AI/Data Informatics: Applications and Uncertainty Quantification at Atomistics and Mesoscales: Session II
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
Monday 2:00 PM
March 15, 2021
Room: RM 33
Location: TMS2021 Virtual
Session Chair: Vahid Attari, Texas A&M University; Wei Chen, Illinois Institute of Technology
2:00 PM
Uncertainty Quantification in Computational Thermodynamics - From the Atomistic to the Continuum Scale: Noah Paulson1; Joshua Gabriel1; Thien Duong1; Marius Stan1; 1Argonne National Laboratory
The design of materials requires comprehensive descriptions of the thermodynamic properties of materials and their constituents. The CALculation of PHase Diagrams (CALPHAD) approach leverages measured and calculated thermodynamic properties and phase stabilities to calibrate mathematical thermodynamic descriptions and provide critical predictive capabilities. In practice, uncertainties are present in the thermodynamic descriptions and resulting predictions as a result of the distribution of calibration data and differences in error between data sources. We describe several approaches, developed at Argonne National Laboratory, that provide thermodynamic information to CALPHAD through atomistic calculations and machine learning that consider the multiple origins of uncertainty and the propagation to higher length-scale predictions. Specifically, we explore uncertainty of enthalpy and specific heat derived from Density Functional Theory (DFT) including an efficient k-nearest-neighbors based acceleration scheme and machine learned interatomic potentials, and Bayesian methods to automatically weight thermodynamic datasets used in the calibration of CALPHAD models.
2:30 PM
Bayesian Inference and Uncertainty Quantification of Grain Boundary Properties: Sterling Baird1; Brandon Snow1; Alexia Bigelow1; David Fullwood1; Eric Homer1; Oliver Johnson; 1Brigham Young University
The development of structure-property models for grain boundaries (GBs) has been historically challenging due to a combination of the high cost (both monetary and temporal) of bicrystal synthesis/characterization/property measurement, and the high-dimensionality of the space. Consequently there are few models that exist, and most of those that do are restricted to high-symmetry subspaces (i.e. they do not consider the full 5D GB character space). We demonstrate the use of Bayesian inference techniques to infer a fully 5D structure-property model for GB energy using published databases. The approach naturally handles underdetermined systems (i.e. limited data), indirect measurements, and yields quantified uncertainty for the inferred structure-property model.
2:50 PM
A Bayesian Optimization Framework for Exploring the Grain Boundary Manifold: Leila Khalili1; Owen Rettenmaier1; Srikanth Patala1; 1North Carolina State University
Over the past few decades, materials scientists have increasingly come to the realization that the distribution and connectivity of different grain boundary (GB) types contribute to the mechanical and functional properties of polycrystalline materials. Despite the role of GB structure in transport and failure mechanisms having been investigated for more than half a century, few robust GB crystallography-property relationships are yet known; this is at least partly due to the inherent complexity associated with the five-dimensional configuration space in which they reside. In this talk, we will introduce a Bayesian optimization framework for constructing GB crystallography-property relationships by sampling the topologically complex (arising due to the bi-crystallographic symmetries) grain boundary manifold in an efficient manner.
3:10 PM
Machine Learning for Predicting Grain Boundary Properties: Lingxiao Mu1; Elizabeth Holm1; 1Carnegie Mellon University
Grain boundary properties, such as energy and mobility, are always significant in materials science. Grain boundary properties are not only related to crystallography (macroscopic structure), but also related to atomic positions (microscopic structure). However, since the five-dimensional space of grain boundaries is complex, prediction of grain boundary properties requires a large amount of data or computation time. Data science and machine learning offer an alternative methodology for predicting boundary properties. In this project, we develop crystallographic descriptors and use machine learning to train a model to predict the grain boundary energy from macroscopic structure, and we compare the results to models trained to predict energy from microscopic structure information. Even though there are some limitations, the results are reasonable and accurate compared to the results predicted from atomic structure. This method may provide a computationally efficient approach for calculating grain boundary energy for mesoscale simulations.
3:30 PM
Machine Learning Prediction of Defect Formation Energies: Vinit Sharma1; Pankaj Kumar2; Pratibha Dev2; Ghanshyam Pilania3; 1University of Tennessee Knoxville; 2Howard University; 3Los Alamos National Laboratory
The feasibility and the stability of a defect in the host lattice is usually obtained via experiments and/or through detailed quantum mechanical calculations. Both of these conventional routes are expensive and time consuming. An alternative is a data-driven machine learning (ML)-based approach. Here, using ML techniques we identify the factors that influence defect formation energy in two material classes namely perovskites and MXenes. Using elemental properties as features and random forest regression, we demonstrate a systematic approach to down select the important features, establishing a framework for accurate predictions of the defect formation energy. Our work reveals previously unknown Hume-Rothery-like rules for complex material systems, chemical trends, and the interplay between stability and underlying chemistries. Hence, these results showcase the efficacy of ML tools in identifying and quantifying different feature-dependencies and provide a promising route toward dopant selection. The framework itself is general and can be applied to other material classes.
