulation characteristics. However, diverse requirements such as low cycle-to-cycle variation, high linearity/symmetry, and high endurance must still be fulfilled for realizing high-performance hardware accelerators. Here, we demonstrate ferroelectric thin-film transistors (FeTFTs) with nanoscale ferroelectric materials and oxide semiconductors. By using oxide semiconductor as the channel layer to create an interfacial layer-free metal-ferroelectric-semiconductor stack, our FeTFTs exhibit analog conductance modulation with exceptionally high endurance. Furthermore, linear and symmetric conductance modulation characteristics are achieved by precisely controlling the polarization state of the ferroelectric layer. Simulations performed based on measured properties show that a neuromorphic system with FeTFTs achieves high recognition accuracy for handwritten digits, which is close to the recognition accuracy achieved using ideal synapses. This study thus provides method of realizing neuromorphic hardware systems based on FeTFTs as synaptic devices.

Gas sensors for hydrogen, oxygen and hydrocarbons are key components for a broad range of applications including battery-powered smart sensorics. Thin-film technology allows for mass-fabrication of low-cost gas-sensors with small device footprint. However, conventional micro- to nanoscale sensors show insufficient device robustness, as well as degradation and drift-effects. We introduce the concept of memristive gas sensors, in which the critical sensing part can be re-programmed using external electrical stimuli.
In our previous work we already analyzed the interaction of ambient oxygen, moisture and effects of incorporation of hydroxyl-ions with memristive devices. Here, we report on the theoretical and experimental findings of the prototypic memristive gas-sensitive devices. We investigated complex nanoionic phenomena by probing the devices’ electrical and physico-chemical behavior in controlled environments. We further discuss how degradation and drift phenomena are compensated by re-programming schemes. Due to the micro- to nanoscale sensing volume fast response times can be achieved. Further advantages of our approach are ultra-low power consumption, and CMOS-friendly integration.

Multiplexing is essential for technologies that require processing of a large amount of information in real time. Here, we present an artificial synaptic multiplexing unit capable of realizing parallel multi-input control system. Ion-gel was used as a dielectric layer of the artificial synaptic multiplexing unit because of its ionic property, allowing multi-gating for parallel input. A closed-loop control system that enables multi-input-based feedback for actuator bending control was realized by incorporating an ion-gel-based artificial synaptic multiplexing unit, actuator, and a bending angle sensor. The proposed multi-input control system could simultaneously process input and feedback signals, offering a breakthrough in industries in which the processing of vast amounts of streaming data is essential.

The success of artificial neural networks (ANN) in machine vision techniques drives hardware researchers to explore more efficient computing elements for energy-expensive operations like vector-matrix multiplication (VMM). But the energy consumption and preprocessing time required for capturing the digitalized image are seldom considered. In this work, the InP-based floating-gate photo-field-effective transistors (FG-PFETs) are demonstrated as the promising computing element at the sensor level and enable optical signal sensing and processing simultaneously. Simulated optical neural network (ONN) constructed from the measured performance of FG-PFETs indicates the high image recognition accuracy (>94%) for color-mixed MNIST handwritten digits. And the heterogeneous gate dielectric structure allows the devices to work offline after training. Notably, the back-end CMOS compatible processes were implemented for the device integration, which paved the way for FG-PFETs as competitive candidates for energy-efficient machine vision.

Recently, the hardware-based neuromorphic chip is considered as the promising solution to realize the synaptic behaviors for complete intelligent process. In particular, neuromorphic photonics has been researched for promising solutions capable of high bandwidth, low crosstalk, high density and low power consumption compared to the neuromorphic device with an electrical domain, such as complementary metal-oxide-semiconductor (CMOS) integrated circuit. The recent research for neuromorphic photonics mainly underlies the persistent photoconductivity behavior (PPC) effect by using the various methodologies such as Schottky barrier, charge trapping, electrochemical doping, film reflectivity change and point defect ionization. By using these approach, photonic and optoelectronic neuromorphic devices with biological synaptic functions can be achieved. Particularly, two-connected indium-gallium-zinc-oxide (IGZO) devices, one of the amorphous metal oxide semiconductors, showed photonic synaptic behaviors using ionization of oxygen vacancies and PPC effect. However, since the light has only positive power, the optoelectronic neuromorphic device with an optical domain barely demonstrated the complete neuromorphic behaviors. In this regard, optoelectronic neuromorphic device capable of bidirectional synaptic modulation by all-photonic stimulation, high-density scalability, wavelength fidelity and the CMOS process compatibility are considered as the key factors to obtain the viability and efficacy of optoelectronic neuromorphic systems.
Here, we propose a large-area pixelized optoelectronic neuromorphic system realized by multispectral light-stimulated bidirectional optoelectronic neuromorphic pixel circuits and demonstrated the hardware-based pattern training capability in an array level. In particular, the bidirectional synaptic modulation is achieved by the optoelectronic neuromorphic pixel circuit comprising metal-oxide (MO) and metal-chalcogenide (MC)-based synaptic transistor and photovoltaic divider. It is performed by adjusting electron trapping or de-trapping in the MC/MO hetero-interface of the photo transistor with the assistance of a photovoltaic divider. This optoelectronic neuromorphic pixel circuit successfully emulated the biological synaptic functions such as short-term plasticity, long-term potentiation, long-term depression and paired pulse facilitation/depression via multi-spectral light selectivity. To confirm the viability of hardware-based optoelectronic neuromorphic computing in an array level, pattern training in pixelized device array was demonstrated using multi-spectral light pattern irradiation. In both the empirical pixel training in the optoelectronic neuromorphic array and the pattern recognition simulation, the optoelectronic neuromorphic array operated with bidirectional synaptic modulation exhibited higher recognition rates (up to 97%), lower mismatched pixel along with fast speed (less epochs) than those with unidirectional synaptic modulation. Furthermore, optoelectronic neuromorphic device based on the combination of two different transition metal dichalcogenide (TMDC) also implemented bidirectional synaptic modulation.

The ability to form and retain memory storing the input history of external stimuli is an essential characteristic required for synthetic neuromechanical systems. Associative memory or the ability to recall events from partial information can become useful for engineering systems as a mechanism for triggering high-level control actions. Specifically, the formation and retention of spatially distributed inputs on soft material systems constitute a new frontier in the development of engineered materials with potential applications in robotics, morphing structures, and wearables.
We leverage multistable metamaterial architectures for collocating sensing of environmental signals via mechanical coupling with memory formation. Concretely, we collocate piezoresistive sensors with history-dependent deformation with locally bistable units that introduce amplified strains when inverted by an external input. This input-accumulation mechanism and the spatial distribution of such units over an area allow us to form memories of the inputs that cause the unit inversions. The nonlinear nature of the units' deformation results in time-dependent piezoresistance that shows a cumulative nature with the number of dome inversions. This deformation-driven mechanism yields a type of associative memory stored in physically embodied Hopfield networks that use the accumulated piezoresistance changes to form interaction matrices storing successive spatial inputs. This metamaterial architecture thus displays collocated mechanosensing and memory formation of successive mechanical inputs without recourse to electronic devices, offering a blueprint for neuromorphic metamaterials leveraging nonlinear deformation.

Neuromorphic, or event-based cameras, are of great interest for their potential applications ranging from space domain awareness to autonomous driving in which the operation is entirely different from traditional frame-based counterparts. The independent and asynchronous changes that are detected by the pixels in these sensors can afford greater dynamic range while maintaining low latency compared to synchronous frame-rate sensors. These benefits are especially attractive to any object detection and tracking applications by alleviating the motion blur, speed limitations and data requirements for the algorithms normally associated with frame-based sensors. Here, we look at the latest offerings from Prophesee and Inivation commercial-off-the-shelf event-based cameras and report our progress towards developing a suite of standard characteristics to compare the technology to standard frame-based sensors.

The rapid development of artificial intelligence has raised critical challenges for conventional computing platforms based on digital circuits and von Neumann architecture to process the explosively growing amount of data. To addressed those challenges, computing-in-memory (CIM) with memristors has emerged as a promising neuromorphic computing paradigm to build intelligent computing chips with orders of magnitudes higher power efficiency than CPU and GPU. In this talk, the principles of memristor-based CIM technologies will be firstly introduced. Recent works on memristor material exploration and device integration for CIM will be covered, including strategies to improve the analog resistive switching performance. After that, a device-to-system co-design framework along with recent hardware implementations will be presented, including architecture-algorithm co-optimization in consideration of the nonideal device characteristics of memristors. Finally, a perspective on the future development of CIM technologies will be provided.

