Materials Informatics and Modeling for 21st Century Ceramics Research: Machine Learning for Materials Microstructure and Property Predictions
Sponsored by: ACerS Basic Science Division
Program Organizers: Ming Tang, Rice University; Jeffrey Rickman, Lehigh University; Turab Lookman, Los Alamos National Laboratory
Tuesday 2:00 PM
November 3, 2020
Room: Virtual Meeting Room 42
Location: MS&T Virtual
Session Chair: Jeffrey Rickman, Lehigh University; Ming Tang, Rice University
2:00 PM Invited
Coarse-grained Equation-free Time Evolution of Microstructures with Deep Learning: Fei Zhou1; Ming Tang2; 1LLNL; 2Rice University
Microstructures play important roles in many advanced materials, with strong, often decisive, effects on their mechanical and functional properties. For the purpose of modeling microstructure time evolution, direct molecular dynamics (MD) simulation is hindered by very demanding length and time scales involved, while coarse-graining approaches such as the phase field method (PFM) require strong prior assumptions such as the functional form the differential equations. We propose deep neural networks (NN) as a new, systematic paradigm for coarse-grained dynamics of microstructure solidification and solid-solid transitions. Through select case studies for both 2D and 3D microstructure evolution, we show that NN can effectively reach time & length scales beyond current methods, with huge speed-up, systematically improvable accuracy, general applicability, and quick turn-around.
2:30 PM Invited
Using Materials Informatics to Quantify Complex Correlations Linking Structure, Properties and Processing
: Jeffrey Rickman1; 1Lehigh University
I will present several examples in which materials informatics can be used to elucidate and quantify complex correlations linking structure, properties and processing of materials. In the first example, I consider the case of high-entropy (HE) (or multi-principal element) alloys, typically comprising five or more elements. In the second example, I examine the use of a canonical correlation analysis on the ubiquitous phenomenon of grain abnormality in a microstructure, with the resulting bimodal structure often having a deleterious impact on the thermomechanical properties of a system. Finally, I will outline the use of detrended correlation analyses to interpret time series data associated with processing.
3:00 PM
Predicting Stress Hotspots in Polycrystalline Materials from Microstructural Features Using Deep Learning: Ankit Shrivastava1; Hae Young Noh2; Kaushik Dayal1; 1Carnegie Mellon University; 2Stanford University
In polycrystalline materials, induced high stress in certain regions, so-called stress hotspots influence material strength. It is observed that local microstructural features such as grain boundaries and junctions, heavily influence these hotspots. In current work, we propose an algorithm to predict hotspots from microstructural features using a convolutional encoder-decoder model. We use images of 128X128 dimensions containing grain misorientation information of a single-phase cubic microstructure as our input. Output to the model is normalized values of Frobenius-norm of von mises stress obtained by solving linear elastic equations under prescribed strains. Since the input is a very high dimensional image, it makes the convolutional encoder-decoder a suitable choice to capture features and predict the hotspots. Using the output from the model, the hotspot location is estimated by thresholding the values in the outputs. The primary result shows that the model can reconstruct the output image with average accuracy with 77%.