Symposium F.NM07—Progress in Neuromorphic Computing Materials, Devices and Systems (Includes Symposium S.EL08—Neuromorphic Materials and Devices for Bioinspired Computing and Artificial Intelligence)

There have been many demonstrations of nonvolatile memory devices, whose memory states can be continuously tuned, and early experimental results of utilizing such functionality in analog and mixed-signal computing, most notably in neuromorphic circuits. In my talk, I will discuss the major challenges in the quest of building practically useful mixed-signal neuromorphic circuits based on analog-grade memory devices, focusing on the simplest, in terms of device requirements, yet most common neuromorphic inference operation. I will conclude my talk by reviewing experimental results for neuromorphic circuits based on intergrated metal-oxide memristors, one of the most promising types of devices for this applicaiton, in the context of these challenges.

Memristive devices have been receiving an increasing interest for a wide range of applications, such as storage class memory, in-memory computing and neuromorphic computing towards artificial intelligence at the edge. Indeed, neuro-inspired computing supported by memristive devices can pave the way to a revolution for an energy-efficient and area-efficient information processing that transcends von Neumann’s computing. This goal can be achieved by developing computational substrates and engines able to mimicking the adaptation, interconnectivity, noise tolerance and energy efficiency of the human brain. Among the proposed memristive technologies, oxide-based resistance switching memory devices (RRAM) are based on redox reactions and electrochemical phenomena in oxides and are very promising because of low power consumption, fast switching times, scalability down to nm scale and CMOS compatibility. Therefore, they can be used as new hardware building blocks emulating electronic synapses in spiking neural networks (SNN) for real time and low power computation systems. In our work, we focus on the switching dynamics of HfO₂-based RRAMs and on their implementation as electronic synapses in spiking neural networks (SNN). The conductance of the RRAM devices can be tuned either in a digital or in an analog fashion by using proper programming algorithms and by engineering the material stack. The device dynamics is used to emulate the biological potentiation (conductance increase) and depression (conductance decrease) processes, over several cycles. Moreover, the conductance value update can be achieved by a spike timing and rate dependent plasticity mechanism, which is demonstrated at hardware level, and that is exploited as learning rule in a SNN. System level simulation of SNN are implemented by using experimental RRAM characteristics for the synaptic nodes and neural equations derived from hardware CMOS implementation. The effect of device switching dynamics (linear and non-linear) and synaptic resolution on the SNN performances is discussed.

Memristors are considered as promising candidates for realizing synapses for neuromorphic computing because of their reduced complexity and nanoscale foot-print when compared to conventional complementary metal oxide semiconductor (CMOS) technology. Non-volatile switching with multiple conductance states in memristors is employed to emulate synapses. In this work, we present synapses realized by memristors using chemical vapor deposited (CVD) MoS2. Graphene is used as electrode for this device. These synapses exhibit linear and symmetric weight update with identical pulses, considered to be the “holy grail”for unsupervised learning. The graphene/MoS2 (G/M) synaptic device exhibits 4 distinct states in both DC potentiation and depression regimes. Additionally, it exhibits data retention for >104 s. For efficient unsupervised learning, the synapses should exhibit linear and symmetric weight update with identical input pulses and a high maximum/minimum conductance ratio (Gmax/Gmin). Resistive random-access memory (RRAM), phase change memory (PCM) and ferroelectric field effect transistors (FeFET) have been used to emulate a synapse. But these devices exhibit non-linear weight update with identical pulses paired with reduced Gmax/Gmin ratio. Identical input voltage pulses with tON= 500 µs and amplitude of 6 V (-4 V) are applied to observe potentiation (depression) in the G/M synapse. With 25 input pulses applied for each regime, we obtain linear and symmetric weight update with an ideal symmetry of 0. The non-linearity factor (NLF) is 0 for both potentiation and depression regimes, and Gmax/Gmin is 16.5. These device characteristics are then incorporated in a pseudo-crossbar array of a 2-layer multi perceptron network with the help of the circuit macro model Neurosim1 to obtain online learning accuracy of 84.33 %. We benchmark our G/M synapse with other emerging synapse devices with respect to NLF, asymmetry, Gmax/Gmin and learning accuracy with Neurosim and show that our device is superior to all other emerging synaptic devices in all these categories. 1 P.-Y. Chen, X. Peng, and S. Yu, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 37 (12), 3067 (2018).

Resistive random-access memory (ReRAM)/memristive devices based on electrical resistance switching are highly expected to overcome persistent technological limitations in current memory technology and to develop next-generation nonvolatile memory because of its superior characteristics. These characteristics include high speed, high scalability, low power consumption, fast fabrication process, multistate storage, and compatibility with complementary metal-oxide semiconductor (CMOS). Particularly, nanowire-based resistive switching devices offer an ideal platform for a deeper understanding of resistive switching properties due to the confined nanostructure and the opportunity to develop nanoscale-memristive device for next-generation memory and neuromorphic computing. In addition, copper oxide is one of the attractive candidates for ReRAM/memristor due to the existing multiple oxidation states from “+2” to metal state (CuO, Cu4O3, Cu2O, Cu) for inducing resistive switching, which is believed to involve the phase transition with multiple oxidation state induced by redox reaction. In this work, we report the volatile and non-volatile resistive switching effect in p-type non-stoichiometric CuOx nanowires, which are fabricated using insulator CuO nanowires via a thermal-annealing technique with hydrogen atmosphere. These resistive switching modes were effectively controlled by the stability of conductive filament via the compliance current in the SET process. The non-volatile resistive switching exhibited low electric-field-induced switching, high ON/OFF ratio of 107, and long retention time of 104 seconds for ReRAM application. On the other hand, the volatile threshold switching was operated with low compliance current for selector application, and the device showed bidirectional operation, selectivity (104), low OFF-current (<100 pA), and stable reliability with endurance over 105 cycles under pulse operation. The presented work will discuss a nanoscale resistive switching mechanism based on hole-based conducting filament formation originated from the local redox reaction.

Neuromorphic computing, inspired by the human brain, provides efficient energy utilization, massive parallelism, flexible adaptive capability, and information processing and transmission. It has the potential to realize computational techniques requiring recognition, learning and decision making; fundamentally overcoming the bottleneck in von Neumann architectures.[1] Artificial synapses form the backbone of neuromorphic computing; emulating controlled functions related to learning and memory. Furthermore, to functionalize devices beyond their synaptic functions, it is essential to merge them with biometric sensing elements such as vision, auditory, touch and olfactory sensors. Optical stimuli in a synaptic device renders ultrafast computational speed due to high bandwidth, low crosstalk, and ultralow power consumption.[2] A photonic synaptic transistor based on 2D materials provides ultra-low power consumption along with flexibility and thickness scalability. Graphene, the first 2D material leads to a short photocarrier lifetime and high dark current due to its zero bandgap and is therefore unfavorable for use as an efficient synapse material. Unlike graphene, monolayer MoS2 having a direct bandgap would allow a high absorption coefficient and efficient electron–hole pair generation under photoexcitation, becomes suitable in various neuromorphic applications. We demonstrate the concept of artificial cogni-retina; an optic-neural synaptic device that features synaptic and optical-sensing functions. It has the potential to replace bulky systems comprising of video camera and processing units. A MoS2 based photonic synaptic transistor for emulating the superior and complex features of artificial cogni-retina is reported. The device finds application in prosthetic human eyes and robotic eyes in artificial intelligence based platforms of autonomous vehicles. The device shows synaptic behavior with high conductance ratio of 103 facilitated by gate tunability and a long-term memory of 104 s with an extremely low energy consumption (15 pJ/event). High quality large area (2cm×2cm) monolayer MoS2 (grain size of 50 ?m) is grown on SiO2/Si substrate via CVD technique by maintaining the precursor at 950°C. As grown MoS2 coated with PMMA is immersed in BOE to release the MoS2 film by removing the underlying SiO2. The PMMA supported MoS2 film is transferred to DI water bath and finally to SiO2/ p+ silicon substrate with nickel source/drain contacts. The final sample is heated for increasing the adhesion of the transferred film, and the PMMA layer is etched by acetone followed by a forming gas annealing. MoS2 is patterned to form the channels of the transistor by reactive ion etching in CF4. We explore the spectral response of our device, expressed through its photoresponsivity, over a wavelength of 400-700 nm for varying intensity. The negligible photoresponsivity obtained for wavelengths above 680 nm clearly depicts the bandgap of monolayer MoS2 corresponding to 1.8 eV. The synaptic transistor exhibits excellent responsivity of 2×103 AW-1 (EQE of 6×105 %) under illumination (450 nm, 10 ?W/cm2) for VG = 0 V and VD = 5 V. Both the responsivity and EQE of the device increase with gate voltage due to lowering of the barriers at the contacts. The device shows n-type behavior (ON/OFF current ratio of 105) and under illumination, an enhancement in current is observed for both ON (VG>VTH) and OFF (VGTH) states for all gate voltages. This indicates that the photocurrent dominates over thermionic and tunneling currents. Gate dependent potentiation and data retention of the device is observed by varying the gate bias. The device also demonstrates other essential synaptic behaviors, such as short-term potentiation (paired-pulse facilitation of 150%), spike-frequency dependent plasticity, spike-duration dependent plasticity. References [1] Q. Xia et al., Nature materials 2019, 18. [2] Z. Cheng et al., Science advances 2017, 3.

Physical neural networks based on analog resistive switching processor arrays are promising for in-memory computing hardware which can implement machine learning algorithms at higher operating speed and lower energy cost. State-of-the-art analog resistive switches rely on the mechanism of either forming conductive filaments or inducing phase change. These processes suffer from poor repeatability or high energy consumption, respectively. The first part of this talk focuses on reducing the variability of the set/reset voltages using model systems of HfO2 grown on Nb:SrTiO3 and Si/TiN substrates. We control the oxide microstructure via growth paramters and thermal treatment, to compare amorphous HfO2 films with nanocrystalline columnar-grained films with both single (-111) texture and multiple orientations This work advances our understanding of the mechanisms behind device variability and achieving devices that meet the strict requirements of neuromorphic computing. The second part of the talk focuses on a deterministic charge controlled switching mechanism, that is small cation intercalation into active swithing materials. Herein we demonstrate a three-terminal resistive switching device that consists of a channel of active material probed by two terminals (source and drain), a proton reservoir layer as gate terminal, and a proton conducting solid electrolyte separating the two. The device is made of entirely inorganic and CMOS compatible materials. The electrical conductivity of the active material can be modulated by the precise control of proton intercalation, providing a highly reproducible resistive switching behavior. More than 100 states with an on/off ratio exceeding 200 have been demonstrated by a 5 ms wide gate current pulse train of +/- 1?A. As the smallest cation, the shuffling of protons between the reservoir and the active material requires minimal energy. The energy consumption is only 3.5 fJ/?m2 per state, comparable with state-of-the-art demonstrations. The switching mechanism is probed by a combination of synchrotron X-ray spectroscopy methods. The advantages of this switching mechanism, namely, low energy dissipation and multi-state capability, are demonstrated.

Efficiency bottlenecks inherent to conventional computing in executing neural algorithms have spurred the development of novel devices capable of “in-memory”computing. Commonly known as “memristors”, a variety of device concepts including conducting bridge, vacancy filament, phase change and other types have been proposed as promising elements in artificial neural networks for executing inference and learning algorithms. I will review the recent advances in memristor technology for neuromorphic computing and discuss strategies for addressing the most significant performance challenges, including 'write' asymmetry, high read/write currents, and endurance. As an alternative to two-terminal memristors, I will introduce the three-terminal electrochemical memory based on the redox transistor (RT) which uses a gate to tune the redox state of the channel. Charge-compensated redox reactions in the bulk of the transistor enable high-density synaptic weight storage. Decoupling the 'read' and 'write' operations with a third terminal enable low-energy operation without compromising analog performance. I will discuss the RT operating mechanisms using organic and inorganic materials, approaches for array integration, and prospects for achieving the device density and switching speeds necessary to make electrochemical memory competitive with established digital technology.

Volatile resistive switching, as observed in Mott materials, is a promising phenomenon which may allow for the implementation of artificial spiking neurons. Despite the technological potential and intense research efforts, many fundamental questions about the nature of this transition remain unsolved. The mechanism driving the insulator-metal (IMT) transition, the nature of the percolation process and the time dynamics still need to be addressed. In this talk, new insights about the resistive switching process in Mott insulators [1,2] will be shown:
Underlying mechanism. Two mechanisms may trigger the IMT: Joule heating and E-field destabilization of the Mott insulator. Which of these mechanisms in dominant depends on the material and defect concentration.
Filamentary percolation. Direct nanoscale imaging of the domain distribution during resistive switching, reveals the inhomogeneous nature of the percolation process.
Dynamics. We observe, for the first time, the switching process with temporal and spatial resolution. The presence of strain leads to slow dynamics both during filament formation and relaxation, with important consequences for the dynamic properties of Mott-based devices.
Mimicking neuron functionalities. We demonstrate that both charge build-up and heat transfer can be used to implement basic biological neuronal functions.
These results provide a deeper understanding of the voltage triggered IMT and the technical challenges to solve in order to implement it in a neuromorphic computing scheme.
This work was supported as part of the Quantum Materials for Energy Efficient Neuromorphic Computing (Q-MEEN-C) Energy Frontier Research Center (EFRC), funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences under Award # DE-SC0019273.
References:
Javier del Valle, Pavel Salev, Federico Tesler, Nicolas Vargas, Yoav Kalcheim, Paul Wang, Minhan Lee, Juan Trastoy, George Kassabian, Juan Gabriel Ramírez, Marcelo J. Rozenberg and Ivan K. Schuller, “Subthreshold firing in Mott nanodevices”, Nature 569, 388 (2019)
Javier del Valle, Pavel Salev, Yoav Kalcheim and Ivan K. Schuller. “A caloritronics-based Mott neuristor”, Scientific Reports-Nature 10, 4292 (2020)

Redox-based memristive devices are one of the most attractive candidates for future non-volatile memory applications and neuromorphic circuits. Many recent studies reported that resistive switching phenomena in transition metal oxides (TMO) by an external bias are associated with the migration of oxygen ions. Here, we study Sr2Fe2O5 brownmillerite structure on the SrRuO3/SrTiO3 substrate and Nb: SrTiO3 substrate as a bottom electrode in order to understand the role of oxygen migration in resistive switching. Topotactic phase changes between the brownmillerite structure and perovskite structure appear by the oxygen ion migration. Using the fast ion path on the brownmillerite structure, it is possible to impede the random movements of oxygen vacancies. In-situ electric biasing holder was used to investigate the changes in atomic configurations as well as electronic structure caused by the transition of the oxygen vacancies. Electric biases were applied to the TEM samples directly in real time which makes it possible to compare the structure changes in the same TMO regions through SET and RESET cycles. To understand the mechanism of whether the oxygen ions are coming from the top or the bottom electrode in the topotactic phase change, the experiment was carried out with different kinds of bottom electrode. This provides a deeper understanding to control the oxygen ions of the redox-based memristive devices to improve and optimize performances.

