Symposium CH04—Accelerating Materials Characterization, Modeling, and Discovery by Physics-Informed Machine Learning

The morning session will cover:
ML to theory (force fields, generic property predictions, visualization, accessing the different databases, etc.)
Deep learning introduction and use to segment images from electron or scanning probe microscopy

The afternoon session will cover:
Spectral processing with matrix/tensor factorization and deep learning methods (including autoencoders) for extracting data from large multidimensional datasets
Reinforcement learning and Bayesian optimization for efficient and smart experiments

Data continues to be a fundamental ingredient for accelerating and optimizing materials design and synthesis. Advances in applying natural language processing (NLP) to material science text has greatly increased the size and acquisition speed of materials science data from the published literature. This presentation will describe work to extract information from peer reviewed academic literature across a range of materials and explicit links between text-derived data and physics-based simulation data to inform materials discovery. We will present cases from microelectronics, catalysis and solid-state electrolytes.

Fabrication of CMLs with variations in deposition parameters, composition, thickness, oxidation state, and even nanostructure morphology have been demonstrated, but measuring and analyzing their properties is not straightforward. Typically, the measurements entail serial methods that measure samples point-by-point on pre-determined positions on the CMLs. While these measurements are often termed ‘high-throughput’ they are only so in the sense of processing large amounts of data. Especially, in the case of electrochemical measurements, the analysis can take a very long time, even days, to complete for one CML.

Here, we present a real high-throughput parallel method for determination of electrochemical properties. We do this using an array of ion-sensing field effect transistors (ISFETs) that was specifically designed for high-throughput measurements. ISFET sensors are used to detect changes in concentrations of specific ions in a solution, and as such can follow changes in electrochemical reactions. For this work we chose to examine CMLs of electrolyzers for oxygen evolution reaction (OER), which may form OH^- or H⁺ intermediates during the reaction. In our setup, the sensor array is placed in solution directly above the studied CML, and we examine the changes in H⁺ concentration as a result of the OER. The ISFET array allows us to quickly screen for material combinations of interest for OER, and generate large amounts of data in a short period of time, enabling parallel high-throughput CML characterization. We will share our most recent results in the generation of high-throughput CMLs [1] and analyzing these with similarly high-throughput parallel electrochemical measurements.

Reference:
[1] Hannah-Noa Barad, Mariana Alarcón-Correa, Gerardo Salinas, Eran Oren, Florian Peter, Alexander Kuhn and Peer Fischer; Combinatorial growth of multinary nanostructured thin functional films, Accepted to Materials Today, 2021.

Understanding, and predicting new materials with relevant electronic properties is a key task in materials science. This task is complicated in the case of materials with complex electronic interactions, particularly in the case of materials with an open d-shell of a transition metal ion known as correlated materials. This is due to the limitations of current theoretical methods, as well as the scarcity and heterogeneity of known materials. Of these materials, materials that exhibit a thermally driven metal-insulator transition (MIT) are particularly important for technological applications, yet only around 60 are known. To accelerate materials discovery and understanding, we have built, with the assistance of natural language processing techniques, the first materials database of MIT and related compounds which is now available for scientists’ use (1,2). We have trained a machine learning model on this database, allowing us to extract new relevant features for this class of materials, including generalized tolerance factors, the Ewald energy, as well as confirmed the importance of well-known ones such as the Hubbard U (1). Finally, we provide this model in a user-friendly binder format that scientists can run in their browser and obtain new predictions on whether a material is a metal, insulator, or MIT in seconds based on .cif structure files - without installing any software on their machine (1,3). Our work shows how machine learning can be used to accelerate materials discovery and understanding at every step of the research process. If time permits, I will present results on newly identified metal-insulator transition compounds.

References:
1) https://arxiv.org/abs/2010.13306 ; in press at Chemistry of Materials
2) https://mtd.mccormick.northwestern.edu/mit-classification-dataset/
2) https://tinyurl.com/mit-classifiers

The information, data, or work presented herein was also funded in part by the Advanced Research Projects Agency-Energy (ARPA-E), U.S. Department of Energy, under Award Number DE-AR0001209. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.

Fundamental atomic-level understanding of the active site structure is a central requirement of rational catalyst design. Due to the large configurational space inherent to many catalytic systems, it is very challenging to resolve the active sites directly. Moreover, these structures often exhibit dynamical behavior in response to the reactive environment. Here, we propose a multipronged strategy of decoding reactive structures at the atomic level, combining catalysis, characterization, machine learning, and first-principles based kinetic modeling. We unambiguously demonstrate the effect of catalyst treatment on its nanostructure and reactivity toward the prototypical hydrogen-deuterium (HD) exchange reaction, using a dilute alloy Pd₈Au₉₂ supported on highly stable raspberry colloid template SiO₂. Remarkably, the Sabatier optimum is established with small Pd_n ensembles, where the reactivity is tuned by modulating the active site on the order of only a few atoms (n = 1-3) through catalyst pretreatment.

The catalyst is activated by O₂ treatment and decreases in activity upon H₂ treatment. These observations are explained in terms of catalyst restructuring induced by the treatments, as evidenced by X-ray absorption spectroscopy and coordination numbers extracted from neural network inversion of the spectra. A quasi-random alloy is revealed: the majority of Pd remains dispersed inside Au, with a small amount of Pd segregating toward the surface upon O₂ treatment and dissolving into the bulk upon H₂ treatment. Extensive theoretical modeling of the HD exchange reaction network on several model surfaces establishes Pd monomers, dimers, and trimers as decisive candidate structures for the active site. Very good agreement is found between theory and experiment in terms of the apparent activation energy: O₂-treated samples are dominated by desorption-limited Pd trimers, whereas H₂-treated samples are characterized by dissociation-limited Pd dimers & monomers. These dilute Pd ensembles numerically satisfy the observed coordination numbers for all of our samples, thereby fulfilling both the kinetic and structural criteria of the active site. Our multidisciplinary approach enables precise identification of the active site at atomic resolution with fully consistent structure-activity relationship.

Understanding the protein structure-function relationship is essential for broad biological fields, with important applications in the design of new functional biomaterials, drugs, and antibiotics for biotechnology and pharmaceutical industries. Recent advances in protein function prediction take advantage of graph-based deep learning approaches to correlate protein 3D structure and topological features with molecular functions. The latest end-to-end model PersGNN had achieved a boost in Gene Ontology classification compared with baseline graph neural networks (GNNs). However, proteins in nature are not static but dynamic molecules interacting with the environment, constantly alternating conformation, and even forming assemblies of quaternary complexes. This talk will present the expressiveness and robustness of GNNs to encode information from spatial proximity, persistence homology, together with normal mode analysis across protein residues. The learned representation aggregates node-level features from local to global hierarchy and provides graph-level embedding with inter-residue dynamical couplings for downstream function prediction. High throughput classification tasks for over 100,000 proteins and 1,000 function tags from Protein Data Bank and Gene Ontology demonstrate remarkable performance gain in the discriminatory power based on such dynamics-informed representation. The proposed method can be readily extended to wide crystalline and amorphous materials that can be represented as graph-structured data and be incorporated with particle-based dynamics.

Disordered multicomponent systems have immense engineering design flexibility and a subsequently rich space of properties. A lattice cluster expansion is one powerful computational tool to obtain precise, detailed configurational and energetic resolution, but in practice has primarily been used in binary systems or ternary alloys. In this presentation we will discuss approaches to significantly increase the complexity of systems to which the cluster expansion can be applied, including a high number of species, different valence states of ions, complex underlying lattices, and large off-lattice relaxations. We demonstrate this approach to a multicomponent, multi-sublattice lithium manganese oxyfluoride electrode system, setting up the largest cluster expansion ever attempted to our knowledge. In order to bridge the vastly under-determined configurational space observed by neutron diffraction with computation, we apply the sparse group lasso statistical method to reduce model complexity from over 4500 features to fewer than 200. We also use Bayesian optimization via Gaussian Processes in oxidation state assignments during data cleaning, demonstrating its success in high dimensional spaces with limited training data. Lastly, we extend existing theoretical cluster expansion methods and derive how to analytically calculate and characterize local structural and chemical orderings in any state of disorder. Our developments make the fitting and analysis of complex systems tractable and provide guidance for thorough, high-throughput computational studies of multi-component, disordered systems.

Although machine learning offers a unique, largely untapped opportunity to accelerate the discovery of novel glasses with exotic functionalities, it faces several challenges. Since they are usually only driven by data and do not embed any mechanistic knowledge, machine learning models can sometimes violate physical and chemical laws. For these reasons, machine learning techniques are usually good at “interpolating” data but have thus far a limited potential for “extrapolating” predictions far from their initial training set, which prevents the efficient exploration of new unknown compositional domains. Here, we present a new machine learning framework aiming to predict the properties of oxide glasses. As expected, we show that “blind machine learning” (i.e., which does not embed any physical or chemical knowledge) can successfully interpolate data but fails at extrapolating predictions far from its training set. As an alternative route, we report a new “topology-informed machine learning” framework that addresses this limitation. This framework relies on artificial neural networks and embeds a topological description of the connectivity of the atomic network. We show that our “topology-informed machine learning” model allows us to extrapolate predictions far from the training set. More generally, topology-informed machine learning offers a promising route to overcome the tradeoff between accuracy, simplicity, and interpretability—which are otherwise often mutually exclusive in traditional “blind machine learning” models.

Molecular crystals play an important role in various fields of science and industry, with applications in the pharmaceutical, electronics, and food industries. Predicting the stability of crystal structures formed from molecular components is non-trivial, given the complexity inherent to the multiscale nature of their structure and the subtle balance of weak intermolecular interactions governing the structure-property relations. We have curated a dataset containing over 3’000 molecular crystals, their constituent molecules, and their related properties. We use it to demonstrate the use of machine-learning techniques to estimate the stability of the crystalline structures. We use general-purpose descriptors of atomic structure and modify them to explicitly incorporate information on the molecular nature and thermodynamics of the material. In combination with a principal covariates regression analysis that explicitly determines the best low-dimensional representation of structure-property relations, we demonstrate how to identify the atomic motifs associated with a strong stabilizing effect on molecular packing.

The space of possible grain boundary structures is vast, with 5 macroscopic, crystallographic degrees of freedom that define the character of a grain boundary. While numerous datasets of grain boundaries have been created to examine this space, none has systematically examined the full range of possibilities. We will present a dataset of more than 4200 unique grain boundaries in the 5D crystallographic space. Our sampling includes a range of possible microscopic, atomic configurations for each unique 5D crystallographic structure, which we refer to as metastable grain boundary structures. In all, the number of metastable structures associated with the 4200 unique grain boundaries is over 31 million. We will present an overview of the methods used to generate this dataset, an initial examination of the raw data, as well as methods and insights gained in machine learning of grain boundary energy structure-property relationships.

Defect dynamics in materials are of central importance to a broad range of technologies from catalysis to energy storage systems to microelectronics. Material functionality depends strongly on the nature and organization of defects – their arrangements often involve intermediate or transient states that present a high barrier for transformation. The lack of knowledge of these intermediate states and the presence of this energy barrier presents a serious challenge for inverse defect design, especially for gradient-based approaches. Here, we present a reinforcement learning (Monte Carlo Tree Search) based on delayed rewards that allow for efficient search of the defect configurational space and allows us to identify optimal defect arrangements in low dimensional materials. Using a representative case of 2D MoS₂, we demonstrate that the use of delayed rewards allows us to efficiently sample the defect configurational space and overcome the energy barrier for a wide range of defect concentrations (from 1.5% to 8% S vacancies) – the system evolves from an initial randomly distributed S vacancies to one with extended S line defects consistent with previous experimental studies. Detailed analysis in the feature space allows us to identify the optimal pathways for this defect transformation and arrangement. Comparison with other global optimization schemes like genetic algorithms suggests that the MCTS with delayed rewards takes fewer evaluations and arrives at a better quality of the solution. The implications of the various sampled defect configurations on the 2H to 1T phase transitions in MoS₂ are discussed. Overall, we introduce a Reinforcement Learning (RL) strategy employing delayed rewards that can accelerate the inverse design of defects in materials for achieving targeted functionality.

Solid-state synthesis from powder precursors is the primary processing route to multicomponent ceramic materials. The black-box nature of the reaction vessel typically precludes an understanding of the solid-state reaction mechanisms that govern phase evolution, leading to trial-and-error approaches to synthesis. In the age of autonomous laboratories, better theories of materials synthesis are needed to predict and guide robotic synthesis approaches to novel materials. Here, we combine ab initio thermodynamics with in situ synchrotron X-ray scattering and TEM to build predictive theories for which non-equilibrium phases form during solid-state ceramic synthesis, and why. I will present examples from the ceramic synthesis of NaxMO2 (M = Co, Mn) layered oxides [1] and also the classic high-temperature superconductor YBa2Cu3O6+x (YBCO) [2]. We derive an interfacial reaction model to predict the first-phase-to-form between two powder precursors, which can consume a large fraction of the reaction energy and can topotactically template structural transformations to other non-equilibrium intermediates. Our insights can help guide the choice of precursors and parameters employed in the solid-state synthesis of ceramic materials, and constitutes a step forward in building predictive thermodynamics and kinetics-informed machine-learning strategies for the autonomous synthesis of novel materials.


[1] M. Bianchini, J. Wang, W. Sun, G. Ceder et al., "The interplay between thermodynamics and kinetics in the solid-state synthesis of layered oxides" Nature Materials 2020

[2] A. Miura, C. Bartel, W. Sun et al., "Observing and modeling the sequential pairwise reactions that drive solid-state ceramic synthesis" Advanced Materials (2021)

Machine learning is playing an increasing role in the physical sciences and significant progress has been made towards embedding domain knowledge into models. Less explored is its use to discover interpretable physical laws from data. We propose parsimonious neural networks (PNNs) that combine neural networks with evolutionary optimization to find models that balance accuracy with parsimony. The power and versatility of the approach is demonstrated by developing models for classical mechanics and to predict melting temperature of materials, and performance of energetic materials from fundamental properties. In the first example, the resulting PNNs are easily interpretable as Newton’s second law, expressed as a non-trivial time integrator that exhibits time-reversibility and conserves energy, where the parsimony is critical to extract underlying symmetries from the data. In the case of melting, the PNNs not only find the celebrated Lindemann melting law, but also new relationships that outperform it in the pareto sense of parsimony vs. accuracy.

High-resolution Transmission Electron Microscopy (HRTEM) is a powerful technique for examining matter at the atomic scale. In particular, the ability to acquire single projection images of the sample volume in a parallel fashion employing low electron doses and dose-rates makes HRTEM attractive to detect the atomic-scale realm of matter [1]. However, interpreting the image is not always straightforward, as images are formed by phase contrast.

