Thermodynamically Consistent Machine Learning Potentials for Constant-Potential Simulations

Simulating electrochemical interfaces under an applied voltage requires a grand-canonical treatment of electrons, yet standard machine-learning potentials (MLPs) are canonical or rely on electronic-property supervision. We introduce a thermodynamically consistent constant-potential MLP (TCCP-MLP) trained solely on energies and forces. The work function emerges as the exact derivative of a unified energy functional, capturing long-range electrostatics and charge transfer. We demonstrate TCCP-MLP by predicting voltage-dependent activation barriers and charge transfer for C-C coupling on the Cu (100) surface.