Analogue neuromorphic systems are attracting considerable attention for their potential in novel bio-inspired computing architectures. For this purpose, circuits presenting advantageous features such as low power dissipation and cognitive capabilities need to be developed. Double-barrier memristive devices (DBMDs) with the layer sequence Nb/AlO_y/HfO_��/Au are promising candidates for such applications. These devices can, e.g., be exploited to emulate the plastic behaviour of biological synapses in analogue electronic circuits. Their resistance can be varied by applying suitable voltage bias to the device electrodes, an effect known as resistive switching. While kinetic Monte Carlo simulations indicate that the drift of charged defects leads to a modulation of the Schottky barrier height and, hence, to resistive switching, experimental verification has been missing so far. Clearly, a deeper understanding of the resistive switching mechanism could pave the way towards tailoring of application specific devices. Here, we show that resistive switching in DBMDs is due to changes in the chemical composition of the HfO_x layer. This is demonstrated using soft and hard x-ray photoelectron spectroscopy at photon energies of 1550 eV and 7000 eV, respectively. Corresponding experiments were performed at beamlines P04 and P22 of PETRA III (DESY, Hamburg, Germany). Using soft x-rays, Hf 4f core-level emissions from the buried HfO_x layer were measured in situ for two different resistance states. Hard x-rays, on the other hand, were used to probe resistance state-dependent changes in the emissions from the buried metal oxides (Hf 3d_5/2, Al 1s) under varying emission angle (15°, 41.4°, and 60°). Spectral shifts between bulk emissions and emissions from Au/HfO_x and HfO_x/AlO_y interfaces were observed for the different resistance states and were analysed in terms of varying chemical composition and defect concentration in the device. The results suggest that the applied electric field induces oxygen vacancy motion in the HfO_x layer, resulting in local oxygen concentration changes and a modification of the Schottky barrier height at the Au/HfO_x interface. In addition, the AlO_y composition is revealed to vary in the bulk, whereas the AlO_y composition at the HfO_x/AlO_y interface appears not to be affected by the resistive switching. The experimental results therefore provide evidence that oxygen vacancy motion within the HfO_x layer is directly related to the resistive switching in DBMDs. Engineering of the oxygen content in the HfO_x layer is foreseen to enable a tailored device response.

Neuromorphic in-memory computing can significantly decrease the energy consumption of important algorithms such as inference and backpropagation by eliminating the need to shuttle information from memory to processor. The key challenge is finding materials and devices that not only contain a large number of analogue non-volatile states but also have the ability to reliably switch between these states. We show how redox transistors, operating on the basis of electrochemical ion insertion, can be used to dynamically tune the concentration of dopants in transition metal oxides. Such dynamic dopants are able to tune the electronic and optical properties of materials in a linear and analogue fashion and can be used to emulate synaptic weights in neural networks. We study the effect of lithium insertion into TiO2 and WO3. In both cases, lithium acts as an n-type donor that adds an electron which localizes on the transition metal. This extra dopant changes the electronic and optical properties of the transition metal host. In the case of TiO2, we harness changes in electronic properties to create a low-voltage synaptic transistor based on electrochemical redox. This synaptic transistor has over 200 distinguishable analogue non-volatile states based on the lithium concentration; these were used to emulate the synaptic weights in a neural network. More importantly, it is possible to switch between neighboring states in a linear and symmetric fashion with low current and voltages. When implemented into a crossbar, simulations show that this allows for high training and classification accuracy, and significantly lower energy consumption than a digital processor. We also explore the case of lithium insertion into WO3. Here, we utilize changes in the optical refractive index nthat arise during lithium insertion, which alters the bandgap of WO3. These changes in bandgap combined with a resonant plasmonic structure enable a device with tunable, analogue optical response. Such system can be utilized for computing in the optical or the hybrid electronic-optical domains. References: Y. Li et al. "Dynamic tuning of gap plasmon resonances using a solid-state electrochromic device." Nano Lett. DOI: 10.1021/acs.nanolett.9b03143 Y. Li et al. "Low-voltage, CMOS free synaptic memory based on LiXTiO2 redox transistors" ACS Appl. Mater. Interfaces, DOI: 10.1021/acsami.9b14338

Memristive devices for neuromorphic computing schemes for non-volatile memories are getting a rapidly increased amount of attention in the last couple of years. However, the requirements for the devices are sometimes very different for the different fields of applications. In neuromorphic electronics, memristive devices characterized by gradual resistance switch or multi-level resistance states are required, where bi-layer oxide systems have emerged as promising candidates for that field. An example for such a device structure is the double barrier Nb/Al/Al2O3/HfOx/Au memristive device, which possess a filamentary-free, homogenous interfacial resistive switching mechanism. The double layer device has a Al2O3 tunnel barrier of 1.2 nm and an ultra-thin active HfOx layer of 3.5 nm. The key mechanism is due to mobile oxygen ions confined within the 3.5 nm thin HfOx layer which alter the Schottky barrier of the top HfOx/Au interface. To gain a deeper insight into the switching mechanism, impedance spectroscopy was applied to correlate the inner electronic and ionic processes with the observed I-V characteristics. In detail, potentiodynamic impedance spectroscopy (PDIS) and time resolved impedance spectroscopy (TRIS) were used in the course of this investigation. Evidence of ion motions at 1.8 V directly related to changes in Schottky barrier resistance has been found as the active mechanism. The results of the impedance analysis were translated into a more general model for memristive devices based on homogeneous ionic motion to map the physical processes during the switching of the memristive device.

Ferroelectric materials have spontaneous polarization states that can be maintained even without an external electric field. Delicate control of polarization is possible because domains in ferroelectric materials can be controlled by an applied electric field. Each polarization state of the ferroelectric layer can be confirmed by changes in channel conductance; to exploit this characteristic, ferroelectric materials have been evaluated as the dielectric layer in transistors. The conductance of channel material in ferroelectric transistor can be gradually controlled by the polarization of ferroelectric materials. These characteristics can be the solution to solve some problems in current synapse devices, such as a small on/off ratio, non-linear weight updates, and variations in electrical properties. Neuromorphic computing has attracted much attention due to its power-efficient data processing. To realize efficient neuromorphic hardware systems, the development of reliable and robust synaptic devices is essential. However, the limited performance of the synaptic devices remains as a challenge for realization of neuromorphic hardware systems. Analog memory characteristics has been demonstrated in the ferroelectric thin-film transistor (FeTFT) based on oxide semiconductor and ferroelectric hafnium oxide layer. The conductance of the oxide semiconductor channel is modulated by controlling the polarization of the ferroelectric hafnium oxide layer. Conductance modulation characteristics of FeTFT were used to demonstrate potentiation and depression characteristics. The desired synaptic device characteristics, such as high linear weight update, small variation, and multiple states were achieved in FeTFTs. In addition, the possibility of FeTFT as a synaptic device in neural network was evaluated by artificial neural network simulation. This work can offer a possibility for FeTFT as a synaptic device to be used in neuromorphic hardware systems.

Emerging low dimensional materials, including both two-dimensional and quasi-one-dimensional materials, provide new optical and electronic properties attractive for developing novel detection and sensing functionalities in the mid-infrared wavelength, including dynamic bandgap tuning covering a broad spectral range in mid-IR, light polarization sensitivity and the ease for electronics integration. Many of these new features are available in the existing mid-IR devices based on traditional semiconductor materials. In this talk, the recent progress in our study of low-dimensional material based mid-IR detector and sensors will be discussed, focusing on layered materials black phosphorus, tellurene and quasi-one-dimensional material BaTiS₃. The common and distinct characteristics of these emerging materials will be examined relating to their crystal structures, optical properties, and the device functionalities. Key performance metrics including the gain and detectivity will be benchmarked, and the new detection features including the polarization sensitivity are studied. I will also provide perspectives about how these new materials may change the future mid-infrared photonics technology.

The search for devices that can emulate the behavior or neurons and synapses, to be used in brain-inspired, in-memory or on-edge computing, is extending among different research fields with an increasing number of materials scientists involved in the quest. However, the number of materials that have been considered as the key elements in devices that are able to produce action potentials and provide tunable conductivity and/or resistive memory, is so far still restricted to the few materials that are compatible with CMOS integration. However, the need to limit the processing to low temperatures, often produces polycrystalline or amorphous materials that diminishes the control of the material’s properties, reduce the device endurance and device-to-device reproducibility and obscure the understanding of the physical mechanisms at play. Thus, in order to acquire a better control of the final device, we propose that a detour into the world of epitaxial, single-crystalline materials should not be seen as a waste of resources but rather as a way to find a direct access to the key materials issues and their solution. In addition, the emergent efforts towards epitaxial growth of complex oxides by ALD[1] allows us to dream of a future of epitaxial oxide microelectronics for which we want to be prepared. In this talk, we will show two examples of epitaxial materials that show interesting neuromorphic features that, to our knowledge, have not been reported in other materials: the particular self-oscillating behavior of epitaxial TbMnO₃ [2] as a compact neuristor [3] and the synaptic plasticity of ferroelectric memristors and ferroelectric field-effect transistors (FeFETs) fabricated with epitaxial, rhombohedral, Hf_0.5Zr_0.5O₂[4], a phase that is different from the ferroelectric phase of hafnia originally discovered[5], whose outstanding properties have put these materials in the spotlight.
H. H. Sonsteby et al. Nature. Comm. 11, 2872 (2020)
M. Salverda & B. Noheda, https://arxiv.org/abs/2004.09903 https://arxiv.org/abs/2004.09903(under review in Nat. Comm.)
M. D. Pickett et al. Nature Mat. 12, 114 (2013); J. del Valle, Scient. Rep. 10, 4292 (2020)
Y. Wei et al, Nature Mat. 17, 1095 (2018); Phys. Rev. Appl. 12, 031001(2019);npj Quan. Mat. 4, 62 (2019); P. Nukala et al. ACS Appl. Elect. Mat. 1, 2585 (2019); arxiv.org/abs/2005.01809 (to appear in the 50th Golden Anniversary Volume of Ferroelectrics, 2020)
T. S. Boescke et al. Appl. Phys. Lett. 99, 102903 (2011); S. Mueller, et al. Adv. Funct. Mater. 22, 2412 (2012)]

Parallel information processing, energy efficiency and unsupervised learning make the human brain a model computing system for unstructured data handling. Oxide memristors, from filamentary to metal-insulator transition devices have been shown to emulate synaptic and neuronal functionalities in artificial neuromorphic circuits. However, challenges like non-linear conductance update, cycle-to-cycle and device-to-device variability and leakage current related issues in a dense crossbar structure are still not resolved. For neurons, often a circuit with multiple active or passive components are required which can complicate circuit designs in a large-scale neural network. In our earlier work, we reported solution-processable ferroelectric tunnel junctions (FTJs) with P(VDF-TrFE) copolymer barriers on semiconducting bottom electrode can show analog memristive behavior with a broad range of accessible conductance states and low energy dissipation of 100 fJ for the onset of depression and 1 pJ for the onset of potentiation by resetting small tunneling currents on nanosecond timescales. Key synaptic functions like programmable synaptic weight, long and short-term potentiation and depression, paired-pulse facilitation and depression, and Hebbian and anti-Hebbian learning through spike shape and timing-dependent plasticity are demonstrated. 1 In the current work, we demonstrate that by manipulating the carrier concentration of the semiconducting bottom electrode in FTJs, it is possible to control the dynamics of ferroelectric domain rotations in a way so that from synaptic devices they can act similar to neuronal devices. 2 These results offer a promising outlook for the FTJ memristors on semiconducting bottom electrodes as both synapse and neuron devices in artificial neural networks by controlling carrier doping concentrations only. References [1] S. Majumdar, H. Tan, Q. Qin and S. van Dijken. Adv. Electron. Mater. 5, 1800795 (2019). [2] S. Majumdar, H. Tan, I. Pande and S. van Dijken. APL Mater. 7, 091114 (2019).

Redox-based memristive devices are one of the most attractive candidates for future non-volatile memory applications and neuromorphic circuits. Many recent studies reported that resistive switching phenomena in transition metal oxides (TMO) by an external bias are associated with the migration of oxygen ions. Here, we study Sr2Fe2O5 brownmillerite structure on the SrRuO3/SrTiO3 substrate and Nb: SrTiO3 substrate as a bottom electrode in order to understand the role of oxygen migration in resistive switching. Topotactic phase changes between the brownmillerite structure and perovskite structure appear by the oxygen ion migration. Using the fast ion path on the brownmillerite structure, it is possible to impede the random movements of oxygen vacancies. In-situ electric biasing holder was used to investigate the changes in atomic configurations as well as electronic structure caused by the transition of the oxygen vacancies. Electric biases were applied to the TEM samples directly in real time which makes it possible to compare the structure changes in the same TMO regions through SET and RESET cycles. To understand the mechanism of whether the oxygen ions are coming from the top or the bottom electrode in the topotactic phase change, the experiment was carried out with different kinds of bottom electrode. This provides a deeper understanding to control the oxygen ions of the redox-based memristive devices to improve and optimize performances.

Almost 40 years after Chua introduced his theory of memristive devices (nowerdays aka resistive switching devices), they have been identified as promising candidates for non-volatile memory applications due to their distinct key features which are i) low power consumption, ii) passivity, and iii) scalability into the nanometer scale. Beyond these applications as non-volatile memories, memristive devices have turned out to be applicable as artificial synapses in neuromorphic circuits. Many of the available metal-oxide resistive switching devices rely on the stochastic phenomenon of creation and rupture of conductive metal filaments within a solid state electrolyte matrix. Inherent are initial electroforming procedures and distinct “on” and “off ” states. The latter are in fact beneficial for memory applications. However, in the context of neuromorphic applications a more "analog" behavior is needed, rather than a purely “digital” behavior. To overcome these limitations, various kinds of interface-based resistive switching devices have been developed, tested, and applied. Although these devices are quite simple with respect to their structural compositions, the analog switching behavior itself - as a result of multiscale multiphysical phenomena - is in fact rather complex and not fully understood yet. This contribution is intended to provide an overview over our modeling and simulation approach in order to describe memristive devices on the human timescale, yet allowing for physical phenomena on the atomistic scale.

Resistive cross-point devices arrays that can simultaneously store and process data locally and all in parallel, are promising candidates for intensive deep neural network (DNN) training workloads. However, when the training is performed using the conventional state of the art learning algorithm, stochasticity gradient descent (SGD) using backpropagation (BP) algorithm, these array architectures must meet a set of stringent device requirements for the cross-point element. A key requirement is that the resistive devices must change conductance in a symmetrical fashion when subjected to positive and negative pulse stimuli. Here, we present an alternative training algorithm, so-called the Tiki-Taka algorithm, so that this stringent symmetry requirement of the cross-point element can be relaxed to a very large extend. We tested the validity of the Tiki-Taka algorithm on a range of network architectures such as fully connected, convolutional and LSTM networks. For all these networks, the Tiki-Taka algorithm with nonsymmetric devices characteristics results in training accuracies that are in par with the ones achieved with the conventional SGD algorithm with symmetric devices. Thanks to the relaxed device specification, this algorithmic improvement would enable earlier adaptation of crossbar-based hardware accelerators for DNN training workloads.

