Transferable machine learning force fields: from lead halide perovskite interfaces to bulk materials

I. Knysh, A. Charkin-Gorbulin, C. Quarti, D. Beljonne, I. Poltavsky, A. Tkatchenko

Halide perovskites (HPs) are promising materials for next-generation photovoltaic technologies but suffer from phase instabilities and ionic migration that limit their long-term stability. [1,2] State-of-the-art machine learning force fields (MLFFs) trained on ab initio data provide a powerful framework to model such dynamic processes at the atomistic level, offering valuable insights into degradation mechanisms and design pathways for improved device performance. In this work, we develop an MLFF based on the equivariant message-passing architecture MACE, [3] trained exclusively on CsPbI3 surface slab data computed at the PBE+MBDNL level of theory. [4,5] The model qualitatively reproduces the experimental orthorhombic-to-cubic phase transitions of bulk CsPbI3 while maintaining stable surface dynamics. This demonstrates that properly curated slab datasets encode transferable information relevant to bulk behavior, providing guidance for developing more generalizable MLFFs for materials exploration where interface-bulk coupling matters.

1. P. Calado; A. M. Telford; D. Bryant et al., Nat. Commun. 2016 7, 13831.
2. W. Nie; J. C. Blancon; A. J. Neukirch et al., Nat. Commun. 2016 7, 11574.
3. I. Batatia; D. P. Kovacs; G. N. C. Simm et al., Adv. Neural Inf. Process. Syst. 2022 35, 11423–11436.
4. J. P. Perdew; K. Burke; M. Ernzerhof, Phys. Rev. Lett. 1996 77, 3865–3868.
5. J. Hermann; A. Tkatchenko, Phys. Rev. Lett. 2020 124, 146401.