Open Access

ARTICLE

Machine-Learning-Assisted Performance Prediction of a Segmented Bi₂Te₃/PbTe Thermoelectric Generator Based on Multiphysics Simulation

Qingsong Song, Yinuo Di, Guangyu Wang^*, Liufu Yuan, Hongtao Li^*, Yunguang Ji
School of Mechanical Engineering, Hebei University of Science and Technology, Shijiazhuang, China
* Corresponding Author: Guangyu Wang. Email: ; Hongtao Li. Email:

Energy Engineering https://doi.org/10.32604/ee.2026.085746

Received 17 May 2026; Accepted 22 July 2026; Published online 29 July 2026

Abstract

Single-material thermoelectric generators (TEGs) often face performance degradation across broad temperature gradients because the transport properties of thermoelectric materials vary strongly with temperature. To improve temperature-gradient utilization, this study investigates a Bi₂Te₃/PbTe segmented TEG using multiphysics simulation and machine-learning-assisted performance prediction. Rectangular and cylindrical thermoelectric-leg configurations were constructed under identical geometric and thermal-electrical boundary conditions to examine the influence of leg geometry on temperature distribution, electric-potential distribution, output power and conversion efficiency. A two-segment n-type leg was designed by placing Bi₂Te₃ near the cold side and PbTe near the hot side, while the p-type leg, overall geometric constraints, and segment ratio were kept unchanged. A three-dimensional thermoelectric model was established in COMSOL Multiphysics 6.3, and 270 steady-state simulation samples were generated by varying the operating current and hot-side temperature. The simulated output power and conversion efficiency were used to train an eXtreme Gradient Boosting (XGBoost) surrogate model for rapid performance prediction within the investigated current-temperature domain. The segmented rectangular configuration produced more uniform temperature and electric-potential distributions than the segmented cylindrical configuration, indicating a more favorable thermal-electrical transport behavior. Compared with the corresponding non-segmented rectangular model, the segmented rectangular model achieved 18.4% higher output power and 15.7% higher conversion efficiency under the same modelling assumptions. The regression-model comparison indicated that RF achieved the highest in-domain interpolation accuracy for the present current-temperature dataset, whereas XGBoost maintained competitive predictive performance for output power and conversion efficiency, with coefficients of determination of 0.9712 and 0.9697, respectively. Within the prescribed operating domain, the XGBoost surrogate was further examined as a regularized gradient-boosted model through learning-curve analysis, permutation-based feature attribution, and repeated-split stability assessment. These results demonstrate that, within a consistent multiphysics modelling framework and under identical thermal-electrical boundary assumptions, Bi₂Te₃/PbTe segmentation enhances the simulated performance of fixed-geometry TEGs, while XGBoost provides a representative regularized gradient-boosted surrogate prediction approach over the investigated current-temperature operating domain.

---

Keywords

Segmented thermoelectric generators; multiphysics simulation; output power; energy conversion efficiency; XGBoost surrogate prediction

---