Phase change memory is currently considered as one of the promising candidates for the next-generation non-volatile memory devices and Resistive Processing Units for neuromorphic computing. Multilevel functionality of the PCM is a crucial next step in enhancing the bit density per unit cell. We investigated the potential of Tellurium-free group III-Sb alloys, specifically Al-Sb and Ga-Sb thin multilayer films deposited under UHV conditions by molecular beam epitaxy for obtaining multiple resistance states. Bulk films of group III-Sb binary alloys of different compositions showed intermediate resistance plateaus upon heating which is a promising first step to achieve multilevel functionality in PCMs. The resistance versus temperature curves of the multilayer films exhibited up to four distinct intermediate states that have the capability of stabilizing multiple PCM resistance levels and potentially reduce drift in PCM arrays. The calculated activation energy for crystallization ranged from 2.5-5.3 eV. Naturally transitions with the highest activation energies show projected data retention as high as 10 years at 160°C. Transmission Electron Microscopy (TEM) analysis of the samples revealed crystalline grain size of 30-100 nm upon annealing. TEM and X-Ray Diffraction (XRD) analysis of Ga-Sb samples annealed at different temperatures showed growth of crystallites mostly in A7 Antimony phase with further tendency to phase separation of Sb and GaSb phases. Initial pulse measurements were done on binary alloy mushroom PCM cells with TiN heater size of down to 100 nm. Test cells demonstrated reversible switching with a set/reset resistance ratio of ~100 for set voltage <1.5 V and pulse widths down to 200 ns. The multilayer film has the potential to be used similarly to the binary cell showing reversible switching. We believe the results obtained in this study demonstrate a path to enable multiple logic states in PCM materials.

The need to reduce the power consumption of machine learning applications while improving performance has driven the development of specialized hardware accelerators. While digital CMOS-based accelerators have sped-up training and inference in neural networks, they have several limitations. Most notably, a spatially-separated memory and computation, volatile memory, and the lack of connectivity between nodes, all hinder the realization of ubiquitous, low-power deep learning accelerators. Alternatively, crossbar arrays of non-volatile memory (NVM) devices able to perform arithmetic (e.g. bit multiplication), are projected to be six orders of magnitude more energy-efficient than current state-of-the-art GPU accelerators. Borrowing bio-related terminology, these analog “synapses” perform in-memory computation, reducing the distance data travels and overall energy consumption. Unfortunately, current state-of-the-art analog “synaptic” devices based on resistive memories suffer from stochastic, asymmetric, and non-linear weight updates, detrimental to training accuracy. Meanwhile, electrochemical ion-based synaptic devices have been shown to be fast, energy-efficient, and exhibit symmetric, linear weight updates. However, the electrolytes for the electrochemical reaction are often CMOS incompatible and suffers from scalability.

Scalable, CMOS-compatible devices are needed for future large-scale production. In this project, we develop a MOS field-effect transistor(FET)-based analog synapse. Doping the SiO₂ gate oxide with protons by exposing the oxide to hydrogen plasma, allows non-volatile electrostatic control of the semiconductor channel conductance. Applying a short voltage pulse on the gate shifts the distribution of the protons, tuning the artificial neural network node weights. Ion trapping in the oxide maintains non-volatility. Electrostatics of a uniform sheet of ions predicts a linear relationship between the number of voltage pulses and the threshold voltage shift. Linear weight updates are expected, necessary for efficient training with crossbar array accelerators. Initially, we study the electrostatics and proton diffusion via Silvaco Atlas. Simulations predict a threshold voltage shift due to the presence of ions at different locations in the gate oxide. We then fabricate n-Si/ALD SiO₂/Al MOS capacitors with proton doping, and characterized their ionic transport. We observe that the MOS gate stack exhibits hysteretic behavior below 2V, indicating non-volatility and low-voltage operation. C-V measurements estimate the ionic concentration to be 10¹²cm^-2. To demonstrate switching characteristics, we fabricate and measure proton-doped GaN FETs as a quick testbed due to its simple fabrication and high carrier concentration. GaN has been intensively investigated as a BEOL material for future 3D integrated Si CMOS systems. Our results indicate that it could also be used for MOSFET synapses with proton-doped gate-oxides. These results are generalizable beyond GaN and they shed light on the feasibility of scalable, CMOS-compatible ionic devices for the next generation of neural network hardware accelerators.

Hardware implementation for neuromorphic computing aims to overcome the von Neumann bottleneck by co-locating memory and logic. In this context, resistive switches, phase-change devices, electrochemical systems, and complementary metal-oxide semiconductor (CMOS) transistors have been explored extensively for artificial synapses and spiking neurons in artificial neural networks. While each of these strategies has its attributes, none of them has clearly emerged as a universal approach for neuromorphic hardware, which suggests that alternatives should continue to be evaluated. For example, neuromorphic devices based on nanomaterials such as quantum dots, nanotubes, and two-dimensional (2D) monolayers have begun to show promise due to their controlled chemical composition, unique architectures, and compatibility with ultimate scaling and speed limits [1]. Furthermore, low-dimensional materials possess low dielectric screening that provides enhanced electrostatic control, thus allowing functions that better mimic a biological neuron [2]. In this presentation, recent advances in the development of artificial synapses and neurons based on 2D semiconductors and mixed-dimensional van der Waals heterojunctions will be discussed [3]. In particular, dual-gated memtransistors based on polycrystalline MoS₂ show novel synaptic responses in crossbar arrays [4]. The non-volatile memory states of these devices are tuned by both biasing history and the field-effect from the two gate terminals [5]. In this manner, the dual gates allow independent addressing of nodes in the crossbar by minimizing sneak currents in addition to facilitating tunability of synaptic weight update rules that resulted in a 94% recognition rate of hand-written digits using a multi-level perceptron model [4]. Since neuromorphic architectures require both artificial synapses and neurons, this talk will also introduce a Gaussian heterojunction transistor design that realizes the functionality of an integrate-and-fire neuron [6]. Specifically, dual-gated, self-aligned, mixed-dimensional p-n heterojunctions based on carbon nanotubes and MoS₂ operate as anti-ambipolar transistors where all parameters of the Gaussian transfer curves can be controlled by the applied gate potentials. These Gaussian heterojunction transistors enable a wide range of spiking behavior including constant spiking, latency, and bursting [6]. Since atomically thin dual-gated memtransistors and Gaussian heterojunction transistors are prepared using standard materials processing and fabrication schemes, they are well-poised to be integrated together and/or with traditional electronic devices for next-generation neuromorphic computing architectures.

References

[1] V. K. Sangwan and M. C. Hersam, Nature Nanotechnology, DOI: 10.1038/s41565-020-0647-z (2020).
[2] M. E. Beck and M. C. Hersam, ACS Nano, 14, 6498-6518 (2020).
[3] D. Jariwala, T. J. Marks, and M. C. Hersam, Nature Materials, 16, 170-181 (2017).
[4] H.-S. Lee, V. K. Sangwan, H. Bergeron, H. Y. Jeong, K. Su, and M. C. Hersam, submitted, 2020.
[5] 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).
[6] M. E. Beck, A. Shylendra, V. K. Sangwan, S. Guo, W. A. Gaviria Rojas, H. Yoo, H. Bergeron, K. Su, A. R. Trivedi, and M. C. Hersam, Nature Communications, 11, 1565 (2020).

Biological systems have evolved to perform computing tasks while consuming little energy. Taking inspiration from the methods used by biological systems is therefore a promising path for building low energy computers, as demonstrated by the large body of work on neuromorphic computing. Here we propose to take inspiration from ant colonies, and in particular the way they are able to solve the shortest path in a graph problem without supervision, in a massively parallel way. Algorithms inspired from ant colonies have been proven useful for tackling the wide range of optimization problems that map to finding the shortest path between two given nodes in a graph. However, on conventional computers, they are mostly sequential and therefore the time and energy required to find the path scales with the size of the graph. Here propose a different approach: to implement this bio-inspired system directly in hardware, using circuits of memristors. The graph to study is constructed by interconnecting memristors so that the length of each edge of the graph is embodied by its number of memristors. When a voltage is applied across two nodes, the least resistive branch – i.e. the shortest path – between them receives more current than the other branches. This causes the memristors on the shortest path to turn on, attracting even more current. Our method leverages this reinforcement mechanisms intrinsic to memristors physics, in order to obtain a steady state where the memristors on the shortest path have a much higher conductance than the memristors on the rest of the graph. In consequence, the user can find the shortest path by a simple conductance measurement. Using numerical simulations, we show that using a voltage ramp makes this method reliable. We validate it on graphs of different sizes and topologies. It is both valid for an experimentally derived memristor model and robust to device variability. We found that, contrary to conventional methods, the time and energy of the computation scale with the length of the shortest path rather than with the size of the graph. In consequence, this method is particularly attractive for solving large graphs with small path lengths (such as neural networks, social networks and power grids). Applying our method to these graphs opens the door to fast and low energy analysis of complex networks with wide ranges of applications.

Decision-making belongs to one of the most prominent principles in the nerve system of living species. The crucial trade-off between time to decide (speed) and accuracy is essential for a creature to react to environmental stimuli and most importantly, relevant for survival. In-depth studies of coherent motion tasks in primates, based on two-choice decision experiments, led to a profound understanding of neuronal information processing and even bridged the gap to behavioural psychology. Three fundamental processing stages are needed to successfully perform a two-choice decision experiment based on a saccadic eye movement task: the neuronal representation of the sensory signal, the integration of the incoming stimuli and the comparison of the accumulated information to a threshold for a final decision. In accordance to the results of decision-making experiments with primates, an analogue electronic decision-making circuit was developed. Our concept study includes a LED-matrix showing moving light dots (i.e. the tasks screen in primate experiments), an array of photo diodes (i.e. retina of the primate, sensory evidence), a Hassenstein-Reichardt Detector (HRD) based motion detection (i.e. a correlation-type motion detector with a high biological relevance, based on an inhibitory circuit and delay lines) and finally a signal integration circuit. The latter stage represents the decision-making unit and is realised through an inhibitory mutual coupling of the accumulated signals. The entirely analogue circuit is based on integrate and fire relaxation oscillators to mimic firing neurons. We would like to emphasise that the presented real time decision-making system is closely related to biological information pathways in nerve systems. Neither a digital computer nor programming is necessary for decision-making, which might be an interesting approach for real time data processing in autonomous robots.

Biomimetic electronic devices such as electronic synapses that emulate both short-term and long-term plasticity constitute an auspicious building block for neuromorphic computing systems and bioinspired electronic applications. The electronic synapse constitutes an energy efficient and versatile solution for new computing paradigms, however, to date such devices with well-controlled plasticity and synaptic behaviour are scarce. Most of the current fabrication processes are limited to conventional techniques, which are typically time-consuming and use ecologically harmful chemicals with considerable material waste. The present work aims on fully inkjet-printed MAM (metal-active-metal) electronic synapses. The fabrication technique introduces an alternative time- and cost-efficient method of low environmental impact. All of the layers are deposited in an additive manner without any material subtraction or chemical treatment. Active inks consisting of TiO2 nanoparticle suspension developed in-house are presented and characterised in this work. The ink synthesis is based on low temperature and low complexity process that includes a mixture of solvents and acids. The active inks are analysed in terms of nanoparticle and morphological characteristics. Additionally, the resulted Z number [(inertial force x surface tension)1/2 / viscous force], which defines the jettability of an ink, are defined experimentally. The formulations used seem to strongly affect the dispersity and sedimentation along with the nanoparticle size of the solution-ink, indicating a route to optimization for improved jettability, shelf life and in general ink robustness. The developed nanolayers of the fully-printed devices are in the range of 100-200 nm thickness, uniform and crack-free. Electrical characterisation of initial devices indicates low power (?m to 1 mA window of output current) and biomimetic properties of synaptic plasticity when trigger voltage pulses of 1 V are applied. The tuning of the active ink's rheological characteristics along with the investigation of the electrical framework under which the electronic devices respond synaptically present a favourable approach to alternative fabrication methods for versatile and flexible electronics.

The deep learning in medical image processing had a significant impact on computer-aided diagnosis for the last decade. Although the advanced medical image processing on deep learning allows improvements in image detection, recognition, segmentation, and registration, some of these applications still suffer from the low quality of input images directly from analogue sensors like photodetectors. Here we suggest that an energy-efficient neuromorphic computing platform where low-quality bioimages can be directly processed at edge without energy-hungry image signal processor (ISP). Even with a small convolutional kernel operation for image processing, a peak signal to noise ratio (PSNR) of images from GaAs detectors has been subsequently improved from 16.25 to 19.26 in ten random bio cell images with > 30 % energy savings compared to silicon-based ISP technology. In a segmentation task after image processing, a cell detection accuracy increases from 75 % to 95 % and an detection error rate decreases from 10 % to 2 %. We believe that our edge neuromorphic computing solution can offer fast image processing time and low energy consumption, and thus provide quick and reliable data while maintaining intuitive and safe image reading.

Device transport processes in the switching of analog-based memristors rely on device conduction mechanism, whose complexity still cannot be captured by traditional modelling approaches. Here we apply the concept of the unique combination of electronic and thermal considerations to modelling device conduction in such systems. The approach can fully account for the rich conduction behavior, as well as analog-type resistive switching, and is therefore, ideal for device films, composites and other geometries. Modelling of the analog-based memristor agrees well with experimental data and reveals conduction behavior not yet captured by conventional models. Fitting of device measurements even allows conduction properties to be extracted, which are subsequently harnessed to guide the prediction of analog-based memristors with improved device performance. Our hybrid approach, together with a simple calculation tool, now allows researchers with limited computation experience to perform realistic simulations.

The control of the dielectric and conductive properties of device-level systems is important for increasing the efficiency of energy- and information-related technologies. In some cases, such as neuromorphic computing, it is desirable to increase the conductivity of an initially insulating medium by several orders of magnitude, resulting in effective dielectric breakdown. Here, we show that by tuning the value of the applied electric field in systems with variable electric permittivity and conductivity, e.g. ion intercalation materials, we can vary the device-level electrical conductivity by orders of magnitude. We attribute this behavior to the formation of filament-like conductive domains that percolate throughout the system.

Resistive switching cells are one of the strong candidates for synaptic devices thanks to analog controllability of the cell resistance. Many previous studies have reported continuous increase in the cell resistance when a voltage is applied to the cell, so-called gradual reset [1]. Nevertheless, the origin of the gradual reset is not clear because it is extremely difficult to determine an ultrafine region where resistive switching occurs. In this study, we investigated depth distribution of chemical compositions in the entire fabricated sample by high-resolution Rutherford backscattering spectrometry (HRRBS) without sample sputtering. Moreover, we discuss correlation between the depth distribution and electrical characteristics in TaO_x-based resistive switching cells.

We fabricated two types of Metal/tantalum-oxide (Ta₂O₅)/platinum (Pt) stack structures on SiO₂/Si substrates. As common fabrication processes, a Pt bottom electrode was deposited by DC sputtering. Then, a Ta₂O₅ film was deposited by radio-frequency (RF) reactive sputtering. Finally, a metal layer was deposited by RF sputtering. We used nickel (Ni), titanium (Ti), and Ta as materials of the metal layer. As for one type of the Metal/Ta₂O₅/Pt structure, the metal layer covered a 5-nm-thick Ta₂O₅ film. The thickness of the metal layer was 10 nm. For these samples, we investigated depth distribution of chemical compositions by the HRRBS. The depth distribution can be obtained without sample destruction, which means intrinsic data are not hidden by a knock-on effect. Moreover, the HRRBS has a higher depth resolution on the order of nanometers than the normal RBS. Then, the actual depth distributions were obtained by fitting energy spectra simulated from assumed depth distributions to experimental energy spectra. Note that RBS measurements at two scattering angles of 77 and 122 degrees improved accuracy in the depth distributions by a comparison of the obtained data each other in the same sample. As for the other type of the Metal/Ta₂O₅/Pt structure, the metal layer was deposited with a thickness of 20 nm and a diameter of 150 μm by using a mask as a top electrode on a 10-nm-thick Ta₂O₅ film. We examined resistive switching characteristics for these Metal/Ta₂O₅/Pt cells at room temperature with the Pt bottom electrode grounded.

