Construction of a Curated Literature Dataset and Machine Learning-Guided Exploration of Supported Metal Catalysts for Low- to Intermediate-Temperature Reverse Water–Gas Shift Reaction

Literature-derived results on reverse water–gas shift (RWGS) catalysts contain extensive information about catalyst design, yet their fragmentation makes systematic, data-driven exploration difficult. Here, we establish a framework that connects a curated RWGS literature dataset to interpretable machine-learning (ML) models and a flexible catalyst-search algorithm. We compiled 607 experimental records from 42 publications on supported metal and metal oxide catalysts evaluated at 300–450 °C, integrating catalyst compositions, preparation and reaction conditions, and CO yields. By systematically comparing nine regression algorithms and nine input representations, we obtained mean cross-validated test R2 values of 0.80, 0.73, and 0.58 for models optimized for the full, 300 °C, and 400 °C datasets, respectively. SHAP analysis identified compositional and experimental features contributing to model predictions, while leave-one-out validation exposed substantial prediction errors in some literature records. In offline sequential-search simulations repeated 100 times, ML-guided search identified high-yield literature entries more efficiently than random selection. Importantly, our newly developed Local-search algorithm jointly explored elemental combinations, loadings, support materials, and experimental conditions without exhaustive enumeration, generating candidates beyond the compositions represented in the training data. The 300 °C search favored Pt–Na/ZrO2-containing candidates, whereas Pd–Na/CeO2-containing candidates predominated at 400 °C. Although these predictions await experimental validation and some violate RWGS equilibrium constraints, this integrated framework provides a reusable data resource and a practical route from accumulated literature knowledge to testable catalyst-design hypotheses.