Explainable machine learning for low-emission methane tri-reforming: Carbon formation, hydrogen production, and operating-window optimization

Tri-reforming of methane (TRM) offers a flexible route to hydrogen-rich syngas while utilizing CO 2 , but performance is constrained by trade-offs among methane conversion, syngas composition, reactor duty, and carbon formation. A Gibbs-equilibrium Aspen Plus model generated 46,464 operating points spanning temperature, pressure, and H 2 O/CH 4 , CO 2 /CH 4 , and O 2 /CH 4 feed ratios. Decision tree, random forest, XGBoost, and artificial neural network models were compared, followed by SHAP and permutation-importance analysis and multi-objective genetic-algorithm optimization. The neural network provided the best predictive performance. Temperature dominated equilibrium carbon-formation propensity. Controlled Aspen calculations showed that the direct CO 2 -feed effect is temperature dependent, while steam and oxygen suppress equilibrium carbon under the tested conditions. Optimization identified multiple operating windows rather than a single universal optimum. The resulting framework supports rapid, interpretable TRM screening, with carbon-related results interpreted as equilibrium-based guidance rather than predictions of catalyst durability or experimentally guaranteed coke-free operation.