Algorithm Development in Materials Science and Engineering: ML Algorithms and Their Applications I
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; Garritt Tucker, Colorado School of Mines; Ebrahim Asadi, University of Memphis; Bryan Wong, University of California, Riverside; Sam Reeve, Oak Ridge National Laboratory; Enrique Martinez Saez, Clemson University; Adrian Sabau, Oak Ridge National Laboratory
Wednesday 8:30 AM
March 2, 2022
Room: 253A
Location: Anaheim Convention Center
Session Chair: Garritt Tucker, Colorado School of Mines; Enrique Martinez Saez, Clemson University
8:30 AM Invited
Capturing Nanoscale Lattice Variations by Applying AI-powered Computer Vision Techniques on Synthetic X-ray Diffraction Data: Niaz Abdolrahim1; 1University of Rochester
Recent advancements in X-ray scattering techniques allow for real-time probe of physical structure of materials at the molecular and nanoscales. However, real-time x-ray crystallography generates very large data that almost makes it impossible to store and analyze manually by experts. This research combines molecular dynamics (MD) simulations with AI-powered computer vision techniques to mine information-rich x-ray diffraction data to filter and detect lattice-level mechanisms that are responsible for phase transformation and other plastic deformation processes. Large-scale MD simulations will be performed on materials that undergo various physical phenomena (such as nucleation of defects, phase transformation, slip and twinning, etc.) and generate large datasets. AI-powered computer vision techniques will then be developed to automatically learn image features and identify lattice-level mechanisms from synthetic data. The proposed framework will be significantly useful in capturing and analyzing new and unknown phenomena in materials under extreme conditions when no prior knowledge is available.
9:00 AM
Application of Information Theory in Molecular Dynamics Simulations: Khaled Talaat1; Osman Anderoglu1; 1University of New Mexico
Molecular dynamics is evolving as an important tool in nuclear energy applications. Among the uses of molecular dynamics in nuclear energy is studying the thermophysical properties of new alloys under irradiation and high temperatures. Discrepancies in calculated material properties have been reported in many areas in molecular dynamics due to inadequate simulation times and inconsistent assessments of convergence. We adapted information theory methods that were previously used in Monte Carlo neutron transport in nuclear reactors to molecular dynamics allowing for reliable assessments of convergence for different types of non-equilibrium molecular dynamics simulations. In this work, various formulations of information entropy are investigated and applied to radiation damage simulations and lattice thermal conductivity calculations. The method has also been demonstrated to be applicable to both solids and fluids.
9:20 AM
Clustering Algorithms for Nanomechanical Property Mapping and Resultant Microstructural Constituent Quantification: Bryer Sousa1; Christopher Vieira1; Rodica Neamtu1; Danielle Cote1; 1Worcester Polytechnic Institute
Tacit assumptions have been made about the suitability of two primary data-driven deconvolution algorithms concerning large (10,000+) data sets captured using nanoindentation grid array measurements, including (1) probability density function determination and (2) k-means clustering and deconvolution. Recent works have found k-means clustering and probability density function fitting and deconvolution to be applicable; however, little forethought was afforded to algorithmic compatibility for nanoindentation mapping data. The present work highlights how said approaches can be applied, their limitations, the need for data pre-processing before clustering and statistical analysis, and alternatively appropriate clustering algorithms. Equally spaced apart indents (and therefore measured properties) at each recorded nanoindentation location are collectively processed via high-resolution mechanical property mapping algorithms. Clustering and mapping algorithms also explored include k-medoids, agglomerative clustering, spectral clustering, BIRCH clustering, OPTICS clustering, and DBSCAN clustering. Methods for ranking the performance of said clustering approaches against one another are also considered herein.
9:40 AM
Materials Design, Model Calibration, and Multi-fidelity Modeling with Latent Map Gaussian Processes: Ramin Bostanabad1; 1University of California, Irvine
I will introduce latent map Gaussian processes (LMGPs) that inherit the attractive properties of GPs but are also applicable to mixed data that have both quantitative (e.g., pressure) and qualitative (e.g., coating type) inputs. I will elaborate on the core idea of LMGPs which consists of learning a low-dimensional latent space where all qualitative inputs are represented by some latent quantitative features. Through a wide range of analytical and real-world examples, I will demonstrate the advantages of LMGPs in terms of accuracy and versatility. I will show that LMGPs (1) can handle variable-length inputs, (2) have a nice neural network interpretation, (3) dispense with manual featurization in Bayesian optimization. I will also demonstrate that LMGPs can fuse multiple sources of information together without imposing any hard constraints on how information sources, regardless of their fidelity level, should be fused or how the covariance of the errors is structured.
10:00 AM Break
10:20 AM
Comparing Transfer Learning to Feature Optimization in Microstructure Classification: Taylor Sparks1; Debanshu Banerjee2; 1University of Utah; 2Jadavpur University
Human analysis of research data is slow and inefficient. In recent years machine learning tools have advanced our capability to perform tasks normally carried out by humans, such as image segmentation and classification. In this work, we seek to further improve binary classification models for high throughput identification of different microstructural morphologies. We utilize a dataset with limited observations (133 dendritic structures, 444 non-dendritic) and employ data augmentation via rotation and translation to enhance the dataset six-fold. Then, transfer learning is carried out using pre-trained networks VGG16, InceptionV3, and Xception achieving only moderate accuracies (83-86%). We hypothesize that feature engineering could yield better results than transfer learning alone. To test this, we employ a new nature-inspired feature optimization algorithm, the Binary Red Deer Algorithm (BRDA), to carry out binary classification and observe accuracies ranging from 96% to 98%.
10:40 AM
Improving Autonomous Data Collection by Iterative Learning Control as Applied to a Robomet.3D Mechanical Serial-sectioning System: Damian Gallegos-Patterson1; Claus Danielson2; Jonathan Madison1; 1Sandia National Laboratories; 2University of New Mexico
Optimization of automated data collection is gaining increased interest for the purposes of enabling closed-loop self-correcting systems that inherently maximize operational efficiencies and reduce waste. Many data collection systems have several variables which influence data accuracy or consistency, and which can require frequent user interaction to monitor and maintain. Operating upon a RoboMET.3D automated mechanical serial-sectioning system, an iterative learning control algorithm has been developed to accelerate data collection and reduce data inconsistency. Using historical data amassed over a decade of experiments, weighted influence for two system variables were determined and employed to demonstrate how experimental setups, system variables and produced data are received as inputs and optimal iterative variable changes, are provided as outputs. Three example cases will be shown with quantitative metrics reported for the algorithm’s suggested modifications and the benefits realized.
11:00 AM
Chemistry and Processing History Prediction from Materials Microstructure by Deep Learning: Amir Abbas Kazemzadeh1; Mahmood Mamivand1; 1Boise State University
Finding the chemical composition and processing history from a microstructure morphology for heterogeneous materials is desired in many applications. In this work, we develop a framework based on a deep learning approach that enables us to predict the chemistry and processing history just by reading the morphological distribution of one element. We used a dataset from spinodal decomposition simulation of FeCrCo alloy created by the phase-field method as a case study. We develop specific algorithms to manage the mixed dataset, including both images and continuous data. Results show that while shallow networks are efficient in chemistry prediction, deep networks are required to predict the processing temperature accurately. The physical concepts behind these observations will be discussed. The results also show that transfer learning outperforms the in-house trained network when it comes to microstructure feature extraction.