Active Learning Scheme for Accelerated Machine-Learned Interatomic Potentials Training with Enhanced Reliability: Case Studies for Strongly Anharmonic Thermal Insulators

Ab initio molecular dynamics (aiMD) has been complemented with machine-learned interatomic potentials (MLIPs) motivated by their efficiency and scalability. However, the MLIP reliability for rare events, such as defect creation, has been questioned as MLIP training is often incomplete due to their deficiency in training data and by their possibly smoothing away upon regularization. This study adopts an active learning (AL) scheme to verify how it improves MLIP reliability for rare events and accelerates training [1]. The configurational space is explored by MD employing MLIPs, NequIP [2] and SO3KRATES [3], while the unfamiliarity of generated data is evaluated by ensemble uncertainty. From 112 materials applications, two erroneous cases are identified: missing genuine events and predicting false events. We demonstrate how these flaws deteriorate MLIP prediction for dynamical properties, exemplified by the over-(under)estimation of phonon lifetimes in CuI(AgGaSe₂) and how AL rectifies these. Finally, we implement our AL scheme to investigate potential ultra-low thermal conducting materials indicated by a symbolic regression ML model [4]. AL enables accelerated, reliable training for MLIPs that calculate thermal conductivity via Green-Kubo formalism.

[1] K. Kang, et al., arXiv:2409.11808 [2] S. Batzner, et al., Nat. Commun. 13, 2453 (2022) [3] J.T. Frank, et al., Nat. Commun. 15, 6539 (2024) [4] T.A.R. Purcell, et al., Npj Comput. Mater. 9, 112 (2023)