As semiconductor technology enters the more than Moore era, there exists an apparent contradiction between the rapidly growing demands for data processing and the visible inefficiency rooted in traditional computing architecture. Neuromorphic systems hold great prospect in enabling a new-generation of computing paradigm that can address this issue, which demands device components with rich dynamics and nonlinearity. In this talk, we demonstrate optoelectronic synapse based on van der Waals ferroelectric α-In₂Se₃ with controllable temporal dynamics under both electrical and optical stimuli, and the tight coupling between the two distinct physical processes gives rise to heterosynaptic plasticity with light tunable relaxation timescale. A multimode reservoir computing system with adjustable nonlinear transformation function is in turn demonstrated, showing enhanced network performance, and the tunable timescale in α-In₂Se₃ synapse enables multiscale signal processing in parallel sub-reservoirs with significantly improved computing accuracy. The inherent conductance drift of phase change memory is exploited as physical decay function to realize in-memory eligibility trace, demonstrating excellent performance during reinforcement learning in various tasks. The spontaneous in-memory decay computing and storage of policy in the same phase change memory give rise to significantly enhanced energy efficiency compared with traditional graphics processing unit platforms. Incorporating dynamics in memristors represents the future way to empower computing.

Memristors are one of the strong candidates for synaptic devices. Transition metal oxides such as titanium oxide (TiO₂) and tantalum oxide (Ta₂O₅) are broadly used as materials of memristors. However, these transition-metal-oxide-based memristors often show an abrupt decrease in the cell resistance (set process), resulting in the difficulty of the analog control of the cell resistance [1]. Although some studies have reported gradual set processes, the mechanism of the gradual set processes has not been clarified [2]. Therefore, any guidelines to improve the analog resistive switching (RS) characteristics, including reduction of power consumption, have not been established. In our previous study, analog RS characteristics were demonstrated at both set and reset processes after a semi-forming phenomenon, which is defined as an initial transition process to a higher resistance state (~ 1 kΩ) than the conventional low resistance state (< 100 Ω) [3], in Ta₂O₅-based RS cells with tantalum sub-oxide (TaO_x) as an oxygen vacancy (V_O) reservoir. However, there is an inevitable trade-off relationship between the current range of analog RS operation and the occurrence frequency of the semi-forming phenomenon, as long as the oxygen composition of the TaO_x layer is widely varied. In this study, the trade-off relationship was optimized by adjusting the oxygen composition and the TaO_x thickness.
Platinum (Pt)/TaO_x/Ta₂O₅/Pt stack cells were fabricated on a SiO₂/Si substrate. A 60-nm-thick Pt bottom electrode (BE) was deposited by DC sputtering. Then, a 10-nm-thick Ta₂O₅ layer and a TaO_x layer with different TaO_x thicknesses (10 nm, 30 nm, and 60 nm) were deposited by reactive RF sputtering. The oxygen compositions of the TaO_x and Ta₂O₅ layers were controlled by appropriately adjusting the oxygen gas flow rate. (TaO_x: 0.8–1.2 sccm, Ta₂O₅: 2.0 sccm) Finally, 25-nm-thick Pt top electrodes (TE) with a diameter of 100 μm were formed by electron beam evaporation. Forming and RS operations in the Pt/TaO_x/Ta₂O₅/Pt cells at room temperature were characterized.
We investigated the dependences of initial resistance (R_ini) and relative occurrence frequency of the semi-forming phenomenon (F_semi) on the oxygen composition x and the TaO_x thickness. When the TaO_x thickness is 30 nm, as x linearly increases from 1.3 to 2.0, R_ini exponentially increases from 1 kΩ to 1 GΩ and F_semi tends to decrease from 0.8 to 0. The increase in x means the decrease in the density of V_Os supplied from the TaO_x layer to the Ta₂O₅ layer. In other words, while the semi-forming phenomenon is more likely to occur as more V_Os are supplied to the Ta₂O₅ layer, the more supplied V_Os cause the decrease in R_ini, which means a trade-off relationship. Moreover, compared with the analog RS characteristics when the oxygen composition x is 1.3 and 1.7, it was revealed that the decrease in R_ini leads to lowering the resistance range of RS operation from 6–40 kΩ to 0.1–1 kΩ. Furthermore, the dependence of R_ini and F_semi on the TaO_x thickness was investigated. Under the oxygen gas flow rate of 1.0 sccm, corresponding to the oxygen composition x of 1.7, when the TaO_x thickness is 10 nm, 30 nm, and 60 nm, R_ini is 30 MΩ, 2 MΩ, and 1 MΩ, and F_semi is 0.1, 0.6, 0.9, respectively. These results indicate that F_semi increases while keeping high R_ini and resistance range of analog RS operations by increasing the TaO_x thickness, that is, the trade-off relationship can be improved. Analog RS characteristics could be demonstrated in the resistance range of 3–10 kΩ also when x is 1.7 and the TaO_x thickness is 60 nm.
In summary, by adjusting the TaO_x thickness and the oxygen composition x appropriately, the current range of analog RS after the semi-forming can be lowered for reduction of power consumption.
[1] D. Ielmini, Microelectronic Engineering, 190, 44 (2018).
[2] R. Waser et al., Faraday Discuss. 213, 11 (2019).
[3] T. Miyatani et al., 2020 virtual MRS spring/fall meeting, F. NM07. 06. 06.

Resistive random-access memory is one of the next-generation nonvolatile memories. A resistive switching (RS) phenomenon offers promising applications for not only high-performance nonvolatile memories but also synaptic devices for neuromorphic computing. The RS phenomenon in a RS cell composed of a transition-metal-oxide is often regarded as nonvolatile transitions in the cell resistance between high- and low- resistance states by formation and dissolution of localized conductive filaments. The trigger required for the RS phenomenon in RS cells using a binary oxide such as a nickel oxide (NiO) and titanium dioxide (TiO₂) is an abrupt reduction in resistance called forming. We reported correlation between a crystalline structure of the oxide layer and forming characteristics in the oxide-based RS cells [1,2]. In this study, dependence of the forming characteristics on ambient temperature was investigated under constant voltage stresses to NiO-based RS cells.
A 60-nm-thick platinum (Pt) bottom electrode was deposited by sputtering on a silicon-dioxide/p-type silicon substrate. Next, a polycrystalline 60-nm-thick NiO with a columnar structure was deposited by radio-frequency reactive sputtering under oxygen gas flow rate precisely regulated. Pt top electrodes with thickness of 25 nm were subsequently deposited on the NiO layer through a metal mask. The Pt/NiO/Pt cell area is defined as the top electrode area in the range of 100-300 μm. In all electrical measurements, the Pt bottom electrode was grounded, and the constant voltage was applied to the Pt top electrode. Time to forming in the initial state was measured under constant voltages using an oscilloscope and external resistor of 50 Ω.
Weibull slopes of time to forming were similar with each other in the cells with different cell areas at 290 K. In addition, these Weibull distributions normalized according to the area scaling law fell in the same line, which agrees with our previous reports [1,2]. The results indicate that the formation of conductive filaments at the forming follows a weakest link theory, and that weakest spots are randomly distributed in the NiO film according to the Poisson statistics. Moreover, time dependences of cell resistance with a diameter of 150 μm were investigated at several ambient temperature. Although cell resistance remained unchanged before the forming at 290 K under constant voltage stress of 4.5 V, gradual reduction phenomena in resistance were observed before the forming at a high temperature of 400 K. We also found the voltage dependence of this gradual reduction, which means gradual formation of conductive filaments. The slope of the resistance reduction decreased, and the gradual reduction became suppressed as the voltage decreased. From this result, a percolation model is proposed for breaking the insulating oxide layer under stress due to not only higher temperature but also larger voltage. The formation of defects, that is, the formation of conductive filaments occurs continuously rather than abruptly.
The formation mechanism of the filaments is a key to uncover an origin of the RS phenomenon. The restriction of positive feedback to create the filament by elevating temperature is quite interesting. We also observed multi-step formation of the filament by a thermal effect, which will be shown in our presentation.
[1] Y. Nishi et al., J. Appl. Phys. 120, 115308 (2016).
[2] M. Arahata et al., AIP Adv. 8, 125010 (2018).

