Accelerating Materials Science with Big Data and Machine Learning: On-Demand Poster Presentations
Program Organizers: Huan Tran, Georgia Institute Of Technology; Muratahan Aykol, Toyota Research Institute
Friday 8:00 AM
October 22, 2021
Room: On-Demand Poster Hall
Location: MS&T On Demand
Poster
Developing Physics-based Descriptors for Property Prediction in Oxide Glasses: Suresh Bishnoi1; Ravinder Ravinder1; N. M. Anoop Krishnan1; 1Indian Institute of Technology, Delhi
Data-driven regression methods are becoming popular tools for predicting and designing novel materials. In glass, learning properties directly from glass composition is very common. However, these composition-based models are restricted to a particular set of compositions as an input for which they are trained. Herein, we develop physics-based descriptors that can predict the property for any given composition by transforming composition space into twelve universal descriptors space. To this extent, we trained ML models using XGBoost (Extreme Gradient Boosting) algorithm to learn the descriptor–property relationships for density, Young’s, shear, bulk moduli, thermal expansion coefficient, Vickers’ hardness, refractive index, glass transition temperature, liquidus temperature and abbe number having twelve universal descriptors as an input feature. Further, we interpreted these models using SHAP value analysis to understand contribution of descriptors in property value. Overall, these physics-based descriptors prove to be advanced, reliable, and global data-driven models to predict novel glasses' property.
Machine Learning to Predict Mechanical Properties of Steel Alloys Based on Chemical Composition and Heat Treatment Process: Yutao Wang1; 1WPI
As-cast steel alloys often have a non-homogeneous microstructure, which includes large grains and other unwanted microstructures. Post-heat treatment is needed to adjust the mechanical properties of the as-cast parts. In this study, Artificial Neural Network (ANN) models are created to predict the mechanical properties of wide steel grades. The independent variables of the ANN model are the chemical compositions of the steel alloys and the corresponding heat-treatment process. Over 30,000 data entries of processing, composition, and property of different steel alloys are collected and included by SFSA from steel foundries all over the United States and are particularly useful for understanding non-standard steel alloys. The ANN prediction results are compared with traditional linear regression results. High accuracy of predictions on Hardness, Yield strength, ultimate tensile strength, and elongation is demonstrated. The prediction results are found to be in good correspondence with the microstructure development during heat treatment.
Topology Optimization for Two-phase Composites Using Active Learning Based Gaussian Process Regression: Tanu Pittie1; Suresh Bishnoi1; N. M. Anoop Krishnan1; 1Indian Institute of Technology (IIT), Delhi
Designing new composite materials is a challenging problem consisting of composition and property prediction, and structure optimization. Recently, machine learning (ML) has emerged as a promising solution for developing design optimization algorithms. However, extensive input data, which is computationally and experimentally expensive, is required to train a reliable ML model for composites. Herein, we use active learning to develop a robust design process for target-specific two-phase composites using sparse data. Finite element analysis is carried out on a handful of randomly generated microstructures to establish the ‘ground truth’. Active learning is used to train a Gaussian process regression model with selective points from the sample space to predict the effective elastic modulus and fracture energy. Promising results close to the ground truth are obtained using very sparse data. It is also shown that the model can learn underlying physics appropriately and hence, can be used for topology optimization of composites.