Explainable Machine Learning Modeling for Predicting Carbon Yield in Catalytic Methane Decomposition Using Multi‑Component Catalyst Systems

This is the authors' abstract. We don't add key points for this paper.

Accurate prediction of carbon yield during catalytic methane decomposition (CMD) can support the analysis of complex interactions among catalyst composition, preparation, reduction, and reaction conditions. The novelty of this work is methodological rather than the introduction of a new catalyst: we establish a leakage-controlled, explainable machine-learning framework that harmonizes 372 literature-derived CMD observations described by 17 compositional and operating variables, separates hyperparameter tuning from final testing, evaluates catalyst-formulation-level generalization, and quantifies prediction uncertainty. Fourteen regression approaches were compared. Hyperparameters were selected only by five-fold cross-validation within the 90% development partition, while the remaining 10% was retained as an untouched test set. Extra Trees (ET) gave the strongest independent-test performance (R² = 0.9972, RMSE = 0.1972 g g⁻¹, MAE = 0.0964 g g⁻¹). Repeated five-fold cross-validation confirmed high interpolation accuracy (R² = 0.987 ± 0.007), whereas catalyst-formulation-grouped validation was more demanding (R² = 0.878 ± 0.117), demonstrating that random-split accuracy overestimates generalization to unseen catalyst formulations. A 90% split-conformal prediction interval achieved 92.1% empirical coverage on the untouched test set. Pearson analysis identified reduction time (r = 0.952) and reduction temperature (r = 0.729) as the strongest positive linear associations, while GHSV was negatively associated (r = −0.595). Leakage-controlled SHAP analysis of the ET model ranked reduction time, reduction temperature, reaction temperature, reduction H₂ concentration, and reaction time as the leading contributors to model predictions; these SHAP values are interpreted as model-attribution measures rather than causal effects. The principal finding is that ET provides excellent within-domain predictive accuracy, but the measurable performance reduction under formulation-grouped validation means that the framework should be used for screening and hypothesis prioritization rather than as a substitute for prospective catalyst experiments.