A paradox in neuroscience is the large number of oscillations of small neural networks compared with the few oscillations observed in the conscious brain. It remains unclear what is the maximum number of synchronized oscillations a network can support and whether all or some of these oscillations would survive in noisy heterogenous circuits. Here, we attempt to answer these questions through a comprehensive study of local inhibitory networks of Hodgkin-Huxley neurons. We use a neuromorphic platform combining electronic noise and device-specific heterogeneity with tuneable extrinsic noise, tuneable network connectivity, and controlled initial conditions. As in the brain, stimuli are instantaneously integrated by analog circuits. This gives us the computing power needed to map the network dynamics over the entire phase space and demonstrate the full complement of limit cycles, basins of attraction, and dependence on network parameters. Our main finding is that the maximum number of limit cycles is equal to the combinatorial number of activation pathways through the network, allowing for coincident action potentials, and that all limit cycles are equally robust to noise and mild heterogeneity but highly dependent on inhibition delay and the timing of stimuli. We established the robustness of individual limit cycles by computing the detailed balance of bifurcations between attractors. This accounts for all transitions in a system where Lyapunov exponents are both positive and negative depending on phase space coordinates and noise intensity. Another interesting finding is the unexpected resilience of limit cycles to mild network heterogeneity. This occurs as stochastic processes recruit quiescent neurons whose subthreshold periodic oscillations help maintain the synchronization of limit cycles against heterogeneity.
A. S. Chauhan, J. D. Taylor, A. Nogaret, Phys. Rev. Res. 3, 043097 (2021)
Abu-Hassan et al., Nature Communications 10, 5309 (2019)
A. S. Chauhan et al, Sci. Rep. 8, 11431 (2018)

Non-volatile memory-based synaptic device and array technologies came into the spotlight due to its feasibility for accelerated AI computations in analog domain. Vigorous research has been conducted to implement analog neural network training accelerators with resistive cross-point arrays. Despite the theoretically promised benefits, such as computation speed-up, area- and power-efficiency, non-ideal switching characteristics of resistive memory devices hinder the realization of such analog accelerators. Especially in neural network training applications, non-idealities such as conductance update asymmetry and retention characteristic are known to critically impact the convergence of the training and performance. To resolve the issue, Tiki-Taka algorithm, in which an auxiliary array, in addition to the main array, is introduced to form a coupled dynamical system, was proposed to minimize the negative impact of asymmetric conductance update [1]. In this work, we study the impact of update asymmetry and retention in two coupled arrays when training neural network with Tiki-Taka algorithm. We carefully inspect the training process and show that these non-idealities in each array affect differently on how the history of gradient information is stored and evolved during the training. By exploring the asymmetry, retention and other network hyper-parameters, we reveal the optimal range of device parameters for two arrays, which guarantees near-software performance with Tiki-Taka algorithm. Our study demonstrates the possibility of significant relaxation in synaptic device specifications to realize high-performance analog neural network training accelerators with practical resistive switching devices [2].

[1] Gokmen, T., & Haensch, W. (2020). Algorithm for training neural networks on resistive device arrays. Frontiers in neuroscience, 14, 103. [2] Lee, C. et al. (2021). Impact of Asymmetric Weight Update on Neural Network Training with Tiki-Taka Algorithm. Frontiers in neuroscience, in press.

Next generation of energy storage may largely benefit from fast Li⁺ ceramic electrolyte conductors to allow for safe and efficient batteries. With recent discoveries in thin film processing solid state lithium ion conductors, such as Li-garnets and LIPON or LiSICON-based solids, have been recently considered as candidate materials not only for next-generation solid-state batteries but also for neuromorphic computing via memristors owing to the fast ionic transport in the solid state electrolyte.
In the first part of this talk, we review various Li solid-state conductor materials and reflect on opportunities of thin film processing, being a requirement to define precisely the Lithium stoichiometries and related electronic state changes for transition metal ions, miniaturize the device, and reach high energy/information densities for neuromorphic computation via resistive switching.
In the second part, we focus on Li-film processing and controlling Lithium stoichiometries to reach fast conductive phases for Li garnets and Li titanates as solid state battery components, and memristive neuromorphic computing units. Insights on structure-phase-transport interaction and implications on performances will be exemplified for energy storage aiming high energy densities, and modulations of synaptic artificial weights through lithium induced metal-to-insulator transitions in lithium titanate memristors.

References for further read:

1. Balaish*, M.; Gonzalez-Rosillo*, J.C.; Kim, K.J.; Zhu, Y.; Hood, Z.D.; Rupp, J.L.M.,
“Processing thin but robust electrolytes for solid state batteries”,
Nature Energy, 6, 227–239 (2021)

2. Zhu, Y.; Gonzalez-Rosillo, J.C.; Balaish, M.; Hood, Z.D.; Kim, K.J.; Rupp, J.L.M.,
“Lithium-film ceramics for solid state lithionic devices”,
Nature Review Materials, 6, 313–331 (2020)

4. Kim K.J.; Rupp, J.L.M.,
“All ceramic cathode composite design and manufacturing towards low interfacial resistance for garnet-based solid state lithium batteries”,
Energy & Environmental Science, 13, 4930-4945 (2020)

5. Kim, K.J.; Henricher, J.J.; Rupp, J.L.M.,
“High energy and long cycles”,
Nature Energy, 5, 278, (2020)

6. Gonzalez-Rosillo, J.C.; Balaish, M.; Hood, Z.D.; Nadkarni, N.; Fraggedakis, D.; Kim, K.J.; Mullin, K.M.; Pfenninger, R.; Bazant, M.Z.; Rupp, J.L.M.,
“Lithium-Battery Anode Gains Additional Functionality for Neuromorphic Computing through Metal-Insulator Phase Separation”,
Advanced Materials, 1907465, (2020)*

7. Mahbub, R.; Huang, K.; Jensen, Z.; Hood, Z.D.; Rupp, J.L.M.; Olivetti, E.A.,
“Text Mining for Processing Conditions of Solid State Battery Electrolytes”,
Electrochemistry Communications, 121, 106860 (2020)

8. K.J. Kim*, M. Balaish*, M. Wadaguchi, L. Kong, J.L.M Rupp
“Solid State Li–Metal Batteries: Challenges and Horizons of Oxide and Sulfide Solid Electrolytes and Their Interfaces”,
Advanced Energy Materials, 202002689 (2020)

9. R. Pfenninger, M. Struzik, I. Garbayo, E. Stilp, J.L.M. Rupp
“A low ride on processing temperature for fast lithium conduction in garnet solid state battery films”,
Nature Energy, 4, 475–4832019 (2019)

Iono-electronic materials and devices are suscitating lots of interest from both bio-electronics and neuromorphic research communities. In the one hand, iono-electronic materials are offering attractive features such as bio-compatibility, water environment operation and efficient ionic to electronic signals transduction. In the other hand, functional devices based on such materials (organic electrochemical transistors, for instance) have shown multiple neuromorphic features from synaptic plasticity to dendritic integration. This talk will present how electropolymerization of PEDOT:PSS materials can be used in both a bio-electronic and a neuromorphic perspective. Electropolymerization of OECT sensors can indeed be advantageously used for optimizing / tuning the iono-electronic responses of organic electrochemical transistors, thus paving the way to plastic electrophysiological sensors. Notably, we will show how transconductance and volumetric capacitance are evolving with potenstiostatic electropolymerization. Secondly, bipolar AC electropolymerization can be used to engineer dendritic-like fibers of PEDOT:PSS. Such unconventional structures can implement various neuromorphic concept such as structural plasticity and synaptic plasticity. Finally, computing task taking advantage of both electropolymerized sensors and neuromorphic computing will present temporal patterns classification with reservoir computing approach.

