Machine learning models representing catalytic activity for direct catalytic CO2 hydrogenation to methanol

Direct catalytic hydrogenation of CO2 to methanol is one of the most attractive ways to meet increasing fuel demand and reduce anthropogenic CO2 emission. An efficient catalyst is required for this highly complex structure-based reaction that simultaneously activates the CO2 molecule along with increased selectivity at feasible operating conditions. Application of machine learning models in catalysis research enables researchers to estimate and develop insights on the catalyst performances. This work focuses on development of machine learning (ML) models that include Multi Linear Regression, Least Absolute Shrinkage Selection Operator, Ridge Regression, Support Vector Regression, Gaussian Process Regression (GPR), Random Forest Regression, Gradient Boost Random Forest Regression (GBRT) and Artificial Neural Network (ANN) using published experimental data (698 datapoints) generated in a fixed bed reactor during the years 2010–2020. CO2 conversion and methanol selectivity were considered as catalytic activity performance indicators. Compared to other ML models, GBRT and ANN model predictions outperformed with R2 ∼ 0.95 and R2 ∼ 0.94 for CO2 conversion and with R2 ∼ 0.95 and R2 ∼ 0.95 for methanol selectivity respectively. Further, the input contributions using ANN models reveal that catalyst composition and calcination temperature are the significant inputs for CO2 conversion and methanol selectivity.