3:50 PM
Accuracy, Uncertainty, Inspectability: The Benefits of Compositionally-restricted Attention-based Networks: Taylor Sparks1; Steven Kauwe1; Ryan Murdock1; Anthony Wang2; 1University of Utah; 2Technische Universitat Berlin
We describe a new model architecture, the Compositionally-Restricted Attention-Based Network (CrabNet). CrabNet generates high-fidelity predictions based on the self-attention mechanism, a fundamental component of the transformer architecture which revolutionized natural language processing. The transformer encoder uses self-attention to encode the context-dependent behavior for the components within a system. In physical environments, elements contribute differently to a material's property based on the materials system itself. For example, boron behaving as an electrical dopant in one system while behaving as a mechanical strengthening bond modification in another. CrabNet's ability to potentially capture this type of context-dependent behavior allows for highly accurate model predictions. Importantly, CrabNet generates simple and inspectable self-attention maps. These attention maps govern the learned material property by representing element importance and interactions. The visualization and analysis of these attention maps are available during training and inference periods.
4:10 PM
A Probabilistic Approach with Built-in Uncertainty Quantification for the Calibration of a Superelastic Constitutive Model from Full-field Strain Data: Harshad Paranjape1; Kenneth Aycock2; Craig Bonsignore1; Jason Weaver2; Brent Craven2; Thomas Duerig1; 1Confluent Medical; 2U.S. Food and Drug Administration
We implement an approach using Bayesian inference and machine learning to calibrate the material parameters of a finite element constitutive model for the superelastic deformation of NiTi shape memory alloy. The calibration scheme uses full-field surface strain measurements obtained using digital image correlation and global load data from tensile tests as the inputs for calibration. We use machine learning to create a surrogate model for the finite element constitutive model. We use the surrogate model to perform the Monte Carlo sampling as part of the calibration process. We demonstrate, verify, and partially validate the calibration results through various examples. We also demonstrate how the uncertainty in the calibrated superelastic material parameters can propagate to a subsequent simulation of fatigue loading. The machine learning surrogate model improves the computational efficiency of the calibration scheme and calibration using the full-field strain data improves the accuracy of subsequent simulations of local deformation.
4:30 PM
Uncertainty Quantification of Microstructures with a New Technique: Shape Moment Invariants: Arulmurugan Senthilnathan1; Pinar Acar1; 1Virginia Tech
Microstructure reconstruction is an efficient strategy to predict the microstructure evolution over large domains given small-scale experimental data. However, such prediction is influenced by the effects of the uncertainties in computations and experiments. While the uncertainty quantification (UQ) of crystallographic texture and grain size is addressed with state-of-the-art methods, the UQ of grain shapes is still an unexplored research challenge. We present a novel UQ methodology for metallic microstructures that utilizes the shape moment invariants in physics to mathematically quantify the uncertainty in grain shapes. The experimental data of the aerospace alloy, Titanium-7wt%- Aluminum (Ti-7Al), is used to generate synthetic microstructures using Markov Random Field (MRF). The UQ formulation is first applied to predict the variations in the texture, grain sizes and shapes of the synthetic microstructures. Next, the propagation of the microstructural uncertainty on the elasto-plastic material properties is studied by utilizing the UQ formulation and crystal plasticity simulations.
4:50 PM
Predicting Adsorption Energies and Surface Pourbaix Diagram of Metal NPs by GCNN Method: Kihoon Bang1; Youngtae Park1; Donghun Kim2; Sang Soo Han2; Hyuck Mo Lee1; 1KAIST; 2KIST
A surface Pourbaix diagram is useful in investigating stability of materials, especially, for catalysts. However, to build it, numerous DFT calculations are needed to obtain adsorption energies and their computational costs are quite high for nanoparticles (NPs) with a large number of atoms. To overcome it, we used a graph based convolutional neural network (GCNN) model to predict adsorption energies of adsorbates on NPs and build the surface Pourbaix diagram of NPs from predicted values. By our GCNN model, we could predict adsorption energies on Pt NPs with a reasonable accuracy for not only a single adsorbate but also multiple adsorptions. Then, we constructed a surface Pourbaix diagram of Pt NPs from predicted adsorption energies by our model and it is similar to the diagram by DFT calculated values. We also predicted adsorption energies on large NPs, which are not used for training, and build their surface Pourbaix diagram.