Group-cross-validated machine learning benchmarking of a Pd-CsW_1.6O₆/g-C₃N₅ photocatalyst demonstrates the need for...

Area of Science:

• Environmental Chemistry
• Materials Science
• Chemical Engineering

Background:

• Semiconductor photocatalysis is a viable method for pharmaceutical contaminant degradation.
• Traditional one-factor-at-a-time (OFAT) experimental designs limit the application of machine learning (ML) for multi-objective optimization in this field.

Purpose of the Study:

• To benchmark five machine learning architectures for optimizing photocatalyst performance.
• To evaluate the impact of OFAT design on ML model generalizability and optimization outcomes.
• To recommend appropriate experimental designs for reliable ML-driven optimization of photocatalytic systems.

Main Methods:

• Benchmarking five ML architectures (including ANN) on a 53-sample Pd-CsW1.6O6/g-C3N5 Z-scheme photocatalyst dataset.
• Utilizing leave-one-group-out cross-validation (LOGO-CV) for robust model evaluation.
• Analyzing stepwise response surface methodology (RSM) and permutation importance to assess factor significance.

Main Results:

• The best ANN model showed poor generalization (test R² = -1.184) due to OFAT limitations, failing to predict variations in pH, catalyst dose, and anion type.
• OFAT design was identified as the root cause of ML model inadequacy, not catalyst performance, which remained high (96% metronidazole degradation).
• Single-objective optimization yielded artifactual Pd concentrations, and multi-objective optimization produced extrapolated solutions, highlighting the failure of OFAT-based ML.

Conclusions:

• OFAT experimental designs are fundamentally inadequate for multi-factor regression and multi-objective optimization in photocatalysis using ML.
• Designed experiments are essential prerequisites for developing reliable and generalizable ML models for optimizing semiconductor photocatalytic systems.
• The study underscores the critical need for appropriate experimental design strategies to unlock the full potential of ML in environmental remediation.