In our information-centric era, flooded by data, self-guided inference without human intervention is essential in vital fields such as intelligent healthcare or smart industry. However, machine-learning approaches based on algorithms fed by larger and larger amounts of data and longer and longer training routines can devour a prohibitively high amount of energy. This is in contrast with the current trend of energy saving for greener electronics and sustainable exploitation of Earth resources. Traditional computing machines rely on the von Neumann paradigm, where data storage and data processing are physically separated. This entails movement of data when analysis tasks are required, which is energetically costly. Thus, to reduce the power consumption of computing urges the design for beyond von Neumann computing approaches. In this respect, the human brain is a source of inspiration for smart solutions, being an example of how sophisticated computation may be accomplished with reduced power consumption. Neural systems can be represented as elemental units yielding periodic signals, interconnected into networks. The functions of such networks are intimately correlated to the emerging of collective behaviors, where oscillatory synchronization is one of the most important examples of collective dynamics. Thus, it is critical to assess the relationships between the features of the elemental units, the interaction strengths and the network structure to achieve synchronization, if any [1].
In this contribution, we propose the implementation of oscillatory neural networks (ONNs) with Beyond Complementary-Metal-Oxide-Semiconductor (CMOS) devices. We show experimental and simulation results of devices and oscillators based on vanadium dioxide (VO₂). VO₂ undergoes a transition from a high-resistive monoclinic crystal structure to a low-resistive tetragonal rutile-like one, triggered by temperature. However, a switch in resistance is observed as well in two-terminal devices when the VO₂ channel is flown by current, in this case the Joule effect being thought to drive the transition. This resistive switching is pivotal for implementing highly compact and scalable oscillators. We use dedicated technology computer-aided design (TCAD [2]) 3D electrothermal simulations of VO₂ oscillators [3, 4]. Finally, exploiting the mixed-mode approach [2], we combine electrothermal TCAD device simulations with SPICE circuit simulations to investigate the dynamics of coupled VO₂ oscillators. It is worth noting that this is the first time that mixed-mode simulations of the circuit dynamics of interacting VO₂ oscillators are shown, to the best of our knowledge. Such simulations are crucial for realistically associating the device’s behavior at the neural network, which is at the core of the ONN function as an analogue computing engine. Our findings give insights on the entangled thermal and electrical behavior of a single VO₂ oscillator as well as the synchronization behavior of networks of oscillators and help provide guidelines for the successful implementation of ONN technology.
Acknowledgments. Authors wish to thank Dr. S. Karg, IBM Research Europe, Zurich, Switzerland, for providing the experimental data used for the calibration of the TCAD model and the useful discussions about the experimental devices. Authors also wish to thank Dr. A. Nejim and Dr. A. Plews, of Silvaco Europe Ltd., Cambridgeshire, United Kingdom, for providing the customized version of PCM model [2] used to simulate the VO₂ material as well as for the useful discussions about the TCAD and mixed mode simulations.
[1] A. Todri-Sanial et al., IEEE Trans. Neural Netw. Learn. Syst., 2021. DOI: 10.1109/TNNLS.2021.3107771
[2] ”Victory Device User Manual”, version 1.19.1.C, Silvaco Inc
[3] E. Corti et al., Front. Neurosci., vol. 15, 2021. DOI: 10.3389/fnins.2021.628254
[4] S. Carapezzi et al., IEEE J. Emerg. Sel. Topics Circuits Syst. 11, 4, 2021. DOI: 10.1109/JETCAS.2021.3128756.

The (GeTe)_m(Sb₂Te₃)_n alloy (Ge-Sb-Te - GST) is a key Phase Change Material, widely studied for its cutting-edge technological applications. It is a prominent material for phase change memories, and, very recently, it has attracted new interest as the most advanced emerging non-volatile memory technology for neuromorphic applications [1-2]. The ability of growing high-quality epitaxial GST represents a key-point to tailor the material properties, such as a large programming window [3], improved cycling endurance and faster crystallization [4]. Depending on n and m, prototypical crystalline GST can have various compositions along the pseudobinary tie line and, in its trigonal structure, exhibits a stacked structure along the [111] direction that involve weak van der Waals (vdW) interactions between adjacent blocks.
vdW epitaxy represents a powerful way for growing heterostructures made of stacked sequences of 2D crystals, potentially exhibiting new phenomena and peculiar properties. In this study we aim at identifying the key parameters to predict the interaction between GST layered materials and substrate surface. In particular, we present the case of GST alloys grown by Molecular Beam Epitaxy on InAs(111). In this system the substrate can be efficiently prepared into self- and un-passivated surfaces to clarify the role of the surface interaction. Furthermore low-lattice mismatch conditions are fulfilled. Those are necessary to avoid relaxation due to formation of misfit dislocations and allow to correctly identify vdW epitaxy. It is known that GST exhibits two different highly ordered 2D structures and a three-dimensional disordered structure, allowing to directly infer the nature of the epitaxy. We will give evidence of the key role played by the substrate surface in vdW epitaxy. We will show that substrate symmetry influences the symmetry of the growing film and, for extremely well passivated covalent or pure 2D surfaces, vdW epitaxy can be fully achieved. When this requisite is relaxed an interaction with the substrate surface arises, which can lead to the formation of coincidence lattices or induce a quasi-pseudomorphic growth. In general, our study paves the way for mastering and design of vdW epitaxial growth of 2D heterostructures as well as hybrid 2D and non-layered materials.
This project has received funding from the European Union's Horizon 2020 research and innovation program under Grant Agreement No. 824957 (“BeforeHand:” Boosting Performance of Phase Change Devices by Hetero- and Nanostructure Material Design).
[1] A.Sebastian, M. Le Gallo, G.W.Burr, S.Kim, M.BrightSky, and E.Eleftheriou, J. Appl. Phys. 124, 111101 (2018)..
[2] W.Zhang, R.Mazzarello, M.Wuttig, and E.Ma, Nat. Rev. Mater. 4, 150 (2019).
[3] V.Bragaglia, F.Arciprete, W.Zhang, A.M.Mio, E.Zallo, K.Perumal, A.Giussani, S.Cecchi, J.E.Boschker, H.Riechert, S.Privitera, E.Rimini, R.Mazzarello, and R.Calarco, Sci. Rep. 6, 23843 (2016).
[4] K.Ding, J.Wang, Y.Zhou, H.Tian, L.Lu, R.Mazzarello, C.Jia, W.Zhang, F.Rao, and E.Ma, Science 366, 210 (2019).

Deep-Neural-Networks accelerators aim at implementing the Vector-Matrix-Multiplication in the analog domain, by a parallel voltage drop through a cross-bar array of reprogrammable resistances. Among the technologies available for the fabrication of the “synaptic” weights, ferroelectric-based devices show analog resistive switching, excellent retention properties and small stochasticity. In this talk, we will discuss the development of HfZrO₄ ferroelectric synaptic weights from an industry perspective and how we constrain our processes to CMOS compatible technologies, by adhering to a moderate thermal budget¹.
Two technologies will be presented: two² and three³ terminals devices, respectively Ferroelectric Tunnel Junctions (FTJs) and Ferroelectric Field-Effect Transistors (FeFETs). Both use a transition metal oxide (WO_x<3) to avoid the formation of an interfacial dielectric layer and maximize the ferroelectric field-effect as well as the retention. They exhibit analog potentiation and depression, as well as an Ohmic behavior around the read voltage of 100 mV, required for the analog multiplication. FeFETs and FTJs exhibit excellent cycle-to-cycle variation, with a standard deviation less than 2%, and excellent endurance: 10⁶ full switching cycles before breakdown for FeFETs and 10⁸ for FTJs.
The processes are developed towards a smaller device footprint: the current density is increased by 10⁴ for FTJs; as well as channel dimensions of FeFETs scaled from 10^-10 m² to 10^-15 m² compared to the state-of-the-art^{2 ,3}. The synaptic functionalities of the devices are evaluated in terms of their scalability: scaled FTJs see their On/Off divided by 3, but allow sub-Volt operation and endurance up to 10¹⁰ cycles for FTJs and 10⁹ for FeFETs. In FeFETs, the maximum On/Off ratio is obtained for gate-first devices with optimized channel lengths of 300 nm. The material parameters are obtained by structural and temperature dependent electrical characterization: in FTJs, the resistive switching is not described by a Schottky height modulation, but by a change in the resistance of the Ohmic contact at the HZO/WO_x interface upon polarization switching. In FeFETs, the experiments show a change in the resistance of the Ohmic contacts to the channel, as well as a modulation by two orders of magnitude in the mobility-carrier density product of the WO_x.
¹O’Connor, É. et al. APL Mater. 6, 121103 (2018).
²Bégon-Lours, L. et al. IEEE JEDS (2021)
³Halter, M., et al. ACS A. M. Int. 12, 17725–17732 (2020).
This work is supported by FREEMIND (840903), BeFerroSynaptic (871737), ULPEC (732642) and the BRNC.

The interest in Phase change Memory (PCM) devices with improved resistance drift is mandatory for neuromorphic applications. Among the different possibilities, our approach combines different phase change materials with opposite properties (Sb₂Te₃ and Ge₂Sb₂Te₅), also introducing a confinement Ge layer to realize a novel PCM test cell. The Sb₂Te₃–Ge₂Sb₂Te₅–Ge heterostructures with a total thickness of 110 nm were deposited by RF sputtering and their thermal stability was analyzed by X ray diffraction and Raman spectra, evidencing a slower crystallization dynamic (by 50°C) than in Ge₂Sb₂Te₅. After optimization, the heterostructures were deposited onto proper single cell vehicles prepared on Si(001) substrates and metal contacts were nanolithographically (EBL) defined. The subsequent technical analysis (I–V, R–V, R–cycles graphs for the SET/RESET states) showed that the cell has a programming current of 1.2 mA and an endurance of 2x10⁵ cycles.
This project has received funding from the European Union's Horizon 2020 research and innovation program under Grant Agreement No. 824957 (“BeforeHand:” Boosting Performance of Phase Change Devices by Hetero and Nanostructure Material Design).

