Iterative fine-tuning of MACE foundation models

The MACE equivariant message-passing graph-neural-network machine-learning interatomic potential (MLIP) framework has enabled the development of pre-trained foundation models, fit to density-functional theory calculation of materials spanning essentially the entire periodic table. Taking advantage of the reasonable as-is accuracy of the foundation models, and multi-head stabilization of the fits, we describe a simple workflow for producing fine-tuned MACE models using an iterative molecular dynamics/fitting procedure. We show how this procedure produces potentials that are more accurate for specific systems. The first example is a potential, based on the MACE-OMAT-0 foundation model, for simulating the temperature-composition phase diagram of a structural metallic alloy. 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 foundation 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 to automatically parallelize the molecular dynamics, DFT evaluation, and MACE fine-tuning fit tasks, and present a VASP-like interface for controlling the simulations.