Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 5: Machine Learning in Dynamics and Statistical Physics I
DY 5.1: Vortrag
Montag, 18. März 2024, 09:30–09:45, BH-N 243
Active-Learning Training of Accurate Machine-Learned Interatomic Potentials for Strongly Anharmonic Materials — •Kisung Kang, Thomas A. R. Purcell, Christian Carbogno, and Matthias Scheffler — The NOMAD Laboratory at the FHI of the Max-Planck-Gesellschaft and IRIS-Adlershof of the Humboldt-Universität zu Berlin
Machine-learned interatomic potentials (MLIP) promise numerically efficient access to long time and large length scales in molecular dynamics (MD) simulations while retaining an accuracy that is on par with ab initio MD. To this end, it is necessary that MLIPs provide reliable predictions even for geometries that are largely unaccounted for in the original training, e.g., for rare dynamic events. For instance, this is required for thermal transport calculations, for which the creation of defects and phase transition precursors can profoundly affect anharmonic effects [1]. To this end, we propose an active learning (AL) technique, in which uncertainty estimates are used to iteratively incorporate strongly anharmonic configurations into the MLIP training. At variance with traditional approaches, this AL method is thereby able to accurately capture those (meta-stable) configurations that are only seldom explored, as demonstrated in the cases of CuI and AgGaSe2. Eventually, we show that this approach results in improved training and data acquisition efficiency for strongly anharmonic materials, whereas virtually no overhead is needed for more harmonic compounds.
[1] F. Knoop, et al., Phys. Rev. Lett. 130, 236301 (2023).
Keywords: Machine learning; Interatomic potential; Active learning; Uncertainty; Rare events