Algorithm Development in Materials Science and Engineering: Machine Learning and Atomistic Algorithms to Accelerate Materials Study and Design
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee, TMS: Integrated Computational Materials Engineering Committee, TMS: Phase Transformations Committee, TMS: Solidification Committee
Program Organizers: Mohsen Asle Zaeem, Colorado School of Mines; Mikhail Mendelev, NASA ARC; Bryan Wong, University of California, Riverside; Ebrahim Asadi, University of Memphis; Garritt Tucker, Colorado School of Mines; Charudatta Phatak, Argonne National Laboratory; Bryce Meredig, Travertine Labs LLC
Monday 8:30 AM
March 15, 2021
Room: RM 36
Location: TMS2021 Virtual
Session Chair: Mohsen Asle Zaeem, Colorado School of Mines; Mikhail Mendelev, Ames Laboratory
8:30 AM Invited
Theory-infused Machine Learning Algorithms of Chemisorption at Metal Surfaces: Hongliang Xin1; 1Virginia Polytechnic Institute and State University
The formation and breakage of chemical bonds at active sites is the molecular basis of catalysis. Being able to rapidly compute interaction strengths between bonding entities and understand their trends holds the key to the design of improved catalysts. Despite recent advances, machine learning (ML) faces a tremendous challenge for catalysis applications due to its poor transferability and explainability. Here we present theory-infused machine learning (TIML) algorithms that integrates convolutional neural networks with the d-band theory of chemisorption for predicting the chemical reactivity of metal surfaces. With *OH and *CO as two representative adsorbates, we demonstrated that the hybrid ML models outperform the purely data-driven ones in both data scarce and rich regions, especially for out-of-sample systems. More importantly, the architecture design enables its physical interpretability, shedding light on the nature of chemicalbonding at metal surfaces.
9:00 AM Invited
Accelerating Atomistic Monte Carlo Simulations with Autoregressive Models: Rafael Gomezbombarelli1; James Damewood1; 1Massachusetts Institute of Technology
In order for high-throughput virtual screening platforms to overcome the challenges of modeling the diversity and complexity of metal systems, modern algorithms must efficiently leverage computational resources. While Monte Carlo simulations are frequently used to predict phase stability or material properties at equilibrium, these calculations can be prohibitively expensive when exploring phase transitions or systems involving many components. New neural network-based generative models have developed into powerful sampling methods that can be integrated into existing Monte Carlo workflows to accelerate traditional approaches. We will present predictions of these machine learning led simulations on tasks relevant for computational discovery in metals. We will examine the sampling of ground states of a Copper-Gold system and the miscibility gap in Nickel-Gold. We will discuss the extension of the model to study multicomponent systems and analyze the scaling of the method with system size and evaluate the potential for usage within high-throughput screening strategies.
9:30 AM
Application of a Shape Moment Descriptor Set Towards a Robust and Transferable Description of Local Atomic Environments: Jacob Tavenner1; Edward Kober2; Garritt Tucker1; 1Colorado School of Mines; 2Los Alamos National Laboratory
In the study of atomistic behavior, mathematical descriptions of atomic structure are critical for robust scientific analysis of both static and dynamic systems. A robust descriptor which improves upon prior methods, requiring no a priori knowledge of the system being analyzed, has been developed. In evidence of the improved performance of these descriptors, a small number of potential applications will be examined. These areas include grain boundary structure, atomic motion, and segregation potential, among others. Improvement of current understanding or analysis methods will be demonstrated using these novel descriptors of local atomic environments. By leveraging these descriptors, the relationship between atomic environments and their underlying physics which drive system behavior can be better understood. Machine learning techniques are utilized to elucidate these complex relationships, demonstrating the applicability of this approach with modern data-driven techniques for processing the substantial volume of data generated through many modern computational studies.LA-UR-20-25125
9:50 AM Invited
High Speed Artificial Neural Network Implementation of Interatomic Force Fields in Metals: Doyl Dickel1; Christopher Barrett1; Mashroor Nitol1; 1Mississippi State University
Machine learning techniques, particularly artificial neural networks (ANNs), have proven to be effective at reproducing DFT and first-principles calculations at accelerated timescales. These tools are capable of sub meV/atom accuracy while operating scaling linearly with the size of the system. However, many of the popular implementations are still orders of magnitude slower than traditional forcefield models such as EAM and MEAM. Overcoming this performance gap is essential to the production of ANNs which are useful to solve modern molecular dynamics problems. Here, we demonstrate our ANN formalism, inspired by existing semi-empirical methods. It is shown that using a physically motivated fingerprint and other innovations from classical methods, the computation time of these force fields can rival MEAM. For several metals, traditionally difficult to model at the atomic scale, we demonstrate the ability of this formalism to produce force fields that can lead to new physical insights.
10:20 AM
Machine Learning and Supercomputing to Accelerate the Development of ReaxFF Interatomic Potentials: Naga Sri Harsha Gunda1; Jian Peng1; Yun Kyung Shin2; Sangkeun Lee1; Adri C. T. Van Duin2; Dongwon Shin1; 1Oak Ridge National Laboratory; 2Pennsylvania State University
We demonstrate a workflow that can significantly accelerate the devolvement of high-fidelity interatomic potentials for atomistic simulations, such as molecular dynamics and reactive force field (ReaxFF). An example we use in this presentation is the development of ReaxFF interatomic potentials of bcc Cr with 18 parameters. We use supercomputing to rapidly populate a large volume of training datasets for parameterization study in the context of machine learning (ML). We start with training multiple ML models that can predict the ReaxFF simulation results in several properties. Then we use Markov Chain Monte Carlo to optimize individual model parameters to replicate temperature- and composition-dependent experimental data. The procedure we have developed will provide an effective methodology that can be applied for traversing a high dimensional space for global optimization of modeling parameters. This research was sponsored by the Department of Energy, Vehicle Technologies Office, Propulsion Materials Program.
10:40 AM
Development of Machine Learned SNAP Potentials for Studying Radiation Damage in Materials: Mary Alice Cusentino1; Mitchell Wood1; Aidan Thompson1; 1Sandia National Laboratories
Molecular dynamics (MD) plays a key role in the multi-scale modeling of materials and is particularly well suited to studying radiation effects in materials. However, the accuracy of MD simulations is limited by the interatomic potential (IAP) used. One method to improve the accuracy of IAPs is to use machine learning (ML-IAP) where the ML-IAP can be trained on a large dataset of highly accurate quantum data, typically generated using density functional theory. One such ML-IAP, the Spectral Neighbor Analysis Potential (SNAP), has been applied to study radiation damage in materials with improved accuracy compared to empirical potentials. Development of SNAP potentials for simulating radiation damage in materials will be presented. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.
11:00 AM
Computational Synthesis of Substrates by Crystal Cleavage: Joshua Paul1; Alice Galdi2; Richard Hennig1; 1University of Florida; 2Cornell University
The discovery of novel substrate materials has been dominated by trial and error, opening the opportunity for a systematic search. To identify stable crystal surfaces, we generate bonding networks for materials from the Materials Project database with few-atom primitive cells. For three-dimensional crystals, we systematically break up to three bonds in the atomic bonding network to generate cleaved surfaces. We identify 4,708 unique surfaces across 2,136 bulk crystals. We create monolayers of these surfaces and calculate the work of adhesion and the partially-relaxed surface energy using DFT to discover over 3,500 potential substrates. Following, we relate the work of adhesion and the bonding character of the precursors. The resulting substrate database is epitaxially matched to various thin-film systems to demonstrate the potential impact of the database. The open-source database of substrates and their properties is being made available at MaterialsWeb.org.