| About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Bridging Scale Gaps in Multiscale Materials Modeling in the Age of Artificial Intelligence
|
| Presentation Title |
Material-Agnostic Training Data Generation for Machine-Learning Interatomic Potentials |
| Author(s) |
Aparna P. A. Subramanyam, Danny Perez |
| On-Site Speaker (Planned) |
Aparna P. A. Subramanyam |
| Abstract Scope |
Machine-learning interatomic potentials (MLIAPs) make it feasible to aim for both accuracy and transferability, something earlier generations of potentials struggled to achieve. However, MLIAPs often fail at extrapolating to properties beyond the configuration space spanned by training data, which makes the quality of the training dataset one of the most important factors determining their performance. The need for human intervention in the curation of training datasets exerts bias and makes their generation labor-intensive and time-consuming. Here, we present an automated approach to generate a material agnostic dataset based on maximization of information entropy of the descriptor distribution. In addition to structural diversity, the approach is extended to include chemical diversity by utilizing chemistry-sensitive descriptors of the local atomic environment. The diversity of this dataset is demonstrated by training MLIAPs for a wide range of elements and binary combinations, highlighting the desirable characteristics of optimal training data, irrespective of material chemistry. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Machine Learning, Modeling and Simulation |