Analysis of photocatalytic CO2 reduction over MOFs using machine learning

Abstract

Photocatalytic CO₂ reduction over metal–organic frameworks (MOFs) is investigated by constructing a database from published articles and analyzed using machine learning tools to predict the total gas product yield (random forest regression) and predominant product types under various conditions (decision tree classification). Hyperparameters of the random forest model, ntree (120) and mtry (14) are optimized by 5-fold cross validation leading to R² values of 0.96, 0.94 and 0.60 for training, validation and testing, respectively indicating the predictive power of the model developed. Reactor volume, sacrificial agent and amount of catalyst per reaction volume were the most important variables for total gas production rate prediction. Decision tree models, developed for gas phase and liquid phase systems separately, to determine the predominant product types (CO or CH₄ in the gas phase, and one of CH₃OH, CO, H₂ and HCOOH in the liquid phase) depending on the photocatalyst properties and reaction conditions, were also successful with an overall testing accuracy of 87% and 77% for gas-phase and liquid-phase processes, respectively.