Group-cross-validated machine learning benchmarking of a Pd-CsW1.6O6/g-C3N5 photocatalyst demonstrates the need for designed experiments in multi-objective optimisation

Semiconductor photocatalysis offers a promising strategy for degrading pharmaceutical contaminants; however, one-factor-at-a-time (OFAT) experimental designs fundamentally limit machine learning (ML) applicability for multi-objective optimisation. We benchmark five ML architectures on a 53-sample Pd-CsW 1.6 O 6 /g-C 3 N 5 Z-scheme photocatalyst dataset using leave-one-group-out cross-validation (LOGO-CV). The best ANN (8-5-1) achieves LOGO-CV RMSE of 30.44 pp, 2.4× higher than random-split estimates, and test R 2 of -1.184 on an independent holdout, confirming that the model cannot generalise to pH, catalyst dose, and anion-type variations, a structural consequence of OFAT design, not a limitation of the catalyst, which retains 96% metronidazole degradation under its validated conditions. Stepwise RSM was excluded by rank deficiency (rank = 11 of 44 estimable terms), the mathematical root of OFAT inadequacy for multi-factor regression. Permutation importance confirms zero signal for pH, dose, and anion type. Single-objective optimisation converges to Pd ≈ 3.35 wt%, a model artifact; the experimentally validated optimum remains 5 wt% Pd. Multi-objective NSGA-II produces a Pareto front comprising 100% extrapolated solutions. No H 2 evolution data exist in the source study; kinetic proxy objectives are theoretical. Langmuir-Hinshelwood kinetics are best-supported (AICc weight = 0.6075) but tentative at n = 9. Designed experiments are recommended as a prerequisite for reliable multi-objective ML optimisation of photocatalytic systems.