AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Material Design III
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Saurabh Puri, VulcanForms Inc; Dennis Dimiduk, BlueQuartz Software LLC; Darren Pagan, Pennsylvania State University; Anthony Rollett, Carnegie Mellon University; Francesca Tavazza, National Institute of Standards and Technology; Christopher Woodward, Air Force Research Laboratory
Wednesday 8:30 AM
March 2, 2022
Room: 256A
Location: Anaheim Convention Center
Session Chair: Tiankai Yao, Idaho National Laboratory
8:30 AM
Neural Message Passing for Prediction of Abnormal Grain Growth in Monte Carlo Simulations of Polycrystalline Materials: Ryan Cohn1; Elizabeth Holm1; 1Carnegie Mellon University
Abnormal grain growth (AGG) significantly influences the properties of materials but is difficult to observe experimentally. Researchers have replicated this phenomenon through Monte Carlo simulations, but have not generated a predictive model that can infer which initial microstructures will undergo AGG during simulated processing. One challenge with modeling grain structures is determining an effective representation to use as an input to a model. Neural message passing allows models to operate on irregular graph structures, which are useful for modeling networks of connected grains but incompatible with traditional neural network architectures. In this study, we apply neural message passing to predicting the occurrence of AGG in Monte Carlo simulations using only the initial state of the system as input. Preliminary results indicate that even a simple graph-based model achieves 75% prediction accuracy and outperforms a comparable computer vision approach.
8:50 AM
Application of Compositionally-restricted Attention-based Network (CrabNet) for Screening Candidate Dispersed Phases for Designing High Strength Alloys: Trupti Mohanty1; Fan Zhang2; K. S. Ravi Chandran1; Taylor D. Sparks1; 1University of Utah; 2CompuTherm,LLC
Dispersion of hard phase particles in the parent phase matrix has been one of the effective methods in designing high strength alloys for numerous advanced applications. The screening of potential dispersed phases with desired hardness from huge data space has been a challenging task. Machine learning tools can be quite useful for rapid screening of suitable dispersed phases by predicting their hardness values, and thereby enabling accelerated development of high strength alloys. The present work focuses on applicability of compositionally restricted attention-based network (CrabNet) for screening some of the potential alloy phases based on predicted hardness values leveraging their elastic moduli, which can be utilized as dispersed phases in designing of alloys. Equilibrium phase composition was derived using CALPHAD and subsequently hardness values of the phases was predicted invoking the trained CrabNet. The training of CrabNet for prediction of elastic moduli on benchmark datasets and its performance have been discussed.
9:10 AM
Feature Selection and Interpretation for Machine Learning Models: Reducing the Dimensionality of Complex Concentrated Alloys: Zachary Mcclure1; Austin Hernandez1; Michael Titus1; Alejandro Strachan1; 1Purdue University
The inherent high dimensionality of complex concentrated alloy design prohibits full exploration of the material space via experimental means. Therefore, large efforts to model properties and phenomena of the design space coupled with validation is critical for efficient procedure. Since available datasets are limited, we often turn to machine learning models with carefully engineered features. With increased feature count is the reward of a more complex and accurate model. However, this is often at the cost of interpretability of individual features. In this study we develop random forest regression models with quantified uncertainties to predict the yield strength of CCAs, followed by an analysis of our selected features using game theory approximations. We use the methods of Shapely coefficients to score and evaluate the impact of our features, and offer explanations for individual feature impact on model predictions.
9:30 AM
Efficient Optimization of Variable and Uncertain Additive Manufacturing Processes Using Machine Learning: Maher Alghalayini1; Ali Khosravani2; Surya Kalidindi3; Chris Paredis1; Fadi Abdeljawad1; 1Clemson University; 2Multiscale Technologies; 3Georgia Institute of Technology
Recent experimental studies have shown considerable variations in the properties of additively manufactured (AM) materials, even under the same processing conditions. Accounting for such variabilities in materials design is of critical importance. However, probing the AM process parameters is an experimentally intensive task due to the large AM design space. In this study, we develop a sequential design approach that learns from previously tested AM design points and surveys the AM process parameter design space in order to adaptively select the next AM process parameters that result in maximum information gain. The novelty lies in the use of Value-of-Information Theory to define the optimization criteria, and the flexibility in the number of design points and AM samples added in each iteration. Method performance is tested on synthetic data. This work is expected to result in a novel efficient sequential method that optimizes variable and uncertain processes.
9:50 AM Break
10:10 AM
NOW ON-DEMAND ONLY – Uncertainty Quantification and Propagation in Prediction of Solid-liquid Interfacial Properties and Solidification Microstructures: Sepideh Kavousi1; Mohsen Asle Zaeem1; 1Colorado School of Mines
Molecular Dynamics (MD) calculations are often associated with aleatoric uncertainty arising from its intrinsic chaotic nature. In this study, we will first perform quantitative estimations of the uncertainty for different high-temperature material properties of Al-Cu binary system. We quantify how various parameters such as equilibration/simulation time, system size, and presence of defects alter different high-temperature material properties such as melting point, phase diagram, interface free energy, kinetic coefficient, and associated anisotropy parameters. These properties are essential for performing atomistic-informed quantitative phase-field modeling of solidification. By integrating the MD results with a quantitative phase-field model of solidification (Acta Materialia 211 (2021) 116885), we investigate how the uncertainty propagates through the modeling hierarchy. For each property, we select multiple samples from the input-space and perform phase-field simulations under different solidification conditions (temperature gradient, solidification velocity) to quantify how the uncertainty in each input property affect the model outputs.
10:30 AM
Understanding Fission Gas Bubble Distribution, Lanthanide Transportation, and Thermal Conductivity Degradation in Neutron-irradiated α-U Using Machine Learning: Tiankai Yao1; Lu Cai1; Fei Xu2; Fidelma Dilemma1; Michael Benson1; Daniel Murray1; Cynthia Adkins1; Joshua Kane1; Min Xian3; Luca Capriotti1; 1Idaho National Laboratory; 2Grand View University; 3University of Idaho
U-10Zr based metallic nuclear fuel is the leading candidate for next-generation sodium-cooled fast reactors in the United States. US research reactors have used and tested this fuel type since the 1960s and accumulated considerable experience and knowledge about fuel performance. Most of the knowledge, however, remains empirical. This paper proposes an image data-driven machine learning approach, coupled with domain knowledge provided by advanced post-irradiation examination, to provide unprecedented quantified insights into the morphology, size, density, and the connectivity of fission gas bubbles and their effect on the fission product transportation and thermal conductivity. The approach can be modified to study other irradiation effects, such as secondary phase redistribution and gaseous fuel swelling in other irradiated nuclear fuels.