Interpretable machine learning analysis of catalyst and reaction parameters governing CO₂ hydrogenation to light hydrocarbons
Abstract
Abstract Carbon dioxide hydrogenation over iron-based catalysts is governed by a complex interplay between thermodynamic driving forces, kinetic constraints, dynamic phase evolution, and promoter-induced electronic effects. Despite extensive mechanistic studies, quantitative separation of activity- and selectivity-controlling domains remains unresolved due to multicollinearity in experimental datasets. Here, we develop an interpretable machine learning framework to statistically disentangle thermodynamic and compositional control in Fe-based CO₂ hydrogenation toward light hydrocarbons. A curated dataset of 184 fixed-bed reactor experiments (2022–2025) was analyzed using ensemble learning algorithms with SHAP-based interpretability. XGBoost achieved robust predictive performance ( $$R_{test}^2 = 0.78$$ for selectivity; 0.59 for conversion). Importantly, SHAP analysis reveals that CO₂ conversion is predominantly governed by thermokinetic parameters (temperature, residence time), whereas hydrocarbon selectivity is primarily dictated by catalyst composition, especially cobalt and alkali promoter loadings. This quantitative separation provides statistical evidence for partial decoupling between RWGS-driven conversion and Fischer–Tropsch chain-growth probability. Optimal operating windows (300–320 °C, ≤3 MPa) emerge from multivariate interactions rather than monotonic trends. The study establishes interpretable machine learning as a hypothesis-generating tool for multiscale catalytic systems and provides a transferable framework for disentangling structure–process coupling in (CO₂) utilization technologies.