As we are rapidly adopting deep learning to all technology that surrounds us, there is a need to look beyond energy-accuracy tradeoff and bring in robustness into the design space exploration. Research has started in earnest to understand the role of efficiency driven optimization techniques towards lending adversarial robustness to deep neural networks (DNNs). In this talk, we will explore some hardware centric techniques (in the context of non-volatile memory (NVM) based crossbars) that can effectively improve the adversarial robustness of neural systems while maintaining the performance and computational efficiency. Specifically, we will look at techniques that allow you to control the intrinsic non-idealities (device, circuit, parasitics, transistor related non-linearity) in NVM crossbars and boost the adversarial robustness of different DNNs mapped onto them by >10% than their standard software implementation. Further, we will talk about a technique called, DetectX, that allows adversarial detection on hardware using current signatures during real-time inference. In the second half of the talk, we will understand the role of event-driven computations and spiking dynamics for tackling robustness for image classification and complex segmentation tasks as well as to preserve privacy during distributed learning. While a considerable amount of attention has been given to optimizing different training algorithms for spiking neural networks (SNNs), the development of explainability still is at its infancy. I will talk about our recent work on a bio-plausible visualization tool for SNNs, called Spike Activation Map (SAM). With SAM, I will highlight the differentiating factors between ANNs vs. SNNs and demonstrate why SNNs tend to be more robust.

Due to their fast and reversible switching between the high resistive amorphous and low resistive metastable crystalline phases upon the application of SET/RESET current pulses, phase change materials (PCM) based on chalcogenides alloys are currently used for the realization of non-volatile memories. Furthermore, partially crystallized intermediate states can be obtained by tuning the height and width of the pulses, making PCM suitable to emulate a synaptic behavior for in-memory computing applications [1], where reliable and reproducible multilevel resistance values are needed [2].
The ternary (GeTe)_m(Sb₂Te₃)_n alloys (GST) are widely used as active materials for the fabrication of both optical and electrical memories, commonly in the composition Ge₂Sb₂Te₂. In view of applications in neuromorphic devices, improved material properties such as reduced resistance drift and a wide programming window are still mandatory. For this reason, thermally stable PCM with higher crystallization temperature have been identified and intensively investigated [3, 4], among which Ge-Sb-Te with Ge-rich compositions (Ge-rich GST) is one of the most promising materials. The combination of PCM layers with different physical properties might help in tailoring the material requirements for applications in embedded devices finding the best compromise between reducing latency and power consumption and increasing retention and cycling endurance. In this context, we present here experimental investigations on composition, chemical state, electronic and structural properties of different chalcogenides combined in double layered heterostructures. Two heterostructures based on Ge-rich GST are fabricated by physical vapor deposition from solid source Knudsen cells on Silicon substrates in ultra-high vacuum, varying the composition of the first PCM layer upon which a second Ge-rich GST film is grown by successive partial deposition steps. The samples are characterized in-situ by a combination of X-ray and Ultraviolet photoemission spectroscopy and ex-situ by X-ray diffraction. Information on intermixing and composition of the alloys across the heterostructures are obtained and discussed as well as the effects of thermal annealing at increasing temperatures. The evolution of the electronic properties of the heterostructures during the formation of the interface will be compared with Density Functional Theory calculations. The role of the interface on the crystallization of the Ge-rich GST second layer will be discussed.
This project has received funding from the European Union's Horizon 2020 research and innovation program under Grant Agreement No. 824957 (“BeforeHand:” Boosting Performance of Phase Change Devices by Hetero- and Nanostructure Material Design).
[1] A. Sebastian et al., Nat. Nanotech. 15, 529 (2020)
[2] D.Kuzum et al., Nano Lett. 11, 693 (2012)
[3] P. Zuliani et al., Solid-State Electron. 111, 27 (2015)
[4] N. Saxena et al, Sci. Rep. 9, 19251 (2019)

The major focus of artificial intelligence (AI) research is made on biomimetic synaptic processes that are mimicked by functional memory devices in the computer industry [1]. It is urgent to find a memory technology for suiting with Brain-Inspired Computing to break the von Neumann bottleneck which limits the efficiency of conventional computer architectures [2].
Silicon-based flash memory, which currently dominates the market for data storage devices, is facing challenging issues to meet the needs of future data storage device development due to the limitations, such as high-power consumption, high operation voltage, and low retention capacity [1]. The emerging resistive random-access memory (RRAM) has elicited intense research as its simple sandwiched structure, including top electrode (TE) layer, bottom electrode (BE) layer, and an intermediate resistive switching (RS) layer, can store data using RS phenomenon between the high resistance state (HRS) and the low resistance state (LRS). This class of emerging devices is expected to outperform conventional memory devices [3]. Specifically, the advantages of RRAM include low-voltage operation, short programming time, great cyclic stability, and good scalability [4]. Among the materials for RS layer, indium gallium zinc oxide (IGZO) has attracted attention because of its abundance and high atomic diffusion property of oxygen atoms, transparency, and its easily modulated electrical properties by controlling the stoichiometric ratio of indium and gallium as well as oxygen potential in the sputter gas [5, 6]. Moreover, since the IGZO can be applied to both the thin-film transistor (TFT) channel and RS layer, the IGZO-based fully integrated transparent electronics are very promising [5].
In this work, we proposed transparent IGZO-based RRAMs. First, we chose ITO to serve as both TE and BE to achieve high transmittance in the visible regime of light. All three layers (TE, RS, BE layers) were deposited using a multi-target magnetron sputtering system on glass substrates to demonstrate fully transparent oxide-based devices. I-V characteristics were evaluated by a semiconductor parameter analyzer, and our devices showed typical butterfly curves indicating the bipolar RS property. And the IGZO-based RRAM can survive more than 50 continuous sweeping cycles. The optical transmission analysis was carried out via a UV-Vis spectrometer and the average transmittance was around 80% out of entire devices in the visible-light wavelength range, implying high transparency. To investigate the thickness dependence on the properties of RS layer, 50nm, 100nm, and 150nm RS layers of IGZO RRAM were fabricated. Also, the oxygen partial pressure during the sputtering of IGZO was varied to optimize the property because the oxygen vacancy concentration governs the RS and RRAM performance. Electrode selection is crucial and can impact the performance of the whole device [7]. Thus, Cu TE was chosen for our second type of device because the diffusion of Cu ions can be beneficial for the formation of the conductive filament (CF). Finally, a ~5 nm SiO₂ barrier layer was employed between TE and RS layers to confine the diffusion of Cu into the RS layer. At the same time, this SiO₂ inserting layer can provide an additional interfacial series resistance in the device to lower the off current, consequently, improve the on/off ratio and whole performance.
In conclusion, the transparent IGZO-based RRAMs were established. To tune the property of RS layer, the thickness of RS layer and sputtering conditions of RS were adjusted. In order to engineer the diffusion capability of the TE material to the RS layer and the BE, a set of TE materials and a barrier layer were integrated into IGZO-based RRAM and the performance was compared. Our encouraging results clearly demonstrate that IGZO is a promising material in RRAM applications and overcoming the bottleneck of current memory technologies.

Chalcogenide based phase change materials have become popular choice for the implementation of the high speed, high endurance, non-volatile phase change memory (PCM) [1], which has shown great potential for brain-inspired computing [2]. However, after years of research, there still are gaps of knowledge in electrical transport mechanisms in amorphous phase of the chalcogenide materials [3,4]. In this work, we performed a systematic study on electric field (0 to 40 MV/m) and temperature (80 K to 350 K) dependent electrical charge transport in amorphized Ge₂Sb₂Te₅ (GST) line-cells, stabilized with high field stresses at low temperatures (T ~ 80 K), and observed two distinct exponential I-V characteristics in the low-field (< 30 MV/m) and high-field (> 30 MV/m) regimes [5-7]. We also calculated the electrical conductivity of GST at around one volt as a function of temperature (ranging from ~2×10^-10 S/cm at 80 K to ~6×10^-4 S/cm at 300 K) and found it to follow the drifted thin film conductivity of the as-deposited amorphous GST (a-GST) below the glass transition temperature.

We performed the experiments on two-terminal PCM line cells with length (L) × width (W) × thickness (t) = ~400-1000 nm × ~120-150 nm × ~ 50±5 nm, after amorphizing each cell with a single 100 ns voltage pulse. We used 0V to 40V voltage sweeps with a current compliance set to 50nA to characterize both the low and high field response of the cells starting at ~ 80 K and increasing the temperature up to 350 K.