A focal series of HRTEM images enables reconstruction of the exit wave function, which provides the most informative measure of the sample. A series of typically 20-50 images acquired with varying defocus is used to numerically reconstruct the most likely exit wave, which is then interpreted as it corresponds closely to the structure of the sample [2,3].

Here we show that a convolutional neural network is able to reconstruct the exit wave function from a focal series of only two or three images – but not from a single image. The convolutional neural network is based on the U-Net architecture [4] and is almost the same as the one we have previously used to identify the positions of atomic columns in single images of metallic nanoparticles [5].

We train the neural networks on simulated images. The simulated images are produced with the multislice algorithm, using pyqstem [6], both the exit wave function and images produced with three different values of the defocus are saved. The exit waves are then blurred slightly by folding with a gaussian, and a neural network is trained to reconstruct the exit wave from the images. The network is validated on a different set of simulated images.

In all cases we work with thin flakes of 2D materials, although the methods should generalize to other classes of materials. We train and validate networks based on three different types of samples of increasing difficulty. The first two are unsupported nanoflakes of molybdenum disulphide (MoS₂), and molybdenum disulphide supported on a graphene substrate. MoS₂ is an active catalyst for photooxidation, hydrogen evolution and hydrodesulphurization reactions. The third type is a diverse collection of existing and proposed quasi-two-dimensional materials, taken from the Computational 2D Materials Database (C2DB) [7]. In all cases the neural network is able to learn to reconstruct the exit wave function of the vast majority of the samples in the validation sets.

Finally, we demonstrate that the network trained on simulated data for graphene-supported molybdenum disulphide can also be used to analyze images from an experimental focus series taken on a sample of a realistic model of an industrial MoS₂ based hydrodesulphurization catalyst. The neural network is able to reconstruct the exit wave with such fidelity that the atomic structure of both the MoS₂ nanoparticle and the supporting graphene is clearly recovered.

1. Kisielowski, C., Frei, H., Specht, P., Sharp, I. D., Haber, J. A., Helveg, S., Adv Struct Chem Imag 2, 13 (2016). doi:10.1186/s40679-016-0027-9
2. de Beeck, M.O., Van Dyck, D., Coene, W., Ultramicroscopy 64, 167–183 (1996). doi:10.1016/0304-3991(96)00058-7
3. Tiemeijer, P.C., Bischoff, M., Freitag, B., Kisielowski, C., Ultramicroscopy 118, 35–43 (2012). doi:10.1016/j.ultramic.2012.03.019
4. Ronneberger, O., Fischer, P., Brox, T., in: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pp. 234–241. Springer, Cham (2015). doi:10.1007/978-3-319-24574-4_28
5. Madsen, J., Liu, P., Kling, J., Wagner, J.B., Hansen, T.W., Winther, O., Schiøtz, J., Adv. Theory Simul. 1, 1800037 (2018). doi:10.1002/adts.201800037
6. Madsen, J., Liu, P., Wagner, J. B., Hansen, T. W., Schiøtz, J., Adv. Struct. Chem. Imag. 3, 14 (2017). doi:10.1186/s40679-017-0047-0
7. Haastrup, S. et al., 2D Materials 5, 042002 (2018). doi:10.1088/2053-1583/aacfc1 See also https://cmr.fysik.dtu.dk/c2db/c2db.html

Recent advancements in transmission electron microscopy (TEM) have significantly improved the information-richness and efficiency of acquisition of TEM data. Traditional methods for TEM data analysis based on manual interpretations of TEM images are often found insufficient to take full advantages of the richness of information contained in these data. Here, we will demonstrate how machine learning can be used to provide potential solutions for this problem. First, we will demonstrate an unsupervised machine learning algorithm for classifying the shapes of gold nanoparticles from TEM images. The algorithm is capable of extracting shape parameters of nanoparticles from TEM images, and classify the particles based on their shapes, requiring minimum human input in the process. The algorithm’s performance will be tested on two datasets of gold nanoparticles with different morphologies, and compared to human analysis of the same datasets to validate its accuracy and efficiency. We will show that the algorithm provides a method for efficient extraction and interpretation of statistical morphological information of metal nanoparticles, while introducing minimum human biases into the results. We will then demonstrate the use of machine learning to identify and track atomic movements during the oriented attachment and annealing processes of semiconductor nanocrystals observed by in-situ TEM. We will show how advancements in the automation of electron microscope imaging has allowed the acquisition of this dataset in high-throughput, and how machine learning can provide an efficient method to extract information from such a highly complex dataset. We hope that the two applications we show here will serve as examples of how machine learning can be incorporated into the analysis of large and/or complex TEM datasets, enabling the extraction of information that may not be obtainable otherwise.

X-ray diffraction (XRD) is indispensable in high-throughput materials screening, both for confirming the synthesis of expected phases and for revealing new structures in unexplored material systems. As a case study, samples synthesized from binary and ternary intermetallic systems were screened for A15-type phases based on XRD measurements. These compositions span both known and unknown phase diagram compositions. Many A15 phases are Type-II superconductors at comparatively high temperatures relative to other metallic alloys, motivating a hypothesis that screening novel A15 phases will lead to a high proportion of novel superconductors. A bottleneck in this approach is identification of relevant phases from the XRD patterns, which, even using current software, requires human intervention and is thus not suitable for fully automated workflows. We report results of a convolutional neural network (CNN) trained to classify A15-like phases based on a large database of synthetic data with simulated noise, and show strategies needed to address class imbalance challenges intrinsic to identification of single classes of phases. This enables high-throughput sample screening and allows focusing of human time only on the most interesting materials. These approaches can extend readily to other materials discovery settings where properties are strongly tied to particular phases that are distinguishable based on XRD patterns.

The powerful combination of supercomputing resources, robust algorithms for solving the laws of physics and state-of-the-art software infrastructure are enabling rapid, systematic calculations of real materials properties from quantum mechanics across chemistry and structure. A result of this paradigm change are databases like the Materials Project (www.materialsproject.org) which is charting the properties of all known inorganic materials and beyond, designing novel materials and offering the data free of charge to the community together with online analysis and design algorithms. The growing body of available, reliable data has reached the stage where automated learning algorithms can be effectively trained and utilized to accelerate all aspects of the materials design cycle: from property prediction to materials synthesis and characterization. In 2021, the Materials Project has rolled out a completely revamped infrastructure to meet growing demands on data accessibility and user experience. A stand-alone data portal has been launched for community data contributions, which is designed to automatically connect contributed materials data with the organization and users of the Materials Project. We will present the new infrastructure and exemplify the approach of data-driven materials design using in-house case studies as well as community-driven research. The data spans design, characterization and synthesis - closing the loop and showcasing rapid iteration between ideas, computations, insight and new materials development.

Materials processed under conditions that are far from equilibrium can yield new metastable materials with uncommon and valuable properties. Exhaustive exploration of the material space is often intractable due to the many synthesis degrees of freedom, e.g., processing time, temperature, and multiple composition variables. To explore this complex space most efficiently, our workflow employs high-throughput experimentation along with powerful machine learning algorithms. Our experiments consist of lateral gradient laser spike annealing of thin films characterized by micron scale analytics. To rapidly interpret these data, we use a combination of complementary active learning schemes: Gaussian process regression to map out phase boundaries and NMF/compressive sensing to label phases where possible. For a wide range of material systems (oxides, nitrides, and chalcogenides), we demonstrate the richness of the processing phase diagrams and identify methods to address the challenges that arise when executing completely autonomous interpretation.

The most common and popular methods for structure search and optimization are based on evolutionary design. This can often be cumbersome, limited to few tens of parameters, and fails for large structural configurations or design problems with high degrees of freedom. Reinforcement learning approaches mostly operate in discrete action space such as in Go game but the applications of that to inverse problems is limited since most inverse problems deal with continuous action space. There are a large number of inverse structural search problems ranging from crystal structure search in material sciences to topology design in Quantum information, where it is highly desirable to optimize structure/configuration to target desired properties or functionalities. This talk will provide an overview of our current efforts to perform scalable crystal structure search to discover and design metastable or non-equilibrium phases with desired functionality. We will demonstrate select use cases that highlight the efficacy of CASTING for molecular, crystal structure, defects, topology, device, and component design.

The idea of material discovery has excited and perplexed scientists for centuries. Several methods have been employed to find new types of materials, ranging from replacement of atoms in a crystal structure to advanced machine learning methods for predicting entirely new crystal structures. In this work, we investigate the performance of various Generative Adversarial Network (GAN) architectures to find innovativee ways of generating theoretical crystal structures that are synthesizable and stable. Over 300,000 entries from Pearson’s Crystal Database are used for the training of each GAN. The space group number, atomic positions, and lattice parameters are parsed and used to construct an input tensor for each of the network architectures. Several different GAN layer configurations are designed and analyzed, including Wasserstein GANs with weight clipping and gradient penalty, in order to identify a model that can adequately discern symmetry patterns that are present in known material crystal structures.

Gold nanorods (AuNRs) have found applications in cancer imaging and treatment, catalysis, and solar technology due to their size dependent and highly tunable optical properties. The utilization of these applications requires the synthesis of AuNRs of specific sizes and shapes. The determination of the sizes and shapes of AuNRs produced by a colloidal synthesis is commonly achieved using TEM analysis. This procedure requires a significant time investment and expensive instrumentation to attain the volume and quality of images necessary to produce statistically significant distributions. Our method extracts quantitative information about AuNR size distributions from absorption spectra. We treat an absorption spectrum as a linear combination of single particle spectra, and use numerically simulated spectra to determine the length, diameter, and aspect ratio distributions of simulated AuNRs which reproduce the experimental spectrum. This model has been used as an automated tool on AuNRs synthesized using robotic high throughput synthesis, which utilizes a modified one step seedless AuNR synthesis procedure. The accuracy of this model is verified using TEM analysis of samples of this dataset and by the extraction of well labeled spectra from the AuNR synthesis literature. This project demonstrates an effective method to extract population level size distribution information from a colloidal nanocrystal sample. Through using this model on high throughput samples and literature spectra where quantitative information about the size distributions is unknown, the quantitative impact of synthetic procedure on the size distributions of AuNRs can be determined, further developing the fundamental knowledge of AuNR synthesis. In principle, this approach can be extended to any other nanocrystal system where the absorption spectra are size dependent and accurate numerical simulation of the absorption spectra is possible.

Three-dimensional (3D) characterization across the nanoscale is now possible using scanning / transmission electron microscopes (S/TEM). Unfortunately, tomographic reconstructions can take one to several days to complete depending upon the dataset size or algorithm(s) employed. Even worse, the reconstruction occurs offline, long after all the data has been collected, preventing immediate interpretation during an ongoing experiment. While advancements in detector hardware have boosted throughput with digital data collection, substantial human effort and computational resources are still required to process the deluge of raw data. Thus, it has been a longstanding goal to begin 3D analysis of specimens in real-time allowing immediate assessment of nanoscale structure. Continuous 3D feedback provides early diagnosis of specimens to bridge material synthesis with accelerated characterization. Achieving high-throughput electron tomography requires an integrated pipeline that links the microscope hardware to optimized reconstruction algorithms and efficient 3D volumetric visualization. Moreover, this pipeline should be multi-threaded to run dynamic visualizations that updates as new data is collected or reconstruction algorithms proceed.

Here we present interactive 3D visualization that seamlessly runs while experimental projections are collected in an electron microscope using the tomviz platform (www.tomviz.org). Tomviz presents 3D materials structure to scientists in real-time enabling high-throughput specimen interpretation with immediate tomogram visualization. During the experimental acquisition, tomviz monitors for new projections which are continuously appended into the reconstruction process. After a reconstruction updates, tomviz immediately renders the 3D volume. We demonstrate real-time tomography on a helical nanoparticle comprised of a Cysteine amino-acid coordinated with Cadmium (Cys/Cd). The specimen’s left-handed chirality is discerned in the first twenty minutes and a high-resolution volume is available within 50% of the experiment (20-30 minutes). Real-time tomography demands maximally efficient computational speed to ensure the optimization runs faster than the experimental data acquisition rate. Tomograms are reconstructed in parallel with data acquisition and a high-quality 3D reconstruction is available before the experiment ends. Tomviz contains novel real-time algorithms that are compatible with high-performance computing clusters to produce and visualize large 3D volumes (>1024 voxels).

The multithreaded pipeline in tomviz enables interactive 3D analysis of the current reconstruction state with minimal impact on performance. A robust graphical interface allows for 3D objects to be rendered as shaded contours or volumetric projections and these objects can be rotated, cropped, or sliced as the reconstruction evolves. Thus, scientists can go beyond superficial inspection to quantify specimen features or internal structure while simultaneously operating the microscope. This immediate feedback can save researchers days of effort as reconstructions are no longer processed offline.

Tomviz offers unparalleled electron tomography throughput with real-time 3D analysis — it is an open-source repository that is available to all institutions for download at www.tomviz.org

The use of functionalized tips in Atomic Force Microscopy (AFM) has enabled the investigation of organic molecules with atomic resolution [1]. To more easily interpret the resulting images - which is often a non-trivial task, even for experts - a machine-learned approach based on Convolutional Neural Networks (CNN) has been recently proposed [2]. This method associates unique descriptors to a high-resolution AFM image, allowing automatic extraction of the spatial configuration of the molecule being probed, even for highly three-dimensional cases.
However, if no previous knowledge is available about the identity of the molecule in question, it is very difficult to draw meaningful conclusions.
Here, we approach the problem of recognizing both the identity and the configuration of unknown molecules in high-resolution AFM by using the CNN-extracted descriptors to guide the latent space optimization of a deep generative model for molecular graphs.
These models have seen a rapid development over the past few years, as a way to generate plausible small-molecule drug candidates [3]. Usually, once a generative architecture has been trained to reproduce sufficiently well the distribution of the training data (e.g. the QM9 dataset), optimization of the model’s latent space can then be carried out to maximize a property of interest, such as drug-likeness, obtaining a list of best candidates.
In the context of AFM, we optimize the latent space of a state of the art generative model using a score which gives the alignment of each generated molecule with the original CNN-extracted descriptors. As a result, a ranked list of best-fit, chemically valid molecules is proposed as possibly underlying the starting AFM image.

