A segmented thermoelectric generator optimization method based on interpretable XGBoost and deep generative model

The implementation of segmented thermoelectric generators (STEGs) can enhance the energy conversion efficiency of radioisotope thermoelectric generators (RTGs). This study proposes a method integrating interpretable XGBoost and Conditional Variational Autoencoder (CVAE) to optimize STEG. An XGBoost regression model trained on validated finite element data predicts electrical performance while Shapley Additive Explanations (SHAP) analysis quantifies parameter impacts. Simultaneously, CVAE was implemented to construct mappings between design parameters and temperature field images to provide visual feedback during the design process. The results demonstrate the XGBoost model achieves exceptional regression performance, and enabling rapid prediction and multi-objective optimization. Based on the optimal design parameters, the CVAE predicts the temperature field image within 2 s, with a structural similarity index (SSIM) of 0.9676. SHAP-based interpretation reveals the key factors affecting electrical performance and provides decision support for optimized parameter selection.