We model the electronic transport at low electric fields assuming hopping transport between traps that limit conduction, where the trapped carriers overcome a potential barrier to hop from a ‘source-trap’ to a ‘receiving-trap’ and the barrier height is modulated by the applied voltage. We allow the activation energy, trap-to-trap distance and relative location of the barrier peak to vary in the model. We fit the experimental electrical characteristics to this model and observe extremely good agreement in a wide voltage range, yielding a temperature dependent activation energy ranging from 0.2 eV at T ~ 80 K to 0.37 eV at T ~ 300 K, hopping distances increasing from ~ 2 nm at T ~ 80 K to ~ 8 nm at T ~ 300 K, in agreement with the literature [8]. The results also indicate that the relative location of the barrier peak moves from the location of the receiving-trap at ~ 80 K towards the midpoint between the source-trap and the receiving-trap with increasing temperature.

The distinct change we observe from the low-field regime to the high-field regime, at a critical electric field of ~30 MV/m, maybe explained by (i) increased emission rate from the traps due to extreme reduction in the barrier height, while transport is still limited by hopping between rate limiting trap sites, (ii) impact ionization. Assuming (i), we calculate rate limiting trap-to-trap distance to increase with voltage while the relative location of the peak to remain independent of applied voltage in the high-field regime. Assuming (ii), we expect the increase in current to be due to impact ionization of trapped charges, rather than band-to-band impact ionization and avalanche breakdown, as the mean free paths and the expected kinetic energy of the accelerated carriers are relatively small.

Acknowledgments: This work is supported by US NSF grant # 1711626.

References:
1. S. W. Fong et al. IEEE Transactions on Electron Devices 64.11 (2017): 4374.
2. A. Sebastian et al. Journal of Applied Physics 124.11 (2018): 111101.
3. D. Ielmini et al. Journal of Applied Physics 102.5 (2007): 054517.
4. M. Le Gallo et al. New Journal of Physics 17.9 (2015): 093035.
5. R. S. Khan, Doctoral dissertation, University of Connecticut, 2021.
6. R. S. Khan et al. Applied Physics Letters 116.25 (2020): 253501.
7. R. S. Khan et al. arXiv preprint arXiv:2002.12487 (2020).
8. D. Ielmini et al. 2007 IEEE International Electron Devices Meeting. IEEE, 2007.

As one of the most promising candidates for developing neuromorphic architectures for non-von Neumann computing and information storage, phase-change memory materials in both photonic and electronic devices show non-volatile, fast, and multi-level switching between the amorphous and metastable states. With the assistance of machine learning methodology, in particular, a closed-loop autonomous system and a clustering method, we systematically perform combinatorial exploration on a broad composition range of M-Sb-Te ternary systems, where M is a transition metal. The composition spreads are fabricated by the co-sputtering method, and their composition ranges are measured by wavelength dispersive spectroscopy across the Si spread wafers. To identify optimized compositions out of these spreads, X-ray diffraction measurements (synchrotron radiation), Raman spectroscopy, resistance and optical mapping are carried out in order to determine the evolution of structure and other properties. One composition we identified is Ti_0.3SbTe₂, which shows a higher phase-change temperature and a lower melting point compared to the widely-used Ge₂Sb₂Te₅ (GST225). It also has the largest bandgap contrast between the amorphous and crystalline phases within the Ti-Sb-Te (TST) spread. Scanning transmission electron microscope (STEM) measurements indicate that Ti_0.3SbTe₂ is an epitaxial nanocomposite where TiTe₂ nanograins have grown homogenously inside the TST matrix. We believe TiTe₂ nanoprecipitates can play the role of a precursor in phase transformation. Both photonic and electrical devices fabricated using this nanocomposite phase-change memory material show clear multi-level symmetric switching with low resistance drift and low quenching energy. These results suggest that this composition can be promising for neuromorphic devices. This work is funded by an ONR MURI (Award No. N00014-17-1-2661) and nCORE (Task 2966.010).

Phase change memory (PCM) is a high speed, high density non-volatile memory technology that utilizes the fast and reversible transition between resistive amorphous phase and conductive crystalline phase of chalcogenide materials such as Ge₂Sb₂Te₅ (GST). However, the amorphous phase shows ‘resistance drift’ behavior (the spontaneous increase in resistance with time) which can cause erroneous mixing of intermediate states in multi-level cells, limit denser storage capability and hinder the application of PCM in neuromorphic computing [1,2]. Many studies have been reported so far explaining the origin of the drift [3,4] and describing strategies to suppress this phenomenon [5,6]. In this work, we determined the current-voltage (I-V) characteristics of melt-quenched amorphized GST line cells with widths ~120-140 nm, lengths ~390-500 nm, and thickness ~50 nm using a parameter analyzer in 80 K to 350 K range [7, 8]. When we continuously swept voltage across the cell from 0 V to 25 V with 0.1 V steps and back to 0 V with a current compliance set at 50 nA, we observed a hysteresis loop in the I-V characteristics of the first measurement on each device whereas the successive measurements on the same device were very stable. Measurements performed in the low field regime, which do not disturb the state of the cell, were used to determine the drift before and after the high voltage stress. We found that at low temperatures (T < 250 K) high voltage sweeps (> 15 MV/m), dramatically increase the rate of resistance drift and cells stabilize within a few minutes. Since the current levels are extremely low, this acceleration and stoppage of the resistance drift are attributed to field induced emission of the trapped charges rather than thermally activated processes [9]. I-V characteristics of the stabilized cells under light (for T < 250 K) are distinctly different in the low-voltage regime compared to measurements in dark, while there is practically no difference between the two cases in the high-voltage regime. The difference becomes smaller with the increase in temperature as the thermally activated carriers dominate the conduction process over the photo-excited carriers.

Acknowledgments: This work is supported by US NSF grant # 1711626.

References:
1. A. Pirovano et al. "Electronic switching in phase-change memories." IEEE Transactions on Electron Devices 51.3 (2004): 452-459.
2. W. Zhang et al. "Unveiling the structural origin to control resistance drift in phase-change memory materials." Materials Today (2020).
3. D. Ielmini et al. "Physical interpretation, modeling and impact on phase change memory (PCM) reliability of resistance drift due to chalcogenide structural relaxation." 2007 IEEE International Electron Devices Meeting. IEEE, 2007.
4. S. R. Elliott, "Electronic mechanism for resistance drift in phase-change memory materials: link to persistent photoconductivity." Journal of Physics D: Applied Physics 53.21 (2020): 214002.
5. S. Kim et al. "A phase change memory cell with metal nitride liner as a resistance stabilizer to reduce read current noise for MLC optimization." IEEE Transactions on Electron Devices 63.10 (2016): 3922-3927.
6. W. W. Koelmans et al. "Projected phase-change memory devices." Nature communications 6.1 (2015): 1-7.
7. R. S. Khan, "Characterization and Mitigation of Resistance Drift in Amorphous Ge2Sb2Te5 Devices at Cryogenic Temperatures." Doctoral dissertation, University of Connecticut, 2021.
8. R. S. Khan et al. "Accelerating and Stopping Resistance Drift in Phase Change Memory Cells via High Electric Field Stress." arXiv preprint arXiv:2002.12487 (2020).
9. R. S. Khan et al. "Resistance drift in Ge2Sb2Te5 phase change memory line cells at low temperatures and its response to photoexcitation." Applied Physics Letters 116.25 (2020): 253501.

Computers based on the von-Neumann architecture have become extremely inefficient to process the huge amount of information collected with big data. The continuous data transfer between memory and processor results in energy and speed inefficiency, also known as von-Neumann bottleneck. A transition to novel computing paradigms and architectures, such as bioinspired analog in-memory computing concept is therefore paramount. The crossbar array is an architecture enabling the colocation of memory and processing. It consists of an interconnected array of memory devices with tunable conductance, also known as memristors. This configuration allows to perform vector-matrix-multiplications in a single operation. VMMs are extensively used in AI algorithms of training and inference of neural networks (NNs). The fully connected NN training requires the mapping of the network weights into the array conductance. It also requires adjusting the weights in a monotonic and symmetric multi-step manner by applying identical pulses[1].
Memristors based on HfO2 ReRAM are promising candidates for in-memory computing. The scalable 2-terminal structure is key to integration into highly dense crossbars, in addition to their CMOS compatibility. Conventional ReRAM are based on HfO2 stacked in between a bottom inert electrode and a top electrode with an oxygen scavenging function. After an electroforming operation, a filament of oxygen vacancies allows current to flow through the top and bottom electrodes. The transition from the high to the low resistive state (set) occurs abruptly during programming and it is due to a self-accelerated oxygen exchange between the HfO2 filament and the top electrode[2].
In literature it has been shown that conventional HfO2 technology performances improve with a bilayer ReRAM stack. In this concept, the metallic top scavenging layer is replaced by a properly designed conductive-metal-oxide (CMO) material and engineered to be part of the device layer stack. Although bilayer ReRAM have already shown remarkable capabilities to implement neuromorphic systems, the resistive switching mechanism, as well as the role of the top electrode, is not fully understood[3].
In our work we developed multiple CMO/HfO2 ReRAM concepts which show gradual bidirectional resistive switching and improved symmetry, which approaches the required specs for training a network of memristive devices.
To explain the switching properties of our CMO/HfO2 based devices we modelled with an equivalent electrical circuit the forming, set and reset operations within a dedicated Impedance Spectroscopy experiment.
Our bilayer ReRAMs represents an attractive and flexible memory element for large-scale integration, considering its excellent granular switching properties and CMOS compatibility.
References
[1] Gokmen T. and Vlasov Y. Front. Neurosci., 10 p333 (2016).
[2] J. Woo, S. Yu, IEEE Nanotechnology Magazine, vol. 12, pp. 36-44, 2018.
[3] S. Kim, Y. Abbas, Y. Jeon, A. Sokolov, B. Ku, C. Choi, vol. 29, p. 415204, 2018.