[1] - Gross L., Mohn F., Moll N., Liljeroth P. and Meyer G. 2009 Science 325 1110
[2] - Alldritt B., Hapala P., Oinonen N., Urtev F., et al 2020 Science Adv 6 6913
[3] - Gómez-Bombarelli R. et al., 2018 ACS Cent Sci 4 268–276

Advances in automation and data analytics can aid exploration of the complex chemistry of nanoparticles. Lead halide perovskite colloidal nanocrystals provide an interesting proving ground: there are reports of many different phases and transformations, which has made it hard to form a coherent conceptual framework for their controlled formation through traditional methods. In this work, we systematically explore the portion of Cs–Pb–Br synthesis space in which many optically distinguishable species are formed using high-throughput robotic synthesis to understand their formation reactions. We deploy an automated method that allows us to determine the relative amount of absorbance that can be attributed to each species in order to create maps of the synthetic space. These in turn facilitate improved understanding of the interplay between kinetic and thermodynamic factors that underlie which combination of species are likely to be prevalent under a given set of conditions. Based on these maps, we test potential transformation routes between perovskite nanocrystals of different shapes and phases. We find that shape is determined kinetically, but many reactions between different phases show equilibrium behavior. We demonstrate a dynamic equilibrium between complexes, monolayers, and nanocrystals of lead bromide, with substantial impact on the reaction outcomes. This allows us to construct a chemical reaction network that qualitatively explains our results as well as previous reports and can serve as a guide for those seeking to prepare a particular composition and shape. The dataset shown here is openly accessible, has already resulted in several-follow on publications and has been used in a classroom setting to help undergraduate students learn how to explore questions at the intersection of materials chemistry and data analytics.

Double perovskite structure has brought significant attention due to its great potentials for applications of rechargeable batteries, lighting-devices, and energy harvesting materials. Despite of its importance, it is difficult to find the thermodynamically stable structure satisfying target properties due to its enormous chemical space. Conventionally, the tolerance factor is used to predict the stability of the perovskite materials, but its usage has not been fully validated for the double perovskite structures. For example, the halide-type double perovskite structures have demonstrated exceptional physicochemical properties, but its search space is enormous; the chemical formula of the halide-type of double perovskite is ABB’X₃ and the number of atomic elements, which can locate at A and B site, reaches 70 and 4 elements to X site, leading to 1,400,000 structures. In this regard, applying the high-throughput materials screening to this type of material cannot be performed with the conventional method such as experiments or ab initio calculations.
In this study, to overcome above-mentioned issues, machine learning algorithm is employed to search for new stable double perovskite halide materials. First, the inorganic materials properties are adopted from well-established Materials Project database and double perovskite oxide database that we constructed in our previous work and they are used as a training set (around 120,000 structures and 11,000 double perovskite oxide). Then, the chemical features are constructed to describe given materials. We used the regression as well as the classification method for optimizing the surrogate model. Finally, trained machine learning model is applied to the whole chemical space of the double perovskite halide material (13,542 structures; they are chosen based on previous tolerance factor for validation of the current model). Initially, the prediction accuracy for the formation energy was obtained to be r-squared score of 0.77 and RMSE of 0.399 (eV/atom) when trained only with previous data sets. Finally, to further improve the prediction accuracy more efficiently, we employed the optimization method; exploration by quantifying the prediction uncertainty. This method clearly validated that the r-squared value is increased to 0.90, only by adding 6.5% of selected halide data to the training data.
In this study, the actual formation energy and convex hull energy of the Double perovskite candidate material were calculated using DFT. Machine learning models performed using DFT computational results were successfully able to predict formation energy and convex hull energy of double perovskite halide.

Deactivation of supported metal nanocatalysts under harsh reaction conditions limits the feasibility and efficiency of important industrial reactions. High reaction temperatures and reactive environments enable nanoparticle evolution to reduce the particles’ high surface energy, while simultaneously reducing the number of available reaction sites and catalytic activity overall. While research aimed to describe coarsening and degradation processes in nanoscale systems has been undertaken for decades, the difficulty of in situ observation and the analysis of in situ experimental data has inhibited the development of a theory which comprehensively describes the evolution of supported nanoparticles.

With the goal of understanding coarsening and degradation kinetics in model catalyst systems, we used in situ Transmission Electron Microscopy (TEM) to track the evolution of supported gold nanoparticles heated to 900C under vacuum. Using unsupervised machine learning to automatically process TEM images, nanoparticles sizes and positions as a function of time were extracted from experimental images. Using detailed characterization from both ensemble- and particle-scale measurements, we present an analytical model describing nanoparticle degradation as a competition between evaporation into the surrounding vacuum and diffusion between particles along the support. Moreover, we can fit our model to the evolution trajectories of hundreds of individual nanoparticles to determine statistically validated values for fundamental physical properties such as the diffusion coefficient, solid-vapor surface energy, and the equilibrium vapor pressure for gold nanoparticles at high temperature. By comparing these values with literature references, and thereby affirming the validity our model, we can investigate the physical meaning of the remaining parameters which are not measurable by traditional experimental methods. As an example, we are able to use our rich dataset describing nanoparticle evolution to characterize mean diffusion distances for gold adatoms on the silicon nitride support to show that, while nearest neighbor interactions dominate the role of diffusion in nanoparticle growth, there is a measurable contribution for long range interactions across the system of supported particles. This finding demonstrates how a traditional mean-field kinetic model paired with novel data analysis methods can yield results which could not be obtained by either method alone and motivates further study into new physical models which can describe coarsening and evaporation without relying on mean field assumptions.

Gold surfaces have a long history of microscopic studies due to their propensity to reconstruct in complex patterns depending on the terminating surface facet and the environment (temperatures and atmosphere). However, the details of restructuring processes and metastable intermediate structures, which could be critical for catalytic reactions, remain unknown due to the limits of microscopy time resolution and accuracy of classical force field simulations.

This work demonstrates a robust many-body Bayesian potential trained with an on-the-fly active learning framework implemented using Gaussian process regression in the FLARE code [1]. The active learning module uses molecular dynamics (MD) simulations of low-index surfaces to sample important configurations and runs density functional theory (DFT) calculations when the prediction Bayesian uncertainty exceeds a threshold. The trained Bayesian force field is then used to perform a large-scale MD to study surface diffusion and reconstructions of various surface facets, e.g., semi-hexagonal reconstruction on (100), Herringbone reconstruction on (111), and missing-row reconstruction on (110).

[1] J. Vandermause et al, NPJ Computational Materials 6 (2020).

Atomic force microscopy (AFM) with molecule-functionalized tips has emerged as the primary experimental technique for probing the atomic structure of organic molecules on surfaces [1]. Most experiments have been limited to nearly planar aromatic molecules, due to difficulties with interpretation of highly distorted AFM images originating from non-planar molecules [2]. Here we develop a deep learning infrastructure that matches a set of AFM images with a unique descriptor characterizing the molecular configuration, allowing us to predict the molecular structure directly in a few seconds on a laptop [3]. We apply this methodology to resolve several distinct adsorption configurations and conformations of molecules based on low-temperature AFM measurements. In general, we find high success rates in predicting the atomic and chemical structure of molecules. This approach opens the door to apply high-resolution AFM to a large variety of systems for which routine atomic and chemical structural resolution on the level of individual objects/molecules would be a major breakthrough. We also look at applications of similar approaches to the imaging of biomaterials and AFM imaging in liquids.

[1] N. Pavliček and L. Gross, Nat. Rev. Chem. 1, 0005 (2017)
[2] P. Jelinek, J. Phys.:Condens. Mat. 29, 343002 (2017)
[3] B. Alldritt, P. Hapala, N. Oinonen, F. Urtev, O. Krejci, F. F. Canova, J. Kannala, F. Schulz, P. Liljeroth, and A. S. Foster, Sci. Adv. 6 (2020) eaay6913

To enable the development of self-driving laboratories and their application to the synthesis of inorganic materials, we designed a deep learning approach that automates the analysis of diffraction spectra and identifies chemical reaction products. Using a convolutional neural network, our model was trained on simulated spectra that were augmented with physics-informed perturbations to account for possible experiments artifacts left over from sample preparation and synthesis. These diffraction spectra were derived from experimentally reported phases as well as hypothetical solid solutions with off-stoichiometric compositions. For the analysis of spectra obtained from mixtures of materials, we developed a branching algorithm to evaluate suspected mixtures and identify the set of phases that maximize the probabilities given by the neural network. Our model was tested on simulated and experimental data from materials in a five-component system, achieving accuracies up to 92% and 87% for the identification of single- and multi-phase spectra, respectively. These results show an improvement greater than 10% when compared to previously reported methods based on profile matching and deep learning. To extend the application of our approach to arbitrary sets of materials, we provide a software suite that facilitates the training of new models, which we hope may be used to automate the characterization step in high-throughput and autonomous experimental workflows.

GeSn alloys on Si have been an attractive research area given their capability to engineer the bandgap for mid-infrared applications and compatibility with Si-based integrated circuits. Recently, short-range order (SRO) in GeSn alloys has been theoretically predicted [1] especially for >15 at.% Sn composition, which profoundly impacts the band structure of GeSn alloys. In principle, atom probe tomography (APT) is a powerful experimental technique to study SRO (e.g., via radial distribution function, RDF) by offering 3D atomic reconstruction of the alloy. In reality, however, high-fidelity retrieval of lattice information and accurate RDF has been limited to metals [2,3]. Deriving SRO of semiconductors alloys from APT has been a significant challenge due to larger measurement error in atomic positions as well as lower detection efficiency compared to elemental metals. In fact, the APT data of semiconductor alloys resembles a significantly perturbed lattice, and ~50% of the atomic sites are empty due to the limited detection efficiency. Therefore, one can no longer identify the lattice from the raw APT data. Currently, SRO analyses in alloys are limited to relatively crude methods such as artificially cutting a line midway between the theoretical distances of the k th and (k+1) th nearest neighbor shells to accommodate the perturbed atomic positions [4]. To overcome this limit and retrieve SRO at high fidelity, here we demonstrate a new method that only requires the input of crystal structure to derive SRO in GeSn alloys from the nonideal APT data by combining Poisson distribution and k-nearest neighbors (KNN) analysis. This approach establishes a statistical correlation between the nominal KNN directly obtained from the nonideal APT data to the true KNN shell for diamond cubic structure, thereby regrouping the former to reconstruct the latter at high fidelity. Lattice constants can also be derived using an iterative approach given the crystal structure. We demonstrate that our method can retrieve lattice constants with ~1.5% accuracy from the APT data of Ge and GeSn alloys. For GeSn alloys, we define SRO parameter α^KNN_Sn-Sn as the probability for a Sn atom to have another Sn atom in the true KNN shell divided by the Sn at.% in the same region. For example, α^KNN_Sn-Sn>1 indicates more Sn-Sn clustering than a random alloy. This way, we were able to study nanoscale SRO distribution along the depth of the GeSn thin films. Our study shows that α^KNN_Sn-Sn decreases from 1.052±0.051 to 1.026±0.045 as the depth increase from 150 to 800 nm under the surface, with 18 10 nm×10 nm×10 nm blocks examined at each depth. These results indicate more preference towards Sn-Sn clustering near the surface where the Sn composition is higher due to strain relaxation [5]. The discovery of a relatively large fluctuation in SRO in GeSn, together with the theoretical prediction of large band structure variation with SRO, explains why the photoluminescence (PL) peaks are broad even at low T for GeSn (i.e., width >>k_bT). Furthermore, we propose an algorism based on least square fit to rearrange atoms to nearby perfect lattice sites, which show a very good consistency with the aforementioned results before rearrangement. A great advantage of APT-based SRO analyses is that we can map the SRO at nanoscale in a relatively large region with millions of atoms (e.g., 30x30x800 nm³), while other methods either give an average (EXAFS) or only analyze a very small region (STEM). Our method can be extended to other material systems such as high-entropy alloys.

[1] Cao, B., et al. ACS Applied Materials & Interfaces 12, 51, (2020): 57245–57253.
[2] Gault, B., et al. Microscopy and Microanalysis 16.1 (2010): 99-110.
[3] Geiser, B.P., et al. Microscopy and Microanalysis 13.6 (2007): 437-447.
[4] Ceguerra, A. V., et al. Acta Crystallographica Section A: Foundations of Crystallography 68.5 (2012): 547-560.
[5] Dou, W., et al. Scientific reports 8.1 (2018): 1-11.

The global rise of environmental concern has highlighted the need to switch from harmful vapor-compression-based cooling technologies to more friendly options. Solid-state cooling devices based on multicaloric materials are one promising option. Multicaloric materials undergo a reversible thermal change driven by more than one external energy source. One promising multicaloric materials family of interest is the intermetallic MTX compounds.

The MTX materials [where M = transition metal (Mn, Fe, Ni, Co); T = transition metal (Mn, Fe, Ni, Co); X = main group p-block element (Si, Ge, Sn, Al, Ga)] undergo a martensitic phase transition from the low temperature, martensitic Pnma (TiNiSi-type) phase to the high temperature, austenitic P6₃/mmc (Ni₂In-type) phase. Nominally, in the stoichiometric parent compound, MnNiSi, the martensitic phase transition (T_m) occurs at 1200 K. The Pnma phase is also ferromagnetic and has a Curie temperature (T_c) of 662 K. Through alloying with other transition metals and p-block elements, T_m and T_c can be tuned to coincide resulting in a hysteretic magnetostructural phase transformation that is first-order in thermodynamic character There is significant interest in identifying alloy compositions where the magnetostructural transition temperature (T_t) is observed near room temperature with minimal hysteresis so that devices can be built with maximum cooling efficiency.

Here, we develop a data-driven machine learning (ML) approach to predict the T_t and hysteresis of solid solutions of MTX alloys. We compiled a dataset of experimentally synthesized and characterized MTX solid solutions from surveying the published literature. We represented each composition using three sets of features: (1) compositions, (2) elemental descriptors, and (3) those generated from density functional theory calculations. Four separate ML methods were considered for regression (1 linear, 3 non-linear), and the best method was selected based on test set performance quantified by R² goodness-of-fit. The top-performing model was then validated experimentally on new solid solutions.

In addition to prediction, we also implemented a novel post hoc interpretability technique to understand model behavior. Typically, many ML algorithms are treated as black-box models. While ML-based approaches have been demonstrated to optimize materials design and discovery, the lack of model interpretability means that it is very difficult to extract valuable scientific insights from the trained models. To address this problem, we implement a common model interpretability technique, global feature importance, along with less common, local instance-level variable attribution techniques. The results show that the model prioritizes features that are known to impact the T_t in the system. The results of the local analysis also show that the variable attribution varies significantly for different compositions highlighting the limitations of global feature importance.

High-throughput experimentation today explores not only compositional space, but also process space. In our work, we use micro- to milli-second thermal processing to synthesize metastable phases in binary and ternary systems, and a combination of optical spectroscopy and synchrotron x-ray diffraction (XRD) to characterize them. However, post-processing analysis constrains understanding to the terminal structures only, without identifying transient structures that may develop during synthesis. The search for autonomous materials using ML and AI algorithms would be enhanced by incorporating time-resolved information regarding the phase evolution alongside the characterization of the final phases.

