Active and Transfer Learning of High-Dimensional Neural Network Potentials for Transition Metals.

Classical molecular dynamics (MD) simulations represent a very popular and powerful tool for materials modeling and design. The predictive power of MD hinges on the ability of the interatomic potential to capture the underlying physics and chemistry. There have been decades of seminal work on developing interatomic potentials, albeit with a focus predominantly on capturing the properties of bulk materials. Such physics-based models, while extensively deployed for predicting the dynamics and properties of nanoscale systems over the past two decades, tend to perform poorly in predicting nanoscale potential energy surfaces (PESs) when compared to high-fidelity first-principles calculations. These limitations stem from the lack of flexibility in such models, which rely on a predefined functional form. Machine learning (ML) models and approaches have emerged as a viable alternative to capture the diverse size-dependent cluster geometries, nanoscale dynamics, and the complex nanoscale PESs, without sacrificing the bulk properties. Here, we introduce an ML workflow that combines transfer and active learning strategies to develop high-dimensional neural networks (NNs) for capturing the cluster and bulk properties for several different transition metals with applications in catalysis, microelectronics, and energy storage, to name a few. Our NN first learns the bulk PES from the high-quality physics-based models in literature and subsequently augments this learning via retraining with a higher-fidelity first-principles training data set to concurrently capture both the nanoscale and bulk PES. Our workflow departs from status-quo in its ability to learn from a sparsely sampled data set that nonetheless covers a diverse range of cluster configurations from near-equilibrium to highly nonequilibrium as well as learning strategies that iteratively improve the fingerprinting depending on model fidelity. All the developed models are rigorously tested against an extensive first-principles data set of energies and forces of cluster configurations as well as several properties of bulk configurations for 10 different transition metals. Our approach is material agnostic and provides a methodology to transfer and build upon the learnings from decades of seminal work in molecular simulations on to a new generation of ML-trained potentials to accelerate materials discovery and design.

AutoFocus: AI-driven alignment of nanofocusing X-ray mirror systems.

We describe the application of an AI-driven system to autonomously align complex x-ray-focusing mirror systems, including mirrors systems with variable focus spot sizes. The system has been developed and studied on a digital twin of nanofocusing X-ray beamlines, built using advanced optical simulation tools calibrated with wavefront sensing data collected at the beamline.We experimentally demonstrated that the system is reliably capable of positioning a focused beam on the sample, both by simulating the variation of a beamline with random perturbations due to typical changes in the light source and optical elements over time, and by conducting similar tests on an actual focusing mirror system.

Deep learning at the edge enables real-time streaming ptychographic imaging.

Coherent imaging techniques provide an unparalleled multi-scale view of materials across scientific and technological fields, from structural materials to quantum devices, from integrated circuits to biological cells. Driven by the construction of brighter sources and high-rate detectors, coherent imaging methods like ptychography are poised to revolutionize nanoscale materials characterization. However, these advancements are accompanied by significant increase in data and compute needs, which precludes real-time imaging, feedback and decision-making capabilities with conventional approaches. Here, we demonstrate a workflow that leverages artificial intelligence at the edge and high-performance computing to enable real-time inversion on X-ray ptychography data streamed directly from a detector at up to 2 kHz. The proposed AI-enabled workflow eliminates the oversampling constraints, allowing low-dose imaging using orders of magnitude less data than required by traditional methods.

Demonstration of an AI-driven workflow for autonomous high-resolution scanning microscopy.

Modern scanning microscopes can image materials with up to sub-atomic spatial and sub-picosecond time resolutions, but these capabilities come with large volumes of data, which can be difficult to store and analyze. We report the Fast Autonomous Scanning Toolkit (FAST) that addresses this challenge by combining a neural network, route optimization, and efficient hardware controls to enable a self-driving experiment that actively identifies and measures a sparse but representative data subset in lieu of the full dataset. FAST requires no prior information about the sample, is computationally efficient, and uses generic hardware controls with minimal experiment-specific wrapping. We test FAST in simulations and a dark-field X-ray microscopy experiment of a WSe2 film. Our studies show that a FAST scan of <25% is sufficient to accurately image and analyze the sample. FAST is easy to adapt for any scanning microscope; its broad adoption will empower general multi-level studies of materials evolution with respect to time, temperature, or other parameters.

