Systematic fine-tuning of MACE interatomic potentials for catalysis

Once trained, machine-learned interatomic potentials (MLIPs) offer a fast and accurate way to investigate catalytic reaction pathways, but their performance depends strongly on training-set design. We compared nine MLIPs developed using from-scratch (FS) training and fine-tuning (FT) of foundation models. The models were evaluated on reaction energies ( E r ) and activation barriers ( E a ) for 141 reactions, including CO 2 RR to C 2 /C 3 products, propane dehydrogenation, hydrogen intercalation on Pd, and out-of-distribution OER on metal oxides. The FS models containing 5–10% perturbed, high-energy configurations from molecular dynamics or contour exploration reduced errors by more than twofold relative to models trained only on relaxation trajectories. However, FT models were less sensitive to sampling and generalized better to out-of-distribution chemistry. For example, the FT model fine-tuned on metallic catalysts achieved an MAE of 0.30 eV for OER on iridium oxide polymorphs, while the FT model fine-tuned on metal-oxides gave an E a of 0.19 eV for CO 2 RR on Cu. The best model, fine-tuned on 49,860 configurations, performed consistently across metallic and oxide catalysts and screened 90,781 bimetallic alloys with a reaction-energy MAE of 0.15 eV, including unseen high-index surfaces. These results identify the necessary training set configurations for training transferable catalytic MLIPs.