A Generalized Machine-learned Interatomic Potential for Chemical Reaction Pathways and Transition States

Gaining an atomistic understanding of chemical reactions at catalytic surfaces often requires quantum-mechanics based methods, density functional theory (DFT), to compute the adsorption energy, reaction pathways and kinetic barriers. However, the computational cost of DFT prohibits the comprehensive examination of all possible adsorption sites and reaction pathways to fully understanding catalysts.

In this work, a machine learned interatomic potential (MLIP) is trained using the MACE architecture for accurately examining catalysts and reaction pathways, and applied to find the reaction energy of more than 80 unique reactions with surface terminations up to (5,3,2). Furthermore, the MLIP presented here can recover DFT accuracy for kinetic energies based on independently performed ML-NEB calculations for three case studies, (a) CO₂ reduction to C₂ products, (b) CO₂ reduction to C₃ products and (c) Tafel step in hydrogen evolution reaction with the barrier prediction within 0.1-0.3 eV of DFT. The MLIP in this work trained using samples from Molecular Dynamics (MD) and contour mapping (CM).The latter in particular leads to an MLIP that accurate for modeling relaxed states of adsorbates, and in bond breaking events where molecules may dissociate or combine to create larger molecules.

Overall, this study shows that well-trained MLIPs, without being directly trained to DFT NEB calculations, can perform well in a variety of chemical reaction pathways with an estimated speed-up of up to 10⁴ compared to DFT.