Benchmarking Universal Machine-Learning Interatomic Potentials for Accelerated Adsorption Energy Evaluation with DFT Single-Point Calculations

Abstract

Adsorption-energy calculations are essential for understanding molecule–surface interactions, but full DFT optimization remains too expensive for high-throughput screening. Here, we use a DFT single-point workflow as a platform to benchmark universal machine-learning interatomic potentials (MLIPs) for accelerated adsorption-energy evaluation. In this workflow, MLIPs optimize structures, followed by DFT single-point calculations for final adsorption-energy evaluation. Using a data set of 201 organic molecules adsorbed on Zn slabs, we compare M3GNet, eSCN, UMA, and MatterSim on adsorption-energy accuracy, adsorption-configuration fidelity, and computational cost. MatterSim delivers the best overall performance, achieving a success rate of 85.4%, a mean closest molecule–surface distance difference of 0.12 Å relative to DFT-optimized references, and an approximately 4000-fold speedup over DFT for single-point energy evaluation. UMA provides competitive ranking performance and strong binding-site prediction, despite systematically overestimating adsorbate–surface distances. These results highlight that appropriate MLIP selection is critical for enabling reliable and efficient adsorption screening of interfacial materials.