Iterative fine-tuning of the universal MACE-MP0 model

The MACE-MP0 [1] universal machine-learning interatomic potential (MLIP) was fit to the MPtrj dataset of configurations from the Materials Project structure database. This potential, fit to PBE/PBE+U DFT calculations, enables stable and reasonable dynamics across the entire periodic table. We describe a simple fine-tuning workflow for producing improved MACE models using additional configuration from an iterative molecular dynamics/fitting procedure, taking advantage of multi-head stabilization of the fits. We show how this procedure can produce MACE potentials that are more accurate for specific systems. The first example is a potential for simulating the temperature-composition phase diagram of a eutectic ionic liquid. Another is a fit of the HSE hybrid functional PES for Ni(OH)₂ + H₂O. Despite the fact that this functional is quite different from the one used to fit the underlying MACE-MP0 model, the fine-tuning procedure yields a MLIP that reproduces the improved description of bonding of the hybrid functional at a dramatic cost savings compared to training only on hybrid-functional data. We also describe the workflow scripts, which use the wfl python library [2] to automatically parallelize the molecular dynamics, DFT evaluation, and MACE fine tuning fit tasks, and present a VASP-like interface for controlling the simulations.

[1] Batatia, I. et al. "A foundation model for atomistic materials chemistry." http://arxiv.org/abs/2401.00096 (2024).

[2] https://github.com/libAtoms/workflow