DPχ: Constant-Potential Machine-Learning Potentials for Large-Scale Electrocatalysis

We introduce DPχ, a machine-learning potential that enables electrocatalysis simulations at spatiotemporal scales 10^3-10^4 beyond AIMD while retaining near first-principles accuracy. DPχ adopts a Bader-basin-centroid (BC) representation and explicitly decomposes interfacial charge into a neural-predicted chemical component and a conductor component confined to metal sites. The conductor charges are determined on-the-fly by a Siepmann-Sprik-type polarizable-electrode model under an electroneutrality constraint, which adaptively apportions charge between electrodes and yields constant-potential dynamics without imposing the Fermi level. Replacing Wannier-center sites with BCs preserves the far-field accuracy of Wannier-based schemes while permitting fractional charge transfer across electrified metal-water interfaces. On Pt(111)-H2O, DPχ reproduces DFT-level forces, electrode potential, hydrogen-evolution barriers, and vibrational spectra, opening a route to predictive electrocatalysis under realistic conditions.