Regensburg 2022 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 83: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 4
O 83.6: Vortrag
Freitag, 9. September 2022, 11:45–12:00, S054
Indirect learning interatomic potential models for accelerated materials simulations — •Joe D. Morrow and Volker L. Deringer — Department of Chemistry, Inorganic Chemistry Laboratory, University of Oxford, Oxford OX1 3QR, United Kingdom
Machine learning (ML) based interatomic potentials are emerging tools for materials simulations but require a trade-off between accuracy and speed. We show how one ML potential can be used to train another: we use an existing, accurate, but more computationally expensive model to generate reference data (labels and locations) for a series of much faster “indirectly-learned” potentials. Extensive reference datasets can be easily generated without the need for quantum-mechanical reference computations at the indirect learning stage, and we find that the additional data significantly improve the predictions of fast potentials with less flexible functional forms.
We apply the technique to disordered silicon, including a simulation of vitrification and polycrystalline grain formation under pressure with a system size of a million atoms. When comparing indirectly learned potentials to models learned directly from a DFT-labelled database, the latter make unphysical predictions for large systems (105 atoms) that are not apparent in smaller simulations (≤ 104 atoms). This emphasises the importance of carefully validating ML potentials chemically, not only via numerical error measures. Our work provides conceptual insight into the machine learning of interatomic potential models, and it suggests a route toward accelerated simulations of nanostructured materials.