Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 33: Machine Learning in Dynamics and Statistical Physics I

DY 33.12: Vortrag

Donnerstag, 20. März 2025, 12:30–12:45, H47

Accelerating the Training and Improving the Reliability of Machine-Learned Interatomic Potentials for Strongly Anharmonic Materials through Active Learning — •Kisung Kang, Thomas A. R. Purcell, Christian Carbogno, and Matthias Scheffler — The NOMAD Laboratory at the FHI of the Max Planck Society

Machine-learned interatomic potentials (MLIP) can efficiently implement molecular dynamics (MD) simulations with large spatial and long time scales. However, immature training for rare dynamical events, such as defect creation, may happen due to their absence or insufficiency in training data or their fadeout during regularization, leading to the critical deterioration of MLIP predictions regarding dynamical properties like transport phenomena. To improve the MLIP’s reliability and accelerate the whole training process, we adopt a sequential active learning (AL) scheme via MD employing MLIP (MLIP-MD) and uncertainty estimates [1]. In each iterative step, MLIP-MD serves as an efficient exploration tool for configurational space to generate training data, while uncertainty estimates identify unfamiliar data to be sampled for subsequential MLIP models. The representative examples of CuI and AgGaSe₂ among 112 materials display erroneous MLIP predictions of missing and fictitious rare events. We demonstrate how AL addresses these issues, specifically correcting unfamiliar regions for the MLIP potential energy surface. At last, the over(under)estimation of their phonon lifetimes is rectified after the AL steps.

[1] K. Kang, T. A. R. Purcell, et al., arXiv:2409.11808 (2024).

Keywords: Machine-learned interatomic potentials; Active learning; Strongly anharmonic materials