In today's information society, an enormous amount of data is generated daily, more in fact than humans can process, so there is a growing need for high-performance artificial intelligence systems that can quickly learn from such data and make appropriate decisions. One promising type of computer, which is thought to be able to satisfy this requirement, consists of neuromorphic systems that operate by mimicking the functions of the human brain. The development of such brain-like computers, smaller in size and with low-power consumption, would require not only further progress of conventional semiconductor devices, but also the creation of new conceptual devices that operate on analog signals, in much the same way that human neurons operate.
Material properties are greatly affected by slight changes in the arrangement of their constituent atoms. We have been actively exploring interesting nano-phenomena arising from the control of nanostructures, such as the arrangement of atoms. The nanostructure control can be achieved by the use of ionic architectonics method that allow the local control of ion transport and electrochemical chemical reaction occurring at the interfaces and on the surfaces of solids. Our research has enabled us to create nanoionics devices with electrical, magnetic, and optical functions that operate by utilizing such nano phenomena. In this talk, I will introduce an artificial vision device that mimics the human retina to reproduce optical illusions, a decision-making device that learns and makes decisions on its own, and other nanoionics devices for information and communication technology that we have created as part of our research. These ionic devices are be expected to be indispensable in the development of the next generation of hardware-oriented artificial systems, which will complement or even replace conventional software-oriented artificial intelligence systems.

Resistive Random-Access Memory (RRAM) technology is a promising candidate for next-generation non-volatile memory and neuromorphic circuit applications. Many high-k metal-oxides such as HfOx, TiOx and AlOx have been extensively studied to develop RRAM devices.¹ Despite full CMOS compatibility and low cost of fabrication, SiOx RRAMs have received relatively less attention. Previous works on SiOx RRAMs typically report high electroforming and/or high switching voltages.^2,3 In this work, we demonstrate a SiOx RRAM with low switching voltages, high non-linearity and near-gradual set/reset operation. Crossbar devices with dimensions ranging from 2 mm to 10 mm were fabricated using Au (50 nm) as the bottom electrode, SiOx as the switching layer and Ti/Au as the top electrode. Thickness of the Ti top electrode was varied from 5 nm to 15 nm while the thickness of the Au capping layer was fixed at 50 nm. The SiOx films were deposited using inductively coupled plasma enhanced chemical vapor deposition (ICP-PECVD) at a deposition temperature of 100 °C and RF Power of 300 W. Bottom and top electrodes were fabricated using photolithography and electron-beam evaporation. The devices show bipolar resistance switching after an initial electroforming step. The forming voltage can be reduced either by reducing the SiOx thickness or by increasing the Ti electrode thickness. The devices with 10.8 nm SiOx and 10 nm Ti electrode thickness exhibit set and reset operation at ~ 1V and -1.5V, respectively. Low cycle-cycle variability is observed for Vset and Vreset (~ 1% and 4.6% standard error, respectively). The switching characteristics at 100 µA compliance indicate gradual set/reset operation and high (~10) non-linearity in the on-state. However, the abruptness of resistance change during the set operation slightly increases at higher current compliance. Non-linearity observed in the on-state can be beneficial for suppression of sneak-paths in large crossbar arrays. The on-state resistance increases with decreasing device area, indicating an interfacial effect dominating the resistance switching mechanism. Detailed investigation of the switching mechanism by temperature dependent electrical measurements and transmission electron microscopy (TEM) will be presented. Overall, our results offer promise for SiOx RRAM as a low-cost, fully CMOS compatible technology suitable for low-power neuromorphic applications.
REFERENCES:
1. H. S. P. Wong et al., “Metal-Oxide RRAM”, Proceedings of the IEEE, vol. 100, pp. 1951-1970, 2012.
2. A. Bricalli et al., “Resistive Switching Device Technology Based on Silicon Oxide for Improved ON–OFF Ratio—Part I: Memory Devices”, IEEE Transactions on Electron Devices, vol. 65, no. 1, pp. 115-121, 2018.
3. A. Mehonic et al., “Silicon oxide (SiOx): A Promising Material for Resistance Switching?”, Advanced Materials, vol. 30, p. 1801187, 2018.

From the viewpoint of processing speed and safety, edge computing, which operates physically close to the user or data source, has been attracting attention [1]. In order to enable data processing at any edge, a method that reduces computational cost while maintaining high learning performance is required. One of the candidates is reservoir computing (RC), which is a computational framework suitable for processing time series-data. RC transforms time-series data into spatio-temporal patterns with a high-dimensional space of mutually conjugate nonlinear elements called reservoirs, which can be easily analyzed for patterns using a simple learning algorithm at readout time [2]. Physical reservoir computing (PRC) is expected to achieve low power consumption and simple device structures because it uses the physical dynamics of real physical systems as a reservoir. Among PRCs, a tin-doped indium oxide (ITO)/Nb-doped SrTiO₃ (Nb:STO) junction, which works as an artificial optoelectronic synapse that respond to not only electric stimuli but also photonic ones, is interesting because human beings receive much of the information from the environment through visual perception [３].
In this paper, we discuss the mechanism of synaptic behavior of ITO/Nb:STO junctions, focusing on the correlation between photo- and electric- responses. For this purpose, we prepared ITO/Nb:STO junction that shows a large hysteresis in I-V characteristics, meaning that the ITO/Nb:STO junction can work as a resistive random access memory with a large memory window. We also demonstrated the influence of the correlation between optical and electrical on PRC.
We have evaluated the photoresponsive characteristics of ITO/Nb:STO junctions irradiated with ultraviolet (UV) light from the ITO transparent electrode side. The measurement was performed by applying a constant read out voltage (V_read) to the device using a source measure unit (Keysight B2901A), and the current value of the device was read out. As a result, we found that the photoresponsive behavior of the ITO/Nb:STO junction is significantly different depending on whether V_read is applied or not.
First, we irradiated the ITO/Nb:STO junction with UV light and measured the photoinduced current at V_read of 0 V, i.e., with no external voltage application. It was found that the current value quickly increased with UV irradiation, and quickly decayed to the initial current level when UV irradiation was stopped. This characteristic is considered to be suitable for applications such as photo sensors [４].
Secondly, we performed the same experiment, except that a finite V_read was applied. When V_read was set to lower than -0.3 V, it was found that the current value nonlinearly and drastically increased with UV irradiation. Interestingly, when the UV irradiation is stopped, the current kept retaining large value although the current slightly decayed. In this case, the current never relaxed to the initial current value. This behavior is similar to Persistent Photoconductivity (PPC), which has already been observed in STO [５]. We have found for the first time that both photosensor-like mode with quick response and PPC-like mode with very slow but strong response can be controllably switched by an externally applied voltage to the device. This behavior is also interesting from a physical point of view.
In addition, by using the nonlinear photoresponse behavior and the slow relaxation caused by the PPC, we have confirmed that the 4-bit data written by light irradiation ON as “1” and no light irradiation as “0” can be distinguished.
[1] W. Shi et al., IEEE Internet of Things Journal 3, 637 (2016).
[2] G. Tanaka et al., Neural Networks 115, 100 (2019).
[３] S. Gao et al., ACS Nano 13, 2634 (2019).
[４] L. Wu et al., Adv. Funct. Mater 31, 2010439 (2021).
[５] MC. Tarun et al., PRL 111, 187403 (2013).

Deep learning is a hugely successful and powerful algorithm for machine learning applications such as computer vision and natural language processing. However, the training of these neural networks is limited by the traditional von Neumann architecture of our current CPUs and GPUs. Shuttling data back and forth between the separate memory and computation units in such architecture results in significant energy consumption; many orders of magnitude greater than the energy consumption in human brain. Our projects focus on designing materials and hardware that can instead perform data storage and computation in a single architecture using ions, inspired by the human brain. In one project, we have designed a protonic electrochemical synapse that changes conductivity deterministically by current-controlled shuffling of dopant protons across the active device layer; resulting in energy consumption on par with biological synapses in the brain. In our second project, we demonstrate controlled conducting oxygen vacancy filament formation in resistive memory devices via dopant and microstructure engineering. These strategies provide pathways towards neuromorphic hardware that has high yield and consistency, performs data storage and computation in a single device, and uses significantly lesser energy as compared to current computing architectures.