Comparing the depth distributions of the as-deposited Metal/Ta₂O₅/Pt samples, thickness of the oxygen diffusion area in the Metal/Ta₂O₅ interface, which is less than 10 nm, increased in the order of Ni/Ta₂O₅/Pt, Ta/Ta₂O₅/Pt, and Ti/Ta₂O₅/Pt samples. The diffusion of oxygen from the Ta₂O₅ film to the metal layer corresponds to supply of oxygen vacancies from the metal layer to the Ta₂O₅ film. Here, the standard Gibbs energy of oxide formation of the metal layer is −4.39 eV (Ni), −7.92 eV (Ta), and −9.21 eV (Ti) [3], indicating that more oxygen vacancies tend to be supplied to the Ta₂O₅ film as the magnitude of the standard Gibbs energy is higher.

The Ni/Ta₂O₅/Pt cells show relatively high initial resistance at 0.1 V of roughly 10¹² Ω before a forming process. On the other hand, the Ti/Ta₂O₅/Pt cells show lower initial resistance of about 10⁶-10⁹ Ω because more oxygen vacancies are supplied to the Ta₂O₅ film. Furthermore, when a negative voltage is applied to the Ni/Ta₂O₅/Pt cells, the cell resistance abruptly increases, approximately at −1 V from 100 Ω to 10 MΩ. In contrast, in the case of the Ti/Ta₂O₅/Pt cells, the cell resistance gradually increases from 1 kΩ to 100 kΩ when a voltage is applied from −1 V to −3 V.

These results suggest that the analog controllability of the cell resistance in the reset processes can be improved by supplying a larger number of oxygen vacancies to the Ta₂O₅ film.

[1] D. Ielmini, Microelectronic Engineering, 190, 44 (2018).
[2] T. Miyatani et al., MRS Advances, 4, 2601 (2019).
[3] C. Chen et al., J. Appl. Phys., 114, 014502 (2013).

Introduction
Ferroelectric nondoped HfO₂ has advantages compared to the doped HfO₂ in the reduction of crystallization temperature and threshold voltage (V_TH) control to realize the Metal-Ferroelectrics-Si FET (MFSFET) [1,2]. We have reported that the ferroelectric nondoped HfO₂ formation on the Si(100) substrates by controlling the Ar/O₂ flow ratio during the reactive sputtering [3-5]. In this paper, we have investigated the effect of post metallization annealing (PMA) utilizing Pt gate electrode for the MFSFET with ferroelectric nondoped HfO₂ gate insulator.

Experimental Procedure
After the active region patterning, n⁺ source/drain (S/D) ion implantation was carried out followed by the activation annealing at 900^oC/20 min in N₂ ambient. Then, the 1 nm thick Hf interlayer was deposited to improve the film quality followed by the 20 nm thick HfO₂ in-situ deposition at room temperature (RT) by the RF magnetron sputtering with the Ar/O₂-plasma utilizing Hf target. The flow rate of Ar for Hf interlayer and Ar/O₂ for HfO₂ was 3.6 and 2.3/0.5 sccm, respectively, with the sputtering power of 60 W. Then, the 20 nm thick Pt gate electrode (G) was in-situ deposited at RT with the Ar flow rate of 4.4 sccm and the sputtering power of 80 W. Next, the S/D contact region was patterned followed by the Pt layer deposition for S/D. Then, PMA was carried out at 600^oC/30 s in N₂ ambient followed by the G and S/D patterning. The gate length (L) and width (W) was 10/90 μm. The MFS diodes were fabricated in a same process, and the post deposition annealing (PDA) was also examined utilizing evaporated Al gate electrode for the comparison.

Results and Discussion
The remnant polarization (2P_r) of 2.5 μC/cm² was obtained from the P-V characteristics for the MFS diodes formed by the PMA process under the ±8 V sweep which is larger than the 2P_r of 2.3 μC/cm² for the MFS diodes formed by the PDA process. From the C-V characteristics, the memory window (MW) was markedly increased from 0.22 V for the MFS diodes formed by the PDA process to 0.44 V for the MFS diodes formed by the PMA process. This is due to the enhancement of the orthorhombic phase HfO₂ formation by the PMA process which was confirmed by the grazing incident x-ray diffraction (GI-XRD) observation. The physical thicknesses were confirmed as 3.6 nm for SiO₂ interfacial layer (IL) and 20.8 nm for HfO₂ from the cross-sectional transmission electron microscopy (TEM) image. By considering the coercive voltage (V_c) of 1.08 V extracted from the P-V characteristics, the coercive field (E_c) was estimated as 0.52 MV/cm. The ferroelectric HfO₂ formation without doping realized the lower E_c compared to the reported data for the doped HfO₂.
Next, the MFSFET fabricated utilizing PMA process was evaluated. From the I_D-V_G characteristic, MW of 0.41 V was obtained under the V_G sweep from -1 V to 3 V. The program and erase (P/E) characteristics were also examined. The MW obtained from the threshold voltage (V_TH) difference of 0.41 V were also confirmed utilizing the P/E pulses of -5.5 V/100 ms and 6.5 V/100 ms, respectively. The MW of 0.15 V was obtained even at the P/E pulse width of 100 ns.

Conclusions
In this paper, we investigated the effect of PMA for the device characteristics of MFSFET with ferroelectric nondoped HfO₂ formed by the Ar/O₂-plasma sputtering. It was found that the small E_c was obtained for the 20 nm thick ferroelectric nondoped HfO₂. Furthermore, the MW of 0.15 V was realized under the P/E pulse width of 100 ns with P/E voltages of -5.5 V and 6.5 V, respectively.

Acknowledgements
This work was partially supported by JSPS KAKENHI Grant Number 19H00758, NEDO, JST, and CASIO Foundation.

References
[1] T.S. Boscke et al., IEDM Tech. Dig., p. 547 (2011)., [2] F. Winkler et al., DRC Conf. Dig., p. 49, (2019)., [3] M.G. Kim, S. Ohmi et al., Jpn. J. Appl. Phys., 57, 111UF09 (2018)., [4] S. Ohmi et al., Jpn. J. Appl. Phys., 58, SIIB16 (2019)., [5] S. Ohmi et al., DRC Conf. Dig. (2020). [to be presented]

Conductive bridge random access memory (CBRAM) is one of the most promising nanodevices for future information technology with applications not only for memory, but also for logic and neuromorphic computing because of its low power consumption and response times in the nanosecond range, as well as scalability down to the atomic level [1]. Recently, we have reported that several drastic improvements for controllability of memory performance can be achieved by using ionic liquids (ILs) as a tunable solvent supplied into the grain boundaries of metal oxide film [2, 3]. We focused, as a next step, on the electrolyte engineering, in which the electrolyte should not be regarded as a diffusive medium for the migration of metal ions as ever but should be regarded as a porous body retaining ILs which mediates the ion migration in it. Metal-organic frameworks (MOFs), which is three-dimensional nanoporous materials composed of metal ions and organic ligands coordinated together, are the best candidate for the porous body retaining ILs without compromising their features owing to structural and functional flexibility in the design for uniformly sized nanopores, as well as relatively high thermal stability [4]. Although there are some reports so far that investigate the electrical characteristics of MOFs, most of their experiments has been done using pellet or polycrystalline film samples from which it is difficult to extract intrinsic bulk properties due to a lot of grain boundaries [5]. Moreover, ILs-loaded MOFs for the application of CBRAM has not yet been discussed so far.
In this work, basic investigations in terms of electrical properties for HKUST-1 (Cu₃(1,3,5-benzenetricarboxylic acid)₂), which is to be easily synthesized and stable in air, were carried out using single crystalline sample. We clarified the dominant conduction mechanism of the HKUST-1 single crystals which include much less grain boundaries, by measuring the electrical conductivity as functions of a bias voltage, temperature and humidity. Ion conductivities of the IL ([emin][TFSI] (1-ethyl-3-methylimidazolium bis(trifluoromethylsulfonyl)-imide)) loaded into the nanopores of HKUST-1 single crystal were compared with that of bulk IL as a function of temperature using electrochemical impedance spectroscopy. The both activation energies extracted from the temperature dependence of ion conductivities were comparable, suggesting that the HKUST-1 enables to retain the IL without compromising its original property. We, furthermore, demonstrated by means of electromotive force measurement that the ionization of the electrochemically active electrode in a Ag/HKUST-1/Au structure under bias voltage, which is the first important step in CBRAM operation, was drastically promoted by loading the IL into the nanopores of HKUST-1. The ILs-loaded MOFs system provides a novel electrolyte which endow CBRAM with further improved memory performance.
[1] R. Waser et al., Adv. Mater. 21, 2632 (2009).
[2] A. Harada et al., J. Mater. Chem. C 4, 7215 (2016).
[3] K. Kinoshita et al., Jpn. J. Appl. Phys. 56, 04CE13 (2017).
[4] I. Stassen et al., Chem. Soc. Rev. 46, 3185 (2017).
[5] A. A. Talin et al., Science 343, 66 (2014).

Networks of nanoscale objects are subject of increasing interest as resistive switching systems for the fabrication of neuromorphic computing architectures. Nanostructured films of bare gold clusters produced in gas phase with thickness well beyond the electrical percolation threshold, show a non-ohmic electrical behavior and resistive switching resulting in groups of current spikes with irregular temporal organization [1]. Here we report the systematic characterization of the temporal correlations between single spikes and spiking rate power spectrum of nanostructured Au two-terminal devices consisting of a cluster-assembled film deposited between two planar electrodes. By varying the nanostructured film thickness we fabricated two different classes of devices with high and low initial resistance respectively. We show that the switching dynamics can be described by a power law distribution in low resistance devices whereas a bi-exponential behavior is observed in the high resistance ones [2]. The measured resistance of cluster-assembled films shows a 1/f^α scaling behavior in the range of analyzed frequencies. Our results suggest the possibility of using cluster-assembled Au films as components for neuromorphic systems where a certain degree of stochasticity is required.

1 M. Mirigliano, et al., Nanoscale Adv. 1, 3119 (2019)
2 M. Mirigliano, et al., Nanotechnology 31, 234001 (2020)

In pursuit of the goal of mimicking how the human brain learns and functions, emulation of synaptic plasticity, i.e. the ability of synapses to strengthen or weaken over time, is a critical requirement for the implementation of learning in neuromorphic systems. Finding the proper materials to synthesize nanoscale synapses and neurons that are reliable, durable, replicable and scalable remains a challenge. Recent developments on ferroelectric field effect transistor based non-volatile memory are encouraging in this respect; in particular, their multiple threshold voltage states based on the polarization state of a ferroelectric oxide may offer a suitable structure for synaptic applications. This work investigates the ferroelectric behavior of BaTiO₃ (BTO) with and without annealing of the deposited film. 100 nm of BTO was deposited via Pulsed Laser Deposition (PLD) on a Niobium doped Strontium Titanate (Nb:STO) substrate. The process parameters included a temperature of 750 °C, an operating pressure of 10 mTorr under 35 sccm of O₂, and a laser energy of 300 mJ. After deposition, the film was characterized via a coupled 2-theta-omega X-Ray Diffraction (XRD) scan; looking for the peak associated with the tetragonal phase of BTO. This was followed by characterizing the poling/switching characteristics using Piezoresponse Force Microscopy (PFM). Following these initial characterizations, the film was subjected to a 30 minute anneal at 850 °C in 100 Torr of O₂. After this annealing step, the film was once again re-characterized using XRD and PFM. The BTO 2-Theta peak shifted from the annealed vs. non-annealed sample, to around 44.2 °, closer to the expected value for the tetragonal phase. A small region was then poled at +5 V, and the phase of the film monitored for the next 24 minutes. The poling did not result in any damage nor oxidation to the surface, which differed from the non-annealed BTO film. No degradation in the poled region was visible for the duration of measuring, whereas for the non-annealed film, the poled region started relaxing after 8 minutes. Finally, 100 nm of W was sputtered and patterned into 100 µm by 100 µm contact pads, and the film was subjected to electrical testing. This consisted of a forming sweep (PUND – Positive, Up, Negative, Down) of varying pulse lengths and amplitudes, followed by a hysteresis sweep of varying pulse lengths and amplitudes. For each test, an average was taken from 3 different devices. For a given forming sweep amplitude, pulse length, and a hysteresis sweep amplitude, an increasing pulse length of the hysteresis sweep resulted in an increase in the measured remnant polarization from 1.3 µC/cm² to 6 µC/cm². Similarly, for all other variables constant, an increase in the pulse length of the forming sweep resulted in an increase in the remnant polarization from 3.53 µC/cm² to 5.73 µC/cm². And again, all other variables constant, an increase in the forming amplitude resulted in an increased remnant polarization from 2.9E-3 µC/cm² to 5.9 µC/cm². Here we demonstrate multiple, distinct polarization states of BTO as a function of the forming and hysteresis sweep pulse lengths and amplitudes.

Nanogranular golden films present an interesting resistive switching behavior that can be exploited in the designing of components for neuromorphic computing. Its investigation has started only recently [Mirigliano et al. 2019, Mirigliano et al. 2020] and the underlying microscopical physical mechanisms are far from being clear. We address the challenge of its theoretical charaterization from the limited information on the microscopic structure of the films with the tools of Effective Mean Theory, and present a model for thermally regulated local structural rearrangements that may explain the observed data.

Resistance switching memories are promising candidates for low power consumption non-volatile memory applications owing to their dimensional scalability and fast switching speed. Furthermore, the possibility to obtain multiple stable states in such systems is of importance for applications in neuromorphic computing. Among the proposed systems, Cu/SiO₂ based memory devices have demonstrated very promising performances, including quantized conductance with a switching voltage as low as 250 mV and allow easy integration with CMOS technology [1]–[5]. It is known that in such systems the different states originate from the formation of a Cu filament in the SiO₂ solid electrolyte [4]. This offers opportunity to use the different quantum states that occur during the filament growth as different memory states. However, because of its filamentary nature, the devices are very sensitive to local inhomogeneities like defects in the SiO₂ dielectric or interface roughness. As a result, the set/reset voltage can have significant device-to-device variations and the poor repeatability of the different states limits the advantage of such devices for potential applications. Additionally, the retention and endurance of these memories are also a limiting factor and needs to be improved. In order to quantify and understand the origin of the distribution of these characteristics, we present a broad statistical analysis of the current-voltage behavior of Cu/SiO₂/W devices, as such data are missing in the literature. Several batches of samples were prepared, with several dies on each samples and tens of devices on each die. The influence of different voltage sweep rates on the set and reset voltages is investigated. Furthermore, the distribution of the quantum states occurring at different current is also studied by means of voltage-current measurements. We evidence half- and quarter G₀ conductance states. We will discuss the identification of key parameters that can be optimized to reduce device-to-device and sample-to-sample variability.
[1] M. Wang et al., “A novel CuxSiyO resistive memory in logic technology with excellent data retention and resistance distribution for embedded applications,” Dig. Tech. Pap. - Symp. VLSI Technol., pp. 89–90, 2010.
[2] S. P. Thermadam, S. K. Bhagat, T. L. Alford, Y. Sakaguchi, M. N. Kozicki, and M. Mitkova, “Influence of Cu diffusion conditions on the switching of Cu-SiO2-based resistive memory devices,” Thin Solid Films, vol. 518, no. 12, pp. 3293–3298, 2010.
[3] S. Tappertzhofen, S. Menzel, I. Valov, and R. Waser, “Redox processes in silicon dioxide thin films using copper microelectrodes,” Appl. Phys. Lett., vol. 99, no. 20, pp. 1–4, 2011.
[4] S. R. Nandakumar, M. Minvielle, S. Nagar, C. Dubourdieu, and B. Rajendran, “A 250 mV Cu/SiO 2 /W Memristor with Half-Integer Quantum Conductance States,” Nano Lett., vol. 16, no. 3, pp. 1602–1608, Mar. 2016.
[5] A. Mehonic et al., “Silicon Oxide (SiO x ): A Promising Material for Resistance Switching?,” Adv. Mater., vol. 30, no. 43, p. 1801187, Oct. 2018.

