Alloy Development for Energy Technologies: ICME Gap Analysis: Machine Learning and Deformation Modeling
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Integrated Computational Materials Engineering Committee
Program Organizers: Ram Devanathan, Pacific Northwest National Laboratory; Raymundo Arroyave, Texas A & M University; Carelyn Campbell, National Institute of Standards and Technology; James Saal, Citrine Informatics
Monday 2:00 PM
March 20, 2023
Room: Sapphire I
Location: Hilton
Session Chair: Raymundo Arroyave, Texas A&M University; Carelyn Campbell, National Institute of Standards and Technology; James Saal, Citrine Informatics; Ram Devanathan, Pacific Northwest National Laboratory
2:00 PM Invited
Voxelized Representations of Atomic Systems for Machine Learning Applications: Surya Kalidindi1; Matthew Barry1; Pranoy Ray1; 1Georgia Institute of Technology
We will present a novel framework employing voxelized atomic structures (VASt) for extracting structure-property models using emergent machine learning tools. In the VASt framework, the atomic structure is quantified by the two-point spatial correlations of its charge density field to serve as regressors for the prediction of the effective properties. The two-point spatial correlations can be utilized directly as the input to a convolutional neural network for implicit feature engineering or projected to a salient low-dimensional feature-space using principal component (PC) analysis, and then correlated to physical properties using Gaussian process regression (GPR). The uncertainty quantification provided by GPR enables an active learning strategy based on Bayesian experiment design, which minimizes the amount of computationally expensive first-principles simulation data required for training. New atomic structures exhibiting desired or tailored properties can be reverse engineered from the PC space representations. We demonstrate the benefits of VASt framework through multiple case studies.
2:30 PM
Unsupervised Techniques for Outlier Identification in Alloy Datasets: Madison Wenzlick1; Osman Mamun2; M.F.N. Taufique2; Ram Devanathan2; Keerti Kappagantula2; Kelly Rose1; Jeffrey Hawk1; 1National Energy Technology Laboratory; 2Pacific Northwest National Laboratory
An increasing emphasis is being placed on the importance of data processing and quality for improving the trustworthiness of machine learning (ML) models and their relevance to material science challenges. Assessing outliers in a dataset can inform where data are not well represented, and where in the data space the resulting model may be less confident. Further, the presence of outliers may result in model overfitting. In this work, we apply dimensionality reduction and unsupervised clustering to two alloy datasets and explore the presence of outliers across the multi-dimensional alloy space. Outliers are assessed relative to each cluster as well as to the overall dataset, and the characteristics of the outlier points are explored to validate the outlier label. The resulting changes in the performance of a ML regression model are investigated after removing outliers. The effect of adding new data on the outlier identification process is explored.
2:50 PM
VPSC's New Clothes: Developing a Modern MATLAB API for Automating High-throughput VPSC Experiments: Benjamin Begley1; Victoria Miller1; 1University of Florida
Though lower fidelity than full-field crystal plasticity models, the computationally inexpensive viscoplastic self-consistent model (VPSC) should excel in the high-throughput, rapid-iteration ICME paradigm. However, the text-file interface creates a steep learning curve, and lack of published automation tools for VPSC limits its potential value for ICME. The authors discuss development of a modern application programming interface (API) in MATLAB which streamlines user interaction without sacrificing the original nuance, and which includes functionality for automating VPSC experiments. The MTEX toolbox—a library of MATLAB code for representing and transforming crystallographic, microstructural, and deformation data—is used as the exemplar for an easy-to-learn modern API, with a planned integration of the VPSC automation API into the MTEX toolbox for wider accessibility. To demonstrate the high-throughput capabilities, a case study uses the VPSC automation API as part of an ICME strategy to optimize the energy efficiency of titanium alloy processing.
3:10 PM Invited
ExtremeMat: towards Microstructure and Composition Sensitive Models for the Creep Deformation of Engineering Steels: Laurent Capolungo1; Arul Kumar1; Ricardo Lebensohn1; Michael Glazoff2; Michael Gao3; Yuki Yamamoto4; 1Los Alamos National Laboratory; 2Idaho National Laboratory; 3National Energy Technology Laboratory; 4Oak Ridge National Laboratory
Structural steels utilized for power generation applications are subjected to particularly complex loading conditions. A robust knowledge of the evolution of the performance steels during service entails that one can establish firm bridges between microstructure, composition and material performance: such is one of the primary goals of the consortium ExtremeMat. In this presentation, focus will be placed on detailing recent advances made by the consortium in what concerns the development and validation of advanced polycrystalline creep models. Among others, leveraging the elastic viscoplastic Fast Fourier Transform polycrystal mechanical solver, a mechanistic constitutive model sensitive to the dislocation content and arrangement, interstitial and substitutional solutes, and precipitates will be introduced. A detailed analysis of the precipitate formation and evolution processes will be presented and rationalized using density functional theory simulations. By applying the model againstexperimental data, the role played by solute vs precipitate strengthening processes will be discussed.
3:40 PM Break
4:10 PM
Data Quality Evaluation and Influence on the Predictability of
Data-Driven Alloy Design
: Sunyong Kwon1; Yukinori Yamamoto1; Jian Peng1; Michael Brady1; Dongwon Shin1; 1Oak Ridge National Laboratory
The lack of high-quality, consistent ground truth data is an often limiting factor in applying machine learning (ML) for alloy design. Although its importance, general practice to evaluate the quality of materials datasets to be used for data analytics has not been established. We propose a quantitative method to evaluate the dataset quality and its influence on the predictability of physics-augmented data-driven surrogate ML models. We use an example creep-rupture experimental data of alumina-forming austenitic (AFA) alloys, of which data has been collected over a decade. The ML model predictability trained by two different qualities of datasets will be discussed. We also present a workflow to assess new alloy compositions worth experimental validation from the ML-predicted millions of hypothetical alloys. The research was sponsored by the Vehicle Technologies Office, the U.S. Department of Energy.
4:30 PM
Design of Creep-resistant Additively Manufactured Stainless Steels for Nuclear Reactors: Pedro Rivera-Diaz-Del-Castillo1; Wei Wen1; Weiling Wang1; Hossein Eskandari Sabzi1; 1Lancaster University
Additive manufacturing (AM) emerges as prominent technology for nuclear reactor design. AMed austenitic stainless steels display improved strength and fatigue properties; but their creep properties and their comparison to wrought alloys remain relatively unknown. Knowledge gaps result from: (1) Creep data can take decades to collect, and AM datasets are scarce. (2) Wrought and AM grain morphology are very different. (3) Linking electron backscatter diffraction (EBSD) data with creep modelling is a challenge, particularly for AM. In this work, crystal plasticity finite element method (CP-FEM) simulations are carried out based on grain structure mapping imported directly from EBSD, and adopting a new physics-based plasticity model. The local stress-strain response during creep deformation is predicted. The tertiary creep stages are also captured through the classic Gurson-Tvergaard- Needleman damage model. This work shows that grain boundary orientation determines creep failure of AM alloys, providing a tool for their design against creep.