The unique experimental characteristics of the lateral-gradient Laser Spike Annealing technique allows this temporal structural evolution to be mapped onto a spatial dimension, enabling a wide range of characterization techniques. For initial studies, optical imaging has been used to track the phase formation sequence as a function of both time and temperature during lgLSA in materials such as Bi₂O₃ and Ga₂O_3. In single-composition Bi₂O₃, for example, initial nucleation of the δ phase occurs early during heating with transformation to the β phase also observed during the temperature ramp; in this system, no subsequent transformations have been identified during cooling to ambient conditions. Behaviors across a number of binary systems will also be discussed. Ultimately, these data can be integrated into closed-loop experiments to inform the autonomous search for new materials.

Dislocation boundaries dominate the multiscale structure and properties of metals, setting the complex dynamics that determine how they respond to external stresses or heating. Characterization of these intricate multiscale dynamics is extremely difficult at the single-dislocation level, due to experimental and analytical challenges. Dark-field X-ray microscopy (DFXM) was recently demonstrated to resolve the dynamics of populations of deep subsurface dislocations at the single-dislocation level, but challenges persist in assigning the positions and character. We have developed a suite of computer-vision methods to identify and track dislocations in DFXM movies collected on single-crystal aluminum during annealing. This talk will present our new methods and the opportunities they hold to quantify statistically significant defect dynamics in crystalline materials at the mesoscale.

Chiral nano- and micron-scale inorganic structures have potential in many technological fields, including biosensing, chiral catalysis, enantioselective separation, and quantum computing. While they have numerous unique physico-chemical properties, their ability to exhibit high chiroptical activity is the key characteristic that spurs research in this area. To date, many researchers are focused on the development of materials with hierarchical chirality since the combination of nano-, meso-, and micron-scale chiral elements in one structure can allow to rotate light polarization in a wide spectral range. However, top-down approaches, such as ion-beam lithography, extensively used for growing such helices, are limited for the fabrication of nanoscale elements and have a low mass production efficiency. Also, there is a need for novel methods for the structural characterization of chiral nano- and micron-scale inorganic structures.
Here we developed a facile bottom-up approach for the formation of microscale helices of bowtie shape. Depending on the enantiomer of chiral molecules used, bowties have a right- or left-handed twist with 100% enantiomeric excess. Transmission electron microscopy of bowties allowed to establish that bowties are self-assembled from nanoparticles with 3.4 nm diameter that are then arranged into stacked twisted nanosheets. The latter have a thickness of about 100 nm and a 50 nm separation between each other. Careful adjustment of synthesis conditions allows growing bowties with desired length, width, thickness, and twist, which, in turn, lead to the strong chiroptical activity that could be carefully tuned in a from UV-Vis to NIR spectral range. Computations of bowties chiroptical activity have shown that the circular dichroism and g-factor spectra blue-shift as the pitch of bowties decreases.
Varying synthesis conditions, we synthesized more than 600 different samples and obtained more than 1200 scanning electron microscopy (SEM) images to study their nucleation, growth, and final morphology. Analysis and characterization of SEM images can be significantly improved and fastened by using artificial intelligence approaches. We developed a convolutional neural network model for the recognition of bowties on SEM images and the determination of their handedness (right-handed, left-handed). Our next step is to advance the model and make it able to calculate the enantiomeric excess of bowties and their dimensions.
Also, to predict the optical properties based on the concentrations of chemicals used, we developed a multi-layer fully connected neural network. The approach that was chosen is to model target parameters independently to get better accuracy. The model has four hidden layers with the following number of neurons in each - 400, 2000, 1000, 400. The developed neural network allows to predict bowties parameters (length, width, thickness, and twist angle) and their optical properties (circular dichroism peak wavelength and g-factors) with good accuracy.

We believe that these results open new horizons for designing, characterization, and engineering of chiroptical materials and metamaterials with a strong and tunable polarization rotation essential for multiple emerging technologies.

Vapor condensation is of great significance to a plethora of applications such as power generation, thermal management, and water desalination. Over the past century, many hydrophobic coatings have been developed to promote dropwise condensation (DWC) on surfaces of interest. DWC on hydrophobic coatings is preferred as it has heat transfer coefficients (HTC) one order of magnitude higher than filmwise condensation (FWC) on an uncoated surface. Quantification of the heat transfer performance during condensation is normally done in vacuum systems to limit the presence of the non-condensable gases (NCG), and to enable high-fidelity temperature measurements for heat transfer quantification. Numerous difficulties exist with these classical techniques mainly related to the high uncertainties in the measured heat flux and calculated HTC when proper experimental conditions are not used. Here, we demonstrate a visualization-based method for rigorous and high-fidelity heat transfer characterization during DWC. State-of-the-art computer vision techniques are used to extract physical features which are used to predict the heat transfer performance of hydrophobic tubes undergoing DWC. We demonstrate the utility of our visual approach by using it to estimate the local axially-varying heat flux on a tube with spatially varying wettability without using any temperature measurement. Unlike classical methods, we demonstrate that our method has a fixed heat flux estimation uncertainty of 13% of the measured value. In contrast, we show that the measurement uncertainty of classical techniques is a function of sensor uncertainty, geometrical parameters such as length and diameter of the sample, and the heat flux itself, and is generally several times higher than the uncertainty associated with our vision-based method. Our visualization-based method does not require high-speed imaging and can be applied with regular 30 FPS imaging. Our computer-vision method demonstrates significant utility for conditions where conventional temperature-based methods fail such as DWC with short or small diameter tubing, experiments with laminar cooling flows, small heat flux condensation, or obtaining local heat flux measurements.

Accelerating the development of electrochemical devices is pivotal for the societal shift towards a sustainable and de-fossilized energy economy. Imaging is crucial to quantify structural characteristics of materials, as well as to detect and monitor structural changes during fabrication and normal operating cycles of the device (battery, fuel cell, electrolysis cell). The escalating use of imaging infrastructure in the energy materials domain drives a significant growth of data in terms of their amount and complexity. This leaves the applications of standard machinery for image processing often ad hoc, indiscriminate, and empirical which in turn makes the quantification metrics for analysis elusive. Moreover, the use of standard techniques for materials characterization is expensive, slow and often involves several pre-processing steps. In this contribution, we present a deep learning-based approach to automate the analysis of particle size distribution (PSD) from Transmission Electron micrographs (TEM) of carbon supported catalyst electrodes for Polymer Electrolyte Membrane (PEM) Fuel Cells or PEM electorlyzers. The catalyst layer components are fabricated based on catalyst nanoparticles (e.g. Pt) on a high-surface-area carbon support material. The automation was enabled via a multi-step approach based on supervised learning, by 1. manual annotation of TEM images [1], 2. training a U-Net model for instance segmentation [2] of catalyst nanoparticles, followed by 3. employing computer vision techniques for the diameter measurement of individual particles and 4. generating particle size distribution plots. Results on the validation set are in well agreement with PSD analysis by our experimental experts, presenting an efficient use case of ML-enabled materials characterization.

References:

[1] P. Bankhead, et al, Scientific Reports 2017, 7, 16878.
[2] Ronneberger, et al. In International Conference on Medical image computing and computer-assisted intervention, Springer, Cham, 2015, 234.

The ‘lifetime’ of materials is a central topic in the field of corrosion science and is a crucial consideration for any real-world use of emerging materials. Recent years have witnessed a surge in the development of low-dimensional nanomaterials (e.g. 2D materials) towards commercial applications, driven by discoveries of novel, record-breaking material properties. However, systematic studies on their ageing dynamics are comparatively rare and the fundamental processes that dictate their lifetime remain poorly understood [1].
Meanwhile, statistical variations are expected to significantly impact dynamic processes at the nanoscale, including ageing. Despite this, the collection of sufficiently large datasets that permit their statistical analysis remains challenging and is commonly not emphasised when studying the properties of low dimensional nanomaterials. Recent reports which combine high-throughput image processing workflows with large electron microscopy datasets of “static” nanomaterials, illustrate a promising approach for rapidly extracting quantitative datasets of statistically relevant sizes [2,3]. However, such methods have rarely been extended to studies of nanoscale dynamics, given the lack of suitable characterisation platforms that can probe nanomaterials in-situ, under realistic working conditions.
Here, we combine a custom in-situ environmental scanning electron microscope (ESEM), equipped with localised gas injection and heating capabilities, with automated image analysis to observe and collect statistically large datasets that quantify nanoscale dynamics. The formation of corrosion pits during the thermal oxidation of atomically-thin 2D materials is selected as a prototype system to demonstrate the capabilities of our workflow. By leveraging the ESEM’s large field of view and our high-throughput experimental design, the expansion dynamics of 10³ – 10⁴ corrosion pits are tracked and quantified simultaneously in each experiment. Our statistically relevant datasets of pit formation permit conventional corrosion theories to be statistically tested in the 2D limit. The collection and analysis of gigabyte to terabyte levels of kinetic data are semi-automated, with heavy reliance on developing suitable computer image recognition algorithms. Our research paradigm, in which the statistical analysis of large in-situ experimental data is emphasized, promises a more repeatable and reliable experimental analysis of nanoscale processes.

References
[1] Li et al., J. Mater. Chem. A, 2019, 7(9), 4291
[2] Lee et al., ACS Nano, 2020, 14(12), 17125
[3] Laramy et al., ACS Nano, 2015, 9(12), 12488

Artificial intelligence agents offer great opportunities for accelerating materials discovery by autonomous data analysis. With advanced methods of high-throughput experimentation producing increasingly larger datasets, autonomous analysis and decision making by AI is required in collaboration with the deep reasoning provided by expert scientists. While AI should be focused on autonomously analyzing the standard cases, expert scientist resources should be centered around the most interesting or challenging tasks. A persistent issue of AI models is associated with drawing conclusions from uncertain settings, e.g. when the information stored in the data does not allow for a definite decision or the training data does not support aberrations encountered in experimental data.

In this contribution, we demonstrate that variational autoencoders (VAE) provide a powerful toolset for the analysis of large X-ray diffraction datasets. Visualization of the latent vectors can reveal intrinsic features of the dataset that allow for an assessment of possible failures modes during a classification task. The representation can further be used for on-the-fly visualization of a dataset distribution across different structures. Here, latent space proximity can be used as a structural estimator. A strategy based on the evaluation of the reconstruction error and the clustering of latent variables for the detection of out-of-distribution data is reported and demonstrated on experimental data. By inclusion of experiment specific domain knowledge (e.g. chemical composition) the model can be tuned to produce valid solutions under awareness of constraints. Transferability to other experimental sources of spectral data is demonstrated. We envision the VAE working cooperatively together in a federation of specialized AI agents.

In this work, we introduced a shift-invariant variational autoencoder (shift-VAE), which is an unsupervised method developed for analyzing 1D spectra data disentangling the physically-relevant shifts from other latent variables. Using synthetic datasets, we show that the shift-VAE latent variables are linear functions of the ground truth parameters, indicating the strength of shift-VAE in disentangling physically-relevant variables in a model-free unsupervised manner. The application of shift-VAE is illustrated for band-excitation piezoresponse force microscopy (BE-PFM) data. From two BE-PFM data sets, we consistently observed that shift-VAE successfully disentangles ferroelectric polarization and crosstalk-induced resonance frequency peak shift from BE curves. In addition, shift-VAE also shows the strength in denoising raw curves without sacrificing real response. Overall, shift-VAE proves to be a powerful technique for learning spectra, which should have very broad applications in many fields. This approach can also be extended to analysis of X-ray diffraction, photoluminescence, Raman spectra, and other data sets.

Acknowledgments: This work is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences Energy Frontier Research Centers program under Award Number DE-SC0021118. This work is conducted at the Center for Nanophase Materials Sciences, a US Department of Energy Office of Science User Facility.

Electron channeling contrast imaging (ECCI) is a non-destructive scanning electron microscope (SEM) surface imaging technique that is sensitive to distortion of planes in crystal lattices. It detects the diffraction intensity of electrons incident on a sample, where the angles between detector, sample normal, and the SEM electron beam, are carefully chosen so that the orientation of diffracting planes with respect to the incident electron beam is approaching their Bragg angle, creating strong contrast in the image. Local lattice strain induced deviations to the crystallographic orientation or lattice constant will then affect the intensity of the back scattered electrons, and produce a variation in contrast in the resulting surface image that indicates the presence of a defect. These images are over a relatively wide area (~100 μm²), with a spatial resolution of ~20 nm, and allow us to detect and classify threading dislocations, grain boundaries, and other defects at semiconductor surfaces [1].

Of particular interest is the ability to conduct a quantitative analysis of threading dislocations (TDs) present in a GaN thin film, which manifest an ECC image as spots with a dipolar black-white (B-W) contrast. One obstacle to this is the need for the painstaking marking and manual analysis of each individual experimental image by a skilled researcher. Bulk epitaxial GaN thin films often have TD area densities of 10⁸ to 10⁹ cm^-2 [2], adding to the scale of the problem. This creates a need for a fast, automated, and relatively accurate method of extracting defect data from multiple images.

Supervised machine learning techniques have been proven in recent years to be able to achieve a high degree of success in detecting predetermined features, particularly in the medical imaging field [3]. We have utilized previously documented techniques for quickly identifying TDs on III-Nitrides [1,4] to build a training dataset for a U-Net based deep fully convolutional network architecture [5,6]. This includes images of defects both with a strong B-W contrast as well as dislocations at close to the invisibility criteria which are faintly resolved.

The neural network that is trained on this set can then relatively reliably detect the positions of TDs in ECCI of GaN with an accuracy on the order of 75-90% for independent testing data (i.e. marked images unseen by the network during training), with the main source of errors being false positives due to strong non-TD features, and false negatives due to either TD faintness at certain diffraction conditions, or overlapping, closely spaced TDs affecting the predictive binary mask. The recognition process for a 2576×1936px image currently takes around 90 seconds when using a modern 8-core consumer grade CPU. For our current dataset size the network can be trained to convergence in 30±10 minutes when using a single NVIDIA V100 GPU. Fig. 1 shows the marked locations for a cluster of dislocations in GaN.

[1] Trager-Cowan, C. et al. Semicond. Sci. Technol. 35, 054001 (2020).
[2] Kapolnek, D. et al. Appl. Phys. Lett. 67, 1541–1543 (1995).
[3] Arganda-Carreras, I. et al. Front. Neuroanat. 9, (2015).
[4] Naresh-Kumar, G. et al. Phys. Rev. Lett. 108, 135503 (2012).
[5] Wiehman, S. & Villiers, H. de. Semantic segmentation of bioimages using convolutional neural networks. in 2016 International Joint Conference on Neural Networks (IJCNN) 624–631 (2016).
[6] Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. in MICCAI 2015 234–241 (Springer, 2015). doi:10.1007/978-3-319-24574-4_28.

