Data-driven catalyst–process optimization of the water–gas shift reaction using an interpretable TabPFN–ensemble framework

The water–gas shift (WGS) reaction remains a cornerstone for low-carbon hydrogen production, yet achieving stable high CO conversion is hindered by the complex coupling between catalyst formulation and operating parameters. Here, we present an integrated machine learning framework combining a transformer-based TabPFN classifier, a stacking-ensemble local surrogate, and Optuna-guided catalyst–process optimization trained on 8,715 literature-derived experiments. The TabPFN classifier achieved strong classification performance for four CO conversion classes, with a test accuracy of 0.8887 and a macro-F1 score of 0.8692, enabling focused exploration of the high conversion class (75–100 %). Within this high conversion domain, the stacking ensemble predicted continuous CO conversion with R2 = 0.7753, MedAE = 0.0140, and P90AE = 0.0589. Interpretability analyses, including SHAP, partial-dependence plots, Sankey mapping, and decision-tree-derived rules, identified operating temperature, gas composition, and support-related descriptors as dominant factors, while catalyst formulation and process variables acted as secondary, context-dependent modulators. Optuna-guided optimization identified a Na-promoted Pt/TiO2 catalyst as a data-supported low-temperature high conversion candidate, achieving a predicted CO conversion of 92.24 % at 249.5 °C with 0.6008 wt% Pt and 2.3923 wt% Na. Comparison with thermodynamic equilibrium and literature benchmark data further supported the physicochemical plausibility of the optimized condition. This study establishes an interpretable and data-driven pathway for prioritizing experimentally testable WGS catalyst–process candidates before laboratory validation.