Symposium S.CT01—Artificial Intelligence for Material Design, Processing and Characterizations

The dream of autonomous materials discovery systems will only be realized if the entire design/synthesis/characterization process is described in an unambiguous machine-readable format. Specifying the experiment process in a machine-readable way facilitates both remote operation, algorithmic control of experiments, and enhanced experimental replicability. A significant challenge is that materials experiments often require a combination of instructions for human operators and machines. Furthermore, the variety of materials characterization data and experimental performance metadata during a process often require significant extraction/transformation/loading before machine learning methods can be applied. Existing tools require significant development time that is incompatible with rapidly evolving scientific needs.

In this talk, I will describe ESCALATE (Experiment Specification, Capture and Laboratory Automation Technology) an adaptable open source package for experiment description and data collection. As a specific example, I will describe its application to Robotic-Accelerated Perovskite Investigation and Discovery (RAPID). The first generation of RAPID uses inverse temperature crystallization (ITC) to grow halide perovskite single crystals for x-ray structure determination and bulk characterization using commercial liquid handling robots. Experiment plans are contributed remotely, by either human scientists or algorithms trained on the reaction data. Incoming experimental data collected by ESCALATE is used to automatically train machine learning models, evaluate model performance and feature influence on model predictions to gain human-interpretable scientific insight into the formation process, and quantify experimental reproducibility using this complete dataset of all experimental successes and failures. A live web dashboard communicates these insights to the scientist and management in visual form; furthermore experimental data can be exported in the form of interactive visualizations for use in publications. Collected characterizations includes qualitative measurements of crystal quality (including photographic images of the samples), powder X-Ray diffraction, and spectroscopic measurements along with background laboratory conditions. The second generation of RAPID has been extended to include antisolvent precipitation synthesis of perovskite single crystals and perovskite thin film synthesis and characterization including remote monitoring of long-term environmental stability. We have used the ESCALATE/RAPID platform to perform experimental tests benchmarking a variety of data-driven approaches for discovering new compounds against human-expert and random baselines. The comprehensive data collected from these experiments has also allowed us to ascertain the relative importance of descriptive physical models for reaction composition relative to raw experimental descriptions when constructing machine learning models for reaction outcome prediction.

Machine learning has become an integral part of many aspects of materials research. As a branch of machine learning, active learning has attracted much attention which can effectively help navigate experimental sequences in materials research. We are actively incorporating active learning in screening of combinatorial libraries of functiona materials. For some physical properies, each characterization/measurement requires time/resources long/large enough that "high"-throughput measurement is not possible. Examples include detection of martensitic transformation and superconducting transitions in thin film libraries. By incorporating active learning into the protocol of comabinatoal characterization, we can streamline the measurement and the analysis process substantially. We will present examples of mapping of phase change memory materials as well as others functional materials. This work is performed in colaboration with A. Gilad Kusne and V. Stanev. This work is funded by ONR and AFOSR.

Herein, we report the successful discovery of a new hierarchical structure of metal-organic nanocapsules (MONCs) by integrating chemical intuition and machine learning algorithms. By training datasets from a set of both succeeded and failed experiments, we studied the crystallization propensity of metal-organic nanocapsules (MONCs). Among four machine learning models, XGBoost model affords the highest prediction accuracy of 91%. The derived chemical feature scores and chemical hypothesis from the XGBoost model assist to identify proper synthesis parameters showing superior performance to a well-trained chemist. This paper will shed light on the discovery of new crystalline inorganic-organic hybrid materials guided by machine learning algorithms.

Van der Waals heterostructures are comprised of stacked atomically thin two-dimensional crystals and serve as novel materials providing unprecedented properties. However, the random natures in positions and shapes of exfoliated two-dimensional crystals have required the repetitive manual tasks of optical microscopy-based searching and mechanical transferring, thereby severely limiting the complexity of heterostructures.

To solve the problem, we develop a robotic system, powered by artificial intelligence, that automatically searches exfoliated 2D crystals and assembles them into vdW superlattices inside glovebox [1-3]. The system can automatically scan the surface of silicon substrates, analyze optical microscope images, and detect a variety of atomically thin 2D materials without human intervention. The stacking order, positions, and crystallographic orientations of the 2D crystals are designed using the customized CAD software. The robotic assembly system of two-dimensional crystals enabled stacking four cycles of the designated two-dimensional crystals per hour with few minutes of human intervention for each stack cycle.

The system enabled fabrication of the vdW superlattice structures consisting of 29 alternating layers of the graphite and the hexagonal boron nitride (hBN) flakes. Encapsulated graphene devices exhibited high charge carrier mobilities (>1,000,000 cm²/Vs), demonstrating the applicability of the system for prototyping a variety of high-quality vdW superlattices.
The wider material design freedom enabled by our system offers unprecedented opportunities for exploring the potential of vdW heterostructures. By this development, we can reduce the human intervention involved in the vdW heterostructure fabrication by orders. We believe that this work free up researchers from repetitive tasks and letting them focus on more intellectually creative tasks.

The hardware/software designs are available at https://github.com/tdmms/

[1] S. Masubuchi et al., Nature Communications 9, 1413 (2018).
[2] S. Masubuchi et al., npj 2D Materials and Applications 3, 4 (2019).
[3] S. Masubuchi et al., arXiv:1910.12750 (2019).

In my annnual poster, I'll review the MIT PVLab's recent work combining experiment and theory in a closed-loop, machine-learning (ML) driven framework. Closed-loop experimentation driven by ML enables rapid and efficient convergence on optima, within vast search spaces. A review of scientific insights, as well as our latest machine-learning algorithms, will be presented.

Meeting current and future demands for advanced structural materials designed to operate in extreme environments requires rigorous mechanical performance evaluation at such conditions. When considering alloy development for such applications, such as nuclear materials, multiple compositions and processing routes create many permutations of material to be evaluated at different temperatures and radiation damage levels. Where traditional bulk scale mechanical testing is relatively time consuming, high speed nanoindentation mapping techniques can provide large (and highly localized) Hardness and Modulus datasets. This allows for analysis of heterogeneous distributions of properties in multi-phase alloys, which can then be correlated to the microstructure through use of supplementary techniques, such as EBSD, which has similar resolution. However, this requires accompanying analysis methods to also be high throughput, for which machine learning based techniques are of interest. Some datasets, as large as one million indentations in a given map, are beyond the limit of reasonable human analysis.
In this study, a high vacuum, high temperature nanoindenter was utilized to produce temperature controlled mapping of three advanced types reactor materials up to 500°C while protecting the sample surface from oxidation. The materials studied were Fe-based HEAs, with varied Al content to give 3 distinct microstructures. Structure-property correlation was quantitatively explored through clustering to partition the hardness and modulus data to individual microstructural elements and evaluate their thermal stability, with irradiation studies to follow. While qualitatively these methods performed comparably to by hand grouping, there is a need to quantitatively understand their systematic uncertainties without making assumptions regarding the underlying probability distribution. This is further explored for this data set by using a parametric bootstrapping method to resample the probability distribution function from the individual data points. Three different machine learning methods, Kernel Density Estimation, Gaussian Mixture and Multi-Normal, are used to generate a probability distribution function. This is then resampled to produce a simulated dataset 1000 times the original dataset in size. This allows K-means, Spectral and Gaussian Mixture clustering techniques to be applied and compared, but in principal, this can be applied to evaluate other properties as well.

Closed-loop, sequential learning is a key paradigm in autonomous materials development. Within this framework, aspects of the materials system under study are modeled, and such models are used to decide subsequent experiments to be run, results of which are fed-back to update models. In the past, off-the-shelf solutions and algorithms have been employed to optimize material properties. In this talk, we describe how autonomous materials platforms can be used for problems other than materials property optimization. We will specifically highlight work on autonomous phase mapping and real-time control, and how a common framework can be used to capture problem-specific structure and objectives. Relevant to this is the development of tools to perform autonomous materials characterization. We will describe how rich characterization data such as microscopy images or three-dimensional reconstructions from atom probe tomography can be analyzed without human intervention to encode experimental results for subsequent use within the closed loop. Together, the work described demonstrates the broad applicability of autonomous platforms throughout materials science research.

In this talk I will look at how we can open up the percieved 'black box' of machine learning when designing and characterising materials. I will show how being able to look at and understand the inner workings of deep neural networks increases the value of these approaches for problems in materials science. Understaning the workings is important for (i) debugging and building better models, (ii) building confidence in models that perform well, (iii) learning from models that exceed human performance in certain problem areas. I will take examples of our research where we use graph neural networks to infer and predict dielectric properties of solids and where we use deep convolutional neural netowrks to interpret data from inelastic neutron scattering. In the former case a mixture of deep and shallow learning allows for extraction of principles as well as construction of robust accurate models. In the latter case understanding the workings of the neral network allows us to develop a method for guiding experiments to sample the most pertinant areas of signal space.

The next generation of high performance optical materials require precise engineering to transcend the inherent limitations of bulk materials and achieve desired optical performance. These targeted properties originate from the interactions between inherent optical properties of the constituent materials and meticulously engineered sub-wavelength inhomogeneities. However, the design of such metamaterials poses a significant challenge to researchers due to the near-infinite design space to consider, even for relatively simple metamaterials such as the two-dimensional all-dielectric metasurface. This metasurface has been the subject of intense research interest because of its unprecedented reflective performance resulting from low-loss dielectric materials and shape-dependent optical resonances. Here we employ artificial neural networks (ANNs) to invert the design process, conceiving metasurface designs with targeted reflection and transmission spectra. These ANNs are capable of coordinating the complex interactions of the eigenmodes from which these desired properties arise, delivering device designs demonstrating coefficients of reflection greater than 99% over a range of wavelengths more than 450nm.

DNA-templated silver clusters (Ag_N-DNA) are known to exhibit a wide range of fluorescence colors for different choices of the DNA template sequence [1]. While these clusters are promising biosensors and biomarkers [2], rational design of Ag_N-DNA for specific applications is challenged by the huge space of possible DNA template sequences. To understand how DNA sequence selects for silver cluster size and optical properties, we are employing high-throughput experiments and tools from machine learning and data mining. We previously showed that machine learning methods trained on experimental data can be used to design new DNA templates that select for Ag_N-DNA color, for the specific and limited case of 10-base DNA oligomers [3]. An important open question is whether such a design process developed for a specific biopolymer template length is applicable at other lengths, with different numbers and diverse configurations of cluster nucleation sites. More broadly within materials informatics, it is important to determine when models learned on limited training data sets are relevant for prediction within expanded input parameter spaces.

Here, we present a newly developed flexible design approach that utilizes color-correlated DNA base motifs learned from data on > 2,000 ten-base DNA oligomers [4]. We test this motif-based design for templates ranging from 8 bases to 16 bases long, for which the sizes of the sequence spaces differ by nearly 5 orders of magnitude. The experimental data show that designed strands of all lengths are selective for Ag_N-DNA color in the target wavelength band of 600-660 nm, strongly suggesting that color-selective motifs learned for one template length generalize to other lengths. Thus, a motif-based design approach may be broadly suitable for future Ag_N-DNA applications.

[1] E. Gwinn, D. Schultz, S. M. Copp, S. Swasey. Nanotechnology, 5, 180-207 (2015).
[2] Y. Chen, M. L. Phipps, J. H. Werner, S. Chakraborty, J. S. Martinez. Acc. Chem. Res. 51, 756-2763 (2018).
[3] S. M. Copp, A. Gorovits, S. M. Swasey, S. Godibandi, P. Bogdanov, E. G. Gwinn. ACS Nano, 12, 8240-8247 (2018).
[4] S. M. Copp, A. Gorovits, S. M. Swasey, P. Bogdanov, E. G. Gwinn. In revision.

Large-scale data-mining workflows are increasingly able to predict successfully new materials that possess a targeted functionality. The success of such materials discovery approaches is nonetheless contingent upon having the right data source to mine, adequate supercomputing facilities and workflows to enable this mining, and algorithms that suitably encode structure-function relationships as data-mining workflows which progressively short list data toward the prediction of a lead material for experimental validation.

This talk describes how to meet these data science requirements by showing how to auto-generate large material databases of experimental information from documents, using natural language processing and machine learning, via our ChemDataExtractor tool [1-3]. The resulting large databases of chemical structures and their cognate properties are then mined for materials discovery using custom-built algorithms that are encoded forms of structure-function relationships. These molecular design rules progressively filter the parent set of chemicals until a lead candidate appears, which is experimentally validated. Case studies are demonstrated to illustrate the power of data-driven materials discovery for several functional applications.
References
[1] www.chemdataextractor.org
[2] Swain & Cole, J. Chem. Inf. Model. 56 (2016) 1894–1904
[3] Court & Cole, Scientific Data 5 (2018) 180111.

The development of a hydrogen economy has potentially immense benefits for addressing the crisis of global heating. However, storage remains a problem and decades of research have left us empty-handed: no known materials can satisfy all of the DOE targets for vehicular storage. Metal hydrides are among the most promising solid-state storage materials, but their complex chemistry is a major impediment to discovery of game-changing materials. Specifically, their thermodynamics are often ill-matched to hydrogen release at temperatures compatible with fuel cells. This limitation requires an equilibrium H₂ pressure of at least 5 bar at the maximum delivery temperature of 80 °C. Empirical discovery approaches using, for example, high-throughput synthesis, are largely ineffective because measurement of the equilibrium H₂ pressure can take days or even weeks. Machine learning (ML) methods have already shown promise for accelerating the discovery of new hydrogen sorbents and identifying key features that influence their properties. In recent application to metal hydrides, however, the use of “black-box” algorithms prohibited development of crucial vital scientific insight needed to go beyond existing compositions. We employed a gradient-boosting tree regressor, an “explainable” ML that can reveal the feature dependence of material properties. In particular, we find that the equilibrium H₂ pressure of intermetallic metal hydrides strongly depends on a volume-based feature that requires only the elemental composition to be computed. Remarkably, this simple structure-property relationship is valid across a wide range of compositions, metal substitutions, and structural classes found in intermetallic metal hydrides. This result provides a rational basis for targeting novel hydrogen storage materials with thermodynamic properties appropriate for specific applications. As an illustration, we identified a known intermetallic that forms a hydride with very low stability. Although not yet validated by experiment, this result was confirmed by density functional theory modeling.

Nanoporous materials (e.g., zeolite, activated carbon, metal-organic framework, polymeric membranes, etc.) have various technological applications, including gas separation, gas storage, catalytic transformations, etc. The functionalities of nanoporous materials strongly depend on their pore size and shape distribution—which present virtually limitless degrees of freedom. Here, based on high-throughput simulations and a convolutional neural network (CNN) predictor, we present a model allowing us to predict the water sorption isotherm of hydrophilic and hydrophobic nanoporous materials. The training of a CNN generator then enables the inverse design of optimal porous microstructures featuring tailored/unusual sorption isotherms.

Recent successes demonstrated by machine learning (ML) approaches have created great enthusiasm for their potential translational contributions to other fields. Among these are efforts to demonstrate the feasibility of using machine-based approaches to study and speculate on new energetic materials. This talk will discuss the performance of ML models developed to predict the detonation properties of energetic materials. A comprehensive comparison of approaches reveals that models are highly sensitive to feature selection as well as algorithm differences. But using a relatively limited volume of data and careful problem definitions, surprisingly strong cross validation scores suggest that there is promise to both improve and ultimately use ML methods widely in this field. The talk will conclude with an overview and some perspectives of other recent efforts including generative modeling and natural language processing techniques at their intersections with energetics research.

Inorganic scintillator-based detector materials find a wide variety of applications, ranging from medical imaging to radiation detection for global security and high energy physics experiments. These materials essentially convert a fraction of the total energy deposited by incident gamma rays or X-rays into visible or near-visible range of the spectrum. In many applications a scintillator with high light output, fast response time, and emission at suitable wavelengths, is generally required, among many other application-specific desired characteristics. However, owing to several conflicting relationships between the desired properties (e.g., brightness versus speed), no single scintillator is ideal for all uses; there is a need to design custom scintillators optimized for each application. Traditionally, the discovery and design of new detector materials have relied on a laborious, time-intensive, trial-and-error approach; yielding little physical insight and leaving a vast space of potentially revolutionary materials unexplored. To accelerate the discovery of optimal scintillator materials with targeted properties and performance, efforts are ongoing to develop a closed-loop machine learning driven adaptive design framework based on data from literature, in-house experiments and quantum mechanical calculations. This talk will present an overview of this framework, focusing on the screening of complex wide band gaps chemistries with an aim of identifying promising materials that are amenable to band-gap/band-edge engineering to yield custom scintillation properties. The developed framework is general and is expected to prove useful for applications beyond scintillator discovery such as photovoltaics and catalysis.

In recent years the use of various machine-learning (ML) methods to discover new materials with desired properties have emerged as a powerful tool. In this research, we have designed a stacked ensemble model (SEM) by combining several supervised ML models to predict the adsorption of hydrogen in metal-organic frameworks (MOFs). The SEM model was trained by using the hydrogen adsorption data generated by screening of 1000s of MOFs by performing grand canonical monte carlo (GCMC) simulations at 1 atm and 77 K. This SEM was integrated with optimization algorithms and in-house MOF structure generation code to design MOFs with high adsorption of hydrogen. Overall, this new approach of integrating SEM with optimization algorithms accelerated the new MOF discovery process, resulting in a MOF with enhanced hydrogen storage capacity. This new ML framework is transferable and can be used to discover other porous (e.g. zeolites), soft (e.g. polymers), and hard (e.g. alloys) materials for desired applications.

