Constant-Potential Machine Learning Force Field for Electrochemical Interface

Better understanding and prediction of electrochemical interface requires large-scale atomistic simulations. Machine learning force field (MLFF) has proven to be an effective approach. However, current MLFFs typically do not account for the effect of electrode potential, which requires treating interface electrons with grand canonical ensemble. Here we develop a constant potential MLFF (CP-MLFF) based on equivariant graph neural network and implement it into MACE. Specifically, we design an architecture which can take the number of electrons as input and accurately predict the Fermi level. The CP-MLFF allows us to examine the convergency of electrochemical barrier with respect to sampling, which we demonstrate through the example of CO₂ reduction on Ni-N-C catalyst. Our work provides a useful method and tool enabling accurate and efficient large-scale simulation of electrochemical interface.