Multislice forward modeling of coherent surface scattering imaging on surface and interfacial structures.

To study nanostructures on substrates, surface-sensitive reflection-geometry scattering techniques such as grazing incident small angle X-ray scattering are commonly used to yield an averaged statistical structural information of the surface sample. Grazing incidence geometry can probe the absolute three-dimensional structural morphology of the sample if a highly coherent beam is used. Coherent surface scattering imaging (CSSI) is a powerful yet non-invasive technique similar to coherent X-ray diffractive imaging (CDI) but performed at small angles and grazing-incidence reflection geometry. A challenge with CSSI is that conventional CDI reconstruction techniques cannot be directly applied to CSSI because the Fourier-transform-based forward models cannot reproduce the dynamical scattering phenomenon near the critical angle of total external reflection of the substrate-supported samples. To overcome this challenge, we have developed a multislice forward model which can successfully simulate the dynamical or multi-beam scattering generated from surface structures and the underlying substrate. The forward model is also demonstrated to be able to reconstruct an elongated 3D pattern from a single shot scattering image in the CSSI geometry through fast-performing CUDA-assisted PyTorch optimization with automatic differentiation.

X-ray nanodiffraction imaging reveals distinct nanoscopic dynamics of an ultrafast phase transition.

SignificancePhase transitions, the changes between states of matter with distinct electronic, magnetic, or structural properties, are at the center of condensed matter physics and underlie valuable technologies. First-order phase transitions are intrinsically heterogeneous. When driven by ultrashort excitation, nanoscale phase regions evolve rapidly, which has posed a significant experimental challenge to characterize. The newly developed laser-pumped X-ray nanodiffraction imaging technique reported here has simultaneous 100-ps temporal and 25-nm spatial resolutions. This approach reveals pathways of the nanoscale structural rearrangement upon ultrafast optical excitation, different from those transitions under slowly varying parameters. The spatiotemporally resolved structural characterization provides crucial nanoscopic insights into ultrafast phase transitions and opens opportunities for controlling nanoscale phases on ultrafast time scales.

Multi-reward Reinforcement Learning Based Bond-Order Potential to Study Strain-Assisted Phase Transitions in Phosphorene.

We introduce a multi-reward reinforcement learning (RL) approach to train a flexible bond-order potential (BOP) for 2D phosphorene based on ab initio training data sets. Our approach is based on a continuous action space Monte Carlo tree search algorithm that is general and scalable and presents an efficient multiobjective optimization scheme for high-dimensional materials design problems. As a proof-of-concept, we deploy this scheme to parametrize multiple structural and dynamical properties of 2D phosphorene polymorphs. Our RL-trained BOP model adequately captures the structure, energetics, transformation barriers, equation of state, elastic constants, and phonon dispersions of various 2D P polymorphs. We use this model to probe the impact of temperature and strain rate on the phase transition from black (α-P) to blue phosphorene (ß-P) through molecular dynamics simulations. A decrease in critical strain for this phase transition with increase in temperature is observed, and the underlying atomistic mechanisms are discussed.

Learning in continuous action space for developing high dimensional potential energy models.

Reinforcement learning (RL) approaches that combine a tree search with deep learning have found remarkable success in searching exorbitantly large, albeit discrete action spaces, as in chess, Shogi and Go. Many real-world materials discovery and design applications, however, involve multi-dimensional search problems and learning domains that have continuous action spaces. Exploring high-dimensional potential energy models of materials is an example. Traditionally, these searches are time consuming (often several years for a single bulk system) and driven by human intuition and/or expertise and more recently by global/local optimization searches that have issues with convergence and/or do not scale well with the search dimensionality. Here, in a departure from discrete action and other gradient-based approaches, we introduce a RL strategy based on decision trees that incorporates modified rewards for improved exploration, efficient sampling during playouts and a "window scaling scheme" for enhanced exploitation, to enable efficient and scalable search for continuous action space problems. Using high-dimensional artificial landscapes and control RL problems, we successfully benchmark our approach against popular global optimization schemes and state of the art policy gradient methods, respectively. We demonstrate its efficacy to parameterize potential models (physics based and high-dimensional neural networks) for 54 different elemental systems across the periodic table as well as alloys. We analyze error trends across different elements in the latent space and trace their origin to elemental structural diversity and the smoothness of the element energy surface. Broadly, our RL strategy will be applicable to many other physical science problems involving search over continuous action spaces.

