Algorithm Development in Materials Science and Engineering: Machine Learning Algorithms and Computational Modeling for Study and Design Materials
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee, TMS: Integrated Computational Materials Engineering Committee, TMS: Phase Transformations Committee, TMS: Solidification Committee
Program Organizers: Mohsen Asle Zaeem, Colorado School of Mines; Mikhail Mendelev, NASA ARC; Bryan Wong, University of California, Riverside; Ebrahim Asadi, University of Memphis; Garritt Tucker, Colorado School of Mines; Charudatta Phatak, Argonne National Laboratory; Bryce Meredig, Travertine Labs LLC
Monday 2:00 PM
March 15, 2021
Room: RM 36
Location: TMS2021 Virtual
Session Chair: Bryan Wong, University of California, Riverside; Bryce Meredig, Travertine Labs LLC
2:00 PM Invited
Multi-Information Source Bayesian Optimization Applied to Materials Design: Raymundo Arroyave1; Danial Khatamsaz1; Richard Couperthwaite1; Abhilash Molkeri1; Douglass Allaire1; Ankit Srivastava1; 1Texas A&M University
Materials design involves the solution to an inverse problem connecting desired performance to a target microstructure and ultimately to a required chemistry and processing combinations. ICME prescribes the use of simulations along the PSPP chain as a way to solve this problem. An implicit assumption of ICME is the existence of one single model per element of the PSPP chain, even though in principle there are potentially many sources (physics-based models, ML predictions, expert opinion) that could be used at each stage of the process. In this talk, I will present some recent work in which we have developed, deployed, and tested advanced Bayesian Optimization (BO) frameworks that are capable of fusing multiple information sources by exploiting the correlations among them and with the ground truth. We show how these schemes are superior to conventional BO-based materials design and showcase some recent developments including active subspace and batch mode optimization.
2:30 PM
Understanding Grain Boundary Metastability Using the SOAP Descriptor and Unsupervised Machine Learning Techniques: Lydia Serafin1; Derek Hensley1; Jay Spendlove1; Gus Hart1; Eric Homer1; 1Brigham Young University
While it has long been known that grain boundaries (GBs) exhibit a diversity of atomic configurations, our ability to quantify these microscopic configurations and link them to observed properties remains a challenge. Recent work has shown the importance of the metastable GB states in addition to the minimum energy configurations commonly studied. Unfortunately, the added complexity of these configurations makes exploration of this space difficult. We use unsupervised machine learning techniques to examine this space and group GB states into unique clusters, that may correspond to potential energy basins in the GB configuration space. Since annotated datasets of this type are not readily available, this makes validation of the unsupervised machine learning techniques difficult. We will present our efforts to validate these approaches using an annotated dataset, where success of unsupervised machine learning techniques in characterizing GB structure would present an important step forward in grain boundary design.
2:50 PM
Grain Boundary Network Optimization through Human Computation and Machine Learning: Christopher Adair1; Oliver Johnson1; 1Brigham Young University
Grain boundary defect networks are inherently complex, high-dimensional features that influence macroscopic material properties. While steps have been taken to model these influences and describe the mesoscale behavior of grain boundary networks, the dimensionality of the space makes design and optimization of the configurational space computationally prohibitive for large 3D simulations. We show that through human computational units in a video game setting, the overall dimensionality of the grain boundary network design problem can be reduced, reproduced through machine learning, and from their strategies draw insights into optimal grain boundary network features.
3:10 PM Invited
Deep Learning for Characterization of Deformation Induced Damage: Ulrich Kerzel1; Setareh Medghalchi2; Carl Kusche2; Talal Al-Samman2; Sandra Korte-Kerzel2; 1IUBH; 2RWTH Aachen University
Machine learning and artificial intelligence have made enormous progress in recent years, in particular in computer vision, opening new frontiers in materials research. One of the main challenges is the analysis of the statistical effects, requiring the automated analysis of many thousands of large scale images at high resolution. Using the example of DP800 steel, we show that a combination of multiple deep neural networks can reliably identify the most prevalent damage mechanisms at a level comparable to human experts. Using a range of data augmentation as well as regularization techniques, we also demonstrate that this approach can be made more general and flexible to allow the analysis of specimens that have been subjected to different damage and strain paths, without having to obtain a large sample of manually labelled ground truth data to re-train the networks for each new use-case.
