Modeling the inner electric double layer using short-range machine learning potentials: A critical assessment

Atomistic simulations of electrochemical interfaces typically require large system sizes and long simulation times to obtain reliable thermodynamic properties. In this context, machine learning interatomic potentials (MLIPs) allow for computationally efficient sampling compared to density functional theory-based molecular dynamics simulations. Although conventional short-range MLIPs are recognized as inadequate for modeling the (potentially nanometer-wide) diffuse layer, they may still be capable of describing the inner electric double layer that is relevant for electrochemical reactions. Here, we evaluate the performance of common (short-ranged) MLIP architectures [the deep potential DP, the (graph) atomic cluster expansion-based models GRACE and MACE, and the pre-trained eSEN-OC25] for charged Au/electrolyte interfaces. In the evaluation, we focus on the effect of message passing and on the potentially adverse effect of data interference when training for a broad range of surface charges. Our results suggest that both local and message-passing models are able to reliably learn the interfacial structure when training on a single charge state. However, when trained on datasets that span multiple charge states, the MLIPs tested yield inconsistent predictions of interfacial water orientation and ion distributions. This behavior can be traced to the MLIPs being "blind" to the actual surface charge. Message-passing models show somewhat greater robustness to mixed-charge training data, likely due to their larger receptive field, but are still prone to artifacts and errors in the description of the interfacial structure. Overall, our results provide practical guidance on the expected accuracy and cross-charge-state transferability of short-ranged MLIPs for the description of the inner layer of electrode/electrolyte interfaces that will be relevant to their use in computational electrocatalysis.