Any sensor material, even if proved as an excellent performer at the research level, it can be put to less use if it does not meet the product criteria. The salient parameters include sensitivity, selectivity, recovery and re-usability. In order to thoroughly evaluate these parameters, huge amount of data is required; using which definite conclusions can be drawn. Calibration, life of the sensor and its scope of selectivity can be well defined. The most crucial part of any successful sensor system is its selectivity and close to 100% accuracy. In this regard statistical analysis along with the sensor design play a crucial role. It can decide how many sensors are required in an array of sensors to determine the presence of a particular analyte and its concentration with highest accuracy. In this context, through this abstract, we present the work done on a metamaterial-inspired ring resonator circuit, which works as an excellent sensor for chemically hazardous gases and liquids, which are propellant materials. Since the molecules in this category mostly have SOx / NOx groups, their chemical differentiation becomes difficult for sensing. We have already shown that metamaterial-inspired ring resonator works on the principle of sensing the changes in the dielectric media in its vicinity, thereby generating a change in the resonant frequency.
In this work we further discuss the approach towards sensor optimization for sensing fixed library of propellant molecules by first. Machine learning algorithms like, Linear Discriminant analysis, Random Forrest and so on, would be optimized to achieve maximum selectivity and accuracy in the sensor. The data would be collected for the sensor operating at 430 MHz, and Vector Network Analyser is used to measure the sensor response. Sensing time, power shift, frequency shift and the recovery time are the basic parameters which would be measured. Sensor output is calibrated for the ppm concentration varying from 10 ppm to 1000 ppm. The data is submitted to the machine learning algorithm and the algorithm is subsequently trained for the unknown sample of the propellant material with unknown concentration, to establish the correctness of the device developed. The emerging data-driven techniques based on statistics collected from the actual sensor testing, and use of machine learning, and artificial intelligence (AI) to unlock predictions of the performance of resonator-based sensor, will be discussed elaborately in this presentation.

The combination of microscopic techniques with time-resolved luminescence detection is a valuable and powerful toolbox to investigate photophysical properties not only of different classes molecular systems, but also semiconductors, nanoparticles and QDs, polymers, solid-states as well as nano structures. Understanding of both photophysical processes as well as structure-property relationships are important steps toward optimization of properties and efficiencies in practical applications.
Investigating fast processes with short lifetimes as well as really long lived excited states (in the µs to ms range) do impose vastly different requirements. On one hand, monitoring fast dynamic processes necessitate high time resolution that can be achieved by using picosecond pulsed lasers and fast detectors along with appropriate time-correlated single photon counting (TCSPC) units. On the other hand, studying long-lived excited states normally require pulsed lasers with low repetition rates, detectors with good efficiency and sensitivity up to NIR, and counting electronics capable of multi-channel scaling mode for acquiring signal on the µs / ms time scale.
Here we demonstrate another way to investigate long lifetime decays (e.g., phosphorescence or delayed fluorescence) of samples by using adapted excitation in combination with a microscope for good spatial resolution, better signal quality, and faster measurement time. This PLIM setup uses a special laser driver with appropriate laser heads and suited photon counting electronics which enables experiments with higher photon flux, even with low laser repetition rates. Resulting intensity images clearly show an almost 7 fold higher intensity using the adapted excitation instead of excitation with a low repetition rate. Therefore we can reduce the acquisition time significantly because of the increased emission intensities and by recording lifetime images with a focus on detecting of photons with longer lifetimes in combination with fast electronics, that are not limited by dead-time effects.

Large volumes of data generated during high-throughput synthesis and rapid characterizations call for rapid and automatic data analysis to accelerate material discovery. Herein, we report a convolution neural network (CNN) that was trained based on theoretic data and very limited experimental data for fast identification of experimental X-ray diffraction (XRD) spectra of metal-organic frameworks (MOFs). To augment the theoretical spectra as the training datasets, noise extracted from experimental spectra was applied to the theoretic spectra. The optimized CNN model can successfully identify all individual spectrum within Top 2 choice. Data dimension reduction analysis on the experimental XRD spectra by the local linear embedding (LLE) shows that the spectra from the same material are clustered in groups in the LLE component map. Some groups are highly distinguishable, while others are very close or overlapped. The reported CNN model together with the data-augmentation technique and procedure for interpreting the black box would not only open the numerous potential applications for identifying XRD spectra for different materials, but also pave avenues to autonomously analyze data by other characterization tools such as FTIR, Raman, and NMR.

We suggest a practical and scientific method to select the optimal piezoelectric material to generate electricity from vibration induced by sound. The proposed method begins by modeling deflections caused by given voltages using partial differential equations (PDE), then implementing machine learning (shallow neural network) to estimate the generated electricity from varying deflections. During the PDE computations, while the geometry such as width and length is fixed, material properties such as the Poisson’s Ratio, Shear Modulus, and Elastic modulus are changed according to the selected piezoelectric material. By running a machine learning simulation trained by these data, we obtain the relationship between voltage changes and deflections which provides the way to select optimal materials. Among the 33 easily accessible, mostly commercialized piezoelectric materials, the trained learning model predicts voltage from given sound/deflection. Upon analysis, it is observed that the Steel Stainless Austenitic 304, Nickel Alloys, and Cork performed highest for our application by showing higher sensitivity between induced voltage to the deflections. This result demonstrates the versatile applicability of our scheme. It shows its transferability and, in other words, how readily the simulation can be applied to other materials with different properties and deflection variations.

Heterogenous catalysts are an important class of materials that receive considerable research attention due to their large impact on energy and the environment. Aberration-corrected in situ environmental transmission electron microscopy (ETEM) is a powerful catalyst characterization tool capable of providing atomic-scale structural information from active catalysts under reaction conditions. Recent advancements in the realization of highly efficient direct electron detectors now enable atomically-resolved ETEM images to be acquired with a temporal resolution in the millisecond or sub-millisecond regime where catalytically relevant structural reconfigurations are thought to occur [1,2,3]. While there is potentially much to be gained from applying these new detectors to catalytic nanomaterials characterization, acquiring ETEM image series with high temporal resolution necessarily produces datasets that can be unwieldly in size and severely degraded by noise, rendering traditional image processing approaches impractical and ineffective at extracting useful scientific information. For example, recording a 64 megapixel in situ ETEM movie at 2.5 millisecond time resolution would produce around 100 gigabytes and many hundreds of frames of image data every second. We are interested in exploring innovative approaches to image processing that leverage new advances in data science on large, noisy, atomic-resolution in situ ETEM data sets.

Currently our efforts have focused on simple and scalable real-space processing techniques that involve fitting atomic columns with two-dimensional Gaussian functions [4]. This method yields information on the time-resolved position and intensity of atomic columns throughout the ETEM image series. Additionally, the results of this method serve as a starting point for comparison to results produced from advanced image processing done via artificial intelligence and machine learning-based approaches. We are also working toward the generation of an extremely large dataset of simulated ETEM images (10,000+), which could serve as the training set for a convolutional neural network that identifies the best possible match between an experimentally acquired image and a given structural model that was used in the image simulation. Developing a methodology that assigns a confidence level to the structural model that best matches the experimental image is important to assess the reliability of the results produced through such approaches, which could ultimately provide highly desirable information on the atomic scale structural evolution of catalyst nanomaterials [5].

[1] Somorjai, G.A., Langmuir, 1991. 7(12): p. 3176-3182.
[2] Crozier, P.A., Lawrence, E.L., Vincent, J.L., and B.D.A. Levin, Microsc. and Microanal., 2019. 25(S2): p. 1464-1465.
[3] Faruqi, A. and G. McMullan, Nuclear Instruments and Methods in Physics Research Section A, 2018. 878: p. 180-190.
[4] B.D.A. Levin, et al., 2019. Tracking the Picoscale Motion of Atomic Columns During Dynamic Structural Change, preprint available at https://arxiv.org/abs/1909.07477
[5] We gratefully acknowledge support of NSF grants CBET-1604971 and OAC-1940263, and the use of facilities at Arizona State University’s John M. Cowley Center for High Resolution Electron Microscopy.

Neural network potentials based on atomic embeddings have demonstrated accurate force and energy evaluations to simulate physical systems [1-5]. Moreover, it also shows good force field transferability across the chemical space of interests, which are critical for efficient computational screening. We propose an active learning strategy [6] that accelerates the sampling and training of transferable neural networks, particularly to capture supramolecular interactions that are critical for applications like designing chelating agents and electrolytes. The proposed method combines deep neural network training and active sampling of both chemical and configurational space in an end-to-end fashion to leverage the transferable configurational information across different but chemically similar species. This workflow has also enabled a high throughput virtual screening over the Ether-ion chemical space, which consists of 10^3 ~ 10^4 molecules, and successfully discovered candidate highly ion-conducting Ether molecules for organic electrolyte applications[7-10].


[1]Zhang, L., Han, J., Wang, H., Car, R. & E, W. Deep Potential Molecular Dynamics: a scalable model with the accuracy of quantum mechanics. (2017).

[2]Schütt, K. T., Sauceda, H. E., Kindermans, P.-J., Tkatchenko, A. & Müller, K.-R. SchNet–A deep learning architecture for molecules and materials. J. Chem. Phys. 148, 241722 (2018).

[3]Behler, J. & Parrinello, M. Generalized neural-network representation of high-dimensional potential-energy surfaces. Phys. Rev. Lett. 98, 1–4 (2007).

[4]Yao, K., Herr, J. E., Toth, D. W., Mckintyre, R. & Parkhill, J. The TensorMol-0.1 model chemistry: a neural network augmented with long-range physics. Chem. Sci. 9, 2261–2269 (2018).

[5]Smith, J. S., Isayev, O. & Roitberg, A. E. ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost. Chem. Sci. 8, 3192–3203 (2017).

[6]Smith, J. S., Nebgen, B., Lubbers, N., Isayev, O. & Roitberg, A. E. Less is more: Sampling chemical space with active learning. J. Chem. Phys. 148, 241733 (2018).48, 241733 (2018).


[7]Seki, S., Serizawa, N., Takei, K., Dokko, K. & Watanabe, M. Charge/discharge performances of glyme-lithium salt equimolar complex electrolyte for lithium secondary batteries. J. Power Sources 243, 323–327 (2013).

[8]Ueno, K. et al. Glyme-lithium salt equimolar molten mixtures: Concentrated solutions or solvate ionic liquids? J. Phys. Chem. B 116, 11323–11331 (2012).

[9]Ueno, K. et al. Li+ solvation in glyme-Li salt solvate ionic liquids. Phys. Chem. Chem. Phys. 17, 8248–8257 (2015).

[10]Westman, K. et al. Diglyme Based Electrolytes for Sodium-Ion Batteries. ACS Appl. Energy Mater. 1, 2671–2680 (2018).

Recently Machine Learning (ML) is becoming an increasingly important tool for material discovery, thanks to its ability to predict results of time and cost-intensive calculations like first-principles calculations based on the density functional theory (DFT)^[1]. In simple terms, each ML model predicts some property and is composed of three equally important elements: a database, a descriptor, and a ML technique.

In this project, we are predicting formation enthalpies (ΔH_f) of arbitrarily defined compounds. To create the model, we have taken an open-source DFT database (OQMD), state of the art Voronoi tesselation-based descriptor developed by the Wolverton group^[1], and then designed a new neural network which allowed us to improve the predictability of formation energy significantly.

The neural network architecture design was done in an empirical fashion. We started with a simple neural network and made iterative modifications to structure and hyper-parameters until we reached Mean Average Error (MAE) around uncertainty in the data. Our best performing network achieved MAE of 35 meV/atom when cross-validated on the test set of randomly selected 5% of the entire OQMD (21,800 entries).
When compared to competing models based on the Random Forest technique, this MAE value represents a two-fold improvement compared with the model using the same descriptor (≈70 meV/atom)^[1] and a three-fold improvement compared to the Coulomb Matrix descriptor (≈100 meV/atom)^[1,2].

Furthermore, we have used our model to investigate the OQMD data set and have found a small number of OQMD entries that had non-physical ΔH_f. After the removal of three such entries from our test set (3/21,800), the cross-validation MAE has dropped to 28 meV/atom.

Our model provides excellent predictability, takes milliseconds to run, and is available ready-to-use in open-source MXNet format, that can be easily deployed on any device ranging from a supercomputer to an iPhone 5S.

References:
[1] L. Ward et al., "Including crystal structure attributes in machine learning models of formation energies via Voronoi tessellations", Physical Review B, vol. 96, no. 2, 2017
[2] F. Faber et al., "Crystal structure representations for machine learning models of formation energies", International Journal of Quantum Chemistry, vol. 115, no. 16, pp. 1094-1101, 2015.

Superconductivity has been an area of intense research for more than a century, yet the intentional design of new superconductors remains a daunting challenge. The massive databases of experimental and calculated materials properties that have developed over recent years makes the use of a data-driven search for new superconducting materials possible. In this study, machine learning is used to model the T_c for 21,000+ known metallic and oxide superconductors in the SuperCon database [1]. The model uses 60 features that are generated based on a compound’s chemical composition. The accuracy of a random forest model with 10-fold cross validation is found to be 92%. Using this model, the T_c of ~4,500 new theoretically stable ternary compounds, with unknown structure, is predicted. Of these compounds, 571 are predicted to have a T_c within 30-77 K and 8 with T_c expected in the range 85-100 K. Examples of identified materials include chalcogen-containing compounds such as BeSe₅Cs₈, BiCs₂Se_2.5, and TeBa₄O₇, as well as hydrogen-rich compounds such as TeO₅H₈, CuO₅H₉, and AgO₅H₉. Evolutionary structural prediction of identified promising compounds will enable ab-initio studies and provide guidance for corresponding material synthesis and characterization.

[1] Superconducting Material Database (SuperCon), National Institute of Materials Science [https://supercon.nims.go.jp/index_en.html] (accessed October 21, 2019).

Layered perovskite structures show high performance as an air electrode in solid oxide fuel cells and proton ceramic fuel cells [1]. Radii difference among metal elements at site A of ABO₃ is a prerequisite for conversion of perovskite structure from complex to layered [2]. In this study, we study energetically favorable structures according to change in composition of site A elements (Pr, Ba, Sr) and site B elements (Co, Fe).
25 compositions were determined by changing x and y in PrBa_1-xSr_xCo_2-yFe_yO_5.5 (x = 0, 0.25, 0.5, 0.75, 1, and y = 0, 0.5, 1, 1.5, 2). The 2×2×2 supercells with the compositions were studied using DFT. Since the number of possible atomic arrangements for each composition is very large, however, a genetic algorithm was used to search energetically favorable structure efficiently [3]. The highest energy difference between complex and layered structures among the 25 compositions was obtained at x=0 and y=1.
[1] S. Choi, et al., Sci. Rep., 2013, 3, 2426; S. Choi, et al., Nat. Energy, 2018, 3, 202
[2] A. A. Taskin, et al., Prog. Solid State Chem., 2007, 35, 481.
[3] J. Kim, et al., Comput. Mat. Sci., 2017, 138, 219.

Autonomous Research Systems like ARES™ are disrupting the research process by using AI and Machine Learning to drive closed-loop iterative research. ARES™ is our autonomous research robot capable of designing, executing and evaluating its own experiments at a rate of up to 100 iterations per day. Previously ARES taught itself to grow carbon nanotubes at controlled rates (NPJ Comp Mat 2016). Here we discuss recent research campaigns on maximizing carbon nanotube growth rates using a Bayesian optimization planner. We also use HOLMES and knowledge gradient descent to introduce advanced decision policies with local parametric models to control nanotube diameter. Implications for nanotube materials development will be discussed. Finally, we have developed a new research robot for additive manufacturing, AM ARES™, which is at the early stages of teaching itself to print structures with unknown inks. We plan to make the AM ARES™ Robot available online so that a broad community of researchers can test concepts and approaches for AI/ML and experimental design as applied to 3D printing, thus building to the larger goal of enhancing citizen science.

Due to recent improvements in image resolution and acquisition speed, materials microscopy is experiencing an explosion of published imaging data. Intensive human efforts, however, are required to digest such published images. In this talk, we will discuss the MaterialEyes project, in which deep learning and natural language processing tools are used to extract microscopy images from scientific literature. The process is accomplished via a suite of scientifically-informed software tools in the EXSCLAIM python package (automatic EXtraction, Separation, and Caption-based natural Language Annotation of Images from scientific figures), which is ultimately focused on the acceleration of inductive materials insights by collecting, structuring, and interpreting an aggregate of materials microscopy imaging data. In this talk, we will discuss how deep learning and NLP is used in EXSCLAIM to: (1) extract figures and caption data from open source scientific documents, (2) automate the construction of self-annotated materials imaging datasets, and then (3) leverage literature-based insights to assist in image classification and segmentation tasks. In addition, we will discuss a new pipeline for crowdsourced image labeling that was used for training some of the deep learning models.

