Machine learning and multi-model mechanistic analysis of CuTaS3/AgTaS3 heterostructures for photocatalytic and electrocatalytic hydrogen evolution

Abstract

Machine learning and mechanistic kinetic-thermodynamic analyses were applied to CuTaS 3 /AgTaS 3 heterostructures for renewable energy applications involving photocatalytic and electrocatalytic H 2 evolution, using existing experimental data without new experimentation. Gaussian Process Regression identified an optimal pH of ~9.2, consistent with the experimental optimum (pH 9.02, H 2 = 1430 µmol/g), with leave-one-out cross-validation R 2 = 0.902. Four independent computational models converged on the same alkaline pH range, constituting a multi-model consistency check rather than independent validation, since all models share the same n = 4–5 experimental dataset. Surface speciation analysis indicates HS − and SO 3 2− as the most probable hole-scavenging species, though their strong co-variation with pH precludes definitive statistical identification. Thermodynamic analysis showed a trend consistent with expectations but did not reach conventional statistical significance. The Na 2 S/Na 2 SO 3 sacrificial mixture exhibited substantial positive synergy relative to its individual components. Stability modelling predicted catalyst half-life of approximately 17.6 cycles, extrapolated beyond the experimentally tested range. Multi-objective optimization confirmed the 5 wt% heterostructure as the best-performing composition across photocatalytic and electrocatalytic metrics simultaneously. This work supports ML-assisted computational analysis as a cost-effective strategy for extracting mechanistic insight from limited experimental datasets, while underscoring the retrospective, interpolative nature of such analysis at small sample sizes.