AI/Data Informatics: Applications and Uncertainty Quantification at Atomistics and Mesoscales: Session III
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Kamal Choudhary, National Institute of Standards and Technology; Garvit Agarwal, Argonne National Laboratory; Wei Chen, University At Buffalo; Mitchell Wood, Sandia National Laboratories; Vahid Attari, Texas A&M University; Oliver Johnson, Brigham Young University; Richard Hennig, University of Florida
Tuesday 8:30 AM
March 16, 2021
Room: RM 33
Location: TMS2021 Virtual
Session Chair: Mitchell Wood, Sandia national lab; Oliver Johnson, Brigham Young University
8:30 AM
AI Guided High-throughput Exploration of Potential Energy Surfaces: Subramanian Sankaranarayanan1; 1University of Illinois Chicago
Molecular dynamics (MD) is a powerful simulation technique for materials modeling. Nevertheless, a substantial gap persists between AIMD, which is highly accurate but restricted to extremely small sizes, and those based on classical force fields (atomistic and CG) with limited accuracy but access to larger length/time scales.In this talk, I will present some of our recent work on the use of decision trees operating in continuous action space to seamlessly bridge the electronic, atomistic and mesoscopic scales for materials modeling. Our automated ML framework allows for high-throughput exploration of potential energy surfaces and aims to bridge the significant gulf that exists between the handful of research groups that develop new interatomic potential models (often requiring several years of effort) and the increasingly large user community from academia and industry that applies these models. We will present success stories for dozens of different elemental systems across the periodic Table.
9:00 AM
Decision Trees in Continuous Action Space for High-throughput Exploration of Potential Energy Surfaces: Sukriti Manna1; Troy Loeffler1; Rohit Batra1; Suvo Banik1; Henry Chan1; Subramanian Sankaranarayanan1; 1Argonne National Laboratory
The dynamic evolution of nanoclusters and temperature-dependent stability and properties remain unexplored due to lack of their available force field. To mitigate this issue, we use a Monte Carlo Tree Search (MCTS) with reinforcement learning to develop potential models in a high throughput manner for nanoclusters of different elements across the periodic table. To ensure the transferability of these parameters across different size regimes, we used an extensive training data set that encompasses the structural and energetic properties of nanoclusters over a wide range of energy windows. Our parameterized BOP model can accurately capture the structure, energetics, forces, and dynamics of several different elemental clusters across the periodic table. This makes our newly developed scheme and the resulting models to be computationally robust but inexpensive tool for investigating a wide range of materials phenomena across a broad range of nanoclusters.
9:20 AM
Building a Better Database to Learn From; Application to Interatomic Potentials: Mitchell Wood1; Nicholas Lubbers2; Danny Perez2; Charles Sievers1; 1Sandia National Laboratories; 2Los Alamos National Lab
While many research challenges in material physics, chemistry and biology lie just out of reach on current peta-scale machines due to length and time restrictions inherent to Molecular Dynamics, questions of the accuracy of our simulations will continue to linger. A recent trend is to develop machine learned(ML) IAP that demonstrate ab initio levels of accuracy, but are far more computationally efficient. The starting point for all ML applications is a collection of training data which constrains the learned parameters of a model form, but most often which data is included as training is skewed by user-defined heuristics and is liable to fail when extrapolating beyond where training was supplied. In order to provide accurate and transferrable ML-IAP, we have focused our efforts on developing methods to generate training sets that can be used by common computing clusters as well as leadership computing platforms.
9:40 AM
Neural Network Reactive Force Field for C, H, N, O Systems: Pilsun Yoo1; Michael Sakano1; Saaketh Desai1; Mahbubul Islam2; Peilin Liao1; Alejandro Strachan1; 1Purdue University; 2Wayne State University
Reactive force fields with physics-based functions parametrized from first-principles data have been in active development and use for nearly two decades in a wide range of phenomena including chemistry at extreme conditions to operations of electrochemical devices and catalysis. These methods have provided invaluable insights and semi-quantitative understanding. However, reactive force fields result in inaccurate quantitative predictions due to intrinsic limitations of functional forms. Machine learning force fields try to address this shortcoming using physics-agnostic but flexible models trained with extensive data. We developed a neural network reactive force field(NNRF) for C,H,N,O systems to study chemical reactions of high energy nitramine 1,3,5-Trinitroperhydro-1,3,5-triazine(RDX). Training data was collected using a semi-automated iterative procedure to include relevant chemical reaction paths until NNRF predicts energy and forces in a desired accuracy. The predictions of NNRF for vibrational properties, and kinetics of thermal decomposition were quantitatively closer to experimental data than available reactive force fields.
