| About this Abstract |
| Meeting |
2024 TMS Annual Meeting & Exhibition
|
| Symposium
|
AI/Data Informatics: Computational Model Development, Verification, Validation, and Uncertainty Quantification
|
| Presentation Title |
Uncertainty Quantification for Accelerated Production of ChIMES ML Force-fields |
| Author(s) |
Jared Stimac, Nir Goldman |
| On-Site Speaker (Planned) |
Jared Stimac |
| Abstract Scope |
Machine-learned potential energy surfaces for atomistic simulations have demonstrated accuracies comparable to quantum methods, while yielding orders of magnitude improvement in computational efficiency. However, to date there do not exist systematic approaches to training data selection and to prevent over-fitting/poor extrapolation for configurations that differ from a model’s training set. In this regard, we report on several uncertainty quantification (UQ) schemes that can enable active learning approaches. We apply our methods to the Chebyshev Interaction Model for Efficient Simulations (ChIMES), a force-matching framework which defines potentials as linear combination of Chebyshev polynomials evaluated on N-body atomic clusters. Our results indicate that training set sizes can be significantly reduced for our test cases, creating high-fidelity molecular dynamics simulations that we benchmark against DFT simulations. Our workflow can accelerate potential development for materials with complex chemical reactivity or for which generation of quantum training data is computationally cumbersome. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, |