Machine learning analysis of photocatalytic CO2 reduction on perovskite materials

A dataset containing 328 samples extracted from 66 experimental articles on photocatalytic CO₂ reduction over perovskite materials was constructed and analyzed using machine learning. Random forest algorithm was used to predict total product yield in gas and liquid phase separately; decision tree algorithm was also utilized to deduce heuristic rules for high performance. Unavailable band gaps were also predicted using a linear regression trained by available data. Random forest models for both phases were quite successful. R² and RMSE for liquid phase were 0.96 and 0.21, respectively for training (0.84 and 0.36 respectively for testing); for the gas phase, R² and RMSE were 0.91 and 0.22 respectively for training (0.87 and 0.24 respectively for testing). The testing accuracy of decision tree models (0.88 % for gas and 0.73 % for liquid phases) were also reasonably high. The perovskite synthesis method was the most important descriptors for both RF and DT models.