10:00 AM
Accelerating Phase-field Predictions via Machine Learning Trained Surrogate Models: David Montes de Oca Zapiain1; James Stewart1; Remi Dingreville1; 1Sandia National Laboratories
The phase-field method is a powerful and versatile computational approach for modeling the evolution of the microstructure and properties of a wide variety of physical, chemical and biological systems. However, existing high-fidelity phase-field models are inherently computationally expensive, requiring high-performance computing resources and sophisticated numerical integration schemes to achieve a useful degree of accuracy. In this presentation, we present a computationally inexpensive, accurate, data-driven surrogate model that directly learns the microstructural evolution of targeted systems by combining phase-field and history-dependent machine learning techniques. We integrate a statistically-representative, low-dimensional description of the microstructure, obtained directly from phase-field simulations, with either a Time-Series Multivariate Adaptive Regression Splines (TSMARS) autoregressive algorithm or a Long Short-Term Memory (LSTM) neural network. Our machine-learning-trained surrogate model shows the best performance and accurately predicts the non-linear microstructure evolution of a two-phase mixture during spinodal decomposition in seconds, without the need for "on-the-fly" solutions of the phase-field equations-of-motion.
10:20 AM
Simultaneous Development and Robust Optimization of a Microstructure Dependent Material Model: Leveraging Sequential Monte-Carlo Methods to Enhance Symbolic Regression Analysis: Karl Garbrecht1; Nolan Strauss1; Geoffrey Bomarito2; Patrick Leser2; Jacob Hochhalter1; 1University of Utah; 2NASA
Recent microstructure characterization techniques combined with Symbolic Regression (SR) analysis has been proven to generate white box plasticity models well suited for incorporation into FEA software. The current work builds upon those efforts and demonstrates the applicability of Sequential Monte-Carlo (SMC) methods within SR analysis to condense model development and robust optimization into a single, co-dependent process. In this project, SMC methods provide a mechanism through which the observed microstructure features and associated variability can be incorporated into the discovery phase of model development and simultaneously recover approximate parameter distributions. The demonstration utilized a data set of tensile test results from sample specimens with corresponding EBSD data. Synthetic volume elements with statistically equivalent microstructure to the sample specimens combined with simulated tensile test data for each volume element was used as training data for a SMC-SR algorithm. The resulting model was validated with data from the original empirical data.
10:40 AM
Exploring Metastability and Mapping Metastable Phase Diagrams Using Machine Learning: Srilok Srinivasan1; Rohit Batra1; Duan Luo1; Troy Loeffler1; Sukriti Manna1; Henry Chan1; Liuxiang Yang2; Wenge Yang2; Jianguo Wen1; Pierre Darancet1; Subramanian Sankaranarayanan1; 1Argonne National Laboratory; 2Center for High Pressure Science and Technology Advanced Research
We introduce an automated workflow that integrates first-principles physics and atomistic simulations with machine learning (ML), and high-performance computing to allow for exploration of the metastable phases of a given elemental composition and construct "metastable" phase diagrams for materials. We demonstrate our workflow on a pure carbon system which exhibits a vast number of metastable phases without parent in equilibrium. Moreover, we build a neural network model which learns to predict the equation of state of the metastable phases given only the structural information. We identify domains of relative stability and synthesizability of metastable materials using our metastable phase diagram. The predictions of our workflow are confirmed using high-resolution transmission electron microscopy (HRTEM) after high temperature high pressure treatment of graphite in a diamond anvil cell. Our introduced approach for constructing the metastable phase diagram is general and broadly applicable to single and multi-component systems.
11:00 AM
Machine Learning Guided Discovery of Novel Oxide Perovskites for Scintillator Applications: Anjana Talapatra1; Blas Uberuaga1; Christopher Stanek1; Ghanshyam Pilania1; 1Los Alamos National Laboratory
Scintillators have wide-ranging applications, from medical imaging to radiation. Despite a pressing need for improved scintillators, the discovery of new scintillators relies on a laborious, time-intensive, trial-and-error approach; yielding little physical insight and leaving a vast space of potentially revolutionary materials unexplored. To accelerate the discovery of optimal scintillators, we are developing an adaptive design framework that couples high-throughput experiments, first-principles computations and machine learning to (1) screen a large chemical space of probable scintillator chemistries and (2) identify chemistries enabling further tuning of the underlying electronic structure for band edge and defect engineering. This talk focuses on the details of the screening strategy applied to the class of single and double oxide perovskites. Specifically, we present a novel hierarchical down-selection approach that employs non-traditional structure maps, DFT-based stability analysis, machine learning models for bandgap predictions and physics-based classification to efficiently predict minimal favorable electronic structure for a viable scintillator.