About this Abstract |
Meeting |
MS&T22: Materials Science & Technology
|
Symposium
|
Materials Processing and Fundamental Understanding Based on Machine Learning and Data Informatics
|
Presentation Title |
High-throughput Machine Learning Experiments with Graph Neural Networks for Predicting Abnormal Grain Growth in Polycrystalline Materials |
Author(s) |
Ryan Cohn, Elizabeth Holm |
On-Site Speaker (Planned) |
Ryan Cohn |
Abstract Scope |
Abnormal grain growth (AGG) significantly affects the properties of materials, but is not well understood for many processes. In this study, Monte Carlo was applied to generate a large dataset of AGG simulations. After representing the microstructure as a graph of connected grains, graph neural networks were trained to predict the occurrence of AGG using only the initial state of the system as input. The preliminary results indicated that a simple graph network outperforms a standard computer vision approach used for comparison. Further exploration required high-throughput experiments, leading to the development of a flexible, containerized, and cloud-native approach for running experiments. Extensive parameter sweeps were conducted to interpret the relative feature importance of the inputs, providing physical insight to the mechanism of AGG. The results motivate ongoing efforts to replace the binary classification approach with statistical predictions, and using a more sophisticated set of features generated through deep learning. |