Machine Learning Potentials for Electrified Interfaces

Abstract

Accurate atomistic modelling of electrochemical interfaces requires accounting for applied electrode potential and performing molecular dynamics at constant potential. The computational cost of grandcanonical density functional theory restricts simulations to small system sizes and short timescales. Machine learning interatomic potentials (MLIPs) have been proven useful for long time and large length scale simulations, but their generalisation to and applicability at electrified interfaces have yet to be demonstrated. In this perspective, we review emerging approaches for developing MLIPs for systems under applied potential, organising them into two categories: (i) linear response methods that learn response properties (electronegativities, Born charges, polarisabilities) and compute energies through physics-based expansions, and (ii) direct learning methods that take the applied electrochemical potential as input variable and learn the potential-dependent energy surface end-to-end. We identify key challenges including training data scarcity, lack of standardisation, and architectural limitations in current MLIPs. To address these issues, we propose flexible templates for incorporating global electrochemical information into MLIP architectures, and data-efficient strategies such as transfer learning and multi-head architectures that leverage abundant existing zero-potential data. We conclude by calling for community initiatives to establish shared datasets and reporting standards for electrified interfaces.