About this Abstract |
Meeting |
2024 TMS Annual Meeting & Exhibition
|
Symposium
|
Algorithm Development in Materials Science and Engineering
|
Presentation Title |
Investigating the Uncertainty in Multi-fidelity Machine Learning Interatomic Potentials |
Author(s) |
Ilgar Baghishov, Jan Janssen, Aparna Subramanyam, Graeme Henkelman, Danny Perez |
On-Site Speaker (Planned) |
Ilgar Baghishov |
Abstract Scope |
Simulating materials under extreme conditions through long-time Molecular Dynamics (MD) is valuable for discovering novel materials applicable in various fields, including fusion reactors. Classical MD combined with Machine Learning (ML) potentials provides an alternative to expensive ab-initio MD limited to short time scales. However, training ML potentials that are transferable across a wide range of conditions remains challenging and computationally costly due to the extensive configuration space.
This study presents a cost-effective approach to parameterize highly transferable ML potentials using active learning that leverages Density Functional Theory (DFT) calculations at different fidelities. We assess the influence of DFT precision on the errors of different ML potentials and propose an active learning method that utilizes mixed DFT fidelity. Our procedure effectively and autonomously balances the potential's accuracy, transferability and the computational cost of generating training data by simultaneously identifying new training configurations and determining the required DFT precision for their characterization. |
Proceedings Inclusion? |
Planned: |
Keywords |
Machine Learning, Computational Materials Science & Engineering, |