Characterizing crystallographic quantities in crystalline materials—including crystal structure, grain orientation, and defects character—is the realm of diffraction-based techniques. We have developed a simple and affordable method—which we call directional reflectance microscopy (DRM)—to capture this information using an optical microscope. DRM relies on producing orientation-dependent features on the surface of crystalline solids and measuring the reflectance they generate when illuminated from different angles. Point-by-point indexing of optical reflectance across the sample surface enables mapping the underlying crystallographic orientation in a similar fashion to electron backscatter diffraction. Here we discuss the advantages of indexing the optical signal using a forward model approach. Our method consists of producing a dictionary of reflectance patterns generated by a discrete sampling of orientation space using a physics-based reflection model. We then compare these simulated patterns with the experimental one to find the best match, which is selected as the nominal orientation. This dictionary indexing method overcomes limitations that are inherent to previous approaches, which relied on fitting good quality patterns. We demonstrate the improved accuracy of the forward-model method on a centimeter-scale aluminum sample.

Optical methods are appealing for application to microstructure analysis due to their simplicity, direct observation of details, and availability of versatile elemental base. However, diffraction of light blurs the collected images and sets the fundamental limit on the achievable resolution [1]. Recently, a number of super-resolving techniques based on nano-particles fluorescence were developed for overcoming the classical resolution limit, for instance stimulated-emission depletion (STED) microscopy [2], photoactivated localization microscopy (PALM) [3], and structured illumination microscopy (SIM) [4]. Super-resolution optical fluctuation imaging (SOFI) is also among promising quantum-inspired techniques [5]. The resolution enhancement in SOFI stems from special statistical analysis of the intensity correlations of randomly fluctuating fluorescent sources: quantum dots [5], organic dyes [6], or other fluorescent particles. Namely, a cumulant image is formed instead of usual intensity images and provides a narrower point-spread function (weaker blurring of the image). The approach is applicable for both investigation of inorganic samples and imaging of living cells and subcellular structures.

Initially, SOFI was believed to provide infinite growth of resolution with the increase of cumulants order. Recently, it has been shown that the inevitable statistical noise imposes fundamental limits on the achievable resolution; the optimal cumulant order is finite (and depends on the imaged object); and combined analysis of several cumulant orders may be profitable [7,8]. The research was based on maximization of Fisher information and included some simplifying limitations: either just two point emitters were considered [8], or the statistical switching noise of the fluorescent sources themselves was neglected [7]. In the current contribution, we follow that line of Fisher information-based research, but consider a general case of several emitters, finite observation time, and realistic fluctuations model.

A measurement in SOFI is fast and provides an image with a large number of pixels (up to 1 million). To benefit from those advantages of the technique (especially, when fast living-cell processes are analyzed), one needs to be able to perform fast on-the-fly processing of the acquired data. For that reason, it is desirable to know and process the optimal cumulant order soon after the dataset is collected. While Fisher information approach is much faster and more reliable than direct Monte-Carlo simulations, it is still computationally expensive for taking such real-time decisions. Here, we show that machine learning is capable of solving that issue. An artificial neural network (a 5-layer perceptron) takes an intensity image of the object (easily generated from a SOFI dataset) as an input and decides, which cumulant order or combination of cumulant orders would yield the maximal information gain for the same computational resources. The training datasets are prepared with the help of Fisher information approach. The preliminary results show that the probability of taking the correct decision among 4 options with approximately equal probabilities exceeds 80%. The proposed tandem approach of Fisher information and machine learning is universal and applicable to many other techniques for optimization of on-the-fly analysis or taking real-time decisions which measurement to perform next.

1. Abbe, E., Archiv für mikroskopische Anatomie, 1873, 9, 413.
2. Hell, S. W. and Wichmann, J., Opt. Lett., 1994, 19, 780.
3. Hess, S. T. et al., Biophys. J., 2006, 91, 4258.
4. Classen, A. et al., Optica, 2017, 4, 580.
5. Dertinger, T. et al., Proc. Natl. Acad. Sci. U.S.A., 2009, 106, 22287.
6. Dertinger, T. et al., M., Angew. Chem. Int. Ed., 2010, 49, 9441.
7. Vlasenko, S. et al., Phys. Rev. A, 2020, 102, 063507.
8. Kurdzialek, S. and Demkowicz-Dobrzanski, R., J. Opt., 2021, 23, 075701.

Multicrystalline materials can be relatively easily fabricated with a large crystal growth furnace leading to cost-effective yield, whereas non-uniformity of crystal properties is a problem for large ingots. Therefore, it is important to extract the ingot-scaled features of the multicrystalline structure. The multicrystalline structure grown by the unidirectional solidification method develops from the nucleus at the bottom of the crucible into a columnar structure, and successive twinning inside the columnar structure causes various grain boundaries. The orientation relationship between the individual nuclei is random, while the orientation relationship due to twinning is crystallographically determined, which consequently contains information about the grain development relationship. We have developed a method to analyze multicrystalline structures using network graphs where edges represent successive twinning from crystal orientations of grains in 30×40 mm² multicrystalline silicon, obtained by electron backscatter diffraction pattern (EBSD) method [1-3]. In this study, we attempted to extract ingot-scaled features of multicrystalline structures by performing network analysis on crystal orientations obtained by the Laue scanner method [4] on 156×156 mm² substrates sliced from the bottom, middle, and top parts of a multicrystalline silicon ingot. The network analysis was performed by coding with a Python package NetworkX [5].
The crystal orientations of 1918, 1490, and 1390 grains (at the center of gravity of each grain) were obtained from the bottom, center, and top of the ingot, respectively. The orientation relationship between grains was attributed to an nth-order (0≤n≤3) twinning relationship (Σ3ⁿ) that gives the minimum residual angle θ. A graph G was generated with edges for the twinning relation satisfying θ/θ_th+d/d_th≤1, where θ_th=2° and d_th=40 mm were thresholds for θ and the distance d between grains. Connected subgraphs contained four or more nodes were plotted on the grain boundary mapping extracted from the optical image, where grains in the same subgraph are considered to have grown from the same nucleus. As the crystal growth proceeded, the number of nodes in a subgraph and the area occupied by a subgraph with a large number of nodes also tended to increase. The analysis using network graphs made it possible to extract ingot-scaled features of multicrystalline structures.

References
[1] T. Kojima et al., Abstr. 66th JSAP Spring Meeting, 9a-W611-3, 2019.
[2] T. Kojima et al., Abstr. 80th JSAP Autumn Meeting, 20a-E314-4, 2019.
[3] T. Kojima et al., Abstr. 67th JSAP Spring Meeting, 14a-A205-3, 2020.
[4] T. Lehman et al., Acta Materialia 69, 1 (2014).
[5] A. A. Hagberg et al., Proc. SciPy2008, pp. 11-15.

Materials informatics (MI) is a data-driven approach to discover novel materials. In this approach, structure information such as crystal structure is one of important information because it has strong relationship to inorganic materials’ properties. In conventional MI approaches, crystal structure is often represented in form of a categorical variable of corresponding space group, but this limits applicable machine learning (ML) methods compared to a continuous variable. Therefore, other representations of structure, such as crystal graph [1], have been proposed to convert structure information into quasi-continuous variable. To apply this kind of methods, crystal structure must be known, this means that structure of sample is analyzed by experts in advance. However, this is time-consuming and will be bottleneck in data-driven materials discovery.
We propose to use ML results of direct application to measurement data as crystal structure representations instead of experts-analyzed data. To proof our concept, we used non-negative matrix factorization (NMF) [2]. NMF approximates non-negative matrix V, a set of measurement data, with multiplication of two non-negative matrices W and H. This method is already studied with one-dimensional X-ray diffraction (1D-XRD) patterns from single composition spread [3]. We studied under more practical situation: we used 2D-XRD patterns from multiple samples with variety in substrate types and fabrication conditions. 2D-XRD is convenient in thin film research compared for 1D-XRD because it includes reflections from various out of planes.
2D-XRD patterns of In-Ga-O thin films were studied. We first analyzed NMF results in terms of generative model to confirm that NMF converts 2D-XRD patterns into variables which follows a continuous probability distribution. NMF results can be recognized that W is a set of factor vectors and H is that of feature vectors. The feature vectors were evaluated whether they represent propensities in original 2D-XRD patterns. We also compared results of samples fabricated by pulsed laser decomposition and those by sputtering. The detail will be discussed in the presentation.

References
[1] T. Xie and J. C. Grossman, “Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties,” Physical Review Letters, vol. 120, no. 14, p. 145301, Apr. 2018, doi: 10.1103/PhysRevLett.120.145301.
[2] D. D. Lee and H. S. Seung, “Learning the parts of objects by non-negative matrix factorization,” Nature, vol. 401, no. 6755, pp. 788–791, Oct. 1999, doi: 10.1038/44565.
[3] C. J. Long, D. Bunker, X. Li, V. L. Karen, and I. Takeuchi, “Rapid identification of structural phases in combinatorial thin-film libraries using x-ray diffraction and non-negative matrix factorization,” Review of Scientific Instruments, vol. 80, no. 10, p. 103902, Oct. 2009, doi: 10.1063/1.3216809.

Atomistic modeling of chemically reactive systems has traditionally relied on either expensive ab initio methods or bond-order force fields requiring arduous parametrization. Here, we introduce FLARE++, a Bayesian active learning method for training reactive many-body force fields “on the fly” during molecular dynamics (MD) simulations. At each timestep, the predictive uncertainties of a sparse Gaussian process (SGP) are evaluated to automatically determine whether additional ab initio data are needed. The resulting SGP is mapped onto a polynomial model whose prediction cost is independent of the training set size. Using this method, we perform large-scale MD simulations of a prototypical system in heterogeneous catalysis---H2 chemisorption on Pt(111)---at chemical accuracy. The model, trained within three days of wall time, performs at twice the simulation speed of an available ReaxFF model while maintaining ab initio accuracy to a much higher fidelity. Our method enables efficient automated training of fast and accurate reactive force fields for chemically complex systems.

2000 series aluminum alloys or Al-Cu-Mg series are high-strength alloys, but it is known that their strength decreases rapidly at high temperatures above 150°C. Therefore, improving the strength at high temperatures is essential for industrial applications, especially automotive and aircraft applications. In this study, we perform a Bayesian inverse design of high-strength aluminum alloys at high temperatures.
We build a neural network to predict mechanical properties from process parameters, including additive elements and heat treatment. By choosing the appropriate activation function and features, neural networks show high prediction accuracy even at high temperatures. Specifically, we show that the hyperbolic tangent or sigmoidal activation functions well reproduce the rapid decrease in strength of aluminum alloys at high temperatures. We also show that by treating the holding temperature and holding time as input nodes in the neural network, we can predict mechanical properties measured at various temperatures. The neural network constructed with this approach can even perform extrapolation prediction, learning the mechanical properties at lower temperatures and predicting at higher temperatures.
In this study, Bayesian inference of neural network parameters is performed using the replica-exchange Monte Carlo method, an extended Markov chain Monte Carlo (MCMC) method. This approach allows us to evaluate the probability distribution and uncertainty of mechanical property predictions. By maximizing the probability of satisfying the desired properties, we explore combinations of process parameters for designing high-strength aluminum alloys at high temperatures.
The thermodynamic calculations are also performed using the Thermo-Calc software to analyze the strengthening mechanism of the alloy designs proposed by the neural network. We investigate how a fine grain structure caused by the Mn addition and the competition between the Al-Cu, Al-Cu-Mg, and Mg-Si compounds formation affect the alloy strength.

Progress in any scientific research area in physical and life sciences is primarily driven by assessments produced by numerous experimental and simulation-based techniques. Computational approaches ranging from lattice models and molecular dynamics to advanced density functional theory and quantum Monte Carlo methods provides a wealth of information on electronic, magnetic, thermodynamics properties, phase transitions of materials, to name a few. Electron and scanning probe microscopies and neutron and X-ray scattering methods provide huge volumes of structural and functional data leading to fundamental insights into structural properties and high-temperature superconductivity, visualization of the light elements, and many others.
Both theoretical and experimental aspects of research require search over broad parameter spaces to discover optimal experimental conditions, regions of interest in the image space, or parameter space of computational models. The direct grid search of the parameter space tends to be exceedingly time consuming, leading towards advancing strategies to balance between exploration of unknown parameter spaces and exploitation towards required performance metrics. The classical approach for this is Gaussian process (GP)-based Bayesian optimization (BO).
However, classical BO strategies do not readily allow for the incorporation of the known physical behaviors or past knowledge.
Here, we explore hybrid approaches combining unstructured (non-parametric) and structured (semi-parametric and parametric) BO approaches. The former utilizes Gaussian process (GP) as the surrogate model while the latter one uses GP augmented with a probabilistic model of the system's physical behavior based on partially known analytical, hierarchical, or numerical models. Hence, this approach seeks to guide the exploration of parameter space when refining the physical understandings describing functional behavior over this space and using combined aleatoric and estimated epistemic uncertainty to guide the exploration. This approach is demonstrated for Ising model followed by looking at magnetism in 2D materials. We further discuss extending this capability towards perform automated experiment in the context of automated materials synthesis, scanning probe microscopy, and electron microscopy by injecting prior knowledge in the form of physical models and past data in this BO framework.

This effort (machine learning) is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Office of Basic Energy Sciences Data, Artificial Intelligence and Machine Learning at DOE Scientific User Facilities (A.G., S.V.K.) and was also supported (STEM experiment) by the DOE, Office of Science, Basic Energy Sciences (BES), Materials Sciences and Engineering Division (O.D.), and was performed and partially supported (M.Z.) at the Oak Ridge National Laboratory’s Center for Nanophase Materials Sciences (CNMS), a
DOE Office of Science User Facility.

Redox flow batteries (RFBs) offer the capability to store large amounts of energy cheaply and efficiently, however, there is a need for fast and accurate models of the charge-discharge curve of a RFB to potentially improve the battery capacity and performance. We develop a multifidelity model for predicting the charge-discharge curve of a RFB. In the multifidelity model, we use the Physics-informed CoKriging (CoPhIK) machine learning method trained on experimental data and constrained by the so-called ``zero-dimensional'' physics-based model. In this talk, we demonstrate that the CoPhIK model shows good agreement with experimental results and significant improvements over existing zero-dimensional models. We show that the proposed model is robust as it is not sensitive to the input parameters in the zero-dimensional model. We also show that only a small amount of high-fidelity experimental datasets are needed for accurate predictions for the range of considered input parameters, which include current density, flow rate, and initial concentrations.

Granular materials are widely used in various fields such as sintering process, tablet manufacturing, and geotechnical engineering. To predict its constitutive behavior in the finite element method (FEM), the modified Drucker-Prager/Cap model (DPC model) is commonly adopted. However, the conventional method requires multiple triaxial compression experiments and even heuristic approaches to determine a great number of parameters in the DPC model. In this study, we propose an alternative method for the DPC model that utilizes the entire stress-strain curves from triaxial compression testing and the Bayesian optimization algorithm to extract the parameters. The method was applied to acquire the constitutive relation of FE simulation for unsaturated sand-bentonite mixtures which successfully predicted experimental depth-sensing indentation results. Since our method dramatically reduces the number of experiments and even eliminates heuristic determination of some parameters, we believe that it will facilitate efficient exploration of the mechanical behavior of granular materials.