Characterization of nanomaterial morphologies with advanced microscopy and/or spectroscopy tools plays an indispensable role in nanoscience and nanotechnology research, as rich information about the chemical compositions, crystallography, other physical and chemical properties, as well as the growth mechanism can be extracted from morphology analysis. However, the interpretation of imaging data heavily relies on the “intuition” of experienced researchers. As a result, many of the deep graphical features are often unused because of difficulties in processing the data and finding the correlations. Such difficulties can be well addressed by deep learning. In this work, we use the optical characterization of two-dimensional (2D) materials as a case study.
Currently, the most widely used approach to obtain high-quality 2D crystals in laboratories is mechanical exfoliation, followed by 2D crystal “hunting” under an optical microscope. This task is time-consuming and difficult especially for inexperienced researchers. Very recently, there has been growing interest in automating such a process. However, existing optical contrast based identification methods could not satisfy the needs for a general-use, fully automatic optical identification and/or characterization system, as they are often very specific to types of 2D crystals, conditions and configurations of the microscopes being used, image qualities, etc. In addition, optical contrast based methods would fail for harder problems in which the classes to be differentiated are not separable in the color space, such as identifying the materials or even predicting their physical properties from unlabeled optical images.
In reality, optical microscopic (OM) images contain rich, often unused information other than optical contrast. These deep graphical features can be extracted through deep learning, especially semantic segmentation methods based on convolutional neural networks (CNNs). In this work, an encoder-decoder semantic segmentation network called “SegNet” is configured for pixel-wise identification of OM images of 2D materials, in real time, and regardless of variations in optical setups. A fully automated system utilizing this algorithm can be used to free up tremendous amount of time for researchers. Additionally, we find that the algorithm finds correlations between the OM images and physical properties of the 2D materials and can thereby be used to anticipate the properties of new, as-yet uncharacterized 2D crystals. Finally, a transfer learning technique is applied to adapt the pretrained network to more optical characterization applications such as identifying layer numbers of chemically synthesized graphene domains. The proposed methodology can potentially be extended for identification and understanding other morphological or spectroscopic data of diverse nanomaterials.
Acknowledgement: This work is supported partially by U.S. ARO through the Institute for Soldier Nanotechnologies, under cooperative agreement number W911NF-18-2-0048, AFOSR FATE MURI, grant no. FA9550-15-1-0514, and the STC Center for Integrated Quantum Materials, NSF Grant No. DMR-599 1231319.

We introduce the first systematic database of scanning tunneling microscope (STM) images obtained using density functional theory (DFT) for two-dimensional (2D) materials. Specifically, the database is constructed using the Tersoff-Hamann method. It currently contains data for 681 exfoliable 2D from the JARVIS-DFT database. These computational STM images can directly aid the experimental analysis of experimental STM results as they provide ideal reference images. Examples of the five possible Bravais lattice types for 2D materials are discussed and the images for all the materials are made available on the website. We validated the computational STM technique by comparing our results to a few experimental images. As the analysis of multiple STM images during an experiment could be tedious, we used our DFT-STM image database to train a high-accuracy convolution neural network (CNN) model to accelerate the characterization process.

A major obstacle in applying machine learning to scientific and engineering tasks is the lack of engineering rules for the design of machine learning systems that allow to make these system reproducible and the limits of the systems known. Based on joint research with physicists and material engineers, this talk presents engineering principles and measurement processes[1] that aim at structured engineering of machine learning systems.
While the focus of this talk is on neural networks, the principles and presented measurements are universal and applicable to a variety of machine learning systems.

[1] G. Friedland: "Reproducibility and Experimental Design for Machine Learning on Audio and Multimedia Data", Proceedings of the 2019 ACM International Conference on Multimedia, October 2019. https://dl.acm.org/citation.cfm?id=3350545

Predicting the magnetostriction of a large-scale (mm-mm) polycrystalline Tb_xDy_1-xFe₂ (Terfenol-D) alloy by direct numerical simulations is mathematically tedious and computationally expensive, because it requires solving a set of partial differential equations which describe the mutual coupling between local magnetization dynamics and polycrystal elasticity, and because it requires a large number of simulation cells (tens of millions and more). For accelerating the prediction of magnetostriction, we apply a graph-based deep learning method, namely, the Graph Neural Network (GNN), to discover the correlation between a microstructure dataset of 350 different polygrain structures and a property dataset describing the magnetic-field-induced deformation of each polygrain structure (viz., magnetostriction). In the GNN based approach, each grain is represented as a node, in which all important physical features (e.g., Euler angle, magnetization direction) inside a grain are stored. As a result, the polygrain structure is represented as a network of nodes, in which the identity information of each grain(node)’s neighboring grains(nodes) is also preserved. This in turn allows us to consider the grain-grain interaction which critically influences the magnetostriction of the entire alloy and interpret the contribution of each node(grain) to the magnetostriction. A >90% prediction accuracy has so far been achieved with ample space for improvement. Our results demonstrate the exciting potential of applying GNN to accelerate the discovery of microstructure-property linkage in polycrystalline materials.

Materials are ubiquitous in society and lie at the heart of many technologies including energy, communication, buildings, transportation, the environment and medicine. The constant need to develop new materials with properties that are matched to applications is a major driver of materials research. Understanding how atoms come together, self-assemble, dissociate and undergo phase transformations under different stimuli is a critical question in materials science. Transmission electron microscopy can now routinely achieve a spatial resolution of better than 1 Å allowing the projected spatial distribution of atoms to be observed. In situ methodologies allow these signals to be mapped both spatially and temporally in the presence of different stimuli. Rapid advances in detector and spectrometer technology now allow local imaging, diffraction and spectroscopic signals to be acquired at unprecedented rates, generating terabytes of data in relatively short times.

The availability of large spatially and temporarily resolved datasets offers unprecedented potential to transform our understanding of the fundamental atomic level dynamic processes taking place in materials. The rate at which atomic resolution data can be generated in experiments has now outstripped our ability to process and mine scientifically important information. The complex nature of atomic resolution images necessitates intense supervision and input in most quantitative image processing operations. Such an approach is possible for a few hundred images but becomes completely impractical when millions of data frames are available. This has become a serious bottle-neck to extracting the new knowledge potentially available from advanced characterization systems. We need approaches that can recognize recurring motifs in noisy data sets, provide statistical confidence levels for detection of real structural features, track dynamic changes within the sample, and flag events which are likely to be scientifically significant.

We are focused on applying emerging data-driven techniques based on statistics and machine learning to characterize structural dynamics in nanomaterials with an initial focus on metal and oxide nanoparticles 1 – 3 nm in size. Nanoparticles have wide applications in many areas of materials science and are an ideal system for developing and benchmarking novel data-driven processing techniques because they only contain a few thousand atoms. The datasets are acquired using in situ electron microscopy and structural dynamic processes are stimulated using heat, gas or electron irradiation. We compare and contrast the performance of automated image processing algorithms designed to determine atomic coordinates and structural evolution with more conventional manual methods. The importance of understanding the confidence level that can be associated with derived structural models is critical to assess the reliability of the scientific information extracted from large data sets.

Acknowledgements
We gratefully acknowledge support of NSF OAC-1940263, and the use of facilities at Arizona State University’s John M. Cowley Center for High Resolution Electron Microscopy.

Developed as a Baushinger effect monitoring technique in polycrystalline metallic materials, passive acoustic wave monitoring has been a critical health monitoring technique in: Atomic energy pressurized tank; Power and structure monitoring in aeronautics; Automotive; As well as recent applications in spinning disk data storage (HDD) health monitoring applications [1]. Passive monitoring of acoustic waves generated during the initial contact point has been attracting the attention of material scientists since the inception of nanomechanical test instruments for nanoindentation and nanoscratch applications. The conventional acoustic wave signal treatment via RMS or integrated energy values proves that quantitative acoustic wave properties correlate well with the local contact materials phenomena such as: Yield point initiation for W(100) [2, 3]; Sapphire [4], phase transformation for SMA [2]; and differentiating of thin film fracture modes [5]. Several attempts have been made to look at the differentiative properties of the acoustic signatures via signal decomposition techniques such as wavelets [2, 6]. Even though acoustic wave signatures were reconstructed, the true potential of the method was not investigated from the machine learning perspective.
In this work a deep learning based signal processing of nanoindentation induced acoustic events was investigated. The synergy of wavelet signal decomposition and information theory based signal presorting prepared data for the deep learning step. In the deep learning step convolutive neural networks sort wavelet coefficients by their statistical significance via provided 2D signal topographies. This creates acoustic signature libraries that are typical for the specific materials phenomena during the indent. The hardware consists of a newly developed ultrasonic probe integrated into the nanoindentation tip, thus eliminating boundary effects and ensuring that only waves that pass via the contact are being recorded. Appropriate signal conditioning and fast data acquisition hardware is synchronized with a quasi-static nanoindentation process. The machine learning routine together with wavelet decomposition and presorting algorithms are implemented into the dedicated acoustic data evaluation kernel which resides in the fast access memory of the designated controller.

REFERENCES:
1. A. Daugela, S. Tadepalli, J. of Microsystem Technol., Vol. 18,No. 9-10(2012) pp. 1425 - 1430.
2. A. Daugela, H. Kutomi, T.J., Wyrobek, Zeitschrift fur Metallkunde, Vol. 92, No 9 (2001), pp. 1052-1056.
3. N. Tymiak, A. Daugela, T. J. Wyrobek and O. L. Warren, J. of Mater.Res., Vol. 18, No. 4 (2003), pp. 1 – 13.
4. N.I. Tymiak, A. Daugela, T.J. Wyrobek and O.L. Warren, Acta Materiallia, Vol. 52 (2004) pp. 553-563.
5. N. Faisal, R. Ahmed, R.L Reuben, Int. Mater. Reviews, Vol. 56, No. 2 (2011).
6. N. Tymiak, D. Chrobak, W. W. Gerberich, O. Warren, R. Nowak, Phys. Rev. B - Cond. Matter and Mater. Phys., Vol. 79, No. 17, (2009).

Thermal characterization in high-temperature environments is difficult because of high radiation losses and significant temperature measurement noise and inaccuracy. Ångström’s method can accurately measure thermal diffusivity in the presence of heat losses by observing the propagation of temperature perturbations through a material sample. This work applies a modified Angstrom’s method to measure thermal properties at extreme temperatures (>1000 °C). Previous studies employing Angstrom’s method at high temperatures typically have employed evenly spaced thermocouples on samples for thermal diffusivity measurements. We have recently demonstrated in preliminary room-temperature testing (copper - 1.13 x 10^(-4) m2/s, 1% uncertainty) that the uncertainty in thermal diffusivity within a Bayesian statistical framework is inversely proportional to the square root of the number of concurrent temperature measurements. In this work, we employ IR thermography for non-contact temperature measurement to greatly increase the number of measurements that can be resolved on a sample compared to the conventional thermocouple approach. To subject materials to extreme temperatures, samples are mounted vertically in a vacuum chamber and heated by concentrated thermal radiation from a 10 kW Xenon bulb. A controller is used to precisely modulate power output to produce the periodic temperature oscillation required for Angstrom’s method. Transient temperature measurements are made along the length of the sample using both an infrared camera and optical spectroscopy. Bayesian statistics then quantify the measurement uncertainty associated with the noisy high temperature environment, and this approach is shown to substantially reduce measurement uncertainty for high-temperature material characterization.

Predicting the performance of rechargeable batteries in real time is of great importance to battery research and industrial production, and hence has been a long pursuit. In this talk, a novel end-to-end unsupervised machine learning approach is shown for the battery performance prediction, which is free from feature engineering and uses only the raw images of the charge–discharge voltage profiles [1]. This model enables unsupervised real-time automatic extraction of latent physical factors that control the performance of Na-ion batteries to classify good or bad cycling performance by using only the voltage profile of the first cycle. This model can also monitor the safety of Li-metal battery systems by giving warnings when the battery is approaching a failure. With the beyond expert-level prediction ability, the abovementioned framework can be a promising prototype to further develop and enable high accuracy predictions of battery performance for realworld applications in the future.

[1] Beyond expert-level performance prediction for rechargeable batteries by unsupervised machine learning, Advanced Intelligent Systems, 1900102 (2019)

Materials exhibiting higher mobilities than conventional organic semiconducting materials (e.g. fullerenes and heteroacenes [1]), are in high demand for applications such as printed electronics. In order to explore new molecules in the heteroacene family that might show improved mobility, a massive theoretical screen of hole-conducting properties of molecules were performed using a cloud computing environment. Over 7,000,000 structures of fused furans, thiophenes and selenophenes were generated, and 250,000 structures were randomly chosen for subsequent DFT (Density Functional Theory) calculations of hole reorganization energies (λ_h). Utilizing the cloud-computed DFT dataset of a quarter-million reorganization energies, the de novo design method proposed by Gomez-Bombarelli [2] was applied to find further chemical structures with minimal reorganization energy. This method converts molecular structures into continuous variables by applying the variational autoencoder/decoder technique, which enables optimization of chemical structures in a continuous numerical space. Results of the inverse design showed that the method has the ability to generate exotic chemical matter, such as structures with a seven- and eight-membered rings, with reasonably low calculated reorganization energies.
[1] K. Takimiya et al., Acc. Chem. Res. 47 (2014), 1493.
[2] R. Gomez-Bombarelli et al., ACS Cent. Sci. 4 (2018) 268.

Materials exhibiting higher mobilities than conventional organic semiconducting materials, (e.g. fullerenes and heteroacenes [1]), are in high demand for applications such as printed electronics. In order to explore new molecules in the heteroacene family that might show improved mobility, over 7,000,000 structures of fused furans, thiophenes and selenophenes were generated and 250,000 structures were randomly selected for subsequent DFT (Density Functional Theory) calculations of hole reorganization energies (λ_h). Utilizing the cloud-computed DFT dataset of a quarter-million reorganization energies, inverse materials design methods were applied to find further chemical structures with minimal reorganization energy. The de novo design methods applied were (1) a method based on reinforcement learning of SMILES strings (the REINVENT method [2]), and (2) a junction tree variational autoencoder that directly generates molecular graphs [3]. Both methods succeeded in obtaining chemical structures whose reorganization energy was lower than the lowest energy in the DFT training dataset, suggesting that an extrapolative materials design protocol can be developed by applying two methods. Finally, these results will be compared to the results obtained from the latent-space de novo molecular design method proposed by Gomez-Bombarelli et al. [4].
[1] K. Takimiya et al., Acc. Chem. Res. 47 (2014) 1493.
[2] M. Olivecrona et al., J. Cheminform 9 (2017) 48.
[3] W. Jin et al., 35th Intl. Conf. Machine Learning, ICML 2018, 5, 3632.
[4] R. Gomez-Bombarelli et al., ACS Cent. Sci. 4 (2018) 268.

Solar fuel generation from H₂O, CO₂ and sunlight comprises a promising renewable energy technology whose future proliferation relies on materials advancements, including identification of a photoanode for the oxygen evolution reaction. A water-oxidizing photoanode can be coupled to a (photo)cathode for scalable generation of H₂ and C-containing fuels. Metal oxides are the most extensively studied class of OER photoanode materials due to the availability of a multitude of metal oxide phases that may stably operate in the oxidative electrochemical environment while converting solar to photoelectrochemical (PEC) energy. Solar fuels photoanodes are quite rare due to the combined needs of solar absorption, charge carrier separation and transport, oxidative stability, and catalytic activity for the OER. The compendium of known metal oxide photoanodes that operate in the visible spectrum is being rapidly expanded via theory-guided high throughput experimentation, including our combination of combinatorial materials synthesis coupled with automated UV-vis spectroscopy and PEC experiments. Discovery of dozens of ternary metal oxide photoanodes will be discussed with focus on the lessons learned in terms of the role of theory and the importance of compositionally graded libraries, and a historical analysis of the discovery rate of metal oxide photoanodes reveals increasing chemical diversity, suggesting there may be many more materials to be discovered. The rapidly expanding portfolio of metal oxide photoanodes creates new avenues of inquiry to realize the requisite stability and efficiency for deployment of solar fuels technologies.

A model describing the origin of water-induced degradation of certain Metal-Organic Frameworks (MOFs) is derived, aiming to explain experimental observations of water stability and instability of prototypal MOFs. The proposed model suggests crucial rule of metal-coordinated water molecules, and solution pH, in triggering a cascade of reactions that in several cases results in pronounced irreversible degradation of several MOFs. It is suggested that such degradation pathways can be affected in neutral to basic solution by deprotonation of metal-coordinated water molecules, or bridging hydroxyl species in multinuclear clusters. Calculations of pKa values for the metal-coordinated water molecules, complemented by geometry relaxation calculations, and variable pH infrared spectroscopy techniques were all simultaneousley utilized to draw a comprehensive description for the early stages in MOFs degradation in aqeuous medium. Mostly dependent on the nature of metal nodes, we propose two distinct mechanisms for structural degradation of carboxylate-based MOFs. The MOFs investigated herein included MOF-5, HKUST-1, UiO-66, and MIL-101-Cr. It is suggested that for each type of MOF, a narrow pH stability range can be described, mostly determined by the pKa of metal-coordinated water molecules and the relative strength of metal-carboxylate interactions. Beyound this fundamental pH range, a relativeley wider pH window can exist, where certain MOFs can be kinetically inert. Finally, we present a highly transferrable and facile approach to induce surface passivation of MOFs crystals to effectively extend such aqeuus solution stability limits, as proven by the case of appreciably maintained microporosity of HKUST-1 in neutral, acidic and basic aqeous solution.

