Explainable machine learning reveals orbital design rules for reversible H₂ adsorption on single-atom-doped TiO₂ nanoparticles

Hydrogen is a promising energy carrier for decarbonized technologies, but reversible on-board storage remains challenging because efficient uptake and release under near-ambient conditions require a narrow adsorption-energy window. Here, we combine high-throughput electronic-structure calculations with explainable machine learning to uncover the governing factors of molecular hydrogen adsorption on single-atom-doped TiO2 nanopar ticles. Screening 30 dopants across perpendicular and parallel H-2 adsorption geometries reveals substantial variation in adsorption strength and clear configuration-dependent behavior. From a broad descriptor space, a three-stage feature-selection strategy identifies eight robust predictors, which are further reduced through symbolic regression to an analytical and interpretable model. The resulting descriptor space is governed by H-H activation, dopant-H-2 distance, and d-electron rearrangement. Adsorption within the target window is associated with balanced sigma-donation and pi-back-donation, with Group 4 and 5 transition metals emerging as the most robust candidates across geometries. Combined with Langmuir-based desorption estimates, these results provide a mechanism-based screening rule for reversible hydrogen storage on doped TiO2.