Generative Adversarial Network-Enhanced Heterogeneous Ensemble Learning for Interpretable Prediction of CO2-to-Methanol Catalyst Performance

Accurate prediction of catalyst performance in CO₂-to-methanol (CTM) conversion remains challenging due to the scarcity of experimental data and the complexity of feature interactions. To address these issues, this study proposes an interpretable gene-adversarial network-enhanced heterogeneous ensemble modeling (GAN-HEM) framework. It integrates four synergistic key functional modules: data augmentation, feature interaction analysis, ensemble modeling, and interpretable analysis. Comparing the variational autoencoder-based data augmentation method demonstrates that the GAN has significant advantages in maintaining the global consistency and local geometric features of the data manifold structure. After the multivariate evaluation of feature association within the hybrid data set, it was found that retaining all the identified CTM catalyst features is beneficial for enhancing the predictive accuracy and generalization ability of the prediction model. Therefore, this data set is further applied to develop various homogeneous and heterogeneous ensemble learning models of the CTM process. The hyperparameters of these models are automatically optimized using the Bayesian algorithm-based Optuna approach. Results indicated that the optimized heterogeneous ensemble learning architecture has the highest prediction accuracy (R² = 0.9314 and RMSE = 0.2636), significantly outperforming homogeneous models through its capacity to capture complex nonlinear feature interactions. The Shapley additive explanation-based interpretability analysis identifies reaction temperature as the dominant feature (>39% contribution). Partial dependence plots reveal competitive selectivity: higher temperature favors CO (thermodynamic constraints), while increased pressure and heating rate enhance methanol selectivity (kinetic promotion). This framework accelerates CTM catalyst discovery and optimization, providing high-fidelity prediction and actionable design insights.