The semiconductor photocatalyst β-TaON used for water splitting is of interest due to its promising properties, such as stability, suitable band gap for visible light photocatalysis and carrier mobility. We have implemented a combinatorial, materials discovery approach using pulse-laser deposition (PLD) for thin film growth, X-ray diffraction (XRD) for phase determination and machine learning for data reduction. A lateral compositional gradient of TaO_xN_y was grown across the surface of a α-Al₂O₃ (012) 25-mm-wide wafer by PLD at 675°C in 10 mTorr of N₂. After annealing in N₂ at 1000°C, small-area XRD scattering patterns were collected for 71 positions across the lateral gradient. Unsupervised machine learning separated the XRD data into 4 clusters (phases); one of which turned out to be the desired β-TaON phase. Using high-resolution XRD we determined that the β-TaON region of the film to be single crystal epitaxy with out-of-plane β-TaON (100) // α-Al₂O₃ (012) and in-plane β-TaON (010) // α-Al₂O₃ (2-10). X-ray reflectivity (XRR) analysis of the β-TaON region of the film determined the film thickness to be 260 Å with an electron density matching that expected for β-TaON. Ta 4f_7/2 XPS measurements across the graded film showed Ta⁵⁺ valence state at the β-TaON region of the film. In comparison to the standard method of growing many uniform films at different compositions, this combinatorial method that produces a library of phases on a single wafer is much more efficient for materials discovery.

Based on their structure, non-crystalline phases can fail in a brittle or ductile fashion. However, the nature of the linkages between structure and propensity for ductility in disordered materials has remained elusive. Here, based on molecular dynamics simulations, we investigate the fracture of several disordered phases (metallic glass, glassy silica, colloidal gel, etc.) with varying degrees of disorder. We find that that, in general, structural disorder results in an increase in ductility. By applying the activation-relaxation technique (an accelerated sampling method to identify transition states), we show that the degree of plasticity is controlled by the topography of the energy landscape.

Computing the ground state energy of quantum many particle systems is of fundamental importance in a variety of fields, including chemistry, physics and materials science, amongst others. However, the complexity of such quantum mechanical computations grows rapidly with the number of particles. Machine learning algorithms do not simulate the physical system, but instead estimate solutions by interpolating values provided by a training set of known examples. However, precise interpolations may require a number of examples that is exponential in the system dimension, and are thus intractable. Tractable algorithms compute interpolations in low dimensional approximation spaces, which leverage the underlying physical properties of the system.

In this talk I will give an overview of machine learning algorithms for computing the ground state energy of quantum many particle systems, describing the core principles of such algorithms and illustrating the common themes that emerge. I will then present in more detail our approach to the problem, which is based on a type of multiscale, multilayer convolutional neural network, called a wavelet scattering transform. Through a cascade of multiscale wavelet transforms and nonlinearities, the scattering transform encodes the appropriate invariants and regularity properties of the physical system. Wavelet scattering regressions, computed over databases of organic molecules and amorphous materials, achieve errors on the order of quantum mechanical simulations, but at a fraction of the computational cost.

This multi-disciplinary research proposes a new framework that can infuse scientists' experience and knowledge (named as scientists' eyes) into a constantly evolving, glass-box physics rule learner that can essentially extract hidden physics rules behind complex nano-scale phenomena. Recently, physics principles are gradually integrated into machine learning (ML) methods, but it is at its infancy to discover hidden physics rules. Such pursuits for hidden rules face formidable challenges when the physical experiments involve multifaceted measurements across three-dimensional (3D) specimens or objects. These remaining challenges render even advanced ML methods such as deep learning secondary, not as a primary explorer, in the broad science and engineering domains. Therefore, there exists a deep chasm still remains between materials science and ML community.
To overcome this challenge and unravel hidden physics rule behind nano-scale complex phenomena, we reinterpreted the deep learning's successful notion of multi-layered convolutions and gave rise to a new notion of "convolved information index." The convolved information index can facilely handle physical measurements over 3D nano-scale specimens and measurements without losing experienced perspectives of scientists in the materials science. Also, the framework proposes "link function" as a generator of mathematical expressions about the identified nano-scale physics rules, thereby pursuing "glass-box" (as opposed to the black-box) learning and prediction. Consistent evolution capability is realized by integrating Bayesian update and evolutionary algorithms.
The framework is applied to nano-scale contact electrification (CE) phenomena, which are difficult to elucidate due to length scales and complex physical interactions [1,2]. Three experimental data sets are used. Tribocharged PDMS nanocup array is fabricated by the sequence of (1) prepare a PET mold with a 750 nm-pitch triangular array of nanocones, (2) pour the liquid phase PDMS, Sylgard 184, Dow Corning, mixed with the curing agent, (3) solidify the specimen, and (4) peel off the PDMS from the PET mold (for more details see [1,2]). Atomic force microscopy (AFM) in the tapping mode is used to measure the surface topography and potential. The charge values were measured by Kelvin probe force microscopy over patterned nano-cups represented by 3D point cloud. The experimental data consist of a variety of conditions regarding geometries, demolding directions, and nano-scale patterns (3000 nm by 3000 nm square area with an interval of 5.86 nm; 4000 nm by 4000 nm, with 7.8125 nm). The nano-cup's average heights are 154 nm, 93.5 nm, and 117 nm. The proposed glass-box learning begins from horizontal-demolding test data, and then the best-so-far identified rules are inherited to the next learning with new specimens in a chained way.
The feasibility tests show promising performances of the proposed framework in unraveling transparent, reasonable “expressions” of hidden physics rule along with reasonable reproduction of complex patterned distributions of electric potentials over the 3D point cloud. The proposed framework is believed to establish a promising partnership of scientists’ eyes and the advanced ML methods, thereby enabling data and ML-driven discovery of groundbreaking physics laws in the nano-world and beyond. This presentation will cover how this new framework can facilitate materials scientists’ discovery with actual nano applications.
[1] Qiang Li, In Ho Cho, Rana Biswas, Jaeyoun Kim, Nano Letters 19, 850-856 (2019).
[2] Qiang Li, Akshit Peer, I. Cho, Rana Biswas, Jaeyoun Kim, Nature Communications 9, 974 (2018).

In this talk we will review recent studies of applying atomistic-level modeling to describe fracture processes at different scales, from molecular to macroscale levels, and to use artificial intelligence and machine learning to design fracture resistant materials. Molecular and related multiscale modeling methods can be an effective method to better understand elementary fracture processes, ranging from brittle to ductile fracture and other materials failure phenomena such as creep, and provides the basis for expansive training sets. To illustrate this framework, we report a design approach for optimizing the toughness of nanocomposite materials using artificial intelligence, implemented in a novel “AutoComp Designer” algorithm. The algorithm consists of a machine learning predictor combined with an AI improved genetic algorithm, which is capable of discovering de novo materials designs in a vast space of possible solutions. Facilitated by a deep convolutional neural network that is trained with a dataset of hundreds of thousands of combinations of soft and brittle materials originating from multiscale models. Through the algorithm, we extend the capability of physical simulations beyond property predictions to optimize the fracture toughness by altering the material distribution. The solutions are generated by our AI model at a dramatically lower computational cost compared to brute-force searching methods. We further investigate the physical mechanism for improving material performance behind the AI approach, and demonstrate the ability of AI to search for optimal designs with very limited sampling. Brute-force molecular dynamics simulations of the nanocomposite designs confirm that our AI design improves the performance by effectively decreasing the stress concentration at crack tip. This AI approach can be easily applied to other nanocomposites, biomaterials and other material classes, and provides a transferrable and reliable rapid design approach expanding current capabilities.

In recent years there has been great progress in the use of machine learning algorithms to develop interatomic potential models. Machine-learned potential models are typically orders of magnitude faster than density functional theory but also orders of magnitude slower than physics-derived models such as the embedded atom method. We demonstrate that machine learning, in the form of genetic programming, can be used to develop accurate and transferable many-body potential models that are as fast as the embedded atom method, making them suitable to model materials on extreme time and length scales. The key to our approach is to explore a hypothesis space of models based on fundamental physical principles and to select models from this hypothesis space based on their accuracy, speed, and simplicity. We demonstrate our approach by developing fast and accurate interatomic potential models for copper that generalize well to properties they were not trained on, and we discuss additional applications. Our approach requires relatively small sets of training data, making it possible to generate training data using highly accurate methods at a reasonable computational cost.

Machine learning (ML) has emerged as a powerful tool in the study and design of materials. In this talk, I will discuss the different ways in which materials can be described and their relative merits and limitations. In particular, I will highlight the importance of choosing representations that not only satisfy known physical invariances (e.g., rotation, translation, etc.), but also incorporate known physics and chemistry in a hierarchy of knowledge. I will provide an overview of leading ML models and applications, and discuss how the problems of data scarcity and bias in materials science can be addressed using multi-fidelity, transfer learning and other approaches.

Electronic levels introduced by impurities and defects in the middle of the band gap are critically important in semiconductors for optoelectronic, photovoltaic (PV) and quantum sensing applications. While “deep” defect levels can prove catastrophic for PV performance by causing non-radiative carrier recombination, impurity levels in the band gap could also be entangled for quantum sensing or lead to increased absorption of sub-gap photons which can enhance efficiencies. Predicting formation energies and charge transition levels for defects in semiconductors is thus paramount; density functional theory (DFT) has been widely applied for such studies to overcome experimental bottlenecks and confirm measured defect levels. However, the requirement of large supercells, advanced levels of theory, and inclusion of charge corrections make DFT computation of defect properties very expensive, and trends and knowledge from previous calculations are not exploited in subsequent ones.

In this work, we combine high-throughput DFT and machine learning (ML) to develop general predictive models for the formation enthalpy and charge transition levels of impurities in two broad semiconductor classes: (a) ABX₃ halide perovskites, and (b) group IV, III-V and II-VI semiconductors. DFT data is generated for impurity atoms selected from across the periodic table and simulated in various possible cation, anion or interstitial sites. Any “semiconductor + impurity” combination is converted into a unique feature vector based on the defect site coordination environment, the tabulated elemental properties (eg. ionization energy, electronegativity, etc.) of the impurity atom, and electronic and energetic properties computed using low cost unit cell defect calculations. State-of-the-art neural networks (NN), random forest regression (RFR) and kernel ridge regression (KRR) are applied to train models for (a) impurity formation enthalpy, and (b) impurity charge transition levels, based on the input feature vector. Model performance is evaluated for different ML techniques, ML parameters such as number of hidden layers and network weights in NN, subsets of features, and size and nature of training dataset; the best predictive models thus obtained are deployed for comprehensive prediction and design purposes. It is seen that models trained on defects and impurities in pure cation/anion composition semiconductors are applicable to mixed or alloyed compositions as well: for instance, by training on data from pure canonical AB semiconductors like CdTe, ZnO, GaAs and SiC, the ML models can make accurate predictions for impurities in alloyed compositions of the same compounds, such as CdTe_1-xSe_x, Cd_1-xZn_xO, Al_1-xGa_xAs_1-ySb_y, etc. This versatility of the machine-learned models provides an avenue to access the optoelectronic impact of any atomic impurity in any possible pure or mixed composition semiconductor belonging to the same chemical space, which is useful for compositions that are expensive to model. We use the predictive models to efficiently screen for dominating impurities, that is impurities that shift the equilibrium Fermi level in the semiconductor as determined by dominant native defects, in hundreds of possible compositions in the halide perovskite and groups IV, III-V and II-VI semiconductor spaces. ML-based screening of impurities was tested against DFT screening for CdTe and seen to have an accuracy of 95%. The quick and accurate estimation of interesting dopants/impurities in semiconductors is made possible by this materials design framework powered by quantum mechanical computations and machine learning, which can be easily expanded to other semiconductor classes and improved by the infusion of fresh data, new descriptors and more advanced ML techniques.

References
1. A. Mannodi-Kanakkithodi et al., Chemistry of Materials 31 (10), 3599–3612 (2019).
2. A. Mannodi-Kanakkithodi, et al., under review, preprint: https://arxiv.org/abs/1906.02244.

In material science, we are frequently interested in understanding the properties and design implication of material at interfaces. These interfaces can be manipulated to improve the desired characteristics of the bulk material. In this study, we are interested in understanding and optimize the impact of interfacial atomic defects on the thermal transport across a Cu/Si junction. To that end, we developed a reinforcement learning based framework to optimize over a potentially large parameter search. Using this technique allows us to accumulate knowledge of the system of a given type of atoms and store this information into a neural network. In this study, we present our results on optimizing the thermal transport by varying the fraction and length of the interfacial atomic defects using molecular dynamics (MD) simulations with normal mode analysis (NMA) to investigate thermal transport.

Deforming a material to a large extent without inelastic relaxation can result in unprecedented properties. However, the optimal deformation state is buried within the vast continua of choices available in the strain space. Motivated by this challenge, we hereby advance a powerful strategy to circumvent conventional trial-and-error methods and adopt a unique type of machine learning framework for rationally designing the most energy-efficient pathway to achieve a desirable material property such as the electronic bandgap. This method invokes recent advances in the field of convolutional neural networks and active learning by utilizing a limited amount of ab initio data for the training of a surrogate model, predicting electronic bandgap within an accuracy of sub-10 meV. Our model is capable of discovering the indirect-to-direct bandgap transition and semiconductor-to-metal transition in silicon by scanning the entire strain space [1]. It is also able to identify the most energy-efficient strain pathways that would transform diamond from an ultrawide-bandgap material to a narrow-bandgap semiconductor or even semimetal [2]. A broad framework is presented to tailor any target figure of merit by recourse to deep elastic strain engineering and machine learning for a variety of applications in microelectronics, optoelectronics, photonics, and energy technologies.

[1] Z. Shi, E. Tsymbalov, M. Dao, S. Suresh, A. Shapeev, J. Li. PNAS, (2019) 116, 4117-4122
[2] E. Tsymbalov, Z. Shi, M. Dao, S. Suresh, A. Shapeev, J. Li (2020) in submission

Machine learning has emerged as a powerful tool for the analysis of mesoscopic and atomically resolved images and spectroscopy in electron and scanning probe microscopy. The applications ranging from feature extraction to information compression and elucidation of relevant order parameters to inversion of imaging data to reconstruct structural models have been demonstrated. However, the fundamental limitation of the vast majority of machine learning methods is their correlative nature, leading to extreme susceptibility to confounding factors. While in classical statistical methods the methodology to address phenomena such as Simpson paradox are established, the complex and often non-transparent nature of modern DL tools renders them extremely prone to misinterpretation. We argue that correlative machine learning provides a reliable and powerful tool in cases when the causal links are well established, and discuss several examples of atom finding in SPM and STEM and analysis of 4D STEM data when this condition is satisfied. Notably, ML applications in theory generally fall under this category. Alternatively, ML methods work well when the confounding factors are effectively frozen via the narrowness of experimental conditions or experimental system. However, both these conditions are violated for experimental studies, when causal relationships are known only partially (and are in fact often the target of study) and confounding factors (composition uncertainty, microscope tuning, contaminations) are abundant. In this presentation, I will discuss the several examples of extending machine learning methods towards the analysis of causative physical mechanisms. One such approach is based on the Bayesian methods that allow to take into consideration the prior knowledge the system and evaluate the changes in understanding of the behaviors given new experimental data. The second pathway explores the parsimony of physical laws and aims to extract these from the set of real-world observations. Finally, the Bayesian networks can be used to explore the causative relationships in the multimodal data sets. These concepts will be illustrated using several examples of causal machine learning, including analysis of phase transitions on a single atom level in 2D materials and interplay between physical and chemical effects in the ferroelectric perovskites. Ultimately, we seek to answer the questions such as whether electronic instability due to the average Fermi level guides the development of the local atomic structure, or frozen atomic disorder drives the emergence of the local structural distortions, whether the nucleation spot of phase transition can be predicted based on observations before the transition, and what is the driving forces controlling the emergence of unique functionalities of morphotropic materials and ferroelectric relaxors.

This research is supported by the by the U.S. Department of Energy, Basic Energy Sciences, Materials Sciences and Engineering Division and the Center for Nanophase Materials Sciences, which is sponsored at Oak Ridge National Laboratory by the Scientific User Facilities Division, BES DOE.

Solar energy plays an important role in solving serious environmental problems and meeting the high energy demand. However, the lack of suitable materials hinders further progress of this technology. Here, we present the largest inorganic solar cell material search till date using density functional theory (DFT) and machine-learning approaches. We calculated the spectroscopic limited maximum efficiency (SLME) using the Tran–Blaha-modified Becke–Johnson potential for 5097 nonmetallic materials and identified 1997 candidates with an SLME higher than 10%, including 934 candidates with a suitable convex-hull stability and an effective carrier mass. Screening for two-dimensional-layered cases, we found 58 potential materials and performed G₀W₀ calculations on a subset to estimate the prediction uncertainty. As the above DFT methods are still computationally expensive, we developed a high accuracy machine-learning model to prescreen efficient materials and applied it to over a million materials. Our results provide a general framework and universal strategy for the design of high-efficiency solar cell materials. The data and tools are publicly distributed at: https://www.ctcms.nist.gov/~knc6/JVASP.html, https://www.ctcms.nist.gov/jarvisml/, https://jarvis.nist.gov/, and https://github.com/usnistgov/jarvis.