Many neural networks are either implemented at the software level or in hardware using CMOS technology. While these have been shown to have unique functionalities, they are not easily scalable to large systems due to energetic considerations. Therefore, it is desirable to develop more energy-efficient solutions by implementing them directly in the hardware using the unique properties of quantum materials. In this fashion, neuronal and synaptic properties maybe implemented to emulate the behavior of biological systems. Recently, it has been shown that the volatile resistive switching in Mott Insulators could be used to implement Leaky, Integrate and Fire (LIF) neuron (“neuristor”) models although scaling of these systems maybe an issue. Very recent studies have shown that combining two physical phenomena may provide the basis for a scalable LIF neuristor. These were based on VO2 and V2O3. In this work, we have synthesized and characterized another vanadium oxide, V3O5, which provides a higher transition temperature alternative. Its electrical properties were studied as a function of temperature, between 300 K and 450 K, across the metal-insulator transition. We show that V3O5 presents the required characteristics as a potential hardware based thermal neuron, and therefore is a promising candidate for neuromorphic computing. Acknowledgements: This work was supported as part of the Quantum Materials for Energy Efficient Neuromorphic Computing (Q-MEEN-C) Energy Frontier Research Center (EFRC), funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences under Award # DE-SC0019273.

Oscillatory Neural Networks are a novel, promising concept as hardware accelerator for neuromorphic computing. Systems of coupled oscillators present associative memory capabilities that can be used to implement tasks such as image recognition [1]. We fabricated compact electronic relaxation oscillators using the phase change properties of VO2, which undergoes a temperature driven metal-to-insulator Mott transition. We studied and developed fabrication techniques for the VO2 coupled oscillators on a Si platform, to retain compatibility with CMOS process. VO2 films were deposited with Atomic Layer Deposition on a SiO2/Si substrate. The as-deposited amorphous VO2 films crystallize during a post-annealing treatment under oxygen flow at a temperature of 450° for 20 minutes. Typically, polycrystalline films with an average grain size between 80 and 50 nm [2] form on the SiO2 surface. The annealed films show a phase transition at 340 K with a resistivity change of 2 to 3 orders of magnitude and a hysteresis width of about 10 K. The polycrystalline film was patterned with E-beam lithography and ICP etching to define scaled VO2 oscillator structures, with dimensions between 200 and 500 nm. The oscillators are tested in their switching characteristics, and the variability impact on their coupling dynamics is addressed. In particular, we investigated the phase transition of the polycrystalline devices using electrical characterization and Scanning Thermal Probe Microscopy (SThM) [3] on a single grain level. The temperature maps obtained with the SThM provide nanometer resolution and allow to observe the thermal signature of single grain switching in the scaled oscillator devices. Hence, the current path can be visualized in the insulating and metallic phases of the VO2 device. The microscopic phase-change properties could be correlated with the characteristic waveforms of the oscillating pattern in a coupled oscillator system. In addition, a milli-second flash-anneal technique was tested and optimized to control the crystallization process. The average grain size could be reduced from 55 nm to 20 nm, resulting in more uniform electrical device characteristics and allowing further scaling of the VO2 oscillators. The devices were tested in vacuum and coupled through external electrical components. A network of coupled oscillators locks in frequency and establishes programmable phase relations depending on the strength of the coupling [4,5]. We demonstrate the capability of the oscillator network coupled with an array of tunable resistors to perform image recognition. Experimental results of a three-coupled oscillator system and simulations on larger networks are shown illustrating the associative memory capabilities of the oscillating neural network. This work is supported by the HORIZON2020 PHASE-CHANGE SWITCH Project (Grant No. 737109). [1] F. Hoppensteadt and E. Izhikevich, Phys. Rv. Lett, vol. 82, 2983-2986, 1999. [2] A. P. Peter et al, Adv. Funct. Mater., 25: 679-686, 2015 [3] F. Menges et al., Rev. Sci. Instr., vol.87, no. 7, pp. 074902, 2016 [4] N. Shukla et al., J. Appl. Phys., vol. 117, no. 5, 2015. [5] E. Corti et al. 2018 IEEE International Conference on Rebooting Computing (ICRC), 2018, pp. 1-7. doi: 10.1109/ICRC.2018.8638626

Natural Intelligence (NI) and Artificial Intelligence (AI) differ in their corresponding analog and digital implementations. One of the consequences of that distinction is the ability of NI to perform logic and memory operations at the local level, without spatially and functionally separating NI computing resources. This functional plasticity of basic units of NI, e.g. nerve cells, is attributed to explain NI’s superior energy and time efficacy in task solving, in particular image and speech recognition/interpretation. Previous attempts to mimic NI analog functionality in digital hardware fall short in terms of dimensional scaling and power requirements. The proposed solution is to develop basic computational units that are better suited for NI tasks, e.g. analog. The main requirements for such devices are scalability down to a few nanometers and their ability to process and store data using minimal power consumption. One of the leading candidates for the realization of these goals is the ‘memristive device’. In these two-terminal devices, the resistivity is a temporal function of an applied electrical field. Due to their inherent plasticity and scalability, memristive devices are envisioned to provide long-term solutions for hardware implementation of neuromorphic algorithms and energy efficient Deep Neural Networks (DNNs).
We present simulation and experimental results for a possible implementation of devices with memristive characteristics. The device functionality is based on a non-stoichiometric oxide thin film, distinctively characterized by coupled ionic and electronic conductivity i.e. a mixed-ionic-electronic-conductor (MIEC). The modeling of the memristive device, in crossbar configuration, is based on the previous theoretical and experimental research on the thermodynamics and transport of oxygen vacancies in nonstoichiometric oxides^1-3 in conjunction with well-developed models of voltage pulse induced self-heating, similar to that used in phase change memories. The praseodymium doped cerium dioxide, Pr_xCe_1-xO_2-δ, was chosen as a test memristive material that is electronically conductive via small polarons and ionically conductive via oxygen vacancies with thermal activation energies of 0.4 eV and 0.8 eV, respectively.⁴ The conductivities, in the latter, can be tuned by changing the praseodymium doping, x, and the oxygen deficiency, δ, levels. The device modelling shows that in order to achieve nano-second ‘write’ times and sufficient non-volatility of the ‘read’ resistance states, both electronic conduction by small polarons and high thermal activation energy of ionic transport, are essential. Finally, a new method for characterization of the ionic transport in MIECs over wide temperature limits, relevant to the memristive device operation temperature (from room temperature up to several hundred centigrade) and structure (crossbar device), will be presented.⁴

[1] S. R. Bishop, T. S. Stefanik, H. L. Tuller, Phys. Chem. Chem. Phys., 2011, 13, 10165.
[2] D. Kalaev, A. Rothschild, and I. Riess, RSC Adv., 2017, 7, 38059.
[3] D. Kalaev, H. L. Tuller, and I. Riess, Solid State Ionics, 2018, 319, 291.
[4] D. Kalaev, T. Defferriere, C. Nicollet, T. Kadosh and H. L. Tuller, Adv. Funct. Mater., 2019 (accepted).

Resistive state change exploiting abrupt Mott phase transition in strongly correlated RNiO₃ (R = rare earth elements) has the potential to enable high speed and densely scaled non-volatile memories for emerging neuromorphic computing applications [1]. While the metal-insulator phase transition can be triggered by electrostatic or thermal excitation, ion doping via chemical reactions as an emerging field can completely reconstruct the electronic band structures by introducing extremely high carrier density resulting in giant resistivity modulation. In recent years, the vast resistivity modulation in the RNiO₃ family has been successfully demonstrated through this chemical route. For example, proton-doped SmNiO₃ (SNO) thin film, where protons are dissociated from hydrogen molecules by utilizing Pt catalytic effect and then doped into the thin film, can lead to an ~10⁸ of colossal and reversible resistivity modulation at 300 K [2]. The colossal resistivity modulation is considered to be governed by proton motion in SNO, such as diffusion in the thin film. At the same time, the kinetics of such external control of proton diffusion, namely electric field, has not been clearly revealed. A comprehension of this matter is of relevance to the use of proton-doped correlated RNiO₃ materials in emerging memory and neuromorphic computing pulls.
In this work, we prepared NdNiO₃ (NNO) thin film resistor with different electrode materials (Pt and Au/Ti) and observed the effect of electric field-assisted hydrogenation. Because the hydrogen dissociates on the Pt but not on the Au/Ti electrode, by using this two-terminal resistor configuration, we were able to reveal that the resistivity modulation in NNO film during hydrogenation is associated with the proton diffusion controlled by the temperature (T) and polarity-dependent electric field (V = ±V). To quantitatively investigate the correlation among them, simulations based on diffusion in semi-infinite solid were carried out for fitting of time-dependent diffusion coefficient (D_i). The estimated D_i were plotted against T^-1 and are in satisfactory agreement with the Arrhenius function of temperature. From this plot, we acquired the activation energy of proton diffusion under positive bias +V (E_a(+) = 0.55 eV) and under negative bias -V (E_a(-) = 0.83 eV) application. Then, a kinetic model based on the estimated E_a(+) and E_a(-) was proposed to describe the correlation between resistivity modulation and proton-lateral diffusion along the NNO channel region. Our kinetic model suggests that the polarities of electric field can ease or blockade the proton-lateral diffusion by decreasing or increasing the diffusion energy activation, respectively. From this relation, we extracted the unadulterated activation energy of proton diffusion (E_a(0)) – which was about 0.69 eV and is slightly larger than that reported of 0.27-0.65 eV for SNO using a first-principles study [3]. The slight margin of that to our result is considered due to the discrepancy between the actual NNO film and the ideal SNO crystal structure. Generally, the value of E_a(0) is important to understand and estimate the significance of external and dynamic control effect, such as electric field in the solid-state proton-gated transistor. In this regard, we finally estimated the efficiency of electric field (η_(E)) as a means to externally and dynamically control the proton-lateral diffusion in a-axis of NNO film following the established model. The estimated η_(E) was 20.38%, which means that despite the generally accepted thermal hopping and concentration gradient effect, electric field can simultaneously be a driving force to diffuse the positively charged proton in NNO film - with a contribution rate of about ~20%. Our detail results and analysis will be discussed in the presentation.
[1] D. Ielmini et al., John Wiley & Sons (2015).
[2] J. Shi et al., Nat. Commun. 5.1 (2014) 1-9.
[3] C, Lan et al., J. Comput. Electron. (2020).

Memcomputing [1, 2] is a novel physics-based approach to computation that employs memory to both process and store information on the same physical location. Its digital version [3, 4] is designed to solve combinatorial optimization problems. A practical realization of digital memcomputing machines (DMMs) can be accomplished via circuits of non-linear, point-dissipative dynamical systems engineered so that periodic orbits and chaos can be avoided. A given logic problem is first mapped into this type of dynamical system whose point attractors represent the solutions of the original problem. A DMM then finds the solution via a succession of elementary instantons whose role is to eliminate solitonic configurations of logical inconsistency (‘‘logical defects’’) from the circuit [5, 6]. I will discuss the Physics behind memcomputing and show many examples of its applicability to various combinatorial optimization problems demonstrating its advantages over traditional approaches [7, 8]. Work supported in part by DARPA, DOE, MemComputing, Inc. (http://memcpu.com/), and CMRR.

[1] M. Di Ventra and Y.V. Pershin, Computing: the Parallel Approach, Nature Physics 9, 200 (2013).
[2] F. L. Traversa and M. Di Ventra, Universal Memcomputing Machines, IEEE Transactions on Neural Networks and Learning Systems 26, 2702 (2015).
[3] M. Di Ventra and F.L. Traversa, Memcomputing: leveraging memory and physics to compute efficiently, J. Appl. Phys. 123, 180901 (2018).
[4] F. L. Traversa and M. Di Ventra, Polynomial-time solution of prime factorization and NP-complete problems with digital memcomputing machines, Chaos: An Interdisciplinary Journal of Nonlinear Science 27, 023107 (2017).
[5] M. Di Ventra, F. L. Traversa and I.V. Ovchinnikov, Topological field theory and computing with instantons, Annalen der Physik 529,1700123 (2017).
[6] M. Di Ventra and I.V. Ovchinnikov, Digital memcomputing: from logic to dynamics to topology, Annals of Physics 409, 167935 (2019).
[7] F. L. Traversa, P. Cicotti, F. Sheldon, and M. Di Ventra, Evidence of an exponential speed-up in the solution of hard optimization problems, Complexity 2018, 7982851 (2018).
[8] F. Sheldon, F.L. Traversa, and M. Di Ventra, Taming a non-convex landscape with dynamical long-range order: memcomputing the Ising spin-glass, Phys. Rev. E 100, 053311 (2019).

Compared to all other neural networks, Spiking Neural Networks (SNN) mimic the brain most faithfully. Memristors are considered as promising candidates for realizing these SNNs because of their reduced complexity and nanoscale foot-print when compared to conventional complementary metal oxide semiconductor (CMOS) technology [1]. Volatile switching in memristors is used to emulate neurons in SNNs. In this work, we present neurons realized by memristors using chemical vapor deposited (CVD) MoS2. The neurons realized by volatile MoS2 threshold switching memristors (TSM), facilitated by using Ag electrodes, exhibit the characteristics of a leaky integrate and fire (LIF) neuron. The LIF neurons exhibit temporal response modulated by circuit elements and input voltage pulses. The volatile switching is obtained due to the diffusion of Ag+ ions into Au/MoS2/Ag stack. The device shows exceptional characteristics like high ON/OFF ratio of 106, low threshold voltage (~0.2 V), high compliance current (100 µA), and a pulsed endurance of 107 switching cycles making it an exceptional candidate for artificial neuron application. The volatile switching characteristics are utilized to implement a LIF neuron. The TSM is integrated with a series resistor and parallel capacitor. The device shows continuous spiking when subjected to input pulses of amplitude of 0.5 V and duration of 5 µs. The output spiking frequency of the neuron increases with increasing amplitude and pulse width. The variation in the spiking frequency follows the sigmoid activation function of ANN. The spiking frequency increases with an increase in the input pulse width. The spiking frequency decreases with increasing pulse interval. References: [1] P.Morella et al., IEEE Custom Integrated Circuits Conference, 2011.

