About this Abstract |
Meeting |
2023 TMS Annual Meeting & Exhibition
|
Symposium
|
AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification
|
Presentation Title |
An Information Theory Based Approach for Training Machine Learned Potentials |
Author(s) |
Jason Gibson, Jan Janssen, Laura Lopes, Richard Hennig, Danny Perez |
On-Site Speaker (Planned) |
Jason Gibson |
Abstract Scope |
The promise of obtaining accuracy on par with first-principle calculations at the computational cost of empirical potentials has made machine-learned interatomic potentials attractive alternatives for studying and characterizing materials. However, while many machine-learned potentials have reported accuracy within several meV/atom of first principle reference data, these errors represent only configurations that do not significantly differ from the training. In comparison, performance on truly novel configurations can incur errors orders of magnitude larger.
This work leverages more than 7M atomic environments in tungsten that have been optimized to maximize its informational entropy in feature space, ensuring broad coverage compared to hand-crafted datasets. First, we investigate the dependence of the test errors on the training set size for various popular machine-learned potentials and delineate data-bound and model-bound regimes. We then compare different strategies to optimally sub-select training data from this large dataset to maximize the transferability/cost tradeoff of the resulting potentials. |
Proceedings Inclusion? |
Planned: |
Keywords |
Machine Learning, Computational Materials Science & Engineering, |