Performance Assessment of Catalyst Materials for CO2 Hydrogenation to Methanol Using Explainable Machine Learning

An extensive dataset containing 1547 data points from 84 published papers was constructed and analyzed using machine learning tools to assess the performance of catalyst materials (active metal and support). Random forest (RF) models were developed for the prediction of CO₂ conversion and methanol selectivity; the SHAP (SHapley Additive exPlanations) analysis, accompanying the RF models, was used to determine the contributions of descriptors, including the catalyst materials, to the conversion and selectivity predictions. Association rule mining analysis (ARM) was also utilized to determine the effects of individual catalyst material and active metal–support combinations on methanol selectivity and to improve the explainability of the results further. RF models for both CO₂ conversion and methanol selectivity were quite successful; the RMSE of training and testing were 2.81 (R² = 0.87) and 3.74 (R² = 0.74), respectively, for CO₂ conversion, and they were 7.31 (R² = 0.94) and 12.74 (R² = 0.80) for methanol selectivity. SHAP analysis indicated that the reaction temperature, the support type, the active metal type, and the catalyst preparation methods are the most significant descriptors for both CO₂ conversion and methanol selectivity; the temperature affects the conversion positively, while its effect on methanol selectivity is negative. ARM analysis for the catalyst material and preparation methods revealed that the use of Ga₃Ni₅, Ga, Ir, Ru, and Y improves methanol selectivity, while Nb₂O₅, CuBr₂, In₂O₃–ZrO₂, and ZnO–ZrO₂ are the best performing supports. The use of evaporation-induced self-assembly method and precipitation was found to be better to improve methanol selectivity. The ARM results also indicate that Cu–Nb₂O₅, Ga–ZnO–ZrO₂, Ru–In₂O₃, and Y–ZrO₂ pairs promote high selectivity, while preparing Cu-based catalysts by precipitation and Ru-based catalyst with deposition-precipitation methods appears to be beneficial. The use of bimetallic Cu-InO₂, Y–In₂O₃, La–In₂O₃, and Zn–ZrO₂ catalysts seems to be also beneficial.