DPχ: A Charge-Based Machine-Learning Potential Validated on the Pt(111)-Water Electrochemical Interface.

Electrocatalytic machine-learning potentials must simultaneously describe long-range electrostatics, nonlocal charge redistribution, and electrode-potential-dependent interfacial response, which makes their physical construction and validation particularly demanding. Here, we introduce DPχ, a charge-based machine-learning potential designed for electrified metal-water interfaces. DPχ represents long-range electrostatics through Bader-basin centroids and decomposes interfacial charge into a neural-predicted chemical component and a conductor component determined self-consistently by a Siepmann-Sprik-type polarizable-electrode model under global electroneutrality. Rather than claiming broad transferability across electrocatalytic materials, we test these physical assumptions on the benchmark Pt(111)-water interface. Systematic benchmarking shows that DPχ reproduces DFT-level forces, interfacial potential drops, hydrogen-coverage-dependent electrode potentials, Volmer barriers, and interfacial vibrational signatures, while remaining robust upon system-size enlargement. These results establish DPχ as a physically consistent and reaction-ready framework for large-scale simulations of the Pt(111)-water electrochemical interface beyond AIMD spatiotemporal scales.