Single-atom catalysts (SACs) have become the forefront of energy conversion studies, but unfortunately, the origin of their activity and the interpretation of the synchrotron spectrograms of these materials remain ambiguous. Here, systematic density functional theory computations reveal that the edge sites—zigzag and armchair—are responsible for the activity of the graphene-based Co (cobalt) SACs toward hydrogen evolution reaction (HER). Then, edge-rich (E)-Co single atoms (SAs) were rationally synthesized guided by theoretical results. Supervised learning techniques are applied to interpret the measured synchrotron spectrum of E-Co SAs. The obtained local environments of Co SAs, 65.49% of Co-4N-plane, 13.64% in Co-2N-armchair, and 20.86% in Co-2N-zigzag, are consistent with Athena fitting. Remarkably, E-Co SAs show even better HER electrocatalytic performance than commercial Pt/C at high current density. Using the joint effort of theoretical modeling, thorough characterization of the catalysts aided by supervised learning, and catalytic performance evaluations, this study not only uncovers the activity origin of Co SACs for HER but also lays the cornerstone for the rational design and structural analysis of nanocatalysts.

Graph neural networks (GNN) have been shown to provide much improved performance for representing and modeling atomistic materials compared with descriptor-based machine-learning models. While most existing GNN models for atomistic predictions are based on atomic distance information, they do not explicitly incorporate bond angles, which are critical for distinguishing many atomic structures. Furthermore, many material properties are known to be sensitive to slight changes in bond angles. We develop Atomistic Line Graph Neural Network (ALIGNN) using a GNN architecture that performs message passing on both the interatomic bond graph and its line graph corresponding to bond angles. We demonstrate that angle information can be explicitly and efficiently included for materials to provide much improved performance. We train 44 models for predicting several solid-state material properties available in the JARVIS-DFT and materials-project databases. ALIGNN can outperform some of the previously known GNN models by up to 43.8 %.

Salt hydrates demonstrate promise as heat storage materials as they possess high energy densities and reversibility at moderate temperatures. Despite their promise, a great number of salt hydrate compositions have not been explored. The goal of this work is to identify new salt hydrates that can outperform known materials in terms of energy density and are predicted to be thermodynamically stable. Around ten thousand hypothetical salt hydrates were generated by substituting cations and anions into salt hydrate crystal structures mined from the Inorganic Crystal Structure Database. These hypothetical hydrates were characterized according to their enthalpy of dehydration (from which stability, energy density, and heat-storing temperature are derived) using high-throughput density functional theory calculations. Machine learning models were then trained on this large dataset. Many different compositional and structural feature sets from the Matminer library were attempted, as well as several different machine learning algorithms, such as random forests, support vector machines, and neural networks. Several models were highlighted, either for their ability to predict the heat storage characteristics of other hypothetical salt hydrates or their identification of property-performance relationships in salt hydrates. In addition to this, several promising hypothetical hydrates were identified with higher energy densities than experimentally-known salt hydrates.

In recent years, materials informatics (MI), which is a fusion of materials science and information science, has been attracting attention. Although MI is becoming a mainstream research field such as magnetic materials and thermoelectric materials, there are few reports on MI in materials research in nuclear fields. In this study, we propose a method to comprehensively explore uranium compounds with high thermal conductivity by using machine learning in order to apply MI to nuclear fuel development in the nuclear field.
We constructed a machine learning model using Starrydata, a database of experimental properties developed by our research group. The features were designed using the basic physical properties based on chemical composition, and the target variable is thermal conductivity. The model was trained using Random Forest, and the prediction accuracy of 93% was obtained. Using the trained machine learning model, the thermal conductivity of uranium compounds that could exist in the world was comprehensively predicted. Based on the predicted thermal conductivities, 20 uranium compounds were extracted from the top and bottom of the list, respectively. Uranium compounds experimentally synthesizable are going to be synthesize, and the relevance of the machine learning model is going to evaluate by comparing the measured values with the predicted values. Based on the evaluation, we are going to propose a new uranium compound with high thermal conductivity.

Although by now machine learning is commonplace within virtually all sectors of materials science, the use of reinforcement learning (RL) agents is only now gaining traction. This is partly because of the extensive compute requirements for RL, and the paucity of scalable algorithms. Recent advances in the field have enabled the training of agents to tackle previously difficult challenges in fields such as online gameplay and board games, but their application to the physical sciences has been very limited. Here, we explored the use of RL agents for two particular tasks: one of atomic fabrication by an electron beam, and another of design of ferroelectric materials by defect engineering.

For both situations, we created a virtual environment in which the RL agent can interact and learn by trial and error, and designed reward functions to achieve specific gains. We then trained agents using both policy gradient and implicit (Q learning) methods, in a model-free setting. For the atomic fabrication case, we employed an environment utilizing molecular dynamics simulations, and trained agents to move multiple Si dopants towards each other on a graphene lattice by specifying quantities of momentum transfer at different steps during the episode. For the second case, we explored RL agents for moving and manipulating defects in a ferroelectric sample, simulated via a discrete Landau model, to affect the type and quantity of topological defects generated. We argue that not only are these agents capable of discovering policies that are novel, but also that the policies encode important information regarding the dynamics of the system, and thus inspection of policies can be an important source of physics knowledge discovery. This work was conducted at and supported by the Center for Nanophase Materials Sciences, a US DOE Office of Science User Facility.

High performing devices often rely on multiphase nanostructures with complex interfacial effects. Transmission electron microscopy (TEM) has proven indispensable to understand these effects and reveal the structure-property-relationship of these devices. One example are bulk heterojunctions (BHJs) for organic solar cells, where the excitons are split at the interface between donor and acceptor domains.¹ A precise segmentation of the donor and acceptor domains is crucial for understanding the process of charge separation. However, low contrast between chemically similar donor and acceptor materials makes classical imaging challenging and does not allow for molecular resolution at the interface. The combination of low loss energy loss signal based hyperspectral imaging and machine learning has proven to efficiently overcome this challenge.² However, the spectral difference for modern non-fullerene acceptor (NFA) systems is weak.³ Thus, the segmentation becomes obscured by experimental influences (such as noise, signal delocalization, multiple scattering and non-isochromaticity of the imaging tool). More sophisticated approaches are needed to obtain molecular resolution for non-fullerene acceptor BHJs.
Unsupervised segmentation and spectral unmixing of the hyperspectral datasets are typically approached by linear dimensionality reduction alone — e.g. principal component analysis (PCA), independent component analysis (ICA) and non-negative matrix factorization (NMF)—, by clustering algorithms alone, or the combination of both.^4,5 For the latter, the influence of linear dimensionality reduction on the cluster result is not clear, especially in the context of the given experimental influences. Furthermore, unequally performing cluster algorithms introduce non-negligible uncertainties.
We use an ensemble clustering approach to obtain a more robust unsupervised segmentation of hyperspectral low loss TEM datasets. Furthermore, we systematically investigate experimental influences as well as the effects of linear and nonlinear dimensionality reduction on the clustering results. Nonlinear dimensionality reduction (manifold learning) improves the tolerance to experimental restrictions from multiple scattering and non-isochromaticity. The combination of the manifold learning algorithm UMAP⁶ and ensemble clustering enables segmentations with an unrivaled tolerance against noise.
We applied this combination of manifold learning and ensemble clustering to hyperspectral images in the TEM low loss regime from PBDB-T:ITIC non-fullerene acceptor BHJs. Segmentations at molecular resolution allowed us to exclude a mixed materials phase at the NFA donor acceptor interface. This contrasts with their fullerene acceptor predecessors and sheds new light on the charge separation in modern BHJs. The combination of TEM based hyperspectral imaging and the machine learning workflow with UMAP and ensemble clustering can be applied to other low contrast material combinations. The machine learning workflow itself is unsupervised, fully automated⁷ and has already been successfully applied to non-TEM hyperspectral datasets.

1. Liu, Q. et al. Sci. Bull. 65, 272–275 (2020).
2. Pfannmöller, M. et al. Nano Lett. 11, 3099–3107 (2011).
3. Köntges, W. et al. Energy Environ. Sci. 13, 1259–1268 (2020).
4. de la Peña, F. et al. Ultramicroscopy 111, 169–176 (2011).
5. Torruella, P. et al. Ultramicroscopy 185, 42–48 (2018).
6. McInnes, L. & Healy, J. ArXiv180203426 Cs Stat (2018).
7. Potapov, P. & Lubk, A. Adv. Struct. Chem. Imaging 5, 4 (2019).

The negative Poisson’s ratio (NPR) is a novel property of materials, which enhances the mechanical feature and creates a wide range of application prospects in lots of fields, such as aerospace, electronics, medicine, etc. Fundamental understanding on the mechanism underlying NPR plays an important role in designing advanced mechanical functional materials. However, with different methods used, the origin of NPR is found different and conflicting with each other, for instance, in the representative graphene. In this study, based on machine learning technique, we constructed a moment tensor potential (MTP) for molecular dynamics (MD) simulations of graphene. By analyzing the evolution of key geometries, the increase of bond angle is found to be responsible for the NPR of graphene instead of bond length. The results on the origin of NPR are well consistent with the start-of-art first-principles, which amend the results from MD simulations using classic empirical potentials. Our study facilitates the understanding on the origin of NPR of graphene and paves the way to improve the accuracy of MD simulations being comparable to first-principle calculations. Our study would also promote the applications of machine learning interatomic potentials in multiscale simulations of functional materials.

Disordered multicomponent systems occupying the mostly uncharted centers of phase diagrams, have been studied for the last two decades as innovative materials. By maximizing their entropy (configurational and/or vibrational) and stabilizing (near) equimolar mixtures, it has been possible to achieve robust systems with promising applications. Still, effective computational discovery of useful materials remains challenged by the immense number of configurations: the synthesizability of high-entropy ceramics is typically assessed using ideal entropy along with the formation enthalpies from density functional theory, with simplified descriptors or machine learning methods. With respect to vibrations — even if they may have significant impact on phase stability — the contributions are drastically approximated to reduce the high computational cost, or often avoided with the hope of them being negligible, due to the technical difficulties posed in calculating them for disordered systems. As such, in high-entropy ceramics, the vibrational contributions remain a matter of conjecture — their inclusion left to researchers’ guesswork. In this presentation we illustrate the problem, its implications, and suggest simple solutions.

Grain boundaries (GBs) in multicrystalline Si (mc-Si) are still of interest in research because mc-Si will maintain a certain share in the substrate for solar cells to consume low-grade Si raw materials that cannot be used for Czochralski Si in addition to its importance as fundamental physics. One of the research issues is a quantitative evaluation of spatially local electrical activities to reveal the influence of local structures such as atomically reconstructed structures at curved GBs. Measurable spatial images of electrical activities, such as photoluminescence (PL) and electron beam induced current (EBIC) images, provide qualitative distributions, however, quantitative analysis of the images, in which parameter fitting is required, needs a large calculation cost.
In this study, we developed a fast and continuous evaluation method of carrier recombination velocity (V_GB) along a GB from a photoluminescence image using carrier simulation and its machine learning. A dataset of 3000 PL profiles was made through PL image simulation, which was developed incorporating the influence of inclination angle of GB (θ) [1], using random values of parameters of V_GB, θ, and carrier diffusion length in left and right crystal grains (L_left and L_right). A fully connected feedforward neural network model was constructed with the inputs of the PL profile and θ, which is determined by measuring the position shift of the GB between the front and rear surface [1], and the outputs of V_GB, L_left, and L_right. The coefficient of determination and the root mean squared error of the trained model for V_log, which is the common logarithm of V_GB, for the test dataset were 0.97 and 0.24, respectively. This prediction error was sufficiently low for the practical estimation of V_GB. The calculation time for the prediction is less than 0.01 seconds. By utilizing this fast prediction method, continuous evaluation of V_GB along a GB was demonstrated. V_GB was clearly visualized, and small fluctuations were detected quantitatively. V_GB increases around the curved region and fluctuates even in straight regions. This quantitative and high spatial resolution evaluation of the property will lead to a deep and quantitative understanding of defects by combining it with microscopic structural and computational analysis.
[1] K. Mitamura, K. Kutsukake, T. Kojima, and N. Usami, Journal of Applied Physics 128, 125103 (2020).

Field-induced domain wall dynamics in ferroelectric materials underpins multiple applications ranging from actuators to information technology devices and necessitates a quantitative description of the associated mechanisms including giant electromechanical couplings, controlled non-linearities, or low coercive voltages. While the advances in dynamic Piezoresponse Force Microscopy measurements over the last two decades have rendered visualization of polarization dynamics relatively straightforward, the associated insights into the local mechanisms have been elusive. Here we using a workflow combining deep learning-based segmentation of the domain structures with non-linear dimensionality reduction using multilayer rotationally-invariant autoencoders (rVAE). A Canny filter and ResHED network were utilized to detect domain walls based on the PFM images. Then, sub-images centered at 180^o domain walls were used for rVAE analysis to encode information of domain switch and wall dynamics. Through analyzing the time dynamics of domain switch and interactions between ferroelastic and ferroelectric domain walls by rVAE, we discovered new factors affecting the ferroelectric wall dynamics and pinning mechanism. We believe that the approach developed in this work will be universally useful for time-dependent dynamics of complex materials.

Acknowledgments: This work is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences Energy Frontier Research Centers program under Award Number DE-SC0021118. This work is conducted at the Center for Nanophase Materials Sciences, a US Department of Energy Office of Science User Facility.

We present one of the targeted analysis pipelines, 4D>Crystal, in an aim to generate a series of reduced crystalline diffraction vectors from a given combination of STEM probe in vacuum with typical 4D-STEM diffraction dataset. 4D>Crystal pipeline is composed of three main workflows namely – 1) crystal selection and data scraping (4D-MANIPULATT), 2) training data generation (4D-MAKE) and 3) Bragg disk detection network training and evaluation (4D-DISK). We use python-based pymatgen package and the materials project (MP) API to fetch crystal structures from more than 120,000 materials present in the MP database. Additionally, AFLOW library is levaraged to fetch structure prototypes from over 250 different structure types of crystals based on their structural similarity. Finally, we select a fraction of structures from the MP database, resembling each structure prototype, following the square-root distribution law. Following the crystal structure selection, we employ supercell manipulation algorithm (4D-MANIPULATT) to manipulate unit cell structures to create randomly rotated supercells. Finally, simulated electron diffraction data (CBED) and the simulated atomic potential for each of these supercells are simulated leveraging the Multislice algorithm implemented. For the deep learning training to predict Bragg disks locations and their intensities we implement a modified complex U-Net type encoder-decoder based network. Complex convolution based network in this work showed improved performance of disk predictions for overlapped diffraction disks at larger thickness.