Learning with Delayed Rewards-A Case Study on Inverse Defect Design in 2D Materials.

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 (RL) [Monte Carlo Tree Search (MCTS)] 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 two-dimensional 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 RL strategy employing delayed rewards that can accelerate the inverse design of defects in materials for achieving targeted functionality.

Artificial Intelligence-Guided De Novo Molecular Design Targeting COVID-19.

An extensive search for active therapeutic agents against the SARS-CoV-2 is being conducted across the globe. While computational docking simulations remain a popular method of choice for the in silico ligand design and high-throughput screening of therapeutic agents, it is severely limited in the discovery of new candidate ligands owing to the high computational cost and vast chemical space. Here, we present a de novo molecular design strategy that leverages artificial intelligence (AI) to discover new therapeutic agents against SARS-CoV-2. A Monte Carlo tree search algorithm combined with a multitask neural network surrogate model for expensive docking simulations, and recurrent neural networks for rollouts, is used in an iterative search and retrain strategy. Using Vina scores as the target objective to measure binding to either the isolated spike protein (S-protein) at its host receptor region or to the S-protein/angiotensin converting enzyme 2 receptor interface, we generate several (∼100's) new therapeutic agents that outperform Food and Drug Administration (FDA) (∼1000's) and non-FDA molecules (∼million). Our AI strategy is broadly applicable for accelerated design and discovery of chemical molecules with any user-desired functionality.

Nanoscale Imaging and Control of Volatile and Non-Volatile Resistive Switching in VO2.

Control of the metal-insulator phase transition is vital for emerging neuromorphic and memristive technologies. The ability to alter the electrically driven transition between volatile and non-volatile states is particularly important for quantum-materials-based emulation of neurons and synapses. The major challenge of this implementation is to understand and control the nanoscale mechanisms behind these two fundamental switching modalities. Here, in situ X-ray nanoimaging is used to follow the evolution of the nanostructure and disorder in the archetypal Mott insulator VO2 during an electrically driven transition. Our findings demonstrate selective and reversible stabilization of either the insulating or metallic phases achieved by manipulating the defect concentration. This mechanism enables us to alter the local switching response between volatile and persistent regimes and demonstrates a new possibility for nanoscale control of the resistive switching in Mott materials.

Screening of Therapeutic Agents for COVID-19 Using Machine Learning and Ensemble Docking Studies.

The current pandemic demands a search for therapeutic agents against the novel coronavirus SARS-CoV-2. Here, we present an efficient computational strategy that combines machine learning (ML)-based models and high-fidelity ensemble docking studies to enable rapid screening of possible therapeutic ligands. Targeting the binding affinity of molecules for either the isolated SARS-CoV-2 S-protein at its host receptor region or the S-protein:human ACE2 interface complex, we screen ligands from drug and biomolecule data sets that can potentially limit and/or disrupt the host-virus interactions. Top scoring one hundred eighty-seven ligands (with 75 approved by the Food and Drug Administration) are further validated by all atom docking studies. Important molecular descriptors (2χn, topological surface area, and ring count) and promising chemical fragments (oxolane, hydroxy, and imidazole) are identified to guide future experiments. Overall, this work expands our knowledge of small-molecule treatment against COVID-19 and provides a general screening pathway (combining quick ML models with expensive high-fidelity simulations) for targeting several chemical/biochemical problems.

Perovskite neural trees.

Trees are used by animals, humans and machines to classify information and make decisions. Natural tree structures displayed by synapses of the brain involves potentiation and depression capable of branching and is essential for survival and learning. Demonstration of such features in synthetic matter is challenging due to the need to host a complex energy landscape capable of learning, memory and electrical interrogation. We report experimental realization of tree-like conductance states at room temperature in strongly correlated perovskite nickelates by modulating proton distribution under high speed electric pulses. This demonstration represents physical realization of ultrametric trees, a concept from number theory applied to the study of spin glasses in physics that inspired early neural network theory dating almost forty years ago. We apply the tree-like memory features in spiking neural networks to demonstrate high fidelity object recognition, and in future can open new directions for neuromorphic computing and artificial intelligence.

