An efficient multi-fidelity design optimization framework for a thermoelectric generator system

Abstract

In recent times, energy conservation and environmental protection have become increasingly significant, elevating the prominence of power generation using thermoelectric generator systems as a technology for alternative energy sources. However, the extensive computational time required for system-level thermoelectric generator simulation models, including fluid, heat exchanger, and thermoelectric modules, poses considerable challenges in analyzing and optimizing thermoelectric generator systems. To overcome these issues, this study proposes an efficient and general framework to analyze and optimize a system-level thermoelectric generator using a multi-fidelity surrogate. Firstly, a fully coupled finite element model is employed as a high-fidelity system, while a decoupled model composed of the reduced finite element and thermal resistance models is treated as a low-fidelity system. The fully coupled finite element model solves all partial differential equations of the thermoelectric generator system. On the other hand, the reduced finite element model solves the partial differential equations except for the fluid analysis of water and electrical analysis, and then only the one-dimensional thermal resistance model is solved given boundary conditions obtained from the reduced finite element model. Then, a multi-fidelity surrogate for power generation is developed with respect to operating conditions and geometric parameters. Numerical results show that the proposed low-fidelity system is four times more efficient than the high-fidelity system, with a Pearson correlation coefficient of 0.99, indicating a very similar overall trend. Furthermore, the accuracy of the proposed surrogate is validated to be more than twice as high as other surrogates under the same computational budget. Finally, the influence of each parameter on power generation and efficiency is investigated, and the optimum values under various conditions are obtained utilizing the cheap-to-evaluate multi-fidelity model.