The integration of knowledge, experience, and intuition of researchers with artificial intelligence (AI) and robotics is the new direction to accelerate progress in materials research [1]. Strategies combining high-throughput synthesis with machine learning have already been producing new small organic molecules and bio-materials at ever faster rates. However, the application of the techniques to inorganic materials research is still in its infancy. Therefore, to drastically accelerate inorganic materials science research, establishing the AI and robotics inclusion into inorganic materials research is an urgent issue.

In this study, we demonstrate the autonomous synthesis of inorganic compounds using robotics and Bayesian optimization. This system fully automates sample transfer, thin film deposition, physical-property characterization, and growth condition optimization. The apparatus is equipped with a robot arm that can access each satellite chamber for growth and characterization. Based on the values obtained from the physical-property characterization, the Bayesian-optimization algorithm predicts the next growth condition (closed-loop). Here, we show the process of reducing the resistivity of Nb-doped TiO₂ transparent conducting thin films [2]. The autonomous synthesis can be applied to a wide variety of functional materials including magnetic, optical, electronic, and ionic properties.
References

[1]: R. F. Service, Science 366, 1295 (2019).
[2]: T. Hitosugi et al. Phys. Status Solidi A 207, 1529 (2010).

This presentation introduces Graphene Recipes for Synthesis of High Quality Materials (Gr-ResQ) (https://nanohub.org/tools/gresq). Gr-ResQ is designed to be a one stop solution for analysis of graphene synthesis by Chemical Vapor Deposition (CVD). Gr-ResQ consists of a comprehensive crowd-sourced database of synthesis recipes and results as well as a suite of software tools to analyze the database. It offers tools to analyze scanning electron microscope images and Raman spectra of graphene. It was developed as part of NSF's nanoMFG node within nanoHUB.

Advancing the manufacturing of graphene by CVD requires the ability to make CVD more cost-effective, including synthesizing low-defect density graphene directly on relevant substrates at reasonable costs. This is currently difficult mainly due to the multi-physics nature and the multi-time, and multi-length scales of the synthesis reaction. CVD processes can have over 200 variables, many of which are difficult to precisely control. This leads to poor repeatability and lack of a quantitative distribution of the parameter space. Characterization of synthesized graphene with electron microscopy and Raman spectroscopy gives useful information about the crystallinity, coverage, nucleation density and other such physical and chemical properties. The challenge of making interpretations and correlations from this data can be met by constructing a database that ties together the variables used in a process (the recipe) with the characterization of the synthesized graphene, which allows the application of data science and machine learning techniques to graphene and other 2D materials. One such use is in the estimation of graphene coverage from SEM images, which is generally a manual and time consuming task. We are trying to replace that by training a deep neural network (DNN) image segmentation algorithm with our manual coverage data to produce high quality, automated graphene coverage metrics.

In this presentation, we will discuss the database and the various tools, showing a number of examples that uniquely take advantage of the functionality of the database.

Advances in applying natural language processing (NLP) to material science text has greatly increased the size and acquisition speed of materials science synthesis data. These methods have been successfully applied to well-studied material systems with large amounts of data. However, we need ways to leverage literature data in materials domains without thousands of papers. Applying NLP pipelines to these types of materials science systems can be challenging due to the general schema and the noisiness of automatically extraction data. In this presentation, we demonstrate how to leverage domain knowledge to build upon existing data extraction techniques and improve extraction accuracy using examples in the zeolite, alternative cement, and solid-state electrolyte fields. We present data engineering techniques and discuss an optimal balance between automatic and manual data extraction.

We present an end-to-end system for computational autonomy in materials discovery (CAMD). This system is designed to sequentially augment a given dataset according to user-specified strategies for choosing new experiments and experimental procedures. The decision-making entities, called "agents," can make use of past knowledge, surrogate models, logic, thermodynamic or other physical constructs, heuristic rules, and different exploration-exploitation strategies. We show a series of examples for how discovery campaigns for finding materials satisfying a phase stability objective can be simulated to design new agents, and how those agents can be deployed in real discovery campaigns to control experiments run externally, such as the cloud-based density functional theory simulations in this work. Lastly, we present results from the cloud-based deployment of our tool that have resulted in over 2000 new inorganic crystalline materials in various crystal symmetries and compositions predicted to be within 200 meV of the phase diagram's convex hull.

Recently, machine learning (ML) is becoming an increasingly important tool for material discovery, thanks to its ability to quickly predict results of time and power-intensive calculations. Fundamentally, each ML model predicts some property and is composed of three essential elements: a database, a descriptor, and an ML algorithm. In this work, we have created and assessed a new tool for predicting formation enthalpies (ΔH_f). We take an open DFT database (OQMD), Voronoi tesselation-based descriptor developed by the Wolverton group^[1], and then design new neural networks that allow us to significantly improve the accuracy of predictions and reduce the computation time down to the order of 100 microseconds per prediction.

During the neural network design process, we created and tested around 50 networks belonging to 19 architectures. We used cross-validation on a randomly selected test set comprising 5% of the OQMD database (21,800 test entries) to compare with competing models. We found that our best network achieved a ΔH_f mean average error of 35 meV/atom, compared with ≈70 meV/atom for a model based on the Random Forest algorithm^[1]. In addition to simple cross-validation against a subset of the data for comparison with other models, the performance was also evaluated on more challenging systems. These included Special Quasirandom Structures (SQS) and 243 endmembers of a ternary σ phase (topologically close packed). Using such an approach during design, allowed us to improve accuracy for structures that are significantly different from those in the original dataset, and thus require extrapolation.

In this presentation, we show not only the new model but also new applications. Most notably, we take advantage of the improved prediction speed to screen whole databases for entries that do not fit the discovered patterns, and then automatically compare flagged entries to other sources. This, in turn, combined with more in-depth analysis, allows us to pinpoint entries that are likely to contain poor quality data, as well as those at risk of having a systematic error in the DFT calculation.


References:
[1] L. Ward et al., "Including crystal structure attributes in machine learning models of formation energies via Voronoi tessellations", Physical Review B, vol. 96, no. 2, 2017

Advances in electron imaging and spectroscopy instrumentation now allow for time-resolved data acquisition at rates over 1000 frames per second. However, this acquisition speed comes at the expense of the signal-to-noise ratio. At high frame rates, the low signal-to-noise (SNR) ratio can causes datasets to be too noisy for analysis with the naked eye without significant post-processing. Another challenge (often times exacerbated by poor SNR) is quantification of oxidation state. For example, in transition metal oxides, many methods exist for valence determination, e.g. L₃/L₂ ratio, chemical shift, or O K-edge pre-peak analysis. Unfortunately, these different methods can yield contradictory results, particularly in the presence of noisy data. Here, we seek to address both of these challenges through the application of machine learning techniques. First, we will discuss supervised learning of oxidation state in SrFeO_x thin films. With a convolutional neural network trained on >10,000 spectra, we obtain ~70% classification accuracy in differentiating SrFeO₃ (Fe⁴⁺) from SrFeO_2.5 (Fe³⁺) during an in situ experiment with very noisy data collected at 400 frames / second. Next, we will discuss ongoing efforts in unsupervised learning of EELS data. Unsupervised learning makes it possible to analyze different materials without needing a comprehensive database of labelled EELS data to compare with. Because unsupervised learning works by finding its own patterns or structures within the data, via techniques such as feature extraction or clustering, applying it to real-time in-situ data could help better understand phase morphology of the materials and could expose if unintentional or incomplete transformations of the sample occurred and in turn, the ability to more precisely and predictively tune complex materials. Unsupervised learning will be applied to various in situ EELS experiments.

Materials exhibiting higher mobilities than conventional organic semiconducting materials (e.g. fullerenes and heteroacenes [1]), are in high demand for applications such as printed electronics. In order to explore new molecules in the heteroacene family that might show improved mobility, a massive theoretical screen of hole-conducting properties of molecules were performed using a cloud computing environment. Over 7,000,000 structures of fused furans, thiophenes and selenophenes were generated, and 250,000 structures were randomly chosen for subsequent DFT (Density Functional Theory) calculations of hole reorganization energies (λ_h). Utilizing the cloud-computed DFT dataset of a quarter-million reorganization energies, the de novo design method proposed by Gomez-Bombarelli [2] was applied to find further chemical structures with minimal reorganization energy. This method converts molecular structures into continuous variables by applying the variational autoencoder/decoder technique, which enables optimization of chemical structures in a continuous numerical space. Results of the inverse design showed that the method has the ability to generate exotic chemical matter, such as structures with a seven- and eight-membered rings, with reasonably low calculated reorganization energies.
[1] K. Takimiya et al., Acc. Chem. Res. 47 (2014), 1493.
[2] R. Gomez-Bombarelli et al., ACS Cent. Sci. 4 (2018) 268.

Materials exhibiting higher mobilities than conventional organic semiconducting materials, (e.g. fullerenes and heteroacenes [1]), are in high demand for applications such as printed electronics. In order to explore new molecules in the heteroacene family that might show improved mobility, over 7,000,000 structures of fused furans, thiophenes and selenophenes were generated and 250,000 structures were randomly selected for subsequent DFT (Density Functional Theory) calculations of hole reorganization energies (λ_h). Utilizing the cloud-computed DFT dataset of a quarter-million reorganization energies, inverse materials design methods were applied to find further chemical structures with minimal reorganization energy. The de novo design methods applied were (1) a method based on reinforcement learning of SMILES strings (the REINVENT method [2]), and (2) a junction tree variational autoencoder that directly generates molecular graphs [3]. Both methods succeeded in obtaining chemical structures whose reorganization energy was lower than the lowest energy in the DFT training dataset, suggesting that an extrapolative materials design protocol can be developed by applying two methods. Finally, these results will be compared to the results obtained from the latent-space de novo molecular design method proposed by Gomez-Bombarelli et al. [4].
[1] K. Takimiya et al., Acc. Chem. Res. 47 (2014) 1493.
[2] M. Olivecrona et al., J. Cheminform 9 (2017) 48.
[3] W. Jin et al., 35th Intl. Conf. Machine Learning, ICML 2018, 5, 3632.
[4] R. Gomez-Bombarelli et al., ACS Cent. Sci. 4 (2018) 268.

Solar fuel generation from H₂O, CO₂ and sunlight comprises a promising renewable energy technology whose future proliferation relies on materials advancements, including identification of a photoanode for the oxygen evolution reaction. A water-oxidizing photoanode can be coupled to a (photo)cathode for scalable generation of H₂ and C-containing fuels. Metal oxides are the most extensively studied class of OER photoanode materials due to the availability of a multitude of metal oxide phases that may stably operate in the oxidative electrochemical environment while converting solar to photoelectrochemical (PEC) energy. Solar fuels photoanodes are quite rare due to the combined needs of solar absorption, charge carrier separation and transport, oxidative stability, and catalytic activity for the OER. The compendium of known metal oxide photoanodes that operate in the visible spectrum is being rapidly expanded via theory-guided high throughput experimentation, including our combination of combinatorial materials synthesis coupled with automated UV-vis spectroscopy and PEC experiments. Discovery of dozens of ternary metal oxide photoanodes will be discussed with focus on the lessons learned in terms of the role of theory and the importance of compositionally graded libraries, and a historical analysis of the discovery rate of metal oxide photoanodes reveals increasing chemical diversity, suggesting there may be many more materials to be discovered. The rapidly expanding portfolio of metal oxide photoanodes creates new avenues of inquiry to realize the requisite stability and efficiency for deployment of solar fuels technologies.

A model describing the origin of water-induced degradation of certain Metal-Organic Frameworks (MOFs) is derived, aiming to explain experimental observations of water stability and instability of prototypal MOFs. The proposed model suggests crucial rule of metal-coordinated water molecules, and solution pH, in triggering a cascade of reactions that in several cases results in pronounced irreversible degradation of several MOFs. It is suggested that such degradation pathways can be affected in neutral to basic solution by deprotonation of metal-coordinated water molecules, or bridging hydroxyl species in multinuclear clusters. Calculations of pKa values for the metal-coordinated water molecules, complemented by geometry relaxation calculations, and variable pH infrared spectroscopy techniques were all simultaneousley utilized to draw a comprehensive description for the early stages in MOFs degradation in aqeuous medium. Mostly dependent on the nature of metal nodes, we propose two distinct mechanisms for structural degradation of carboxylate-based MOFs. The MOFs investigated herein included MOF-5, HKUST-1, UiO-66, and MIL-101-Cr. It is suggested that for each type of MOF, a narrow pH stability range can be described, mostly determined by the pKa of metal-coordinated water molecules and the relative strength of metal-carboxylate interactions. Beyound this fundamental pH range, a relativeley wider pH window can exist, where certain MOFs can be kinetically inert. Finally, we present a highly transferrable and facile approach to induce surface passivation of MOFs crystals to effectively extend such aqeuus solution stability limits, as proven by the case of appreciably maintained microporosity of HKUST-1 in neutral, acidic and basic aqeous solution.

The semiconductor photocatalyst β-TaON used for water splitting is of interest due to its promising properties, such as stability, suitable band gap for visible light photocatalysis and carrier mobility. We have implemented a combinatorial, materials discovery approach using pulse-laser deposition (PLD) for thin film growth, X-ray diffraction (XRD) for phase determination and machine learning for data reduction. A lateral compositional gradient of TaO_xN_y was grown across the surface of a α-Al₂O₃ (012) 25-mm-wide wafer by PLD at 675°C in 10 mTorr of N₂. After annealing in N₂ at 1000°C, small-area XRD scattering patterns were collected for 71 positions across the lateral gradient. Unsupervised machine learning separated the XRD data into 4 clusters (phases); one of which turned out to be the desired β-TaON phase. Using high-resolution XRD we determined that the β-TaON region of the film to be single crystal epitaxy with out-of-plane β-TaON (100) // α-Al₂O₃ (012) and in-plane β-TaON (010) // α-Al₂O₃ (2-10). X-ray reflectivity (XRR) analysis of the β-TaON region of the film determined the film thickness to be 260 Å with an electron density matching that expected for β-TaON. Ta 4f_7/2 XPS measurements across the graded film showed Ta⁵⁺ valence state at the β-TaON region of the film. In comparison to the standard method of growing many uniform films at different compositions, this combinatorial method that produces a library of phases on a single wafer is much more efficient for materials discovery.

Based on their structure, non-crystalline phases can fail in a brittle or ductile fashion. However, the nature of the linkages between structure and propensity for ductility in disordered materials has remained elusive. Here, based on molecular dynamics simulations, we investigate the fracture of several disordered phases (metallic glass, glassy silica, colloidal gel, etc.) with varying degrees of disorder. We find that that, in general, structural disorder results in an increase in ductility. By applying the activation-relaxation technique (an accelerated sampling method to identify transition states), we show that the degree of plasticity is controlled by the topography of the energy landscape.

The dream of autonomous materials discovery systems will only be realized if the entire design/synthesis/characterization process is described in an unambiguous machine-readable format. Specifying the experiment process in a machine-readable way facilitates both remote operation, algorithmic control of experiments, and enhanced experimental replicability. A significant challenge is that materials experiments often require a combination of instructions for human operators and machines. Furthermore, the variety of materials characterization data and experimental performance metadata during a process often require significant extraction/transformation/loading before machine learning methods can be applied. Existing tools require significant development time that is incompatible with rapidly evolving scientific needs.

In this talk, I will describe ESCALATE (Experiment Specification, Capture and Laboratory Automation Technology) an adaptable open source package for experiment description and data collection. As a specific example, I will describe its application to Robotic-Accelerated Perovskite Investigation and Discovery (RAPID). The first generation of RAPID uses inverse temperature crystallization (ITC) to grow halide perovskite single crystals for x-ray structure determination and bulk characterization using commercial liquid handling robots. Experiment plans are contributed remotely, by either human scientists or algorithms trained on the reaction data. Incoming experimental data collected by ESCALATE is used to automatically train machine learning models, evaluate model performance and feature influence on model predictions to gain human-interpretable scientific insight into the formation process, and quantify experimental reproducibility using this complete dataset of all experimental successes and failures. A live web dashboard communicates these insights to the scientist and management in visual form; furthermore experimental data can be exported in the form of interactive visualizations for use in publications. Collected characterizations includes qualitative measurements of crystal quality (including photographic images of the samples), powder X-Ray diffraction, and spectroscopic measurements along with background laboratory conditions. The second generation of RAPID has been extended to include antisolvent precipitation synthesis of perovskite single crystals and perovskite thin film synthesis and characterization including remote monitoring of long-term environmental stability. We have used the ESCALATE/RAPID platform to perform experimental tests benchmarking a variety of data-driven approaches for discovering new compounds against human-expert and random baselines. The comprehensive data collected from these experiments has also allowed us to ascertain the relative importance of descriptive physical models for reaction composition relative to raw experimental descriptions when constructing machine learning models for reaction outcome prediction.

Machine learning has become an integral part of many aspects of materials research. As a branch of machine learning, active learning has attracted much attention which can effectively help navigate experimental sequences in materials research. We are actively incorporating active learning in screening of combinatorial libraries of functiona materials. For some physical properies, each characterization/measurement requires time/resources long/large enough that "high"-throughput measurement is not possible. Examples include detection of martensitic transformation and superconducting transitions in thin film libraries. By incorporating active learning into the protocol of comabinatoal characterization, we can streamline the measurement and the analysis process substantially. We will present examples of mapping of phase change memory materials as well as others functional materials. This work is performed in colaboration with A. Gilad Kusne and V. Stanev. This work is funded by ONR and AFOSR.

