| About this Abstract |
| Meeting |
2023 TMS Annual Meeting & Exhibition
|
| Symposium
|
Computational Discovery and Design of Materials
|
| Presentation Title |
Ultra-fast Interpretable Machine-learning Potentials for Accelerated Structure Prediction of Materials |
| Author(s) |
Richard G. Hennig, Stephen R. Xie, Pawan Prakash, Ajinkya C. Hire, Robert Schmid, Hendrik Kraß, Matthias Rupp |
| On-Site Speaker (Planned) |
Richard G. Hennig |
| Abstract Scope |
Crystal structure predictions and all-atom dynamics simulations are indispensable quantitative tools in chemistry, physics, and materials science. Still, large systems and long simulation times remain elusive due to the trade-off between computational efficiency and predictive accuracy. Machine-learning potentials (MLPs) can provide efficient surrogate models for accurate ab-initio electronic structure methods. However, current limitations of MLPs include data inefficiency, instabilities, and lack of interpretability. To address these challenges, we combine effective many-body potentials in a cubic B-spline basis with second-order regularized linear regression (https://arxiv.org/abs/2110.00624). We demonstrate that these ultra-fast potentials are data-efficient, physically interpretable, sufficiently accurate for applications, can be parametrized automatically, and are as fast as the fastest traditional empirical potentials. We illustrate that combining the ultra-fast force field (UF3) method with genetic algorithms can enable large-scale molecular dynamics simulations and accelerate crystal structure predictions of materials. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Machine Learning, Modeling and Simulation |