Memristors are considered as promising candidates for realizing synapses for neuromorphic computing because of their reduced complexity and nanoscale foot-print when compared to conventional complementary metal oxide semiconductor (CMOS) technology. Non-volatile switching with multiple conductance states in memristors is employed to emulate synapses. In this work, we present synapses realized by memristors using chemical vapor deposited (CVD) MoS2. Graphene is used as electrode for this device. These synapses exhibit linear and symmetric weight update with identical pulses, considered to be the “holy grail”for unsupervised learning. The graphene/MoS2 (G/M) synaptic device exhibits 4 distinct states in both DC potentiation and depression regimes. Additionally, it exhibits data retention for >104 s. For efficient unsupervised learning, the synapses should exhibit linear and symmetric weight update with identical input pulses and a high maximum/minimum conductance ratio (Gmax/Gmin). Resistive random-access memory (RRAM), phase change memory (PCM) and ferroelectric field effect transistors (FeFET) have been used to emulate a synapse. But these devices exhibit non-linear weight update with identical pulses paired with reduced Gmax/Gmin ratio. Identical input voltage pulses with tON= 500 µs and amplitude of 6 V (-4 V) are applied to observe potentiation (depression) in the G/M synapse. With 25 input pulses applied for each regime, we obtain linear and symmetric weight update with an ideal symmetry of 0. The non-linearity factor (NLF) is 0 for both potentiation and depression regimes, and Gmax/Gmin is 16.5. These device characteristics are then incorporated in a pseudo-crossbar array of a 2-layer multi perceptron network with the help of the circuit macro model Neurosim1 to obtain online learning accuracy of 84.33 %. We benchmark our G/M synapse with other emerging synapse devices with respect to NLF, asymmetry, Gmax/Gmin and learning accuracy with Neurosim and show that our device is superior to all other emerging synaptic devices in all these categories. 1 P.-Y. Chen, X. Peng, and S. Yu, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 37 (12), 3067 (2018).

Due to the unique physical properties, charge-density-wave (CDW) materials have been vastly investigated in the past half century, including the direct observation of atomic structure, phase transition, transport mechanism, and applications ranging from superconductor to electronic oscillator. As a layered CDW material, the CDW phase transition of 1T-TaS2 can be incurred by external electric field at room temperature, accompanied by the significant decrease of resistance. The electric-field-driven phase transition makes a room temperature oscillator possible. Furthermore, the quench-and-melt dynamics leads to the random reconfiguration of CDW domains during phase transition, which implies the intrinsic stochasticity of the oscillator. Therefore, here we report a CDW stochastic neuronal oscillator based on 1T-TaS2 thin film. Two working regimes, the regular and stochastic regimes, were found in this device. In stochastic regime, the CDW neuronal oscillator can generate spike trains with random interspike intervals following exponential, Gaussian and Gamma distribution with tunable parameters, closely matching the spike trains of biological neurons. Owing to the stochasticity, many key features of the Hodgkin-Huxley description of neurons can be realized in this compact two-terminal neuronal oscillator. Statistical analysis was conducted on the spike train generated by the CDW neuronal oscillator. By establishing a generalized linear model to describe the statistical distribution of interspike intervals and to quantify the correlation, we demonstrated that the artificial neuron resembles the neurons in the superior olivary complex of mammalian nervous system, in terms of its interspike interval distribution, the time-correlation of spiking behavior and its response to acoustic stimuli. The CDW oscillator can hence provide a new pathway, utilizing statistical features and pattern change in the interspike interval to transmit information, to the design of neuronal component for more sophisticated spiking neural networks and brain emulation.

Neuromorphic chips which mimic the human brain working with extremely low power consumption of 20 W are intensively studied, because current artificial intelligence systems require enormous power consumption. Resistive switching (RS) cells as synaptic devices play an important role in neuromorphic chips by controlling the cell resistance continuously. However, conventional transition metal oxide-based RS cells often show an abrupt decrease in the cell resistance (set process) [1], which inhibits the continuous control of the cell resistance. In this paper, we report improvement of RS continuity in Ta2O5-based RS cells by suppressing the abrupt set process. We fabricated Pt/TaOx/Ta2O5/Pt stack cells on SiO2/Si substrates. A Pt layer was deposited by DC sputtering as a bottom electrode (BE). A Ta2O5 layer and an O-deficient TaOx layer were deposited by reactive radio frequency sputtering. Pt top electrodes (TEs) with a diameter of 100 ?m were subsequently formed. The introduction of the TaOx layer caused diffusion of oxygen vacancies (VOs) from the TaOx layer to the Ta2O5 layer. Consequently, the Ta2O5 layer can be considered to consist of a VO-rich layer and a VO-poor layer for simplicity [2]. We measured electrical characteristics of the cells at room temperature with the TE grounded. We reported that two modes of bipolar RS (BRS), which show hysteresis curves of current-voltage (I-V) characteristics in opposite directions each other, were selectively obtained in a single Pt/TaOx/Ta2O5/Pt cell by appropriately adjusting a range of the applied voltage after a unique forming process [3]. One mode of BRS exhibits set and reset processes at positive and negative voltage, respectively, that is, RS characteristics in the direction of drawing the figure-eight. The other mode of BRS conversely presents RS characteristics in the counter direction. In this study, the maximum applied voltage after the forming was adjusted between the voltage of 1.0 V for the figure-eight BRS and that of 2.5 V for the counter figure-eight BRS. As a result, we demonstrated complementary RS (CRS) from ?1.5 V to 1.5 V, in which a set process and a subsequent reset process occur symmetrically in both bias directions as the magnitude of the applied voltage increases. The CRS characteristics have been expected as one of the ways to solve a sneak path problem. The typical CRS cell consist of anti-serially connected two BRS cells, and its characteristics can be generally interpreted by superposition of the two modes of BRS. However, we found that the stack structure of the VO-rich and VO-poor layers could exhibit stable CRS characteristics. This suggests that each of the two modes of BRS is dominated by RS phenomena which occur in the VO-rich and VO-poor layers. Moreover, we could separate two types of BRS operations with different polar for set and reset processes each other by changing the range of the applied voltage from ?0.5 V to 1.0 V or from ?1.0 V to 0.5 V after several CRS operations. Note that these RS characteristics were sufficiently gradual. Therefore, the cell resistance could be continuously controlled approximately between 0.2 k? and 1 k? by applying the appropriate voltage, indicating that analog RS characteristics were improved through CRS operations. The improvement of analog RS characteristics can be understood by the competition of a set process in the VO-rich layer and a reset process in the VO-poor layer, or vice versa, through the CRS operations. In other words, the CRS operations can be useful to not only suppress sneak pass current but to improve the continuity of RS. [1] D. Ielmini, Microelectronic Engineering, 190, 44 (2018). [2] T. Miyatani et al., Jpn. J. Appl. Phys., 58, 090914 (2019). [3] T. Miyatani et al., Mater. Res. Soc. Spring Meeting, EP09.07.02 (2019).

Deep neural network (DNN) based on the backpropagation(BP) has demonstrated outstanding performance in speech recognition, image detection, and natural language processing. Despite the great success, training of large DNNs is computationally intensive task that requires high energy consumption. To achieve massively parallel and energy-efficient DNN algorithms, in-memory computing with resistive memory devices such as RRAM, PCM, CBRAM, FRAM, etc., have been widely studied. These analog synaptic devices, however, have non-ideal characteristics including nonlinearity of conductance response, asymmetry, and variation originated from their operating mechanism or material properties, resulting in time consuming feedback programming scheme. In this contribution, we propose metal oxide based three-terminal electrochemical memory devices as artificial synapses for neuromorphic computing. Various metal oxides are employed as resistive channel material and a solid electrolyte deposited on the channel decouples the update and read operations. By injecting proton or oxygen ions into the channel, the conductance states of the channel can be tuned precisely. We investigate device structure designs and operation conditions for symmetrical programming with a wide range of states and long term stability through the reversible electrochemical intercalation or redox reaction with a charge control through the gate. More importantly, we discuss issues of the metal oxide based three-terminal electrochemical memory devices for CMOS process-compatibility and a large scale array fabrication.

As artificial intelligence (AI) becomes a popular feature in industry, social networks, smart homes and transportation, there is an increasing need for AI accelerators in edge computing, to solve for excessive traffic to the cloud and ensure privacy and security of the processed data. In this frame, in-memory computing (IMC) has risen as low-power processing approach for data intensive problems. Matrix-vector multiplication (MVM) can be accelerated in crosspoint arrays by leveraging the Kirchhoff’s law for current summation and the Ohm’s law for voltage-conductance multiplication, as well as taking advantage of the parallelism in the memory array. Connecting the crosspoint array in a feedback loop also allows to accelerate inverse operations, i.e., one-step solutions of linear systems such as Ax = b where the NxN matrix A is stored as conductance values in the array, vector b represents applied currents and solution vector x is the output voltage. O(1) complexity of eigenvector extraction and O(N) complexity of matrix inversion have been demonstrated, which outperforms classical digital computing and even quantum computing.
In this work, we show that the same circuit concept can serve to accelerate regression and classification problems, which are two key problems of AI. Linear regression is solved within two coupled crosspoint arrays by computing the pseudo-inverse matrix in one step. M-dimensional linear regression allows to make predictions based on an existing data set or temporal sequence, thus enabling fast, energy efficient machine learning on the edge. The same circuit enables logistic regression, hence one-step classification within the crosspoint array, which allows for one-shot training of multiple layer perceptrons (MLP). The implications of these novel concepts for AI acceleration and the best technological implementation for data-intensive computing will be discussed.

Neuromorphic computing with analog memory can accelerate deep neural networks (DNNs) by enabling multiply-accumulate (MAC) operations to occur within memory. Analog memory, however, presents a number of device-level challenges having macro-implications on the achievable accuracy and reliability of these artificial neural networks. This talk focuses on the adverse effects of conductance drift in phase-change memory (PCM) on network reliability. It is shown that conductance drift can be effectively compensated in a variety of networks by applying a ‘slope correction’ technique to the squashing functions to maintain accuracy/reliability for a period of ∼1 year. In addition to conductance drift, PCM poses considerable variability challenges, which impact the accuracy of the initial weights. This talk summarizes recent advances in optimizing initial weight programming, and provides evidence suggesting that the combination of ‘slope correction’ and programming optimization techniques may allow DNN acceleration using analog memory while maintaining software-equivalent accuracy with reasonable reliability.

This talk will present our latest research adventures on 2D atomic nanomaterials towards greater scientific understanding and computing applications. In particular, the talk will highlight our work on non-volatile memory (atomristors) that can be employed in various applications including information storage, brain-inspired computing, and zero-power RF switches. Non-volatile memory device based on atomically-thin materials is an application of defects, and is a rapidly advancing field with rich physics that can be attributed to metal ion binding to native defects. The promising 2D materials for neuromorphic computing include MoS2 and hBN. In particular, hBN memory devices feature very low switching energy, a relatively wide range of stable programmable resistance states, and good endurance and retention. Recent progress has also demonstrated wafer-scale integration of 2D memory devices with a prospect to leverage silicon electronics for fully integrated systems.

Artificial electrochemical synapses based on the charge-driven proton intercalation demonstrate much promise toward brain-inspired neural computing [1]. Proton transport plays a critical role in these devices determining switching speed and reliability. Despite great attention and intensive experimental investigations, even the cutting-edge proton conductors still suffer from sluggish ion transport kinetics at low temperatures or are not compatible with current CMOS fabrication processes. Therefore, there is an urgent need for fast solid-state proton conductors operating below 100 °C, much required for rapidly growing proton-based neuromorphic computing applications.

In this work, we consider a large class of solid-state materials having hydrogen as a part of their nominal structure such as solid acids and inorganic hydrates. By considering different combinations of common acids and metal ions, it is possible to synthesize thousands of materials with widely varying structures and physicochemical properties, and yet many of them remain relatively unexplored for the ability to conduct protons. We performed a high-throughput computational screening of the available materials datasets to identify chemical compositions and structure types that are likely to have high proton mobility. To estimate proton conductivity, we used ab initio molecular dynamics simulations. We further correlated this data with simple structural features and built a classification model to preselect a small set of materials for more detailed investigations. Based on our simulations, we identified promising materials and suggested effective strategies for improving proton conductivity.

References
[1] Xiahui Yao, Konstantin Klyukin, Wenjie Lu, Murat Onen, Seungchan Ryu, Dongha Kim, Nicolas Emond, Iradwikanari Waluyo, Adrain Hunt, Jesus A. del Alamo, Ju Li, Bilge Yildiz, “Protonic Solid-State Electrochemical Synapse for Physical Neural Networks”, Nature Communications, in press, 2020.

Strongly correlated metal oxide (SCMO), such as manganite, nickelate, and so on show first-order metal-insulator phase transition (MIT) with a change in conductivity by several orders magnitude. Since a discovery of the phase separation during MIT process, many researchers have been trying to capture a nanoscale electric domain and investigate its exotic properties. Typical SCMO material, rare earth Ni oxides (ReNiO₃, Re = Nd, Sm, and Eu) having a perovskite structure, show giant resistance change due to MIT at a wide temperature range of 150 K to 500 K. In this family, bulk neodymium nickelate: NdNiO₃ (NNO) shows 1-2 orders resistance change through MIT at around 200 K [1]. Very recently, the coexistence of insulator and metal domains with a few hundred nm size was observed during the MIT process for NNO system [2]. This phase-separated phenomenon is also observed in typical SCMOs, such as VO₂ and manganite. The overall observed MIT property, i.e., the resistance change for a sample, is considered to be dominated by competing nanoscale electronic phases. A step resistance change due to the electronic phase confinement effect: increase in resistance change rate has been reported in nanoscale samples comparable to domain size [3]. Therefore, it is expected that the nanoscale NNO sample shows a step resistance change due to MIT property of single nanoscale electronic phases. In this study, we fabricated NNO nanowire structures with 50-200 nm widths and measured their transport properties.
The NNO thin films were fabricated on NdGaO₃(110)(a_pc = 0.3858 nm), LSAT(100)(a = 0.3868 nm), and SrTiO₃(100)(a = 0.3858 nm) substrates by pulsed laser deposition method. NNO nanowires with different wire widths from 50 nm to 500 nm were fabricated by electron beam (EB) lithography and Ar ion milling process. All films and nanowire samples showed the resistance changes corresponding to MIT, but the RT curves shape were quite different. While the thin film sample showed gradual resistance change, the nanowire sample showed some step resistance changes both in cooling and heating processes. This step changes indicate that the nano electronic phases are confined in the NNO nanowire and a single electric domain shows first-order MIT with a single-step change in electrical transport property. For example, we found that the 50 nm width nanowire on SrTiO₃ substrate showed the intermittent step-resistance change, implying the further confine electronic phases to a nanometer space. Interestingly, the nanowires on NdGaO₃(110) and LSAT(100) substrates showed more prominent step resistance changes than that on SrTiO₃(100) substrate, which imply the nano-confinement condition were affected by the strain. The dR/dT vs. T plot showed that the transition temperature distribution of each electronic phase in the nanowires were different. This is probably because the strains depending on the three different substrate effect where on the Ni-O-Ni angular distribution, which affected the MIT properties of NdNiO₃ nano-electronic phase. In the presentation, we will present the quantitative correlation between nano-spatial size and nano-confinement MIT properties. We will also discuss the revealed the MIT property for a single domain in a NNO nanowire and the strain effect for the nano-confined domains behaviors.
[1] G. Catalan et al., Phys. Rev. B, 62, 7892 (2000), [2] G. Mattoni et al., Nat. Commun., 7, 13141 (2016), [3] A. N. Hattori et al., Nano Lett., 15, 4322 (2015).

