Thermocatalytic CO2 hydrogenation to methanol is an attractive decarbonization technology to combat climate change while producing a valuable platform chemical and energy carrier. However, predicting the performance of catalytic systems for this process remains a challenge. Herein, we present a machine learning framework to predict catalyst performance from experimental descriptors. A database of Cu-, Pd-, In2O3-, and ZnO-ZrO2-based catalysts with 1425 datapoints is compiled from literature and subjected to data mining. Accurate ensemble-tree models (R2 > 0.85) are developed to predict the methanol space-time yield (STY) from 12 descriptors, where the significance of space velocity, pressure, and metal content is revealed. The model prediction and its insights are experimentally validated, with a root mean squared error of 0.11 gMeOH h−1 gcat−1 between the actual and predicted methanol STY. The framework is purely data-driven, interpretable, cross-deployable to other catalytic processes, and serves as an invaluable tool for guided experiments and optimization.