Data-driven discovery of highly efficient 2D photocatalytic materials for water splitting

Efficient discovery of 2D photocatalysts for water splitting is hindered by the high computational cost of accurate electronic structure calculations and the frequent neglect of realistic operational conditions. Here, we address both challenges by combining a machine learning meta-model trained on GW bandgaps (R2 = 0.985, MAE = 0.18 eV) with cost-effective IPA@PBE optical and Mott-Wannier excitonic descriptors to screen over 20,000 2D materials, ultimately identifying six candidates with suitable band alignments, strong solar absorption, and low exciton binding energies. Non-adiabatic molecular dynamics simulations reveal that four Janus candidates (SbSeI, SbTeI, AsSeI, and AsTeI) exhibit ultralong nonradiative recombination lifetimes of 90.50–121.89 ns, arising from remarkably weak electron-phonon coupling. Finite-temperature analysis shows that a 300 K environment enhances solar absorption via a visible red shift, yet thermal bandgap narrowing eliminates PdSe₂ from consideration, highlighting the necessity of moving beyond conventional 0 K screening. Constant-potential implicit solvent calculations further demonstrate that candidates such as AsTeI and SbTeI can overcome the thermodynamic bottleneck of the oxygen evolution reaction under weakly acidic to strongly alkaline conditions. This work provides a scalable, physically grounded strategy for photocatalyst discovery under realistic temperature and solid-liquid interface conditions.