Fine-tuning beats from-scratch training for catalytic machine-learned potentials

A new study comparing nine machine-learned interatomic potentials for catalysis found that fine-tuning foundation models outperforms from-scratch training, particularly for out-of-distribution chemistry. The best fine-tuned model achieved a reaction-energy error of 0.15 eV while screening over 90,000 bimetallic alloys. From-scratch models required 5-10% perturbed high-energy configurations to reduce errors by more than twofold, while fine-tuned models showed less sensitivity to sampling. A fine-tuned model on metallic catalysts reached 0.30 eV MAE for oxygen evolution on iridium oxide, and another fine-tuned on metal-oxides achieved 0.19 eV for CO2 reduction on copper. The optimal model, fine-tuned on 49,860 configurations, performed consistently across metallic and oxide catalysts, including unseen high-index surfaces. The findings identify necessary training-set configurations for developing transferable catalytic MLIPs, with implications for accelerating catalyst discovery across diverse reaction types.