Much progress has been made over the past decade using conductive bridge random access memory (CBRAM) to emulate both artificial synapses and artificial neurons for neuromorphic applications. However, the intrinsically stochastic switching process of metal filament formation and dissolution results in unreliable analog synaptic weight updating, while the non-volatile characteristic of CBRAM prevents the leaky integrate-and-fire function—a hallmark characteristic of an artificial neuron. We have previously reported that silver nanofilament growth dynamics in a polymer electrolyte can be regulated by the addition of ionic liquid (IL) by: (1) reducing stochasticity of the filament switching (shifting its formation time distribution from power-law to Gaussian) and (2) creating a competition between the formation of IL electrical double layers (EDL) and silver electrodeposition. The competition enables the tuning of filament growth dynamics from abrupt to gradual, with gradual growth introducing multiple well-defined resistance states—a feature necessary for emulating the communications between neurons in an artificial neural network. We also observed a linear correspondence (R² = 0.95) between nanofilament resistance states and program pulses, which is essential to improve learning accuracy in Deep Neural Networks (DNNs). Here, we report on silver filament formation and dissolution in response to presynaptic programming biases and focus on benchmarking the performance against the requirements for artificial synapses. The temporal stability of “abrupt growth” silver filaments is also studied. Lastly, we demonstrate a system integrating both gradual and abrupt silver filaments to implement some key artificial neuron functionalities.

Memristive devices have become a promising candidate for unconventional computing(1) due to their attractive properties(2, 3). Computing can be implemented on a Resistive Neural Network (ResNN) with memristor synapses and neurons or a Capacitive Neural Network (CapNN) with memcapacitor synapses and neurons. For ResNNs as computing accelerators, we have built a dot-product engine based on a 128 x 64 1T1R crossbar array using traditional non-volatile memristors with 64 stable analog resistance levels(4). Accurate image compression and filtering have been demonstrated with such analog computing accelerator(4). In addition, we have demonstrated efficient and self-adaptive in-situ learning in a two-layer neural networks using such memristive arrays(5), which is expected to significantly improve the speed and energy efficiency of machine learning. A variety of neural networks based on such accelerators have been demonstrated, including reinforcement learning(6), LSTM(7), recurrent CNN(8) memristive networks. For ResNNs beyond accelerator applications, we developed diffusive memristors(9) with diffusion dynamics that is critical for neuromorphic functions. Based on the diffusive memristors, we have further developed artificial synapses(10) and neurons(11) to more faithfully emulate their bio-counterparts. We then integrated these artificial synapses and neurons into a small neural network, with which pattern classification and unsupervised learning have been demonstrated(11). For CapNNs, we have developed pseudo-memcapacitive devices based on the diffusive memristors. Capacitive synapses and neurons enabled by these memcapacitive devices have been developed and used to form a fully integrated CapNN(12), which has shown spiking signal classification and Hebbian-like learning functions. Moreover, the diffusive memristors can be used for cybersecurity(13, 14) and robotics applications(15). The requirements on device properties, and thus the corresponding challenges and solutions, are different for ResNNs and CapNNs, which will be discussed in detail in this talk. J. J. Yang, D. B. Strukov, D. R. Stewart, Memristive devices for computing. Nature Nanotechnology 8, 13-24 (2013). S. Pi et al., Memristor crossbar arrays with 6-nm half-pitch and 2-nm critical dimension. Nature nanotechnology 14, 35 (2019). Q. Xia, J. J. Yang, Memristive crossbar arrays for brain-inspired computing. Nature materials 18, 309 (2019). C. Li et al., Analogue signal and image processing with large memristor crossbars. Nature Electronics 1, 52 (2017). C. Li et al., Efficient and self-adaptive in-situ learning in multilayer memristive neural networks. Nature communications 9, 2385 (2018). Z. Wang et al., Reinforcement learning with analogue memristor arrays. Nature Electronics 2, 115 (2019). C. Li et al., Long short-term memory networks in memristor crossbar arrays. Nature Machine Intelligence 1, 49 (2019). Z. Wang et al., In situ training of feed-forward and recurrent convolutional memristor networks. Nature Machine Intelligence 1, 434-442 (2019). E. J. Fuller et al., Parallel programming of an ionic floating-gate memory array for scalable neuromorphic computing. Science 364, 570-574 (2019). Z. Wang et al., Memristors with diffusive dynamics as synaptic emulators for neuromorphic computing. Nature Materials 16, 101-108 (2017). Z. Wang et al., Fully memristive neural networks for pattern classification with unsupervised learning. Nature Electronics 1, 137-145 (2018). Z. Wang et al., Capacitive neural network with neuro-transistors. Nature Communications 9, 3208 (2018). H. Jiang et al., A Novel True Random Number Generator Based on a Stochastic Diffusive Memristor. Nature communications 8, 882 (2017). H. Jiang et al., Provable Key Destruction with Large Memristor Crossbars. Nature Electronics 1, 548 (2018). J. H. Yoon et al., An artificial nociceptor based on a diffusive memristor. Nature Communications 9, 417 (2018).

At present days, an enormous number of semiconductor devices are used in information communications equipment. The memory and computational functions resulting from integration of semiconductor devices have progressed as a result of the development of ultrafine processing technologies, but it is said that this progress will slow down in the near future. Therefore, continued improvement of information communication equipment performance requires not only the further development of semiconductor devices, but also the creation of innovative devices that operate on new principles. Furthermore, development of Artificial Intelligence (AI) hardware such as a brain-type computer that operates with low power consumption and small size is expected in recent years. In order to realize such AI hardware, it will be necessary to create devices that achieve various unique functions such as brain ones. These devices will no longer be based on conventional semiconductors, which exhibit characteristics suitable for digital arithmetic processing. They will be based on new operating principles.
One group such promising novel devices is the nanoionic device’s group, these nanoionic devices operate by controlling local ion migration and electrochemical reactions instead of by electron and hole migration. The control of local ion migration allows for any reconstruction of internal and surface nanostructures of device materials, as well as heterointerface one. Namely, it is a nanoarchitecture of materials achieved by the control of ion migration (referred to as ionic nanoarchitectonics). The unique physical and chemical phenomena and functions are caused by the ionic nanoarchitectonics, and we can create interesting nanoionic devices with a variety of novel functions.
So far, we have already made some nanoionic devices with diverse functions by using the ionic nanoarchitectonics. These devices include an atomic switch, decision-making devices, artificial synapse device, solid-state electric-double-layer transistor, multi-functional on-demand-type device, magnetization and magnetoresistance tuning device, and so on [1,2]. We believe that the ionic nanoarchitectonics is promising as a new method for creating AI devices.

[1] K. Terabe et al. Nanoscale, 8, 13873, 2016
[2] T. Tsuchiya et al. Science Adv. 4, eaau2057, 2018

Memristive devices because of their non-volatile memory behavior and their relatively simple structure as an electronic counterpart to synapses for hardware-based artificial neural networks (neuromorphic circuits). An important behavior of devices is their inherent stochastic nature, which applies especially to devices. Here, the change in resistance is due to the formation and dissolution of electrically conductive paths in a solid state electrolyte, which result in changes in the resistance switching process and nondeterministic device behavior.
In this talk, a learning scheme is presented that uses the randomness of the switching process for neuronal-inspired computing. It is shown that the switching probability of ReRAM cells follows a Poisson distribution at a certain voltage and pulse width, the switching probability being controllable by the number, amplitude and width of the programming pulses. To demonstrate the power of stochastic learning for neuromorphic systems, a neural circuit for pattern recognition is presented, and one that mimics perceptual behavior.

There is a recognized need to develop analog neuromorphic hardware to tackle the ever-increasing digital computation time and resources demanded by current artificial neural networks (ANNs). Here, we demonstrate novel neuromorphic architectures based on Ti/HfO2/Pt and Pt/HfO2/Pt memristive systems and discuss their electrical performance and their potential applications as ANNs. Current state-of-the-art systems are built on a crossbar array of memristors for neuromorphic computing. With a limit on the number of neurons and regular connectivity, the networks lack sparsity and randomness which results in high wiring cost and poor functional connectivity. To address these limitations, we have fabricated a random array of core-shell memristive wires to form the connectivity matrix for neuromorphic hardware. In the systems listed above, the conductive core (Ti or Pt) serves as a conductive path or intermediate electrode with a memristive shell (HfO2) that can be electroformed with a set of top electrodes (Pt) deposited on the surface. These random nanowire networks are fabricated via electron-beam lithography on a 200 ?m × 200 ?m write field. Network quantifying simulations revealed a small-world coefficient of 2.89, the shortest path length of 3.61 and a clustering coefficient of 0.057. Remarkably, depending on the forming process, the individual core-shell wires can exhibit eight-wise (8W), counter eight-wise (C8W) bipolar resistive switching (BRS), and complementary resistive switching (CRS). At a low potential ranges (±1.5 V), the wires switch in a bipolar fashion between LRS (~11.2 k?) and HRS (~0.3 M?). Symmetric arrangement of layers in the wires along with 8W and C8W switching implies two memristors connected anti-serially. This configuration results in the wires exhibit CRS at higher potentials (±2 V). This ability to switch the devices via BRS or CRS enables us to dynamically control the read-write process. Moreover, due to the transient ON in CRS mode, both the currents and power consumption during the operation of the devices is comparatively lower. Given the low switching voltages, low processing temperatures and materials utilized, these neuromorphic devices can be considered CMOS compatible. In this presentation, the fabrication process, the memristive characteristics (endurance, resistance retention, pulse measurements, etc.), switching mechanism and synaptic characteristics (LTP, STP, and LTD) of the above-listed systems will be discussed in detail.

At the era of information technology, the growth of data is quickly moving beyond the processing capability of traditional silicon devices. New paradigms for data storage and processing are therefore needed. The answer may be hidden within the structure of your brain. We aim to construct nanoparticle networks inspired by the architecture of actual neural networks to perform brain-like computing. In this work, we show that polymer linked gold nanoparticle (AuNP) networks can function like neuronal networks in information storage and processing. Molecular dynamics (MD) simulations are used to realize their performance. A simple coarse-grained model (AuNP-PAH) is built, which contains one polyallylamine (PAH) chain of 200 repeating units wrapping on a 4 nm diameter citrate capped AuNP. A 2D network model with 100 AuNPs (10*10*1) and a 3D network model with 125 AuNPs (5*5*5) are constructed by replicating the simple AuNP-PAH structure. The network connection is formed by thermally heating and is also manipulated by an external electric field. The effect of the AuNP volume fraction on network formation will also be reported. We find that these AuNP-PAH networks can process through information differently from existing computing architectures.

Networks based on nanoscale resistive switching junctions are considered promising for the fabrication of neuromorphic computing architectures [1]. To date random networks of nanowires, nanoparticles, and metal clusters embedded in a polymeric matrix or passivated by shell of ligands or oxide layers have been used to produce resistive switching systems. The strategies applied to tailor resistive switching behavior are currently based on the careful control of the volume fraction of the nanoscale conducting phase that must be fixed close to the electrical percolation threshold.
Here, by blending laboratory and computer experiments, we demonstrate that metallic nanostructured Au films fabricated by bare gold nanoparticles produced in the gas phase and with thickness well beyond the electrical percolation threshold, show a non-ohmic electrical behavior and complex and reproducible resistive switching. We observe that the nanogranular structure of the Au films does not evolve with thickness: this introduces a huge number of defects and junctions affecting the electrical transport and causing a dynamic evolution of the nanoscale electrical contacts under the current flow. To uncover the origin of the resistive switching behavior in Au cluster-assembled films, we developed a simple computational model for determining the evolution of a model granular film under bias conditions. The model exploits the information provided by experimental investigation about the nanoscale granular morphology of real films.
Our results show that metallic nanogranular materials have functional properties radically different from their bulk counterparts, in particular nanostructured Au films can be fabricated by assembling bare gold clusters which retain their individuality to produce an all-metal resistive switching system [2].

1 C. Minnai, A. Bellacicca, S. A. Brown, P. Milani, Sci. Rep. 7, 7955 (2017)
2 M. Mirigliano, et al., Nanotechnology 31, 234001 (2020)

Resistive random-access memory (ReRAM)/memristive devices based on electrical resistance switching are highly expected to overcome persistent technological limitations in current memory technology and to develop next-generation nonvolatile memory because of its superior characteristics. These characteristics include high speed, high scalability, low power consumption, fast fabrication process, multistate storage, and compatibility with complementary metal-oxide semiconductor (CMOS). Particularly, nanowire-based resistive switching devices offer an ideal platform for a deeper understanding of resistive switching properties due to the confined nanostructure and the opportunity to develop nanoscale-memristive device for next-generation memory and neuromorphic computing. In addition, copper oxide is one of the attractive candidates for ReRAM/memristor due to the existing multiple oxidation states from “+2” to metal state (CuO, Cu4O3, Cu2O, Cu) for inducing resistive switching, which is believed to involve the phase transition with multiple oxidation state induced by redox reaction. In this work, we report the volatile and non-volatile resistive switching effect in p-type non-stoichiometric CuOx nanowires, which are fabricated using insulator CuO nanowires via a thermal-annealing technique with hydrogen atmosphere. These resistive switching modes were effectively controlled by the stability of conductive filament via the compliance current in the SET process. The non-volatile resistive switching exhibited low electric-field-induced switching, high ON/OFF ratio of 107, and long retention time of 104 seconds for ReRAM application. On the other hand, the volatile threshold switching was operated with low compliance current for selector application, and the device showed bidirectional operation, selectivity (104), low OFF-current (<100 pA), and stable reliability with endurance over 105 cycles under pulse operation. The presented work will discuss a nanoscale resistive switching mechanism based on hole-based conducting filament formation originated from the local redox reaction.

Resistive switching (RS) operations offer various useful applications for nonvolatile memories and neuromorphic computing. However, driving forces of the RS operations have remained unclear and controversial yet. We reported that conductance quantization in nickel-oxide (NiO)-based RS cells was observed upon a voltage sweep [1,2]. Moreover, not only conductance quantization but conductance fluctuations appeared in the cells by controlling amount of oxygen vacancies in NiO thin films [3]. In this study, disappearance conditions of conductance filaments are investigated through experiments and simulations.

A platinum (Pt) bottom electrode was deposited by sputtering on a silicon-dioxide/p-silicon substrate inserted by a titanium adhesion layer. An 80-nm-thick NiO layer as a resistance change material was subsequently deposited by radio-frequency reactive sputtering. The substrate temperature during sputtering was kept at 350^°C. Pt top electrodes with a diameter of 100 μm were deposited on the NiO layer. As a result, a host of Pt/NiO/Pt RS cells were fabricated on the substrate. The first abrupt current jump by an initial voltage sweep to the cell made the cell conductance equivalent to the quantized conductance (2e²/h), indicating the formation of a conductive filament with a quantum point contact (QPC). Moreover, conductance fluctuations before and after the abrupt current jump were observed in several cells.