Ferroelectric Domain Wall Motion in Freestanding Single-Crystal Complex Oxide Thin Film.

Ferroelectric domain walls in single-crystal complex oxide thin films are found to be orders of magnitude slower when the interfacial bonds with the heteroepitaxial substrate are broken to create a freestanding film. This drastic change in domain wall kinetics does not originate from the alteration of epitaxial strain; rather, it is correlated with the structural ripples at mesoscopic length scale and associated flexoelectric effects induced in the freestanding films. In contrast, the effects of the bond-breaking on the local static ferroelectric properties of both top and bottom layers of the freestanding films, such as domain wall width and spontaneous polarization, are modest and governed by the change in epitaxy-induced compressive strain.

Machine learning coarse grained models for water.

An accurate and computationally efficient molecular level description of mesoscopic behavior of ice-water systems remains a major challenge. Here, we introduce a set of machine-learned coarse-grained (CG) models (ML-BOP, ML-BOPdih, and ML-mW) that accurately describe the structure and thermodynamic anomalies of both water and ice at mesoscopic scales, all at two orders of magnitude cheaper computational cost than existing atomistic models. In a significant departure from conventional force-field fitting, we use a multilevel evolutionary strategy that trains CG models against not just energetics from first-principles and experiments but also temperature-dependent properties inferred from on-the-fly molecular dynamics (~ 10's of milliseconds of overall trajectories). Our ML BOP models predict both the correct experimental melting point of ice and the temperature of maximum density of liquid water that remained elusive to-date. Our ML workflow navigates efficiently through the high-dimensional parameter space to even improve upon existing high-quality CG models (e.g. mW model).

Real-time coherent diffraction inversion using deep generative networks.

Phase retrieval, or the process of recovering phase information in reciprocal space to reconstruct images from measured intensity alone, is the underlying basis to a variety of imaging applications including coherent diffraction imaging (CDI). Typical phase retrieval algorithms are iterative in nature, and hence, are time-consuming and computationally expensive, making real-time imaging a challenge. Furthermore, iterative phase retrieval algorithms struggle to converge to the correct solution especially in the presence of strong phase structures. In this work, we demonstrate the training and testing of CDI NN, a pair of deep deconvolutional networks trained to predict structure and phase in real space of a 2D object from its corresponding far-field diffraction intensities alone. Once trained, CDI NN can invert a diffraction pattern to an image within a few milliseconds of compute time on a standard desktop machine, opening the door to real-time imaging.

Three-dimensional X-ray diffraction imaging of dislocations in polycrystalline metals under tensile loading.

The nucleation and propagation of dislocations is an ubiquitous process that accompanies the plastic deformation of materials. Consequently, following the first visualization of dislocations over 50 years ago with the advent of the first transmission electron microscopes, significant effort has been invested in tailoring material response through defect engineering and control. To accomplish this more effectively, the ability to identify and characterize defect structure and strain following external stimulus is vital. Here, using X-ray Bragg coherent diffraction imaging, we describe the first direct 3D X-ray imaging of the strain field surrounding a line defect within a grain of free-standing nanocrystalline material following tensile loading. By integrating the observed 3D structure into an atomistic model, we show that the measured strain field corresponds to a screw dislocation.

Active site localization of methane oxidation on Pt nanocrystals.

High catalytic efficiency in metal nanocatalysts is attributed to large surface area to volume ratios and an abundance of under-coordinated atoms that can decrease kinetic barriers. Although overall shape or size changes of nanocatalysts have been observed as a result of catalytic processes, structural changes at low-coordination sites such as edges, remain poorly understood. Here, we report high-lattice distortion at edges of Pt nanocrystals during heterogeneous catalytic methane oxidation based on in situ 3D Bragg coherent X-ray diffraction imaging. We directly observe contraction at edges owing to adsorption of oxygen. This strain increases during methane oxidation and it returns to the original state after completing the reaction process. The results are in good agreement with finite element models that incorporate forces, as determined by reactive molecular dynamics simulations. Reaction mechanisms obtained from in situ strain imaging thus provide important insights for improving catalysts and designing future nanostructured catalytic materials.

