Machine learning-driven optimization of plasma-coupled photocatalytic CO 2 reduction system for syngas production

Abstract

Dielectric barrier discharge (DBD) plasma-coupled photocatalysis offers a promising route for CO2-to-syngas conversion, but rational optimization remains challenging because operating variables are strongly coupled. Here, we develop a small-sample machine-learning framework to predict and optimize a plasma-coupled photocatalytic CO2 conversion system using a Cu-Pd/TiO2 photocatalyst. Using 120 experimental runs, five operating parameters, including discharge power, catalyst dosage, gas flow rate, relative humidity, and light intensity, were evaluated against CO2 conversion, CO yield, and H2 yield. Among the evaluated models, a shared-weight Gradient Boosting Regressor (GBR)-Kernel Ridge Regression (KRR) hybrid model achieved the great predictive performance, with test-set R2 values of 0.947, 0.950, and 0.954 for CO2 conversion, CO yield, and H2 yield, respectively. Permutation importance and SHapley Additive exPlanations (SHAP) analyses identified relative humidity, light intensity, and discharge power as the most influential variables and revealed distinct model-predicted trends across the three outputs. The trained model was further used for constrained optimization under target H2/CO ratios of 1 and 2, followed by experimental validation of the selected operating conditions. Overall, this work establishes a data-driven strategy for interpreting nonlinear plasma-photocatalytic CO2 conversion and provides practical guidance for syngas-ratio regulation under experimentally relevant conditions.