Machine learning and numerical simulations for electrical, thermodynamic, and mechanical assessment of modified solar thermoelectric generators

Abstract

The frustum leg thermoelectric generator (FLTEG) was recently proposed as a high-performing power generator. However, there is no basis for proposing this device since its thermo-mechanical performance was not compared with the trapezoidal and conventional rectangular-leg design. Furthermore, solutions to the problems associated with designing variable area leg TEGs on numerical computational solvers have never been proposed. This work seeks to solve this challenge by introducing neural networks that predict the thermo-mechanical performance of variable area leg thermoelectric generators much faster with high accuracies. Therefore, this work compares the thermo-mechanical performances of the three shapes of TEGs using a numerical approach. The three-dimensional finite element models are set up by employing ANSYS software, and the numerical results are validated with experimental results, showing an almost perfect correlation. The power and efficiency of the FLTEG are 5.7 % and 5.6 % higher than the trapezoidal-leg thermoelectric, respectively, under a solar flux of 100 Suns. The former also reduces the thermal stresses of the latter by 2.7 %. The neural network, with an almost perfect regressive correlation and low mean squared error of 7.2 × 10-4, was 702 times faster than the numerical method in predicting the performance of the TEG when diverse input parameters are being varied. Finally, the minimized thermal stress of the optimized device is 0.67 GPa, which is 5.6 times lower than that of the unoptimized device design.