Screening foundational machine learning models using the adsorption of thiophene on transition metal surfaces

This article presents a computational screening of several popular foundational machine learning models on the adsorption of thiophene, C4H4S, on Ir, Rh, Pt, Pd, Ni, Au, Ag, and Cu(100) surfaces. Specifically, focus is placed on four popular models: UMA's uma-s-1p1model, MACE's MACE-MH-1model finetuned to the omat dataset, MACE's MACE-MH-1model finetuned to the oc20 dataset, and MACE's MACE-MH-1model finetuned to the omat dataset with D3 dispersion corrections. None of these models contain thiophene adsorption calculations in their training dataset, representing a potent test of the models' transferability. Overall, the MACE's MACE-MH-1model finetuned to the omat dataset with D3 dispersion corrections performs the best, imitating dispersion corrected density functional theory results. Moreover, the model trained with surface/adsorbate DFT calculations in its training dataset, MACE's MACE-MH-1model finetuned to the oc20 dataset, is often outclassed by other models which were only trained with bulk DFT calculations due to the former's training dataset lacking sulfur containing adsorbates while the latter's bulk system training dataset containing sulfur atoms, remarkably demonstrating those models can generalize from bulk calculations to surface calculations.