Descriptor-guided thermodynamic screening of $$\hbox {H}_2$$ adsorption on single-atom-doped anatase $$\hbox {TiO}_2$$ nanoparticles with interpretable machine learning

Understanding how surface dopants tune [Formula: see text] adsorption on oxide nanoparticles is important for the design of reversible hydrogen-storage materials and catalytic interfaces. Here, we present a descriptor-guided screening study of molecular [Formula: see text] adsorption on pristine and single-atom-doped anatase [Formula: see text] nanoparticles using density-functional tight-binding calculations, conceptual DFT descriptors, thermodynamic modelling, and interpretable machine learning. The replacement of one surface Ti atom with Al, Fe, Hf, La, Mo, Nb, Sn, V, W, or Zr enables systematic comparison across chemically distinct adsorption environments. Most dopants preserve molecular adsorption, whereas Fe shows incipient dissociative activation, and the adsorption energies span from [Formula: see text] to [Formula: see text] eV, indicating that single-atom doping can tune [Formula: see text] binding over a practically relevant range. Descriptor analysis separates weakly perturbed wide-gap systems from narrow-gap dopants with dopant-derived frontier states, enhanced softness, and higher electrophilicity. Symbolic regression with leave-one-out cross-validation identifies a compact [Formula: see text]-dependent expression (the electron-donating power, a member of the electrophilicity descriptor family) as the most generalizing descriptor-property relationship in this small-data regime, with higher-complexity formulas exhibiting an overfitting cliff; the large in-sample sensitivity to charge-transfer descriptors in those higher-complexity formulas reflects the chosen symbolic form rather than a directly measurable adsorption-energy variation. Gaussian-process modelling is retained as an uncertainty-driven active-learning sampling-design tool. Thermodynamic screening further shows that Nb and Zr provide the most balanced uptake-release behavior, Sn remains borderline viable, and Hf and Mo define a stronger-binding but less balanced regime. Overall, the workflow provides a data-efficient and physically interpretable basis for screening dopant chemistry in oxide nanomaterials.