We investigated temperature dependences of cell conductance upon heating the sample. While the conductance in several cells did not change between before and after heating to 400 K, the conductance in other cells decreased by the quantized conductance at 370 K, which means that the conductive filament with a QPC is possibly ruptured by heat at a temperature of merely 370 K [1]. In the same manner, we found that the conductance fluctuations, which originate from the modification of oxygen-vacancy densities at grain boundaries in the NiO layer, possibly disappeared upon heating to, at least, 470 K [3]. In contrast, a fat filament with higher conductance than the quantized conductance created by the conventional forming remained unchanged upon heating to 600 K during a few days. These results indicate a relatively small heat is sufficient for the disappearance of the conductance quantization and fluctuations.

We estimated the temperature rise owing to Joule heat to destroy the conventional fat filament by rough simulation using solvers of heat transfer, electric current, and (particle) diffusion in commercial software COMSOL Multiphysics. The diameter of a cylindrical conductive filament and the diffusion coefficient of oxygen vacancies are assumed to be 10 nm and 10¹⁰ m²/s, respectively. The simulation revealed that an applied voltage required for reaching the maximum temperature of 600 K in the filament was 0.34 V, independent of the voltage sweep rate in the range from 0.01 to 1000 V/s. The estimated voltages for reset increased from 0.68 to 1.16 V as the sweep rate increased, which agrees with our experiments to reset upon voltage sweeps. Moreover, the estimated temperature at reset was turned out to be over 680 K. These results indicate that the temperature required for the diffusion of oxygen vacancies is higher than 600 K. Comparing the simulation to the above heat experiments, temperature required for dissolution of the conductive filament with a QPC is estimated to be lower than 600 K, irrespective of external heat or Joule heat.

The conductance quantization seemed to be a special form of the conductance fluctuations. Although the fluctuations may have poor thermostability at the present, the fluctuations imply the existence of multilevel resistances, which offers the possibility of multiple discrete weighting of synaptic devices.

[1] H. Sasakura et al., Appl. Phys. Lett. 107, 233510 (2015).
[2] Y. Nishi et al., J. Mater. Res. 32, 2631 (2017).
[3] Y. Nishi et al., J. Appl. Phys. 124, 152134 (2018).

Although in its embryonic state, bio-inspired computing has never been closer to its realization, due to the recent advances in neuromorphic hardware implementation with non-volatile, binary and analog, devices based on established and novel switching oxides. While binary or multibit information storing as synaptic weights, through the stimulated resistance modulation of such devices, enables in-memory computing and non-Von Neumann architectures, synaptic plasticity will enable direct emulation of biological neural activity. Bio-realistic synaptic plasticity can be facilitated through the pulse stimulation of devices with transient switching behaviour, with an internal timing mechanism through their inherent diffusive dynamics. Consequently temporal association of stimulating pulses, governs the triggered response of such devices, in a manner reminiscent of biological neural behaviour. Searching beyond filamentary resistive switching devices, diffusion based ionic switches, both in three-terminal and two-terminal configurations, receive noticeable attention. Robustness of non-filamentary switching mechanisms, along with unique switching characteristics and plasticity attributes, can expedite the integration of these devices in functional neural networks with bio-inspired capabilities. Cation-based ionic switches, and specifically Li ion diffusion driven switching oxides, attract reasonable attention in the neuromorphic community. Electrochemical relevancy of Li cation with cations such as Na, K and Ca, involved in the polarization/depolarization of neuron cells, apart from being philosophically sensational, disclose great potential for the realization of artificial neurons with synaptic plasticity. This work investigates the memristive/synaptic behaviour and underlying mechanism of two terminal, thin film battery-like, ionic switches based on LiCoO2-cathode, SiO2 electrolyte and TiO2 anode on a Si substrate. Acting as a self-doped p-type semiconductor, highly stoichiometric LiCoO2 presents a reversible insulator to metal transition (IMT) upon electrochemical lithium removal. Each Li ion vacancy induces the oxidation of Co3+ to Co4+, injecting an electron hole and consequently increasing the conductivity of Li-deficient LiCoO2. Controlling the Li concentration in the switching oxide, via voltage stimulated electrochemical migration of Li cations, through a solid electrolyte SiO2 layer and storing them in a TiO2 anode layer, the conductance of the battery-like stack can be tuned. The devices have been characterized by means of dc-cycling and non-abrupt, analog, switching with I-V hysteresis was recorded. Furthermore, non-symmetric, diode-like, characteristics with high rectification effect in reverse bias was evident, along with the rise and decay of an electromotive force after each potentiation. Pulse characterization of the devices lead to the observation of gradual conductance enhancement through a paired pulse facilitation mechanism. Diffusive dynamics of this system, were observed in the form of a decaying excitatory post synaptic current. Regardless, non-volatile multistate conductance enhancement was also monitored. Spike timing dependent plasticity functions where generated, through non-overlapping postsynaptic and presynaptic pulse-pair stimulation. Analog switching observed by these devices, in combination with the diffusive dynamics and the paired pulse facilitation mechanism, effectively facilitate the direct emulation of bio-inspired synaptic functionalities like STDP. Spontaneous state decay, in the form of an excitatory post synaptic current can be advantageous for the emulation of habituation processes observed in biological neurons. Finally non-filamentary switching with rectification characteristics can be beneficial upon device integration in dense cross-bar arrays, effectively enhancing durability and suppressing leakage currents respectively.

Deep learning based on neural networks is a promising approach to accelerate machine learning applications. However, state-of-the-art processing units, based on Complementary Metal–Oxide–Semiconductor (CMOS) circuits, and computer systems based on von Neumann architecture, requires large memory space and high power consumption when simulating neural networks. Improvements in computing performance therefore requires designing novel scalable, fast and energy-efficient hardware structures with both processing and storage capabilities.

It has been demonstrated that crossbar-type arrays containing two-terminal resistive switches, whose electrical conductivity is electrically modulated, enable energy-efficient and fast-processing physical neural networks. Two main types of mechanisms have been demonstrated in artificial intelligence applications to change the conductance of the switches, conductive-filament and phase-change mechanisms. A more recent concept, electrochemical cation intercalation-based resistive switches with three-terminal structures have the potential to circumvent issues faced by other mechanisms. Indeed, multistate capability and low energy consumption was recently demonstrated for inorganic Li-based¹ and organic proton-based² electrochemical devices. However, these devices also suffer from a scaling challenge as they are not compatible with current CMOS processes due to presence of Li and organic switching materials, respectively.

Our group is therefore working on the development of an all solid-state inorganic proton (H⁺) electrochemical intercalation resistive switch, which relies on a deterministic charge-controlled mechanism. Our initial design consists of a channel of active material, whose electronic conductivity is modulated through protonation/deprotonation, enabling for a continuum of resistance states³. A solid proton reservoir layer serves as the gate terminal while a proton conducting solid electrolyte separates the channel and the reservoir. Employing protons, the smallest cation, as the doping ion presents several advantages including high operation speed, good compatibility with Si technology and long lifetime.

Herein, I will first present our demonstration of this concept³ for a device that includes tungsten trioxide (WO₃) as an inorganic active channel material, and Pd-hydride as a solid hydrogen reservoir that is also used for gating the device. Hydrated Nafion is used as the polymer-based proton conducting electrolyte. The device has a very low energy consumption, nearly symmetrical protonation/deprotonation behavior and long cycling lifetime. While the channel and gate materials are CMOS-compatible, a challenge is to come up with CMOS-compatible inorganic proton conducting solid electrolytes. Another challenge is to find more insulating channel materials that can switch faster. I will therefore also present our work on the improvement of the device properties and integrability by exploring alternative materials for the active channel and the proton electrolyte. Vanadium pentaoxide (V₂O₅), which is reported to efficiently intercalate small cations in its crystal structure along the (001) orientation (channel-like structure), and tantalum (Ta₂O₅) pentaoxide, a potential two-phase system that could allow to fix the device open-circuit potential, are the two active materials studied within this context. On the other hand, nanocrystalline yttrium-doped barium zirconate (BZY), a state-of-the-art solid-state proton conducting ceramic, is considered as a possible alternative to replace Nafion polymer.

Fuller, E. J. et al. Li-Ion Synaptic Transistor for Low Power Analog Computing. Adv. Mater. 29, 1604310 (2017).
van de Burgt, Y. et al. A Non-Volatile Organic Electrochemical Device as a Low-Voltage Artificial Synapse for Neuromorphic Computing. Nat. Mater. 16, 414 (2017).
Yao, X. et al. Protonic Solid-State Electrochemical Synapse for Physical Neural Networks. Accepted in Nat. Commun. (2020).

With the advent of artificial intelligence(AI), neuromorphic computing inspired by efficient data processing in human brain has widely attracted immense research interest in various fields. Recent development of AI processor optimized for parallel computing including GPU, FPGA and ASIC has advanced AI-based industries including big data, autonomous vehicles, and internet of things. Nevertheless, CMOS based AI processor has theoretical limits of data bandwidth and energy efficient due to data shuttling between memory and computing units in terms of von Neumann bottleneck. Therefore, new concepts of electronic device have been proposed to mimic biological synapse responsible for both information storage and processing in human brain, termed in-memory computing. Memristors, a portmanteau of memory and resistor, is considered as one of the promising candidates for in-memory computing device because of its desirable properties as tunable resistance, non-volatility, low energy consumption, and fast operation. However, the key difference in working mechanism of biological synapses from artificial electron-based synthetic devices is the cooperative and multimodal activation process which includes the transfer of ions or molecules. In this study, we demonstrate bi-modal operation of peptide memristors, which is analogous to synaptic operation, by electronic and environmental inputs. Based on our understanding on coupling of electron and proton conduction, we designed sequence of amino acids in peptide to obtain controllability on resistive switching characteristics through proton injection by humidity control. As a result, the memristor based on Tyr-Tyr-Ala-Cys-Ala-Tyr-Tyr(Y7C) peptide operates not only by applying voltage inputs but also by applying humidity sweep without voltage sweep corresponding to signaling process by extracellular proton concentration in synapse.
Similar to previous reported electrochemical metallization memory, our peptide-based memory showed resistive switching characteristics owing to redox reactions of metal. However, we examined involvement of phenolic hydroxyl group of tyrosine in resistive switching by introducing dehydroxylated tyrosine, phenylalanine, instead of tyrosine, resulting in 6-fold increment of set voltage. We assumed that electron transfer from tyrosine to Ag metal and vice versa through deprotonation and protonation promotes formation of Ag conductive filament. Thus, increasing relative humidity(RH) would promote resistive switching due to increased proton conduction in Y7C peptide. As a result, set voltage of peptide memristor showed linear dependency on RH values and, further, only humidity injection without voltage sweep induced reversible and non-volatile resistive switching, namely humidity mode of memristor was realized. Once transition from high resistive state to low resistive state is occurred by a forward RH sweep, the device maintained its state in wide range of RH condition for over 10⁴ s, meaning data storage. To erase the data, a reverse RH sweep was applied and transition back to high resistive state was occurred in extremely low RH condition. On off resistance ratio reached up to 10⁴. Complementary bimodal operation of the peptide memristor was successfully carried out by writing data using humidity input and erasing data using voltage input in one device. Peptide film was characterized using circular dichroism(CD) analysis, nuclear magnetic resonance(NMR) spectroscopy and transmission electron microscopy(TEM). Proton conductivity of the peptide film was calculated by electrochemical impedance analysis(EIS). Proton involvement in the device was studied by kinetic isotope effect(KIE) of deuterium oxide vapor. This result suggests that proton can act as another control knob for neuromorphic computing devices and multimodal and biomimetic memory operation can be facilitated by material design, especially in tyrosine-rich peptide.

Neuromorphic computing, inspired by the human brain, provides efficient energy utilization, massive parallelism, flexible adaptive capability, and information processing and transmission. It has the potential to realize computational techniques requiring recognition, learning and decision making; fundamentally overcoming the bottleneck in von Neumann architectures.[1] Artificial synapses form the backbone of neuromorphic computing; emulating controlled functions related to learning and memory. Furthermore, to functionalize devices beyond their synaptic functions, it is essential to merge them with biometric sensing elements such as vision, auditory, touch and olfactory sensors. Optical stimuli in a synaptic device renders ultrafast computational speed due to high bandwidth, low crosstalk, and ultralow power consumption.[2] A photonic synaptic transistor based on 2D materials provides ultra-low power consumption along with flexibility and thickness scalability. Graphene, the first 2D material leads to a short photocarrier lifetime and high dark current due to its zero bandgap and is therefore unfavorable for use as an efficient synapse material. Unlike graphene, monolayer MoS2 having a direct bandgap would allow a high absorption coefficient and efficient electron–hole pair generation under photoexcitation, becomes suitable in various neuromorphic applications. We demonstrate the concept of artificial cogni-retina; an optic-neural synaptic device that features synaptic and optical-sensing functions. It has the potential to replace bulky systems comprising of video camera and processing units. A MoS2 based photonic synaptic transistor for emulating the superior and complex features of artificial cogni-retina is reported. The device finds application in prosthetic human eyes and robotic eyes in artificial intelligence based platforms of autonomous vehicles. The device shows synaptic behavior with high conductance ratio of 103 facilitated by gate tunability and a long-term memory of 104 s with an extremely low energy consumption (15 pJ/event). High quality large area (2cm×2cm) monolayer MoS2 (grain size of 50 ?m) is grown on SiO2/Si substrate via CVD technique by maintaining the precursor at 950°C. As grown MoS2 coated with PMMA is immersed in BOE to release the MoS2 film by removing the underlying SiO2. The PMMA supported MoS2 film is transferred to DI water bath and finally to SiO2/ p+ silicon substrate with nickel source/drain contacts. The final sample is heated for increasing the adhesion of the transferred film, and the PMMA layer is etched by acetone followed by a forming gas annealing. MoS2 is patterned to form the channels of the transistor by reactive ion etching in CF4. We explore the spectral response of our device, expressed through its photoresponsivity, over a wavelength of 400-700 nm for varying intensity. The negligible photoresponsivity obtained for wavelengths above 680 nm clearly depicts the bandgap of monolayer MoS2 corresponding to 1.8 eV. The synaptic transistor exhibits excellent responsivity of 2×103 AW-1 (EQE of 6×105 %) under illumination (450 nm, 10 ?W/cm2) for VG = 0 V and VD = 5 V. Both the responsivity and EQE of the device increase with gate voltage due to lowering of the barriers at the contacts. The device shows n-type behavior (ON/OFF current ratio of 105) and under illumination, an enhancement in current is observed for both ON (VG>VTH) and OFF (VGTH) states for all gate voltages. This indicates that the photocurrent dominates over thermionic and tunneling currents. Gate dependent potentiation and data retention of the device is observed by varying the gate bias. The device also demonstrates other essential synaptic behaviors, such as short-term potentiation (paired-pulse facilitation of 150%), spike-frequency dependent plasticity, spike-duration dependent plasticity. References [1] Q. Xia et al., Nature materials 2019, 18. [2] Z. Cheng et al., Science advances 2017, 3.