Herein, we report the successful discovery of a new hierarchical structure of metal-organic nanocapsules (MONCs) by integrating chemical intuition and machine learning algorithms. By training datasets from a set of both succeeded and failed experiments, we studied the crystallization propensity of metal-organic nanocapsules (MONCs). Among four machine learning models, XGBoost model affords the highest prediction accuracy of 91%. The derived chemical feature scores and chemical hypothesis from the XGBoost model assist to identify proper synthesis parameters showing superior performance to a well-trained chemist. This paper will shed light on the discovery of new crystalline inorganic-organic hybrid materials guided by machine learning algorithms.

Van der Waals heterostructures are comprised of stacked atomically thin two-dimensional crystals and serve as novel materials providing unprecedented properties. However, the random natures in positions and shapes of exfoliated two-dimensional crystals have required the repetitive manual tasks of optical microscopy-based searching and mechanical transferring, thereby severely limiting the complexity of heterostructures.

To solve the problem, we develop a robotic system, powered by artificial intelligence, that automatically searches exfoliated 2D crystals and assembles them into vdW superlattices inside glovebox [1-3]. The system can automatically scan the surface of silicon substrates, analyze optical microscope images, and detect a variety of atomically thin 2D materials without human intervention. The stacking order, positions, and crystallographic orientations of the 2D crystals are designed using the customized CAD software. The robotic assembly system of two-dimensional crystals enabled stacking four cycles of the designated two-dimensional crystals per hour with few minutes of human intervention for each stack cycle.

The system enabled fabrication of the vdW superlattice structures consisting of 29 alternating layers of the graphite and the hexagonal boron nitride (hBN) flakes. Encapsulated graphene devices exhibited high charge carrier mobilities (>1,000,000 cm²/Vs), demonstrating the applicability of the system for prototyping a variety of high-quality vdW superlattices.
The wider material design freedom enabled by our system offers unprecedented opportunities for exploring the potential of vdW heterostructures. By this development, we can reduce the human intervention involved in the vdW heterostructure fabrication by orders. We believe that this work free up researchers from repetitive tasks and letting them focus on more intellectually creative tasks.

The hardware/software designs are available at https://github.com/tdmms/

[1] S. Masubuchi et al., Nature Communications 9, 1413 (2018).
[2] S. Masubuchi et al., npj 2D Materials and Applications 3, 4 (2019).
[3] S. Masubuchi et al., arXiv:1910.12750 (2019).

In my annnual poster, I'll review the MIT PVLab's recent work combining experiment and theory in a closed-loop, machine-learning (ML) driven framework. Closed-loop experimentation driven by ML enables rapid and efficient convergence on optima, within vast search spaces. A review of scientific insights, as well as our latest machine-learning algorithms, will be presented.

Meeting current and future demands for advanced structural materials designed to operate in extreme environments requires rigorous mechanical performance evaluation at such conditions. When considering alloy development for such applications, such as nuclear materials, multiple compositions and processing routes create many permutations of material to be evaluated at different temperatures and radiation damage levels. Where traditional bulk scale mechanical testing is relatively time consuming, high speed nanoindentation mapping techniques can provide large (and highly localized) Hardness and Modulus datasets. This allows for analysis of heterogeneous distributions of properties in multi-phase alloys, which can then be correlated to the microstructure through use of supplementary techniques, such as EBSD, which has similar resolution. However, this requires accompanying analysis methods to also be high throughput, for which machine learning based techniques are of interest. Some datasets, as large as one million indentations in a given map, are beyond the limit of reasonable human analysis.
In this study, a high vacuum, high temperature nanoindenter was utilized to produce temperature controlled mapping of three advanced types reactor materials up to 500°C while protecting the sample surface from oxidation. The materials studied were Fe-based HEAs, with varied Al content to give 3 distinct microstructures. Structure-property correlation was quantitatively explored through clustering to partition the hardness and modulus data to individual microstructural elements and evaluate their thermal stability, with irradiation studies to follow. While qualitatively these methods performed comparably to by hand grouping, there is a need to quantitatively understand their systematic uncertainties without making assumptions regarding the underlying probability distribution. This is further explored for this data set by using a parametric bootstrapping method to resample the probability distribution function from the individual data points. Three different machine learning methods, Kernel Density Estimation, Gaussian Mixture and Multi-Normal, are used to generate a probability distribution function. This is then resampled to produce a simulated dataset 1000 times the original dataset in size. This allows K-means, Spectral and Gaussian Mixture clustering techniques to be applied and compared, but in principal, this can be applied to evaluate other properties as well.

Closed-loop, sequential learning is a key paradigm in autonomous materials development. Within this framework, aspects of the materials system under study are modeled, and such models are used to decide subsequent experiments to be run, results of which are fed-back to update models. In the past, off-the-shelf solutions and algorithms have been employed to optimize material properties. In this talk, we describe how autonomous materials platforms can be used for problems other than materials property optimization. We will specifically highlight work on autonomous phase mapping and real-time control, and how a common framework can be used to capture problem-specific structure and objectives. Relevant to this is the development of tools to perform autonomous materials characterization. We will describe how rich characterization data such as microscopy images or three-dimensional reconstructions from atom probe tomography can be analyzed without human intervention to encode experimental results for subsequent use within the closed loop. Together, the work described demonstrates the broad applicability of autonomous platforms throughout materials science research.

In this talk I will look at how we can open up the percieved 'black box' of machine learning when designing and characterising materials. I will show how being able to look at and understand the inner workings of deep neural networks increases the value of these approaches for problems in materials science. Understaning the workings is important for (i) debugging and building better models, (ii) building confidence in models that perform well, (iii) learning from models that exceed human performance in certain problem areas. I will take examples of our research where we use graph neural networks to infer and predict dielectric properties of solids and where we use deep convolutional neural netowrks to interpret data from inelastic neutron scattering. In the former case a mixture of deep and shallow learning allows for extraction of principles as well as construction of robust accurate models. In the latter case understanding the workings of the neral network allows us to develop a method for guiding experiments to sample the most pertinant areas of signal space.

The next generation of high performance optical materials require precise engineering to transcend the inherent limitations of bulk materials and achieve desired optical performance. These targeted properties originate from the interactions between inherent optical properties of the constituent materials and meticulously engineered sub-wavelength inhomogeneities. However, the design of such metamaterials poses a significant challenge to researchers due to the near-infinite design space to consider, even for relatively simple metamaterials such as the two-dimensional all-dielectric metasurface. This metasurface has been the subject of intense research interest because of its unprecedented reflective performance resulting from low-loss dielectric materials and shape-dependent optical resonances. Here we employ artificial neural networks (ANNs) to invert the design process, conceiving metasurface designs with targeted reflection and transmission spectra. These ANNs are capable of coordinating the complex interactions of the eigenmodes from which these desired properties arise, delivering device designs demonstrating coefficients of reflection greater than 99% over a range of wavelengths more than 450nm.

DNA-templated silver clusters (Ag_N-DNA) are known to exhibit a wide range of fluorescence colors for different choices of the DNA template sequence [1]. While these clusters are promising biosensors and biomarkers [2], rational design of Ag_N-DNA for specific applications is challenged by the huge space of possible DNA template sequences. To understand how DNA sequence selects for silver cluster size and optical properties, we are employing high-throughput experiments and tools from machine learning and data mining. We previously showed that machine learning methods trained on experimental data can be used to design new DNA templates that select for Ag_N-DNA color, for the specific and limited case of 10-base DNA oligomers [3]. An important open question is whether such a design process developed for a specific biopolymer template length is applicable at other lengths, with different numbers and diverse configurations of cluster nucleation sites. More broadly within materials informatics, it is important to determine when models learned on limited training data sets are relevant for prediction within expanded input parameter spaces.

Here, we present a newly developed flexible design approach that utilizes color-correlated DNA base motifs learned from data on > 2,000 ten-base DNA oligomers [4]. We test this motif-based design for templates ranging from 8 bases to 16 bases long, for which the sizes of the sequence spaces differ by nearly 5 orders of magnitude. The experimental data show that designed strands of all lengths are selective for Ag_N-DNA color in the target wavelength band of 600-660 nm, strongly suggesting that color-selective motifs learned for one template length generalize to other lengths. Thus, a motif-based design approach may be broadly suitable for future Ag_N-DNA applications.

[1] E. Gwinn, D. Schultz, S. M. Copp, S. Swasey. Nanotechnology, 5, 180-207 (2015).
[2] Y. Chen, M. L. Phipps, J. H. Werner, S. Chakraborty, J. S. Martinez. Acc. Chem. Res. 51, 756-2763 (2018).
[3] S. M. Copp, A. Gorovits, S. M. Swasey, S. Godibandi, P. Bogdanov, E. G. Gwinn. ACS Nano, 12, 8240-8247 (2018).
[4] S. M. Copp, A. Gorovits, S. M. Swasey, P. Bogdanov, E. G. Gwinn. In revision.

Large-scale data-mining workflows are increasingly able to predict successfully new materials that possess a targeted functionality. The success of such materials discovery approaches is nonetheless contingent upon having the right data source to mine, adequate supercomputing facilities and workflows to enable this mining, and algorithms that suitably encode structure-function relationships as data-mining workflows which progressively short list data toward the prediction of a lead material for experimental validation.

This talk describes how to meet these data science requirements by showing how to auto-generate large material databases of experimental information from documents, using natural language processing and machine learning, via our ChemDataExtractor tool [1-3]. The resulting large databases of chemical structures and their cognate properties are then mined for materials discovery using custom-built algorithms that are encoded forms of structure-function relationships. These molecular design rules progressively filter the parent set of chemicals until a lead candidate appears, which is experimentally validated. Case studies are demonstrated to illustrate the power of data-driven materials discovery for several functional applications.
References
[1] www.chemdataextractor.org
[2] Swain & Cole, J. Chem. Inf. Model. 56 (2016) 1894–1904
[3] Court & Cole, Scientific Data 5 (2018) 180111.

The development of a hydrogen economy has potentially immense benefits for addressing the crisis of global heating. However, storage remains a problem and decades of research have left us empty-handed: no known materials can satisfy all of the DOE targets for vehicular storage. Metal hydrides are among the most promising solid-state storage materials, but their complex chemistry is a major impediment to discovery of game-changing materials. Specifically, their thermodynamics are often ill-matched to hydrogen release at temperatures compatible with fuel cells. This limitation requires an equilibrium H₂ pressure of at least 5 bar at the maximum delivery temperature of 80 °C. Empirical discovery approaches using, for example, high-throughput synthesis, are largely ineffective because measurement of the equilibrium H₂ pressure can take days or even weeks. Machine learning (ML) methods have already shown promise for accelerating the discovery of new hydrogen sorbents and identifying key features that influence their properties. In recent application to metal hydrides, however, the use of “black-box” algorithms prohibited development of crucial vital scientific insight needed to go beyond existing compositions. We employed a gradient-boosting tree regressor, an “explainable” ML that can reveal the feature dependence of material properties. In particular, we find that the equilibrium H₂ pressure of intermetallic metal hydrides strongly depends on a volume-based feature that requires only the elemental composition to be computed. Remarkably, this simple structure-property relationship is valid across a wide range of compositions, metal substitutions, and structural classes found in intermetallic metal hydrides. This result provides a rational basis for targeting novel hydrogen storage materials with thermodynamic properties appropriate for specific applications. As an illustration, we identified a known intermetallic that forms a hydride with very low stability. Although not yet validated by experiment, this result was confirmed by density functional theory modeling.

Nanoporous materials (e.g., zeolite, activated carbon, metal-organic framework, polymeric membranes, etc.) have various technological applications, including gas separation, gas storage, catalytic transformations, etc. The functionalities of nanoporous materials strongly depend on their pore size and shape distribution—which present virtually limitless degrees of freedom. Here, based on high-throughput simulations and a convolutional neural network (CNN) predictor, we present a model allowing us to predict the water sorption isotherm of hydrophilic and hydrophobic nanoporous materials. The training of a CNN generator then enables the inverse design of optimal porous microstructures featuring tailored/unusual sorption isotherms.

Recent successes demonstrated by machine learning (ML) approaches have created great enthusiasm for their potential translational contributions to other fields. Among these are efforts to demonstrate the feasibility of using machine-based approaches to study and speculate on new energetic materials. This talk will discuss the performance of ML models developed to predict the detonation properties of energetic materials. A comprehensive comparison of approaches reveals that models are highly sensitive to feature selection as well as algorithm differences. But using a relatively limited volume of data and careful problem definitions, surprisingly strong cross validation scores suggest that there is promise to both improve and ultimately use ML methods widely in this field. The talk will conclude with an overview and some perspectives of other recent efforts including generative modeling and natural language processing techniques at their intersections with energetics research.

Inorganic scintillator-based detector materials find a wide variety of applications, ranging from medical imaging to radiation detection for global security and high energy physics experiments. These materials essentially convert a fraction of the total energy deposited by incident gamma rays or X-rays into visible or near-visible range of the spectrum. In many applications a scintillator with high light output, fast response time, and emission at suitable wavelengths, is generally required, among many other application-specific desired characteristics. However, owing to several conflicting relationships between the desired properties (e.g., brightness versus speed), no single scintillator is ideal for all uses; there is a need to design custom scintillators optimized for each application. Traditionally, the discovery and design of new detector materials have relied on a laborious, time-intensive, trial-and-error approach; yielding little physical insight and leaving a vast space of potentially revolutionary materials unexplored. To accelerate the discovery of optimal scintillator materials with targeted properties and performance, efforts are ongoing to develop a closed-loop machine learning driven adaptive design framework based on data from literature, in-house experiments and quantum mechanical calculations. This talk will present an overview of this framework, focusing on the screening of complex wide band gaps chemistries with an aim of identifying promising materials that are amenable to band-gap/band-edge engineering to yield custom scintillation properties. The developed framework is general and is expected to prove useful for applications beyond scintillator discovery such as photovoltaics and catalysis.

In recent years the use of various machine-learning (ML) methods to discover new materials with desired properties have emerged as a powerful tool. In this research, we have designed a stacked ensemble model (SEM) by combining several supervised ML models to predict the adsorption of hydrogen in metal-organic frameworks (MOFs). The SEM model was trained by using the hydrogen adsorption data generated by screening of 1000s of MOFs by performing grand canonical monte carlo (GCMC) simulations at 1 atm and 77 K. This SEM was integrated with optimization algorithms and in-house MOF structure generation code to design MOFs with high adsorption of hydrogen. Overall, this new approach of integrating SEM with optimization algorithms accelerated the new MOF discovery process, resulting in a MOF with enhanced hydrogen storage capacity. This new ML framework is transferable and can be used to discover other porous (e.g. zeolites), soft (e.g. polymers), and hard (e.g. alloys) materials for desired applications.

Computing the ground state energy of quantum many particle systems is of fundamental importance in a variety of fields, including chemistry, physics and materials science, amongst others. However, the complexity of such quantum mechanical computations grows rapidly with the number of particles. Machine learning algorithms do not simulate the physical system, but instead estimate solutions by interpolating values provided by a training set of known examples. However, precise interpolations may require a number of examples that is exponential in the system dimension, and are thus intractable. Tractable algorithms compute interpolations in low dimensional approximation spaces, which leverage the underlying physical properties of the system.

In this talk I will give an overview of machine learning algorithms for computing the ground state energy of quantum many particle systems, describing the core principles of such algorithms and illustrating the common themes that emerge. I will then present in more detail our approach to the problem, which is based on a type of multiscale, multilayer convolutional neural network, called a wavelet scattering transform. Through a cascade of multiscale wavelet transforms and nonlinearities, the scattering transform encodes the appropriate invariants and regularity properties of the physical system. Wavelet scattering regressions, computed over databases of organic molecules and amorphous materials, achieve errors on the order of quantum mechanical simulations, but at a fraction of the computational cost.

