Machine learning to predict plasma-based CO2 conversion in dielectric barrier discharge reactors

Abstract

Plasma-based CO₂ conversion is an emerging defossilization technology that converts a potent greenhouse gas into valuable chemical feedstocks, yet its optimization is hampered by complex nonlinear behavior and resource-intensive experimentation. In this work, we collected a comprehensive database, comprising 358 data points with six key operational and geometric parameters, published in literature between 2010 and 2025. Leveraging this dataset, we developed a hybrid machine learning (ML) model integrating physics-informed neural network (PINN), random forest (RF) and extreme gradient boost (XGB) algorithms to predict CO₂ conversion and energy efficiency (EE) in dielectric barrier discharge (DBD) reactors. Under a rigorous group 5-fold cross-validation (CV) protocol, the ensemble consistently outperformed all individual models, with the best-fold model achieving an R² of 0.791. Error-correlation analysis revealed that the ensemble weights adapt to the pairwise error correlation structure: PINN consistently provides complementary information, while RF and XGB, being largely interchangeable, are selected according to their individual performance. When applied to prospective experimental validation, the hybrid model achieves an R² of 0.92 on unseen data within the explored domain, and it eliminates unphysical predictions in data-sparse regimes, yielding strictly non-negative CO₂ conversion estimates. SHapley Additive exPlanations (SHAP) analysis further identified flow rate and power as the dominant input features, collectively accounting for 61%–71% of the model's predictions. This work establishes a robust and interpretable framework while quantifying the generalizability of ML models in heterogeneous data environments, offering a practical tool to accelerate plasma-based gas conversion optimization.