The development of new materials requires many steps, such as sample preparation, X-ray diffraction measurement and analysis, property measurement, and performance evaluation, and requires an enormous amount of time. In recent years, high-throughput material synthesis and high-speed X-ray diffraction measurements have made sample preparation and X-ray diffraction measurements more efficient. However, the analysis of crystal structures has not yet become efficient. In this study, we constructed an optimal feature vector based on the X-ray diffraction pattern to predict the crystal structure with high accuracy, and also predicted the mechanical properties that are highly related to the crystal structure.
Various feature vectors based on the intensity, angle, and number of peaks in the X-ray diffraction pattern were constructed and their effects on the prediction accuracy were clarified. Since the number of diffraction peaks varied greatly among the different crystal systems, it was confirmed that the number of peaks made a difference in the prediction accuracy. The prediction accuracy for Cubic among the seven crystal systems was about 97% at maximum. On the other hand, the prediction accuracies of Monoclinic and Triclinic were lower than that of the other crystal systems due to the similarity of their X-ray diffraction patterns. By constructing an optimal feature vector that can predict crystal structures with high accuracy, highly accurate prediction of mechanical properties such as volume modulus can also be expected.

Materials informatics applied to a range of advanced materials applications can identify a range of compositions of complex materials (having multiple cations and anions) within which combinatorial synthesis is needed to obtain materials with ideal compositions for optimized properties. Combinatorial synthesis via a continuous composition spread is an excellent route to develop thin-film libraries from a single experiment on a single wafer, which is both time- and cost-efficient for developing new materials with ideal compositions and having optimized properties for next-generation applications. A reliable, high-throughput, in-situ characterization analysis method is required to meet the crucial need to screen materials libraries. Creating libraries of functional multi-component complex oxide films requires excellent control over the synthesis parameters combined with high-throughput analytical feedback across a given substrate area. Here, we report on the use of a newly developed in-situ compositional analysis technique, called low-angle x-ray spectroscopy (LAXS), to map the chemical composition profile of combinatorial heteroepitaxial complex oxide film deposited in a continuous composition spread using pulsed laser molecular beam epitaxy (MBE). This state-of-the-art combinatorial growth system facilitates a fully synchronized four-axis mechanical substrate stage without shadow masks, alternating acquisition of chemical compositional data using LAXS at various different positions on the ~41mm x 41mm range and sequential deposition of multilayers of SrTiO₃ and SrTi_0.8Ru_0.2O₃ on a 2-inch (50.8mm) LaAlO₃ wafer in a single growth. Reflection high-energy electron diffraction (RHEED) and x-ray diffraction (XRD) are used to confirm the epitaxy of the combinatorial oxide films, SrTi_1-xRu_xO_3-δ, deposited across the wafer. Rutherford backscattering spectrometry (RBS) is used to calibrate and validate the measurement data produced by the LAXS on the composition spread across the substrate. The calibrated LAXS data exhibits a gradual composition elevation, thus showing a good agreement with the RBS results (0≤x≤~0.2). This study shows the feasibility of combinatorial synthesis extended to heteroepitaxial, functional complex oxide films at wafer-scale using a new in-situ high-throughput characterization technique. The combination of laser-MBE growth with high-throughput LAXS for probing composition and RHEED for epitaxy is a powerful route to obtain new materials with optimized compositions, epitaxy, and functional properties.

Spectroscopic characterization techniques, such as Raman spectroscopy (RS), photoluminescence (PL), or X-ray fluorescence (XRF) are becoming widely spread for advanced material characterization due to their easy application and highly informative output about different materials properties in a fast and non-destructive manner. Spectroscopic data is mostly processed through simple analytical procedures that are time-consuming and undermine their potential. This is especially critical in the context of research in new material-based technologies which often present a high degree of complexity (e.g. multilayer, multiscale or multiprocess) and require the use of several characterization techniques in a combinatorial way in order to obtain an advanced understanding of their fundamental properties and failure mechanisms. The use of simple analytical procedures for this complex analysis results in long conception-to-market development times (normally 10-20 years) using large amounts of resources.

Recently, combinatorial analysis (CA) and machine learning (ML) have started to become common methodologies in the analysis of complex systems. It is foreseen that these advanced methodologies will shorten technology development times by a factor of 10. However, they require deep theoretical, analytical, and programming knowledge as well as big-data experimental generation which sets a barrier for their widespread implementation. In this context, the development of practical methodologies is paramount to accelerate the broad adoption of CA and ML and ultimately shorten technology development times.

This work proposes a novel methodology based on ML and CA to accelerate experimentation, analysis, and development. We present a step-by-step explanation of the methodology that includes the following key aspects: i) defining features and targets, ii) pre-processing and scaling experimental data, iii) selecting and using a ML algorithm, iv) analyzing and interpreting the outputs, v) visualizing the results, and vi) combining ML with CA for feature extraction. When applied successfully, this workflow produces consistent results that offer impactful insights into materials research. We illustrate this methodology by presenting its application to two practical examples of complex materials with high relevance for the thin-film PV community: the development of the earth-abundant Cu₂ZnGeSe₄ (CZGSe) at laboratory scale and of mature Cu(In,Ga)Se₂ (CIGSe) PV technologies. These materials exhibit several difficulties in their development due to the high amount of parameters involved in their multi-step fabrication and their impact on material properties. In the case of CZGSe, it is a high band gap absorber material with a high potential for tandem or semitransparent PV applications. The study was carried out in a combinatorial sample with graded composition characterized by XRF, RS, and current-voltage (I-V). In the case of CIGSe, it is a mature PV technology but with high improvement and cost reduction potential. In this study, CA (using RS, PL, and I-V) and ML methodologies were applied to optimize the fabrication of solar modules on flexible substrates in SUNPLUGGED GmbH’s innovative roll-to-roll (RtR) production line. In both cases, the applied CA and ML methodologies are demonstrated to enable defining the critical deposition parameters and conditions that allow reaching the optimum material properties in terms of their application in PV devices.

Finally, we also introduce a free open-source library based on Python for the presented methodologies to facilitate researchers and developers to perform CA and ML analysis of spectroscopic data. We strongly believe that this work represents the foundations for the future research of complex materials, which aims for more robust, faster, and automated work structures. In the particular case of thin-film PV, this will lead to a more efficient, versatile, and cheaper production of PV devices for a greener future.

The open-source signac data management framework (https://signac.io) helps researchers execute reproducible computational studies, scaling from laptops to supercomputers and emphasizing portability and fast prototyping. With signac, users can track, search, and archive data and metadata for file-based workflows and automate workflow submission on high performance computing (HPC) clusters. We will showcase current research by the materials community using signac, and discuss how recent improvements to the software’s feature set, scalability, and usability are enabling new scientific applications of the framework. Newly implemented synced data structures, workflow subgraph execution, and performance optimizations will be covered. Motivated by the needs of computational materials scientists for heterogeneous data processing with bespoke parallel execution patterns on widely varying hardware, we will demonstrate how these new capabilities enhance signac’s integrated approach to materials simulations and facilitate the high level of flexibility required by this domain of research.

On some occasions, a crystal structure that has been prepared and marketed for years is unexpectedly superseded by a thermodynamically more stable polymorph, rendering the initial crystal structure almost impossible to obtain—a so-called “disappearing polymorph” [1]. Despite the relative rarity of such events, famous cases such as the drug compounds Ritonavir and Rotigotine, where costly reformulations were required to manufacture the product again, underlined the commercial need for extensive screening of solid-state forms. Sometimes christened virtual polymorph screening, crystal structure prediction (CSP) emerged as a complement to experimental solid-form screens in order to evaluate the risk of a “disappearing polymorph” event. With the technology achieving maturity, CSP is now increasingly being used in the pharmaceutical industry to predict the relative thermodynamic stability of solid-state forms and to guide small-molecule solid form selection [2, 3].

A key component of the risk assessment and solid form selection workflow is to map the experimentally synthesized solid forms onto the predicted crystal-energy landscape. As CSP is increasingly used early in the drug development cycle, experimental single crystal structures are often unavailable, and X-ray powder diffraction patterns are the best available experimental representation of the synthesized solid form. Furthermore, several factors can affect the pattern quality, such as preferred orientation of crystallites in the powder sample, amorphous background or peak position offset. The available powder sample may also correspond to disordered crystal structures and/or be phase mixtures of multiple solid forms. Hence, it is imperative to have a sophisticated high-throughput matching functionality to place the experimentally observed forms on the predicted landscape.

In this contribution, we present the results of high-throughput refinement of powder diffraction patterns involving phase mixtures of pharmaceutically relevant compounds. Specifically, we compare the powder diffraction pattern against an extensive list of input crystal structures from an early-stage CSP study by refining or optimizing ~1 million parameter-sets that involve different combinations of structures, phase volume fractions and preferred orientation directions. Further, we discuss the ambiguities inherent to structure solution from powder diffraction data and the importance of calculated lattice energies (often also free energies where vibrational contributions are factored in) in the robust bridging of experimental and computational data.

References
[1] D.-K. Bučar, R. W. Lancaster & J. Bernstein (2015), “Disappearing Polymorphs Revisited” Angew. Chem., Int. Ed. 54, 6972-6993.
[2] M. A. Neumann & J. van de Streek (2018), “How Many Ritonavir Cases Are There Still out There?” Faraday Discuss. 211, 441-458.
[3] J. Hoja, Hsin-Yu Ko, M. A. Neumann, R. Car, R. A. DiStasio Jr. & A. Tkatchenko (2019), “Reliable and Practical Computational Description of Molecular Crystal Polymorphs” Sci. Adv. 5, eaau3338.

Data-centric approaches are already complementing our daily research and will even significantly change materials science in the near future. In this respect, data-analytics and machine-learning approaches are being developed and applied to various problems, and high-throughput screening (HTS) is going hand in hand with the establishment of small- and large-scale data collections. These resources allow us finding trends and patterns that cannot be obtained from individual investigations. Moreover, one can search for materials which exhibit features that are similar to those of other materials but are superior with respect to other criteria.

We have developed a platform that implements various similarity measures and employ this tool to explore the materials available in the NOMAD Encyclopedia [1]. Besides finding materials that resemble each other, e.g. in their electronic properties, this toolbox can also be used for assessing data quality. As such, one can compare the performance of different methodologies for one and the same material or the impact of approximations and computational parameters on calculated properties. We also make use of un unsupervised learning to find trends in the data and rationalize their physical origin. We will also discuss how we are currently expanding our efforts towards other materials properties of interest and developments towards inclusion of experimental data.

[1] C. Draxl and M. Scheffler, The NOMAD Laboratory: From Data Sharing to Artificial Intelligence, J. Phys. Mater. 2, 036001 (2019); https://nomad-lab.eu.

Considerable research attention has focused on metal halide perovskites (MHPs) over the recent years because of the combination of exceptional optoelectronic properties and low fabrication cost, making them ideal candidates for a variety of applications^1-4. Even so, the development of MHPs for commercialization must overcome an obstacle, namely stability in the pure or device-integrated form⁵. Overcoming adverse effects stemming from external stimuli can be minimized or avoided by utilizing established encapsulation techniques and device engineering^6,7. Simultaneously, another strategy is to improve the intrinsic stability by cation and/or halide alloying to synthesis multicomponent MHPs^8,9. A multitude of studies have demonstrated how the incorporation of other cations, particularly Cs⁺ and formamidinium (FA⁺) into methylammonium (MA⁺) systems, leads to improve stability in ambient and operational conditions. Mixing halides, has also proven to be an effective strategy toward stable perovskite materials.
Antisolvent crystallization methods are frequently used to fabricate high quality perovskite thin films, to produce sizable single crystals, and to synthesize nanoparticles at room temperature. However, a systematic exploration of the effect of specific antisolvents on the intrinsic stability of multicomponent metal halide perovskites has yet to be demonstrated. We have previously reported the development of a workflow for materials discovery utilizing both automated synthesis and machine learning (ML)^10,11. Similarly in this study, we develop a high-throughput experimental workflow that incorporates robotic synthesis, automated characterization, and ML techniques to explore how the choice of antisolvent affects the intrinsic stability of binary perovskite systems in ambient conditions¹². Different combinations of the endmembers are used to synthesize 15 combinatorial libraries, each with 96 unique combinations. In total, roughly 1100 different compositions are synthesized. Each library is fabricated twice using two different antisolvents: toluene and chloroform. Once synthesized, photoluminescence spectroscopy is automatically performed every 5 minutes for approximately 6 hrs. Non-negative Matrix Factorization (NMF) is then utilized to map the time- and compositional-dependent optoelectronic properties. Through the utilization of this workflow for each library, we demonstrate that the selection of antisolvent is critical to the stability of MHPs in ambient conditions. We explore possible dynamical processes, such as halide segregation, responsible for either the stability or eventual degradation as caused by the choice of antisolvent. Overall, this high-throughput study demonstrates the vital role that antisolvents play in the synthesis of high quality multicomponent MHP systems.
References:
1 Park, N.-G., Grätzel, M., Miyasaka, T., Zhu, K., Emery, K. Nature Energy 1 (2016).
2 Ahmadi, M., Wu, T., Hu, B. Adv Mater 29 (2017).
3 Lukosi, E. et al. Nuclear Instruments and Methods in Physics Research Section A, 927, 401 (2019).
4 Zou, Y., Yuan, Z., Bai, S., Gao, F., Sun, B. Materials Today Nano 5, (2019).
5 Liu, W. W., Wu, T. H., Liu, M. C., Niu, W. J. & Chueh, Y. L. Advanced Materials Interfaces 6, (2019).
6 Lee, J.-W. et al. Advanced Energy Materials 5, (2015).
7 Seo, S., Jeong, S., Bae, C., Park, N. G., Shin, H. Adv Mater, e1801010 (2018).
8 Xu, F., Zhang, T., Li, G., Zhao, Y. Journal of Materials Chemistry A 5, 11450 (2017).
9 David P. McMeekin, G. S., Waqaas Rehman, Giles E. Eperon, Michael Saliba, Maximilian T. Hörantner, Amir Haghighirad, Nobuya Sakai, Lars Korte, Bernd Rech, Michael B. Johnston, Laura M. Herz, Henry J. Snaith. Science 351 (2016).
10 Higgins, K., Valleti, S. M., Ziatdinov, M., Kalinin, S. V., Ahmadi, M. ACS Energy Letters 5, 3426 (2020).
11 Heimbrook, A., Higgins, K., Kalinin, S. V., Ahmadi, M. Nanophotonics, (2021).
12 Higgins, K., Ziatdinov, M., Kalinin, S. V., Ahmadi, M. arXiv:2106.03312 (2021).