This multi-disciplinary research proposes a new framework that can infuse scientists' experience and knowledge (named as scientists' eyes) into a constantly evolving, glass-box physics rule learner that can essentially extract hidden physics rules behind complex nano-scale phenomena. Recently, physics principles are gradually integrated into machine learning (ML) methods, but it is at its infancy to discover hidden physics rules. Such pursuits for hidden rules face formidable challenges when the physical experiments involve multifaceted measurements across three-dimensional (3D) specimens or objects. These remaining challenges render even advanced ML methods such as deep learning secondary, not as a primary explorer, in the broad science and engineering domains. Therefore, there exists a deep chasm still remains between materials science and ML community.
To overcome this challenge and unravel hidden physics rule behind nano-scale complex phenomena, we reinterpreted the deep learning's successful notion of multi-layered convolutions and gave rise to a new notion of "convolved information index." The convolved information index can facilely handle physical measurements over 3D nano-scale specimens and measurements without losing experienced perspectives of scientists in the materials science. Also, the framework proposes "link function" as a generator of mathematical expressions about the identified nano-scale physics rules, thereby pursuing "glass-box" (as opposed to the black-box) learning and prediction. Consistent evolution capability is realized by integrating Bayesian update and evolutionary algorithms.
The framework is applied to nano-scale contact electrification (CE) phenomena, which are difficult to elucidate due to length scales and complex physical interactions [1,2]. Three experimental data sets are used. Tribocharged PDMS nanocup array is fabricated by the sequence of (1) prepare a PET mold with a 750 nm-pitch triangular array of nanocones, (2) pour the liquid phase PDMS, Sylgard 184, Dow Corning, mixed with the curing agent, (3) solidify the specimen, and (4) peel off the PDMS from the PET mold (for more details see [1,2]). Atomic force microscopy (AFM) in the tapping mode is used to measure the surface topography and potential. The charge values were measured by Kelvin probe force microscopy over patterned nano-cups represented by 3D point cloud. The experimental data consist of a variety of conditions regarding geometries, demolding directions, and nano-scale patterns (3000 nm by 3000 nm square area with an interval of 5.86 nm; 4000 nm by 4000 nm, with 7.8125 nm). The nano-cup's average heights are 154 nm, 93.5 nm, and 117 nm. The proposed glass-box learning begins from horizontal-demolding test data, and then the best-so-far identified rules are inherited to the next learning with new specimens in a chained way.
The feasibility tests show promising performances of the proposed framework in unraveling transparent, reasonable “expressions” of hidden physics rule along with reasonable reproduction of complex patterned distributions of electric potentials over the 3D point cloud. The proposed framework is believed to establish a promising partnership of scientists’ eyes and the advanced ML methods, thereby enabling data and ML-driven discovery of groundbreaking physics laws in the nano-world and beyond. This presentation will cover how this new framework can facilitate materials scientists’ discovery with actual nano applications.
[1] Qiang Li, In Ho Cho, Rana Biswas, Jaeyoun Kim, Nano Letters 19, 850-856 (2019).
[2] Qiang Li, Akshit Peer, I. Cho, Rana Biswas, Jaeyoun Kim, Nature Communications 9, 974 (2018).

In this talk we will review recent studies of applying atomistic-level modeling to describe fracture processes at different scales, from molecular to macroscale levels, and to use artificial intelligence and machine learning to design fracture resistant materials. Molecular and related multiscale modeling methods can be an effective method to better understand elementary fracture processes, ranging from brittle to ductile fracture and other materials failure phenomena such as creep, and provides the basis for expansive training sets. To illustrate this framework, we report a design approach for optimizing the toughness of nanocomposite materials using artificial intelligence, implemented in a novel “AutoComp Designer” algorithm. The algorithm consists of a machine learning predictor combined with an AI improved genetic algorithm, which is capable of discovering de novo materials designs in a vast space of possible solutions. Facilitated by a deep convolutional neural network that is trained with a dataset of hundreds of thousands of combinations of soft and brittle materials originating from multiscale models. Through the algorithm, we extend the capability of physical simulations beyond property predictions to optimize the fracture toughness by altering the material distribution. The solutions are generated by our AI model at a dramatically lower computational cost compared to brute-force searching methods. We further investigate the physical mechanism for improving material performance behind the AI approach, and demonstrate the ability of AI to search for optimal designs with very limited sampling. Brute-force molecular dynamics simulations of the nanocomposite designs confirm that our AI design improves the performance by effectively decreasing the stress concentration at crack tip. This AI approach can be easily applied to other nanocomposites, biomaterials and other material classes, and provides a transferrable and reliable rapid design approach expanding current capabilities.

In recent years there has been great progress in the use of machine learning algorithms to develop interatomic potential models. Machine-learned potential models are typically orders of magnitude faster than density functional theory but also orders of magnitude slower than physics-derived models such as the embedded atom method. We demonstrate that machine learning, in the form of genetic programming, can be used to develop accurate and transferable many-body potential models that are as fast as the embedded atom method, making them suitable to model materials on extreme time and length scales. The key to our approach is to explore a hypothesis space of models based on fundamental physical principles and to select models from this hypothesis space based on their accuracy, speed, and simplicity. We demonstrate our approach by developing fast and accurate interatomic potential models for copper that generalize well to properties they were not trained on, and we discuss additional applications. Our approach requires relatively small sets of training data, making it possible to generate training data using highly accurate methods at a reasonable computational cost.

Machine learning (ML) has emerged as a powerful tool in the study and design of materials. In this talk, I will discuss the different ways in which materials can be described and their relative merits and limitations. In particular, I will highlight the importance of choosing representations that not only satisfy known physical invariances (e.g., rotation, translation, etc.), but also incorporate known physics and chemistry in a hierarchy of knowledge. I will provide an overview of leading ML models and applications, and discuss how the problems of data scarcity and bias in materials science can be addressed using multi-fidelity, transfer learning and other approaches.

Electronic levels introduced by impurities and defects in the middle of the band gap are critically important in semiconductors for optoelectronic, photovoltaic (PV) and quantum sensing applications. While “deep” defect levels can prove catastrophic for PV performance by causing non-radiative carrier recombination, impurity levels in the band gap could also be entangled for quantum sensing or lead to increased absorption of sub-gap photons which can enhance efficiencies. Predicting formation energies and charge transition levels for defects in semiconductors is thus paramount; density functional theory (DFT) has been widely applied for such studies to overcome experimental bottlenecks and confirm measured defect levels. However, the requirement of large supercells, advanced levels of theory, and inclusion of charge corrections make DFT computation of defect properties very expensive, and trends and knowledge from previous calculations are not exploited in subsequent ones.

In this work, we combine high-throughput DFT and machine learning (ML) to develop general predictive models for the formation enthalpy and charge transition levels of impurities in two broad semiconductor classes: (a) ABX₃ halide perovskites, and (b) group IV, III-V and II-VI semiconductors. DFT data is generated for impurity atoms selected from across the periodic table and simulated in various possible cation, anion or interstitial sites. Any “semiconductor + impurity” combination is converted into a unique feature vector based on the defect site coordination environment, the tabulated elemental properties (eg. ionization energy, electronegativity, etc.) of the impurity atom, and electronic and energetic properties computed using low cost unit cell defect calculations. State-of-the-art neural networks (NN), random forest regression (RFR) and kernel ridge regression (KRR) are applied to train models for (a) impurity formation enthalpy, and (b) impurity charge transition levels, based on the input feature vector. Model performance is evaluated for different ML techniques, ML parameters such as number of hidden layers and network weights in NN, subsets of features, and size and nature of training dataset; the best predictive models thus obtained are deployed for comprehensive prediction and design purposes. It is seen that models trained on defects and impurities in pure cation/anion composition semiconductors are applicable to mixed or alloyed compositions as well: for instance, by training on data from pure canonical AB semiconductors like CdTe, ZnO, GaAs and SiC, the ML models can make accurate predictions for impurities in alloyed compositions of the same compounds, such as CdTe_1-xSe_x, Cd_1-xZn_xO, Al_1-xGa_xAs_1-ySb_y, etc. This versatility of the machine-learned models provides an avenue to access the optoelectronic impact of any atomic impurity in any possible pure or mixed composition semiconductor belonging to the same chemical space, which is useful for compositions that are expensive to model. We use the predictive models to efficiently screen for dominating impurities, that is impurities that shift the equilibrium Fermi level in the semiconductor as determined by dominant native defects, in hundreds of possible compositions in the halide perovskite and groups IV, III-V and II-VI semiconductor spaces. ML-based screening of impurities was tested against DFT screening for CdTe and seen to have an accuracy of 95%. The quick and accurate estimation of interesting dopants/impurities in semiconductors is made possible by this materials design framework powered by quantum mechanical computations and machine learning, which can be easily expanded to other semiconductor classes and improved by the infusion of fresh data, new descriptors and more advanced ML techniques.

References
1. A. Mannodi-Kanakkithodi et al., Chemistry of Materials 31 (10), 3599–3612 (2019).
2. A. Mannodi-Kanakkithodi, et al., under review, preprint: https://arxiv.org/abs/1906.02244.

In material science, we are frequently interested in understanding the properties and design implication of material at interfaces. These interfaces can be manipulated to improve the desired characteristics of the bulk material. In this study, we are interested in understanding and optimize the impact of interfacial atomic defects on the thermal transport across a Cu/Si junction. To that end, we developed a reinforcement learning based framework to optimize over a potentially large parameter search. Using this technique allows us to accumulate knowledge of the system of a given type of atoms and store this information into a neural network. In this study, we present our results on optimizing the thermal transport by varying the fraction and length of the interfacial atomic defects using molecular dynamics (MD) simulations with normal mode analysis (NMA) to investigate thermal transport.

Deforming a material to a large extent without inelastic relaxation can result in unprecedented properties. However, the optimal deformation state is buried within the vast continua of choices available in the strain space. Motivated by this challenge, we hereby advance a powerful strategy to circumvent conventional trial-and-error methods and adopt a unique type of machine learning framework for rationally designing the most energy-efficient pathway to achieve a desirable material property such as the electronic bandgap. This method invokes recent advances in the field of convolutional neural networks and active learning by utilizing a limited amount of ab initio data for the training of a surrogate model, predicting electronic bandgap within an accuracy of sub-10 meV. Our model is capable of discovering the indirect-to-direct bandgap transition and semiconductor-to-metal transition in silicon by scanning the entire strain space [1]. It is also able to identify the most energy-efficient strain pathways that would transform diamond from an ultrawide-bandgap material to a narrow-bandgap semiconductor or even semimetal [2]. A broad framework is presented to tailor any target figure of merit by recourse to deep elastic strain engineering and machine learning for a variety of applications in microelectronics, optoelectronics, photonics, and energy technologies.

[1] Z. Shi, E. Tsymbalov, M. Dao, S. Suresh, A. Shapeev, J. Li. PNAS, (2019) 116, 4117-4122
[2] E. Tsymbalov, Z. Shi, M. Dao, S. Suresh, A. Shapeev, J. Li (2020) in submission

Autonomous Research Systems like ARES™ are disrupting the research process by using AI and Machine Learning to drive closed-loop iterative research. ARES™ is our autonomous research robot capable of designing, executing and evaluating its own experiments at a rate of up to 100 iterations per day. Previously ARES taught itself to grow carbon nanotubes at controlled rates (NPJ Comp Mat 2016). Here we discuss recent research campaigns on maximizing carbon nanotube growth rates using a Bayesian optimization planner. We also use HOLMES and knowledge gradient descent to introduce advanced decision policies with local parametric models to control nanotube diameter. Implications for nanotube materials development will be discussed. Finally, we have developed a new research robot for additive manufacturing, AM ARES™, which is at the early stages of teaching itself to print structures with unknown inks. We plan to make the AM ARES™ Robot available online so that a broad community of researchers can test concepts and approaches for AI/ML and experimental design as applied to 3D printing, thus building to the larger goal of enhancing citizen science.

Due to recent improvements in image resolution and acquisition speed, materials microscopy is experiencing an explosion of published imaging data. Intensive human efforts, however, are required to digest such published images. In this talk, we will discuss the MaterialEyes project, in which deep learning and natural language processing tools are used to extract microscopy images from scientific literature. The process is accomplished via a suite of scientifically-informed software tools in the EXSCLAIM python package (automatic EXtraction, Separation, and Caption-based natural Language Annotation of Images from scientific figures), which is ultimately focused on the acceleration of inductive materials insights by collecting, structuring, and interpreting an aggregate of materials microscopy imaging data. In this talk, we will discuss how deep learning and NLP is used in EXSCLAIM to: (1) extract figures and caption data from open source scientific documents, (2) automate the construction of self-annotated materials imaging datasets, and then (3) leverage literature-based insights to assist in image classification and segmentation tasks. In addition, we will discuss a new pipeline for crowdsourced image labeling that was used for training some of the deep learning models.

Characterization of nanomaterial morphologies with advanced microscopy and/or spectroscopy tools plays an indispensable role in nanoscience and nanotechnology research, as rich information about the chemical compositions, crystallography, other physical and chemical properties, as well as the growth mechanism can be extracted from morphology analysis. However, the interpretation of imaging data heavily relies on the “intuition” of experienced researchers. As a result, many of the deep graphical features are often unused because of difficulties in processing the data and finding the correlations. Such difficulties can be well addressed by deep learning. In this work, we use the optical characterization of two-dimensional (2D) materials as a case study.
Currently, the most widely used approach to obtain high-quality 2D crystals in laboratories is mechanical exfoliation, followed by 2D crystal “hunting” under an optical microscope. This task is time-consuming and difficult especially for inexperienced researchers. Very recently, there has been growing interest in automating such a process. However, existing optical contrast based identification methods could not satisfy the needs for a general-use, fully automatic optical identification and/or characterization system, as they are often very specific to types of 2D crystals, conditions and configurations of the microscopes being used, image qualities, etc. In addition, optical contrast based methods would fail for harder problems in which the classes to be differentiated are not separable in the color space, such as identifying the materials or even predicting their physical properties from unlabeled optical images.
In reality, optical microscopic (OM) images contain rich, often unused information other than optical contrast. These deep graphical features can be extracted through deep learning, especially semantic segmentation methods based on convolutional neural networks (CNNs). In this work, an encoder-decoder semantic segmentation network called “SegNet” is configured for pixel-wise identification of OM images of 2D materials, in real time, and regardless of variations in optical setups. A fully automated system utilizing this algorithm can be used to free up tremendous amount of time for researchers. Additionally, we find that the algorithm finds correlations between the OM images and physical properties of the 2D materials and can thereby be used to anticipate the properties of new, as-yet uncharacterized 2D crystals. Finally, a transfer learning technique is applied to adapt the pretrained network to more optical characterization applications such as identifying layer numbers of chemically synthesized graphene domains. The proposed methodology can potentially be extended for identification and understanding other morphological or spectroscopic data of diverse nanomaterials.
Acknowledgement: This work is supported partially by U.S. ARO through the Institute for Soldier Nanotechnologies, under cooperative agreement number W911NF-18-2-0048, AFOSR FATE MURI, grant no. FA9550-15-1-0514, and the STC Center for Integrated Quantum Materials, NSF Grant No. DMR-599 1231319.

We introduce the first systematic database of scanning tunneling microscope (STM) images obtained using density functional theory (DFT) for two-dimensional (2D) materials. Specifically, the database is constructed using the Tersoff-Hamann method. It currently contains data for 681 exfoliable 2D from the JARVIS-DFT database. These computational STM images can directly aid the experimental analysis of experimental STM results as they provide ideal reference images. Examples of the five possible Bravais lattice types for 2D materials are discussed and the images for all the materials are made available on the website. We validated the computational STM technique by comparing our results to a few experimental images. As the analysis of multiple STM images during an experiment could be tedious, we used our DFT-STM image database to train a high-accuracy convolution neural network (CNN) model to accelerate the characterization process.

A major obstacle in applying machine learning to scientific and engineering tasks is the lack of engineering rules for the design of machine learning systems that allow to make these system reproducible and the limits of the systems known. Based on joint research with physicists and material engineers, this talk presents engineering principles and measurement processes[1] that aim at structured engineering of machine learning systems.
While the focus of this talk is on neural networks, the principles and presented measurements are universal and applicable to a variety of machine learning systems.

[1] G. Friedland: "Reproducibility and Experimental Design for Machine Learning on Audio and Multimedia Data", Proceedings of the 2019 ACM International Conference on Multimedia, October 2019. https://dl.acm.org/citation.cfm?id=3350545

Predicting the magnetostriction of a large-scale (mm-mm) polycrystalline Tb_xDy_1-xFe₂ (Terfenol-D) alloy by direct numerical simulations is mathematically tedious and computationally expensive, because it requires solving a set of partial differential equations which describe the mutual coupling between local magnetization dynamics and polycrystal elasticity, and because it requires a large number of simulation cells (tens of millions and more). For accelerating the prediction of magnetostriction, we apply a graph-based deep learning method, namely, the Graph Neural Network (GNN), to discover the correlation between a microstructure dataset of 350 different polygrain structures and a property dataset describing the magnetic-field-induced deformation of each polygrain structure (viz., magnetostriction). In the GNN based approach, each grain is represented as a node, in which all important physical features (e.g., Euler angle, magnetization direction) inside a grain are stored. As a result, the polygrain structure is represented as a network of nodes, in which the identity information of each grain(node)’s neighboring grains(nodes) is also preserved. This in turn allows us to consider the grain-grain interaction which critically influences the magnetostriction of the entire alloy and interpret the contribution of each node(grain) to the magnetostriction. A >90% prediction accuracy has so far been achieved with ample space for improvement. Our results demonstrate the exciting potential of applying GNN to accelerate the discovery of microstructure-property linkage in polycrystalline materials.

Materials are ubiquitous in society and lie at the heart of many technologies including energy, communication, buildings, transportation, the environment and medicine. The constant need to develop new materials with properties that are matched to applications is a major driver of materials research. Understanding how atoms come together, self-assemble, dissociate and undergo phase transformations under different stimuli is a critical question in materials science. Transmission electron microscopy can now routinely achieve a spatial resolution of better than 1 Å allowing the projected spatial distribution of atoms to be observed. In situ methodologies allow these signals to be mapped both spatially and temporally in the presence of different stimuli. Rapid advances in detector and spectrometer technology now allow local imaging, diffraction and spectroscopic signals to be acquired at unprecedented rates, generating terabytes of data in relatively short times.

