From Prediction to Insight: Explainable Ensemble Learning for CO2-to-Methanol Conversion

Abstract

This work introduced an interpretable machine learning framework for the predictive modeling of CO2 hydrogenation performance for methanol production using the Extra Trees Regression, Histogram Gradient Boosting, and CatBoost algorithms, in terms of CO2 conversion efficiency, methanol selectivity, and CO selectivity. Among the approaches evaluated, Extra Trees demonstrated superior performance for predicting methanol and CO selectivity, whereas Histogram-based Gradient Boosting regression exhibited slightly better generalization for CO2 conversion efficiency. Furthermore, SHAP analysis showed that temperature was the most critical parameter with the highest absolute mean SHAP values for all targets and high nonlinearity. Dependence plots also showed a positive thermal promotion of CO2 conversion and CO selectivity, with a decrease in methanol selectivity at higher temperatures. More importantly, Monte Carlo simulations were performed to quantify predictive uncertainty, and the uncertainty bands spanning 5 to 95% of the model predictions were found to be 15.1%, 41.5%, and 45.9% for conversion, methanol selectivity, and CO selectivity, respectively. Hold-out validation of the models was confirmed by residuals randomly distributed around a mean of zero. Generally, the integrated modeling and explainability framework provides a reliable, transparent, and uncertainty-aware method for accelerating catalyst optimization and the operational strategy development in CO2 valorization systems.