Graph neural networks (GNNs) are an emerging machine learning methodology for processing irregularly structured data. They have shown promise in learning/predicting properties of atomistic structures due to their intrinsically physically-informed nature, with graph nodes representing atoms and graph edges symbolizing interatomic bonds. In atomistic modelling, the application of GNNs has generally been limited to small-scale molecular structures or crystalline materials represented by small unit cells. In contrast, describing more complex structural and microstructural features such as grain boundaries, voids, and surfaces requires scaling the application space of GNNs to vastly larger atomistic microstructures. To this end, we have developed a GNN training paradigm to characterize microstructural features without the need for GNN to explicitly identify such information during training. Our methodology aims to capture a fundamental aspect of complex atomic environments, in which one can generalize local geometries on a scale from fully ordered (perfect crystal) to completely disordered. This atomic-level classification system allows our GNN to extrapolate widely to configurations that it has not explicitly seen before. Our methodology also overcomes limitations associated with existing methods where the local environments are predefined to a “hard-coded” reference of specific structural motifs, which often leads to incorrect classification of complex structural environments such as grain boundaries. The notion of scaling between order and disorder presents a much smoother classification system of these geometric domains. We highlight the application of our GNN scheme to the automated detection and characterization of aluminum microstructures, derived from polycrystalline structures containing a variable number of grains and temperatures. We show that with a limited training set, our method correctly identifies and characterizes microstructural features such as surfaces, grain boundaries, voids, or other defect-likeenvironments on a much larger scale than that encountered during training. The examples presented here showcase the ability of our GNN methodology to characterize atomistic structures with unprecedented accuracy and transferability, opening the door to understanding structure-property relationships in complex materials in greater detail.

DNA-stabilized silver clusters (Ag_N-DNAs) are a model system for developing machine learning approaches to materials discovery. These clusters of 10-30 Ag atoms are encapsulated in short DNA strands, whose nucleobase sequences tune fluorescence emission from 500 nm to at least 1000 nm by selection of cluster size and morphology. With high quantum yield fluorescence, unique rod-like geometries, compatibility with DNA, and sensitivity to certain analytes, Ag_N-DNAs are promising for applications including sensing, bioimaging, and nanophotonics. However, the complex relationship between DNA sequence and Ag_N-DNA fluorescence color and the enormous combinatorial space of DNA sequences to choose from has hindered rational design of Ag_N-DNAs for such applications. Previously, we applied supervised machine learning classification together with high-throughput experiments to learn the connection between DNA sequence and Ag_N-DNA color, using data mining together with feature selection to engineer features for these models. Here, we present new efforts to engineer physically motivated features, presenting new features hypothesized based on observations from the first reported crystal structures of Ag_N-DNAs. We show that models using these physics-informed features perform comparably well or better than more complex models with features engineered by naive data mining, and the dimensionality of the new models is one order of magnitude less than the previous naive models. Furthermore, we rank features by a metric of importance to further reduce the dimensionality of our learned models and to inform the understanding of how DNA stabilizes silver clusters. This work demonstrates the importance of merging first-principles knowledge with machine learning and high throughput experimentation for the discovery of sequence-encoded biomolecular materials.

As materials discovery efforts increasingly expand in high-order composition spaces and/or far-from-equilibrium syntheses, characterizing the phase behavior of the materials becomes increasingly challenging. The limited ability of algorithms to automate phase mapping – the generation of the composition map of crystalline phases from a set of x-ray diffraction patterns – creates a bottleneck in high throughput and autonomous materials discovery. For properties where only phase-pure materials are of interest one needs to screen out the mixed-phase diffraction patterns, although properties from catalysis to light harvesting materials may be optimized through interfaces and other multi-phase interactions. As a result, mixed-phase x-ray diffraction patterns must be de-mixed, which can be very challenging when candidate structures contain overlapped features and the diffraction patterns of prototype structures vary due to, for example, alloying. In such cases, experts analyze the data by collectively considering a collection of related mixed-phase diffraction patterns and invoking thermodynamic rules that govern phase mixtures in the given materials space. Previous methods have used such rules to “fix” phase mapping solutions from de-mixing patterns, but when dozens of phase mixtures exist within a dataset and the materials space has multiple degrees of freedom that require hundreds of diffraction patterns to characterize, the data complexity exceeds the capabilities of both human experts and state of the art algorithms. In such cases, the prior knowledge about phase prototypes of thermodynamic rules must be integrated into the demixing process. Application of the rules inherently involves computational reasoning and recognition of the phase-pure patterns in the data inherently involves computational learning. The combination of reasoning and learning to solve challenging problems is a hallmark of human intelligence, and we introduce Deep Reasoning Networks (DRNets) to seamlessly integrate reasoning and learning. DRNets are designed with an interpretable latent space for encoding prior-knowledge domain constraints and integrating constraint reasoning into neural network optimization. DRNets are demonstrated for solving complex phase behavior in 3-cation oxide composition spaces.

Understanding the conditions for the formation and reduction of metal oxides at high temperatures is of great industrial relevance, e.g., for corrosion prevention and for metal production. However, the experimental characterization of high-temperature redox chemistry is challenging and requires specialized equipment. Simulations with the Calculation of Phase Diagrams (CALPHAD) method can be an alternative but require assessed phase diagrams for the system of interest. First-principles calculations, on the other hand, can provide robust estimates of redox potentials at zero Kelvin without experimental input, but simulating redox reactions at high temperatures is computationally demanding and often too approximate.

Here, we will discuss initial progress towards the efficient computational prediction of high-temperature redox chemistry using machine-learning augmented first-principles calculations. We show that a combination of results from zero-Kelvin density-functional theory (DFT) calculations and a machine-learning model trained on temperature-dependent reaction free energies allows predicting reduction temperatures of metal oxides that the model was not trained on. Hence, this initial application for crystalline binary and ternary oxides demonstrates that the temperature dependence of the free energy can indeed be cross-learned from other oxides, thereby removing the limitation to compounds that have previously been thermodynamically assessed.

How can we simulate a complex nanoscale phenomena involving coupled physical and chemical processes and interactions that are hard to model? In this work we show that combining in situ environmental transmission electron microscopy (E-TEM) with automated image processing and statistical machine learning uniquely enables formulating interpretable mathematical models as well as creating accurate simulation tools. In particular, there is a need for a better understanding, characterization, and prediction of the proximity effects among dense populations of metal nanocatalysts as they form and evolve over time. Here, we leverage point process theory—a branch of statistical machine learning—to “learn” the spatial dependencies among ensembles of adjacent alumina-supported iron nanoparticles from a time sequence of E-TEM images. We construct a set of point process models to make statistical inferences about the nature of spatial dependencies that govern the rapid formation, or “popping” of nanoparticles during thin film dewetting, concomitant with metal reduction in the presence of acetylene at 750 °C. We show that nanoparticles exhibit strong dispersion behavior, i.e. new nanoparticles pop in dispersed locations at a predictable distance from their existing territorial neighbors. We also show that the Softcore model, a class of Gibbs point processes, adequately describes the pairwise interactions underlying such time-dependent spatial variations. Further, we leverage the probabilistic nature of our statistical models to develop a computational simulation tool capable of producing accurate simulations of the spatio-temporal evolution of formation patterns of nanoparticles at both finer time resolutions and larger spatial domains than experimental observations. This is a much needed capability, wherein the behavior of hard-to-model coupled nanoscale phenomena involving complex chemical and physical processes are accurately captured in machine-learned mathematical formulations; thus overcoming current limitations in computational methods supporting the design and analysis of collective nanocatalyst populations.

Fracture behaviors of brittle materials have been a crucial problem when it comes to safety and reliability. Unlike the one-sided understanding that defects result in a negative effect on materials, a new perspective from recent studies indicates that nano-defects could strengthen the material if rationally designed. Nevertheless, fracture behaviors of graphene under defective conditions remain unclear and need more comprehensive studies. With deep learning (DL) recently drawing attention in the material field, several studies have applied machine learning techniques on predicting crack growth of brittle materials. This study aims to use a deep sequential model to predict the fracture path of graphene under two systems of defective configuration. One involves graphene with different crystalline, and the other involves graphene with various porous defects. First, we built 20 nm 16 nm graphene sheets with different orientations or pores defects. The 20 nm pre-existing crack is set at the edge of each graphene sheet. Subsequently, we performed uniaxial tensile tests by implementing molecular dynamics simulation and process the results into grayscale images. We sliced every single image into pieces so that the ConvLSTM-based model can learn the spatiotemporal features from the sequential images. The input matrices are mapped from the sequential images, and the output matrices are mapped from the crack images of the next timestep. After iterating the prediction process, we can obtain the final crack path of the graphene. We first used 100 crack images (graphene with ten different orientations in 10 sets) to train the model. The model shows 99.45% binary accuracy and can predict the crack path of graphene under different orientations, whether it is single crystal or bicrystal. We used other 300 crack images (arbitrary porous graphene) to train another model. The model shows 94.50% binary accuracy and can predict crack propagation of arbitrary porous graphene. This study applied deep learning on graphene fracture predictions, which work as a more efficient tool than conventional simulation methods, which is potential for next-generation material design.

Machine learning and artificial intelligence (ML/AI) are rapidly becoming an indispensable part of physics research, with domain applications ranging from theory and materials prediction to high-throughput data analysis. In parallel, the recent successes in applying ML/AI methods for autonomous systems from robotics through self-driving cars to organic and inorganic synthesis are generating enthusiasm for the potential of these techniques to enable automated and autonomous experiment (AE) in imaging. The implementation of this vision requires the synergy of three intertwined components, including the engineering controls of the imaging tools, the algorithmic developments that allow to define non-trivial exploratory patterns, and physics-driven interpretation.
In this presentation, I will discuss recent progress in automated experiment in electron and scanning probe microscopy, ranging from feature to physics discovery via active learning. The applications of classical deep learning methods in streaming image analysis are strongly affected by the out of distribution drift effects, and the approaches to minimize though are discussed. We further present invariant variational autoencoders as a method to disentangle affine distortions and rotational degrees of freedom from other latent variables in imaging and spectral data. The analysis of the latent space of autoencoders further allows establishing physically relevant transformation mechanisms. Extension of encoder approach towards establishing structure-property relationships will be illustrated on the example of plasmonic structures. Finally, I illustrate transition from post-experiment data analysis to active learning process. Here, the strategies based on simple Gaussian Processes often tend to produce sub-optimal results due to the lack of prior knowledge and very simplified (via learned kernel function) representation of spatial complexity of the system. Comparatively, deep kernel learning (DKL) methods allow to realize both the exploration of complex systems towards the discovery of structure-property relationship, and enable automated experiment targeting physics (rather than simple spatial feature) discovery. The latter is illustrated via experimental discovery of the edge plasmons in STEM/EELS and ferroelectric domain dynamics in PFM.

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.

In materials science and engineering, one is typically searching for materials that exhibit exceptional performance for a certain function, and all these searches face the following situation[1]

-- The number of possible materials is practically infinite.

-- The electronic and atomistic processes that rule a desired materials function are many, and their concerted action is typically highly complex and intricate, resulting in an immense number of possibly relevant mechanisms.

-- The number of data that are “clean” (i.e. comprehensively characterized and high-quality) and relevant for the function of interest are typically very low.

Under these daunting conditions we aim to identify the rules that govern the rare phenomena corresponding to particularly exceptional materials. Thus, the scientific discovery typically resembles the proverbial hunt for the needle in a haystack.
This talks describes two methods that guide the way towards solutions of the challenge: Subgroup Discovery [2] and Sure Independent Screening and Sparcifying Operator (SISSO) [3]. Both methods enable us to identify the “materials genes”, i.e. the basic physico-chemical parameters that are relevant for the property of interest. In analogy to biology, these “materials genes” are correlated to processes that trigger, actuate, or facilitate, or hinder the property of interest.

In the coordinate systems of these genes we identify maps of materials properties, i.e. (small) regions where the desired high-performance materials can be found.

The talk describes the concepts and illustrates them with specific examples, e.g. crystal-structure prediction, topological semiconductors, materials that may convert green-house gases into useful chemicals and fuels, and more.

1) C. Draxl and M. Scheffler, Big-data-driven materials science and its FAIR data infrastructure. Handbook of Materials Modeling, eds. S. Yip and W. Andreoni (Basel: Springer International Publishing) p. 49 (2020).
2) B. R. Goldsmith, M. Boley, J. Vreeken, M. Scheffler, L. M. Ghiringhelli, Uncovering structure-property relationships of materials by subgroup discovery. New J. Phys. 19, 013031 (2017).
3) R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, and L. M. Ghiringhelli, SISSO: a compressed-sensing method for identifying the best low-dimensional descriptor in an immensity of offered candidates. Phys. Rev. Mat. 2, 083802 (2018).

In this talk I will present some of our recent work on different ways of representing materials for various machine learning applications. Starting with a simple case of using the well-known Coulomb matrix representation for solid-solutions and alloys, I will demonstrate how this significantly outperforms the cluster expansion approach, which has been the standard method for atomistic modelling of alloys for several decades. The Coulomb matrix representation, in combination with neural networks is shown to be highly effective in predicting non-linear relationships, exemplified by the case of band gaps in a solid-solution of (Zn, Mg)O. Next I will consider graph-neural networks (GNNs), which encode the topology of a material as a message passing graph – typically these GNNs are coupled with a standard neural network to predict a given material property. I will demonstrate how GNNs can be used instead to learn a representation of the material as a compressed vector, which can then be used in a Bayesian approach to predict properties, but with added value of a principled estimate of the uncertainty in the estimation. By using this GNN-fed Gaussian process, we are then able to develop an active learning protocol whereby we select the next best experiment (or high-level calculation) to obtain new labelled data. We find that this active learning process leads to model improvement at twice the rate of adding data without guidance. Finally I will demonstrate how principles from natural language processing (NLP) can be applied to learn representations of atoms and then to assemble these atoms into compounds. Using this approach we can leverage existing databases, such as the Materials Project to rapidly learn distributed embeddings and then apply these embeddings to predict compound properties – I will show how this NLP-inspired approach outperforms many existing approaches on a number of standard materials science benchmark tests.

The past decade has witnessed a surge in the development and adoption of machine learning algorithms to solve day-a-day computational tasks. Yet, a solid theoretical understanding of even the most basic tools used in practice is still lacking, as traditional statistical learning methods are unfit to deal with the modern regime in which the number of model parameters are of the same order as the quantity of data - a problem known as the curse of dimensionality. Curiously, this is precisely the regime studied by Physicists since the mid 19th century in the context of interacting many-particle systems. This connection, which was first established in the seminal work of Elisabeth Gardner and Bernard Derrida in the 80s, is the basis of a long and fruitful marriage between these two fields.

In this talk I will motivate and review the connections between Statistical Physics and problems in high-dimensional Statistics, such as the ones stemming from the fields of Machine Learning and Signal Processing. Finally, I will exemplify this discussion with some recent concrete applications of the Statistical Physics toolbox to different problems of interest.