About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Analysis and Reduction of DFT Database for Rapid Artificial Neural Network Interatomic Potentials |
| Author(s) |
Micah Nichols, Christopher Barrett, Doyl Dickel |
| On-Site Speaker (Planned) |
Micah Nichols |
| Abstract Scope |
Machine learning (ML) interatomic potentials for molecular dynamics simulations make it possible to investigate materials at the atomic scale with near density functional theory (DFT) accuracy; however, they often require large training databases. These databases are comprised of DFT simulations meant to capture various atomic environments. DFT simulations come at a substantial computational cost both in database creation and ML potential training. Reducing the size of the database without losing vital information about unique atomic environments can improve computational time and potential development. This work employs Uniform Manifold Approximation and Projection (UMAP) to reduce the dimensionality of the input neurons of a rapid artificial neural network (RANN) interatomic potential to examine and reduce the size of the DFT database. We show that the DFT database can be significantly reduced without affecting the training accuracy, validation accuracy, or the interatomic potential’s prediction capabilities while decreasing training time. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Modeling and Simulation, |