About this Abstract |
Meeting |
2023 TMS Annual Meeting & Exhibition
|
Symposium
|
Bulk Metallic Glasses XX
|
Presentation Title |
Machine Learning versus Human Learning in Complex Materials Discovery and Science: Predicting Glass-forming Ability of Metallic Glasses |
Author(s) |
Guannan Liu, Sungwoo Sohn, Sebastian A. Kube, Arindam Raj, Andrew Mertz, Anna Gilbert, Mark D. Shattuck, Corey S. O’Hern, Jan Schroers |
On-Site Speaker (Planned) |
Guannan Liu |
Abstract Scope |
Glass formation is defined by a large number of atoms, thus often too complex to be solved using first-principles calculations. Machine learning (ML) appears to be able to overcome the limitations of today's approach to such complex materials. To test ML’s ability, we created ML models that predict glass-forming ability. Surprisingly, we discovered that when prediction accuracy is tested using interpolation, a general-material ML model with features constructed using simple statistical functions from elemental features is indistinguishable from models that use unphysical features or do not consider any features. Only when significant separation of training and testing data is performed, the general-material model outperforms the unphysical or composition models, but performs significantly worse than a human learning-based model. The general-material model's limited performance is explained by its inability to accurately represent alloy features through elemental features. We conclude that complex materials problems necessitate physical insights to develop effective ML models. |
Proceedings Inclusion? |
Planned: |
Keywords |
Machine Learning, |