Neuromorphic engineering provides an exciting avenue to mimic the high performance of the human nervous systems and sensors¹. Critically, the energy efficiency of the human neural networks for learning relies on event-driven, temporally encoded action potential streams. In this talk, I will discuss how we can digitize tactile information through inspiration from somatosensory neural science. We have developed an asynchronous protocol for parallel transmission of tactile information in an artificial peripheral nervous system we call ACES: Asynchronously Coded Electronic Skins². The parallel transmission encodes spatial temporal information with very high temporal precision (sub-100ns) even when large numbers > 10,000 sensor nodes are transmitting simultaneously. Such systems can be interfaced with soft sensors or flexible/stretchable electronics to enable more intuitive robotics and healthcare applications.


1. Bartolozzi, C. et al. Neuromorphic Systems. Wiley Encycl. Electr. Electron. Eng. 1–22 (2016) doi:10.1002/047134608X.W8328.
2. Lee, W. W. et al. A neuro-inspired artificial peripheral nervous system for scalable electronic skins. Sci. Robot. 4, eaax2198 (2019).

Transition metal oxides (TMOs) showing several order of magnitude resistance change across a metal-insulator transition (MIT) hold promise for applications in resistive switching and neuromorphic devices and circuits. One of the main aspects that distinguishes MIT materials from other “switching” materials is the fact that the resistance change in these Mott insulators is a consequence of an electronic phase transition (driven by strong electron interactions) and therefore reliable, fast and reproducible device performance can be expected. Besides that, the possibility to tune the different resistance states by temperature, pressure, doping, electric field or strain offers a versatile control of their functionality. However, there are still open questions related, for example, to the intrinsic mechanisms ruling the phase transition or the formation and dynamics of non-equilibrium phases that can play a very important role in the development of practical Mottronic devices.
Vanadium sesquioxide (V₂O₃) is an archetypal Mott insulator featuring a MIT at low temperature accompanied by a structural phase transition from a corundum to a monoclinic phase. In bulk samples, also a purely electronic isostructural MIT close to room temperature is observed when slightly doping the material.
In this talk, I will present results on i) the influence that the nanotextured character of the insulating phase in V₂O₃ has on the stabilization of a non-thermal metallic state and ii) the stabilization of a room temperature Mott MIT by strain engineering in V₂O₃ (even in the absence of doping) with access to intermediate states absent in bulk. I aim to discuss the impact that these results can have in the possible design of Mottronic devices based on V₂O₃.
A. Ronchi et al., Phys. Rev. B 100, 075111 (2019).
P. Homm et al., APL Materials 9, 021116 (2021).
A. Ronchi et al., Phys. Rev. Appl.15, 044023 (2021).

For the most common type of memristive ReRAM devices, the non-volatile change of the resistance is induced by the movement of oxygen vacancies along nanosized filaments. In contrast to this, in area-dependent memristive devices the ionic motion as well as the change of the resistance occurs along the whole devices area. We will give an overview over area-dependent memristive devices based on the p-conducting Pr_0.7Ca_0.3MnO₃ (PCMO) and different tunnel barriers and eludidate the underlying switching mechanisms by operando spectroscopy. We will discuss the differences of these devices to the common filamentary devices with respect to switching kinetics, reliability and analogue operation. Based on the very gradual type of switching, the resistance can be tuned in an analogue fashion by repeated switching with voltage pulses of the same amplitude and polarity. We investigate in detail the impact of different pulse heights and pulse lengths on the shape of the stepwise SET and RESET curves. We use these measurements as input for the simulation of training and inference in a multilayer perceptron for pattern recognition, to show the use of PCMO based ReRAM devices as weights in artificial neural networks which are trained by gradient descent methods. Based on this, we identify certain trends for the impact of the applied voltages and pulse length on the resulting shape of the measured curves and on the learning rate and accuracy of the multilayer perceptron.

Spiking Neural Networks (SNN), provide distributed computational capability and emulate the brain processing principles. They encode the information by modulation of spike frequency and pulse width, compared to conventional Artificial Neural Networks (ANN) that modulate the magnitude of the signal [1]. One of the target applications of hardware implementations of SNN, known as neuromorphic systems [2], are Brain-Computer Interfaces (BCI) [3] due to intrinsic signaling compatibility between spiking neuromorphic systems and biological brains. However, with a handful of exceptions [4], such implementations are typically implemented using physically hard, rigid, and ultimately non-biocompatible silicon technology, resulting in a permanent damage to the brain's soft tissue. Physically soft and flexible, biocompatible organic electronics show promise for direct interfacing with biological tissues [5].

Synapses play a crucial role in neural networks by interfacing between adjacent neurons [1]. They are responsible for tuning the neural signals and, as such, are critical in memory, learning, and cognition. Therefore, artificial synapses are seen as promising bidirectional neuromorphic Brain-Machine interfaces [3].

We report on the fabrication of spiking neuromorphic synaptic circuits implemented with physically flexible organic electronics. Organic Field-Effect Transistors (OFETs) were fabricated on physically flexible, 50 μm thin Polyimide substrates. OFETs’ architecture was bottom-gate top-contact, and fabricated via physical (thermal) or chemical vapor deposition of Cr (3 nm)/Au (50 nm) as the gate electrode, 400 nm of Parylene diX-SR as the dielectric, 30 nm of DNTT as the p-type and PDI8-CN2 (also known as N1200) as the n-type organic semiconductors. Source and the drain electrodes were formed from a 50 nm layer of Au. A layer of Parylene encapsulated the structure. All circuits used fabrication-integrated Parylene-based capacitors.

The results demonstrate that the fabricated, all organic differential-pair integrator (DPI) synaptic circuit [6], is capable of reproducing the exponential dynamics observed in excitatory and inhibitory postsynaptic currents of biological synapses, implementing a low-pass filter with linear transfer function. The circuit provides proportional output current based on continuously tunable synaptic weights. We also show the effects of synaptic weighting and threshold voltages. The use of fabrication-integrated synaptic capacitance result in favorable and more biologically compatible time constants. Additionally, we demonstrate an improvement on our previous work of organic log-domain integrator (LDI) synapse [7], by demonstrating that the DPI synapse addresses the limitations of LDI synapse by multiplexing in time spikes arriving from different sources. Finally, we show that our all-organic artificial DPI synapse can produce proportional synaptic current in response to being stimulated with a train of biologically-relevant somatic spikes from an artificial inorganic soma. All of the results were separately and independently verified in simulation.

We believe that this work shows a promise to future integration of synaptic neuromorphic circuits with biological neurons, for future implantable BMIs. Furthermore, with our previously demonstrated spiking organic electronic Axon-Hillock somatic circuits [8], it paves the way for future all-organic, flexible and biocompatible neuromorphic computing.

[1] F. Rieke and D. Warland, MIT Press, 1999.
[2] R.A. Nawrocki, et al., IEEE Trans. Ele. Dev.,63 (10), 2016.
[3] F. Boi, et al., Front. Neuro., 10 (563), 2016.
[4] R.A. Nawrocki, et al., IEEE Trans. Ele. Dev., 61 (10), 2014.
[5] G. Schwartz, et al., Nat. Comm., 4 (1), 2013.
[6] C. Bartolozzi and G. Indiveri, Neural Comp., 19 (10), 2007.
[7] Hosseini, M.J.M., et al., Adv.Ele.Mat., under review, 2022.
[8] Hosseini, M.J.M., et al., J.Phys.D: Appl.Phys., 54, 104004, 2021.

Sneak path current is a fundamental issue and a major roadblock to the wide application of memristor crossbar arrays. Traditional selectors such as transistors compromise the 2D scalability and 3D stack-ability of the array, while emerging selectors with highly nonlinear current-voltage relations contradict the requirement of a linear current-voltage relation for efficient multiplication by directly using Ohm’s Law. Herein we propose and demonstrate the concept of a timing selector, which addresses the sneak path issue with a voltage-dependent delay time of its transient switching behavior while preserves a linear current-voltage relationship for computation. We built crossbar arrays with silver-based diffusive memristors as the timing selectors and experimentally demonstrated the operation principle and operational windows. The timing selector will enable large memristor crossbar arrays that can be used to solve large-dimension real-world problems in machine intelligence and neuromorphic computing.