AI/Data Informatics: Applications and Uncertainty Quantification at Atomistics and Mesoscales: Poster Session I
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 5:30 PM
March 16, 2021
Room: RM 33
Location: TMS2021 Virtual
A Machine Learning Investigation of Crystallographic Parameters for Abnormal Grain Growth: Meizhong Lyu1; Joseph Pauza1; Ryan Cohn1; Elizabeth Holm1; 1Carnegie Mellon University
Abnormal grain growth is characterized by rapid growth of a few grains in polycrystalline materials. Although theoretical growth advantages and mechanisms have been widely studied over the past decades, researchers cannot precisely predict the occurrence of abnormal growth from an initial microstructure. This study aims at sifting the favorable crystallographic parameters for abnormal grain growth by quantitative analysis using data mining and machine learning. Two-dimensional simulations were based on the Monte Carlo Potts model with initial matrix grains varying in boundary mobilities. For candidate grains with an initial size advantage, abnormal grain growth is related to the geometric, topological, and crystallographic neighborhood of the grain. Given the mobility type of particular nearest neighbor grains, the logistic regression classification accuracy can be quite high. Further improvement in accuracy is achieved by introducing information from the second and third nearest neighbors.
A Sensitivity Analysis of Microstructure-Based Model for U-10Mo Hot Rolling and Annealing: Yucheng Fu1; William E Frazier III1; Kyoo Sil Choi1; Lei Li1; Zhijie Xu1; Vineet V Joshi1; Ayoub Soulami1; 1PNNL
The manufacturing of low-enriched uranium alloyed with 10 wt% molybdenum (U-10Mo) involves multiple hot rolling passes and annealing steps after homogenization. The resultant microstructure depends on various casting attributes such as grain size, uranium carbide (UC) volume fraction, size, and distribution; and rolling reductions. Over 150 numerical simulations had been conducted using an integrated microstructure – based Finite Element – Monte Carlo Potts model framework. A sensitivity analysis using the stragihforward one-factor-at-a-time (OAT) method and multivariate adaptive regression spline (MARS) surrogate model was performed to analyze the influence of the model input parameters on the final grain size, distribution, and the recrystallized percentage.
Machine Learning Approach of Molecular Dynamics Simulations for Body-Centered Cubic Zirconium: Vanessa Meraz1; Bethuel Khamala1; Armando Garcia1; Adrian De La Rocha1; Jorge Munoz1; Tess Smidt2; Wibe de Jong2; 1The University of Texas at El Paso; 2Lawrence Berkeley National Laboratory
Although an extremely capable tool, one of the steadfast shortcomings of density functional theory based molecular dynamics (DFT-MD) simulations have been their computational costs. Because of the ability of these quantum mechanics based calculations to predict thermodynamic properties, there is great interest in a machine learning approach. We built a high-quality DFT-MD dataset for the body-centered cubic structure (bcc) of zirconium (Zr), which is stable at a high temperature. The dataset is used to build a Euclidean symmetry equivariant neural network (E3NN) model to map the energy landscape of the system as a function of the atomic displacements from the ideal lattice. Given that the framework of the network uses physical concepts which allows for the ability to train with less data, we generate predictions, allowing us to compare them with our dataset.
Microstructure-driven Parameter Calibration for Mesoscale Simulation: Theron Rodgers1; Dan Bolintineanu1; Daniel Moser1; Reeju Pokharel2; 1Sandia National Laboratories; 2Los Alamos National Laboratory
Mesoscale microstructure simulation techniques often have parameters that are difficult or impossible to determine experimentally. The growing use of advanced microstructure characterization techniques offers increasing availability of quantitative experimental data. Here, we present an approach to calibrate unknown simulation data through direct comparison of simulated microstructures with experimental data. Two studies utilizing this approach will be presented: calibration of a rules-based model of thermal spray deposition with 3D µCT data and calibration of 3D grain growth simulations with 2D EBSD data. Parameter calibration is performed using optimization methods available in Sandia’s Open Source Dakota software. A range of optimization approaches are explored including Gaussian process and the use of surrogate models.SNL is managed and operated by NTESS under DOE NNSA contract DE-NA000352
Mining Structure-property Linkages in Nonporous Materials Using Interpretative Deep Learning Approach: Haomin Liu1; Niaz Abdolrahim1; 1University of Rochester
Relating the role of the microstructure to mechanical properties of nanoporous (NP) materials is a complicated problem. Deep Learning methods have shown strong performance in the mechanical design of materials by providing high learning capability. In this study, a deep learning approach is designed to model an elastic homogenization structure-property relationship of NP materials. Our model predicts the stiffness of the NP structure with a wide range of microstructures while exhibiting high accuracy and low computational cost. The main drawback of deep learning algorithms is poor interpretability. Thus, a novel interpretation method is developed to unravel the salient features of the microstructure that lead to better stiffness. Our interpretation method identifies the effective microstructure parameters that strongly impact stiffness, including surface area, dangling ligaments, and connectivity. Our interpretative deep learning framework can be transferred to build other structure-property relationships such as chemo-mechanical properties of nanomaterials in the future.