Physics-Driven Optimization of Thermoelectric Generators for Low-Grade Waste Heat Energy Recovery

Low-grade waste heat recovery is important for improving energy efficiency and reducing carbon emissions. Thermoelectric generators offer a solid-state solution, but conventional parameter optimization still relies heavily on time-consuming experimental trial-and-error or simplified simulations. This study develops a machine learning framework that incorporates physical knowledge to model and optimize thermoelectric generator performance using 5,205 real experimental records from the public ESTM dataset. Three progressive feature sets are constructed by gradually embedding classical thermoelectric relations, including original transport properties, physics-informed descriptors, and interaction terms. A Stacking ensemble model is trained on the physics-informed features and achieves strong predictive accuracy on the test set. SHAP analysis shows that the feature contributions align with thermoelectric theory. Bayesian optimization is then performed with explicit physical consistency constraints, yielding a maximum modeled ZT of 2.0815 in the 300–800 K range. The proposed approach provides a practical workflow that combines data-driven modeling with physical rationality and generates engineering-feasible parameter recommendations for low-grade waste heat recovery.