Accelerating Computational Modeling of Reactant Adsorption through a Combined MACE+DFT Approach: Furfural on Cu Surfaces

Machine learning has great potential to accelerate computational discovery of new materials and catalytic reactions but is challenging to implement with quantum chemical accuracy for diverse chemical systems. Here, we explore the pretrained foundation model MACE-MP-0, a machine-learned interatomic potential (MLIP) using the MACE architecture, as a low-cost computational method for structure relaxation prior to density functional theory (DFT) computations of furfural (FUR) adsorption over single-crystal copper surfaces. A data set of 240 adsorption configurations of FUR, including both trans- and cis-FUR rotamers, was created using an automated workflow to randomly orientate the bioderived platform chemical at heuristic adsorption sites on Cu(111), (100), and (110) surfaces. The performances of DFT alone versus a MACE+DFT approach were compared, with common local minima geometries identified through structure relaxation followed by hierarchical agglomerative clustering using the smooth overlap of atomic positions (SOAPs) as a descriptor of the chemical environment. New adsorption configurations were discovered for FUR on Cu(110). Combining MACE-MP-0 prerelaxation with agglomerative clustering to preselect energetically stable configurations for subsequent DFT relaxation eliminated human bias in the starting geometries and reduced the computational cost by 75%.