Dual-atom catalysts on Si(111) for the hydrogen evolution reaction: DFT and machine-learning data

Dual-atom catalysts (DACs) are promising for the hydrogen evolution reaction (HER) because neighboring metal centers can act synergistically. Experimentally screening the large number of possible metal pairs is challenging because stable Si(111)-supported DACs require precise control over metal configurations and atomic dispersion. Here, DFT and interpretable ML are used to systematically screen 59 DACs on Si(111). Four regression algorithms are benchmarked alongside a consensus ensemble to predict hydrogen adsorption free energy (ΔG_H*), using pristine-catalyst electronic and elemental descriptors and excluding adsorption-derived inputs to prevent target leakage. Random forest regression achieves the highest test-set R² of 0.713, while the consensus model yields R² = 0.700, a mean absolute error of 0.165 eV, and a root-mean-square error of 0.227 eV. Thirteen systems exhibit ΔG_H* ≤ 0.10 eV. Five promising candidates Zr–V, Cr–Co, Fe–Pd, Mn–Cu, and Zn–Co show near-thermoneutral ΔG_H* values of +0.00892, −0.00973, −0.01309, +0.01622, and −0.02215 eV, respectively, alongside favorable formation energies (−0.46 to −1.87 eV) and strong anchoring energies (−7.11 to −12.02 eV). Feature attribution identifies valence-electron count and spin magnetization as key descriptors of catalytic activity. Bonding analysis clarifies the nature of metal–hydrogen interactions, while optical calculations and finite-element simulations reveal visible-light absorption and localized energy dissipation within the catalyst overlayers.