3:40 PM
Automatic Segmentation of Microstructures in Steel Using Machine Learning Methods: Hoheok Kim1; Junya Inoue1; Tadashi Kasuya1; 1The University of Tokyo
Microstructure of a material greatly influences its mechanical and chemical properties, so many endeavors have been made to characterize microstructure. Especially, microstructural characterization of steel materials is very challenging due to the existence of various constituent phases such as ferrite, pearlite, bainite, martensite, etc. Recently, researches using machine learning methods have actively conducted and shown good performances in classifying microstructures. However, those approaches are generally based on supervised learning algorithms that require preparation of labeled datasets, which is not only time-consuming but also difficult even for experts. In this study, we propose an unsupervised algorithm that performs microstructure segmentation without the need for labeled datasets. The new method, which is a combination of convolutional neural networks and a superpixel algorithm, is applied to various steel microstructure images and the results show that microstructures are well divided into their constituent phases.
4:00 PM
2D Microstructure Reconstruction for SEM via Non-local Patch-based Image Inpainting: Anh Tran1; Hoang Tran2; 1Sandia National Laboratories; 2Oak Ridge National Laboratory
Microstructure reconstruction problems are usually limited to the representation with finitely many numbers of phases, e.g. binary and ternary. However, images of microstructure obtained through experimental, for example, using a microscope, are often represented as an RGB or grayscale image. In this talk, a microstructure reconstruction method based on image inpainting techniques, which produces statistically equivalent microstructures at the fidelity of experiments, is presented without introducing any physics-based microstructure descriptor. The image texture is fully preserved, while the resulting microstructure images are of high quality. A significant advantage of the proposed method is to remedy the data scarcity problem in materials science, where experimental data is scare and hard to obtain. The method is demonstrated using the UltraHigh Carbon Steel micrograph DataBase (UHCSDB). Tran, A., & Tran, H. (2019). Data-driven high-fidelity 2D microstructure reconstruction via non-local patch-based image inpainting. Acta Materialia, 178, 207-218.
4:20 PM Invited
AI-assisted Analysis of Flame Stability: Marius Stan1; Jessica Pan2; Noah Paulson1; Joseph Libera1; 1Argonne National Laboratory; 2Princeton University
Current limitations in the understanding of flame stability are impeding the reliable production of nanoparticles using methods such as Flame Spray Pyrolysis (FSP). We present a methodology and algorithms to detect and control the stability of flames by using components of Artificial Intelligence such as machine learning, computer vision and reduced-order modeling. The methodology starts with analyzing the brightness of the anchor point of the flame, followed by data analysis via unsupervised and supervised machine learning techniques such as principal component analysis and object detection classifiers. The driving algorithm can track and classify FSP flame conditions in real time and alert users of instabilities that can affect the quality and safety of the material synthesis process.
4:50 PM
Neural Network Model of He Diffusion in W-based High Entropy Alloys: Gustavo Esteban-Manzanares1; Enrique Martínez2; Duc Nguyen2; Javier Llorca3; 1IMDEA Materials Institute; 2Los Alamos National Laboratory; 3IMDEA Materials Institute & Technical University of Madrid
In this work the diffusion of He particles in a W 0.35Ta 0.15Cr 0.15V high entropy alloy is examined. To this end, the formation energy of He interstitial in a randomly distributed W Ta Cr V is computed through density functional theory. Furthermore nudged elastic band calculations are devoted to calculate the energy barrier for He to diffuse inside the lattice. A cluster expansion formalism is used to model the activation energy of He diffusion as a function of the lattice occupation vector obtaining the effective cluster interaction within a neural network framework. First principle calculation results are used to train the model while its accuracy is examined by cross validation. This methodology is suitable to be applied in order to model other metallic alloys.
5:10 PM
Comparison of Correction Schemes for Charged Point Defects in 2D Materials: Preston Vargas1; Anne Marie Tan1; Biswas Rijal1; Richard Hennig1; 1University of Florida
The growing interest in the band engineering of 2D materials using charged defects raises the need for accurate and efficient methods for modeling these systems. A compensating jellium background charge is often employed when modeling charged defects within plane-wave density functional theory, causing the defect formation energy to diverge with vacuum spacing. Freysoldt and Neugebauer developed a method to remove the energy contributions due to incorrect electrostatic boundary conditions using a surrogate model. An alternative approach by Wu, Zhang, and Pantelides models charged defects as excited neutral defects, avoiding jellium altogether. We compare and benchmark both methods by applying them to the well-studied 2D material MoS2. We quantify the effect of supercell size and choice of modeling parameters on each method's accuracy and computational efficiency and discuss the circumstances under which each approach may be suitable for accurate calculations of charged defects in 2D materials.