The availability of large spatially and temporarily resolved datasets offers unprecedented potential to transform our understanding of the fundamental atomic level dynamic processes taking place in materials. The rate at which atomic resolution data can be generated in experiments has now outstripped our ability to process and mine scientifically important information. The complex nature of atomic resolution images necessitates intense supervision and input in most quantitative image processing operations. Such an approach is possible for a few hundred images but becomes completely impractical when millions of data frames are available. This has become a serious bottle-neck to extracting the new knowledge potentially available from advanced characterization systems. We need approaches that can recognize recurring motifs in noisy data sets, provide statistical confidence levels for detection of real structural features, track dynamic changes within the sample, and flag events which are likely to be scientifically significant.

We are focused on applying emerging data-driven techniques based on statistics and machine learning to characterize structural dynamics in nanomaterials with an initial focus on metal and oxide nanoparticles 1 – 3 nm in size. Nanoparticles have wide applications in many areas of materials science and are an ideal system for developing and benchmarking novel data-driven processing techniques because they only contain a few thousand atoms. The datasets are acquired using in situ electron microscopy and structural dynamic processes are stimulated using heat, gas or electron irradiation. We compare and contrast the performance of automated image processing algorithms designed to determine atomic coordinates and structural evolution with more conventional manual methods. The importance of understanding the confidence level that can be associated with derived structural models is critical to assess the reliability of the scientific information extracted from large data sets.

Acknowledgements
We gratefully acknowledge support of NSF OAC-1940263, and the use of facilities at Arizona State University’s John M. Cowley Center for High Resolution Electron Microscopy.

Developed as a Baushinger effect monitoring technique in polycrystalline metallic materials, passive acoustic wave monitoring has been a critical health monitoring technique in: Atomic energy pressurized tank; Power and structure monitoring in aeronautics; Automotive; As well as recent applications in spinning disk data storage (HDD) health monitoring applications [1]. Passive monitoring of acoustic waves generated during the initial contact point has been attracting the attention of material scientists since the inception of nanomechanical test instruments for nanoindentation and nanoscratch applications. The conventional acoustic wave signal treatment via RMS or integrated energy values proves that quantitative acoustic wave properties correlate well with the local contact materials phenomena such as: Yield point initiation for W(100) [2, 3]; Sapphire [4], phase transformation for SMA [2]; and differentiating of thin film fracture modes [5]. Several attempts have been made to look at the differentiative properties of the acoustic signatures via signal decomposition techniques such as wavelets [2, 6]. Even though acoustic wave signatures were reconstructed, the true potential of the method was not investigated from the machine learning perspective.
In this work a deep learning based signal processing of nanoindentation induced acoustic events was investigated. The synergy of wavelet signal decomposition and information theory based signal presorting prepared data for the deep learning step. In the deep learning step convolutive neural networks sort wavelet coefficients by their statistical significance via provided 2D signal topographies. This creates acoustic signature libraries that are typical for the specific materials phenomena during the indent. The hardware consists of a newly developed ultrasonic probe integrated into the nanoindentation tip, thus eliminating boundary effects and ensuring that only waves that pass via the contact are being recorded. Appropriate signal conditioning and fast data acquisition hardware is synchronized with a quasi-static nanoindentation process. The machine learning routine together with wavelet decomposition and presorting algorithms are implemented into the dedicated acoustic data evaluation kernel which resides in the fast access memory of the designated controller.

REFERENCES:
1. A. Daugela, S. Tadepalli, J. of Microsystem Technol., Vol. 18,No. 9-10(2012) pp. 1425 - 1430.
2. A. Daugela, H. Kutomi, T.J., Wyrobek, Zeitschrift fur Metallkunde, Vol. 92, No 9 (2001), pp. 1052-1056.
3. N. Tymiak, A. Daugela, T. J. Wyrobek and O. L. Warren, J. of Mater.Res., Vol. 18, No. 4 (2003), pp. 1 – 13.
4. N.I. Tymiak, A. Daugela, T.J. Wyrobek and O.L. Warren, Acta Materiallia, Vol. 52 (2004) pp. 553-563.
5. N. Faisal, R. Ahmed, R.L Reuben, Int. Mater. Reviews, Vol. 56, No. 2 (2011).
6. N. Tymiak, D. Chrobak, W. W. Gerberich, O. Warren, R. Nowak, Phys. Rev. B - Cond. Matter and Mater. Phys., Vol. 79, No. 17, (2009).

Thermal characterization in high-temperature environments is difficult because of high radiation losses and significant temperature measurement noise and inaccuracy. Ångström’s method can accurately measure thermal diffusivity in the presence of heat losses by observing the propagation of temperature perturbations through a material sample. This work applies a modified Angstrom’s method to measure thermal properties at extreme temperatures (>1000 °C). Previous studies employing Angstrom’s method at high temperatures typically have employed evenly spaced thermocouples on samples for thermal diffusivity measurements. We have recently demonstrated in preliminary room-temperature testing (copper - 1.13 x 10^(-4) m2/s, 1% uncertainty) that the uncertainty in thermal diffusivity within a Bayesian statistical framework is inversely proportional to the square root of the number of concurrent temperature measurements. In this work, we employ IR thermography for non-contact temperature measurement to greatly increase the number of measurements that can be resolved on a sample compared to the conventional thermocouple approach. To subject materials to extreme temperatures, samples are mounted vertically in a vacuum chamber and heated by concentrated thermal radiation from a 10 kW Xenon bulb. A controller is used to precisely modulate power output to produce the periodic temperature oscillation required for Angstrom’s method. Transient temperature measurements are made along the length of the sample using both an infrared camera and optical spectroscopy. Bayesian statistics then quantify the measurement uncertainty associated with the noisy high temperature environment, and this approach is shown to substantially reduce measurement uncertainty for high-temperature material characterization.

Predicting the performance of rechargeable batteries in real time is of great importance to battery research and industrial production, and hence has been a long pursuit. In this talk, a novel end-to-end unsupervised machine learning approach is shown for the battery performance prediction, which is free from feature engineering and uses only the raw images of the charge–discharge voltage profiles [1]. This model enables unsupervised real-time automatic extraction of latent physical factors that control the performance of Na-ion batteries to classify good or bad cycling performance by using only the voltage profile of the first cycle. This model can also monitor the safety of Li-metal battery systems by giving warnings when the battery is approaching a failure. With the beyond expert-level prediction ability, the abovementioned framework can be a promising prototype to further develop and enable high accuracy predictions of battery performance for realworld applications in the future.

[1] Beyond expert-level performance prediction for rechargeable batteries by unsupervised machine learning, Advanced Intelligent Systems, 1900102 (2019)

Machine learning has emerged as a powerful tool for the analysis of mesoscopic and atomically resolved images and spectroscopy in electron and scanning probe microscopy. The applications ranging from feature extraction to information compression and elucidation of relevant order parameters to inversion of imaging data to reconstruct structural models have been demonstrated. However, the fundamental limitation of the vast majority of machine learning methods is their correlative nature, leading to extreme susceptibility to confounding factors. While in classical statistical methods the methodology to address phenomena such as Simpson paradox are established, the complex and often non-transparent nature of modern DL tools renders them extremely prone to misinterpretation. We argue that correlative machine learning provides a reliable and powerful tool in cases when the causal links are well established, and discuss several examples of atom finding in SPM and STEM and analysis of 4D STEM data when this condition is satisfied. Notably, ML applications in theory generally fall under this category. Alternatively, ML methods work well when the confounding factors are effectively frozen via the narrowness of experimental conditions or experimental system. However, both these conditions are violated for experimental studies, when causal relationships are known only partially (and are in fact often the target of study) and confounding factors (composition uncertainty, microscope tuning, contaminations) are abundant. In this presentation, I will discuss the several examples of extending machine learning methods towards the analysis of causative physical mechanisms. One such approach is based on the Bayesian methods that allow to take into consideration the prior knowledge the system and evaluate the changes in understanding of the behaviors given new experimental data. The second pathway explores the parsimony of physical laws and aims to extract these from the set of real-world observations. Finally, the Bayesian networks can be used to explore the causative relationships in the multimodal data sets. These concepts will be illustrated using several examples of causal machine learning, including analysis of phase transitions on a single atom level in 2D materials and interplay between physical and chemical effects in the ferroelectric perovskites. Ultimately, we seek to answer the questions such as whether electronic instability due to the average Fermi level guides the development of the local atomic structure, or frozen atomic disorder drives the emergence of the local structural distortions, whether the nucleation spot of phase transition can be predicted based on observations before the transition, and what is the driving forces controlling the emergence of unique functionalities of morphotropic materials and ferroelectric relaxors.

This research is supported by the by the U.S. Department of Energy, Basic Energy Sciences, Materials Sciences and Engineering Division and the Center for Nanophase Materials Sciences, which is sponsored at Oak Ridge National Laboratory by the Scientific User Facilities Division, BES DOE.

Solar energy plays an important role in solving serious environmental problems and meeting the high energy demand. However, the lack of suitable materials hinders further progress of this technology. Here, we present the largest inorganic solar cell material search till date using density functional theory (DFT) and machine-learning approaches. We calculated the spectroscopic limited maximum efficiency (SLME) using the Tran–Blaha-modified Becke–Johnson potential for 5097 nonmetallic materials and identified 1997 candidates with an SLME higher than 10%, including 934 candidates with a suitable convex-hull stability and an effective carrier mass. Screening for two-dimensional-layered cases, we found 58 potential materials and performed G₀W₀ calculations on a subset to estimate the prediction uncertainty. As the above DFT methods are still computationally expensive, we developed a high accuracy machine-learning model to prescreen efficient materials and applied it to over a million materials. Our results provide a general framework and universal strategy for the design of high-efficiency solar cell materials. The data and tools are publicly distributed at: https://www.ctcms.nist.gov/~knc6/JVASP.html, https://www.ctcms.nist.gov/jarvisml/, https://jarvis.nist.gov/, and https://github.com/usnistgov/jarvis.

The integration of knowledge, experience, and intuition of researchers with artificial intelligence (AI) and robotics is the new direction to accelerate progress in materials research [1]. Strategies combining high-throughput synthesis with machine learning have already been producing new small organic molecules and bio-materials at ever faster rates. However, the application of the techniques to inorganic materials research is still in its infancy. Therefore, to drastically accelerate inorganic materials science research, establishing the AI and robotics inclusion into inorganic materials research is an urgent issue.

In this study, we demonstrate the autonomous synthesis of inorganic compounds using robotics and Bayesian optimization. This system fully automates sample transfer, thin film deposition, physical-property characterization, and growth condition optimization. The apparatus is equipped with a robot arm that can access each satellite chamber for growth and characterization. Based on the values obtained from the physical-property characterization, the Bayesian-optimization algorithm predicts the next growth condition (closed-loop). Here, we show the process of reducing the resistivity of Nb-doped TiO₂ transparent conducting thin films [2]. The autonomous synthesis can be applied to a wide variety of functional materials including magnetic, optical, electronic, and ionic properties.
References

[1]: R. F. Service, Science 366, 1295 (2019).
[2]: T. Hitosugi et al. Phys. Status Solidi A 207, 1529 (2010).

This presentation introduces Graphene Recipes for Synthesis of High Quality Materials (Gr-ResQ) (https://nanohub.org/tools/gresq). Gr-ResQ is designed to be a one stop solution for analysis of graphene synthesis by Chemical Vapor Deposition (CVD). Gr-ResQ consists of a comprehensive crowd-sourced database of synthesis recipes and results as well as a suite of software tools to analyze the database. It offers tools to analyze scanning electron microscope images and Raman spectra of graphene. It was developed as part of NSF's nanoMFG node within nanoHUB.

Advancing the manufacturing of graphene by CVD requires the ability to make CVD more cost-effective, including synthesizing low-defect density graphene directly on relevant substrates at reasonable costs. This is currently difficult mainly due to the multi-physics nature and the multi-time, and multi-length scales of the synthesis reaction. CVD processes can have over 200 variables, many of which are difficult to precisely control. This leads to poor repeatability and lack of a quantitative distribution of the parameter space. Characterization of synthesized graphene with electron microscopy and Raman spectroscopy gives useful information about the crystallinity, coverage, nucleation density and other such physical and chemical properties. The challenge of making interpretations and correlations from this data can be met by constructing a database that ties together the variables used in a process (the recipe) with the characterization of the synthesized graphene, which allows the application of data science and machine learning techniques to graphene and other 2D materials. One such use is in the estimation of graphene coverage from SEM images, which is generally a manual and time consuming task. We are trying to replace that by training a deep neural network (DNN) image segmentation algorithm with our manual coverage data to produce high quality, automated graphene coverage metrics.

In this presentation, we will discuss the database and the various tools, showing a number of examples that uniquely take advantage of the functionality of the database.

Advances in applying natural language processing (NLP) to material science text has greatly increased the size and acquisition speed of materials science synthesis data. These methods have been successfully applied to well-studied material systems with large amounts of data. However, we need ways to leverage literature data in materials domains without thousands of papers. Applying NLP pipelines to these types of materials science systems can be challenging due to the general schema and the noisiness of automatically extraction data. In this presentation, we demonstrate how to leverage domain knowledge to build upon existing data extraction techniques and improve extraction accuracy using examples in the zeolite, alternative cement, and solid-state electrolyte fields. We present data engineering techniques and discuss an optimal balance between automatic and manual data extraction.

We present an end-to-end system for computational autonomy in materials discovery (CAMD). This system is designed to sequentially augment a given dataset according to user-specified strategies for choosing new experiments and experimental procedures. The decision-making entities, called "agents," can make use of past knowledge, surrogate models, logic, thermodynamic or other physical constructs, heuristic rules, and different exploration-exploitation strategies. We show a series of examples for how discovery campaigns for finding materials satisfying a phase stability objective can be simulated to design new agents, and how those agents can be deployed in real discovery campaigns to control experiments run externally, such as the cloud-based density functional theory simulations in this work. Lastly, we present results from the cloud-based deployment of our tool that have resulted in over 2000 new inorganic crystalline materials in various crystal symmetries and compositions predicted to be within 200 meV of the phase diagram's convex hull.

Recently, machine learning (ML) is becoming an increasingly important tool for material discovery, thanks to its ability to quickly predict results of time and power-intensive calculations. Fundamentally, each ML model predicts some property and is composed of three essential elements: a database, a descriptor, and an ML algorithm. In this work, we have created and assessed a new tool for predicting formation enthalpies (ΔH_f). We take an open DFT database (OQMD), Voronoi tesselation-based descriptor developed by the Wolverton group^[1], and then design new neural networks that allow us to significantly improve the accuracy of predictions and reduce the computation time down to the order of 100 microseconds per prediction.

During the neural network design process, we created and tested around 50 networks belonging to 19 architectures. We used cross-validation on a randomly selected test set comprising 5% of the OQMD database (21,800 test entries) to compare with competing models. We found that our best network achieved a ΔH_f mean average error of 35 meV/atom, compared with ≈70 meV/atom for a model based on the Random Forest algorithm^[1]. In addition to simple cross-validation against a subset of the data for comparison with other models, the performance was also evaluated on more challenging systems. These included Special Quasirandom Structures (SQS) and 243 endmembers of a ternary σ phase (topologically close packed). Using such an approach during design, allowed us to improve accuracy for structures that are significantly different from those in the original dataset, and thus require extrapolation.

In this presentation, we show not only the new model but also new applications. Most notably, we take advantage of the improved prediction speed to screen whole databases for entries that do not fit the discovered patterns, and then automatically compare flagged entries to other sources. This, in turn, combined with more in-depth analysis, allows us to pinpoint entries that are likely to contain poor quality data, as well as those at risk of having a systematic error in the DFT calculation.


References:
[1] L. Ward et al., "Including crystal structure attributes in machine learning models of formation energies via Voronoi tessellations", Physical Review B, vol. 96, no. 2, 2017

Advances in electron imaging and spectroscopy instrumentation now allow for time-resolved data acquisition at rates over 1000 frames per second. However, this acquisition speed comes at the expense of the signal-to-noise ratio. At high frame rates, the low signal-to-noise (SNR) ratio can causes datasets to be too noisy for analysis with the naked eye without significant post-processing. Another challenge (often times exacerbated by poor SNR) is quantification of oxidation state. For example, in transition metal oxides, many methods exist for valence determination, e.g. L₃/L₂ ratio, chemical shift, or O K-edge pre-peak analysis. Unfortunately, these different methods can yield contradictory results, particularly in the presence of noisy data. Here, we seek to address both of these challenges through the application of machine learning techniques. First, we will discuss supervised learning of oxidation state in SrFeO_x thin films. With a convolutional neural network trained on >10,000 spectra, we obtain ~70% classification accuracy in differentiating SrFeO₃ (Fe⁴⁺) from SrFeO_2.5 (Fe³⁺) during an in situ experiment with very noisy data collected at 400 frames / second. Next, we will discuss ongoing efforts in unsupervised learning of EELS data. Unsupervised learning makes it possible to analyze different materials without needing a comprehensive database of labelled EELS data to compare with. Because unsupervised learning works by finding its own patterns or structures within the data, via techniques such as feature extraction or clustering, applying it to real-time in-situ data could help better understand phase morphology of the materials and could expose if unintentional or incomplete transformations of the sample occurred and in turn, the ability to more precisely and predictively tune complex materials. Unsupervised learning will be applied to various in situ EELS experiments.