Defect Dynamics in 2-D MoS2 Probed by Using Machine Learning, Atomistic Simulations, and High-Resolution Microscopy.

Structural defects govern various physical, chemical, and optoelectronic properties of two-dimensional transition-metal dichalcogenides (TMDs). A fundamental understanding of the spatial distribution and dynamics of defects in these low-dimensional systems is critical for advances in nanotechnology. However, such understanding has remained elusive primarily due to the inaccessibility of (a) necessary time scales via standard atomistic simulations and (b) required spatiotemporal resolution in experiments. Here, we take advantage of supervised machine learning, in situ high-resolution transmission electron microscopy (HRTEM) and molecular dynamics (MD) simulations to overcome these limitations. We combine genetic algorithms (GA) with MD to investigate the extended structure of point defects, their dynamical evolution, and their role in inducing the phase transition between the semiconducting (2H) and metallic (1T) phase in monolayer MoS2. GA-based structural optimization is used to identify the long-range structure of randomly distributed point defects (sulfur vacancies) for various defect densities. Regardless of the density, we find that organization of sulfur vacancies into extended lines is the most energetically favorable. HRTEM validates these findings and suggests a phase transformation from the 2H-to-1T phase that is localized near these extended defects when exposed to high electron beam doses. MD simulations elucidate the molecular mechanism driving the onset of the 2H to 1T transformation and indicate that finite amounts of 1T phase can be retained by increasing the defect concentration and temperature. This work significantly advances the current understanding of defect structure/evolution and structural transitions in 2D TMDs, which is crucial for designing nanoscale devices with desired functionality.

Configurational-Bias Monte Carlo Back-Mapping Algorithm for Efficient and Rapid Conversion of Coarse-Grained Water Structures into Atomistic Models.

Coarse-grained molecular dynamics (MD) simulations represent a powerful approach to simulate longer time scale and larger length scale phenomena than those accessible to all-atom models. The gain in efficiency, however, comes at the cost of atomistic details. The reverse transformation, also known as back mapping, of coarse-grained beads into their atomistic constituents represents a major challenge. Most existing approaches are limited to specific molecules or specific force fields and often rely on running a long-time atomistic MD of the back-mapped configuration to arrive at an optimal solution. Such approaches are problematic when dealing with systems with high diffusion barriers. Here, we introduce a new extension of the configurational-bias Monte Carlo (CBMC) algorithm, which we term the crystalline-configurational-bias Monte Carlo (C-CBMC) algorithm, which allows rapid and efficient conversion of a coarse-grained model back into its atomistic representation. Although the method is generic, we use a coarse-grained water model as a representative example and demonstrate the back mapping or reverse transformation for model systems ranging from the ice-liquid water interface to amorphous and crystalline ice configurations. A series of simulations using the TIP4P/Ice model are performed to compare the new CBMC method to several other standard Monte Carlo and molecular dynamics-based back-mapping techniques. In all of the cases, the C-CBMC algorithm is able to find optimal hydrogen-bonded configuration many thousand evaluations/steps sooner than the other methods compared within this paper. For crystalline ice structures, such as a hexagonal, cubic, and cubic-hexagonal stacking disorder structures, the C-CBMC was able to find structures that were between 0.05 and 0.1 eV/water molecule lower in energy than the ground-state energies predicted by the other methods. Detailed analysis of the atomistic structures shows a significantly better global hydrogen positioning when contrasted with the existing simpler back-mapping methods. The errors in the radial distribution functions (RDFs) of back-mapped configuration relative to reference configuration for the C-CBMC, MD, and MC were found to be 6.9, 8.7, and 12.9, respectively, for the hexagonal system. For the cubic system, the relative errors of the RDFs for the C-CBMC, MD, and MC were found to be 18.2, 34.6, and 39.0, respectively. Our results demonstrate the efficiency and efficacy of our new back-mapping approach, especially for crystalline systems where simple force-field-based relaxations have a tendency to get trapped in local minima.