Artificial neural network for geometric design and optimization of three-stage segmented thermoelectric generators

The quest for greener power generation is a major driver for the development of thermoelectric generators. Segmented thermoelectric generators are an effective means to significantly improve the efficiency. Given the complexity of multiple geometric parameters in a three-stage segmented thermoelectric generator, the conventional practice is to use software for numerical simulation and geometric optimization before manufacturing. While this method saves time and increases production efficiency, it lacks reproducibility. However, artificial neural networks offer a solution to this problem by further reducing the computational time required. In this work, artificial neural networks are used to replace these processes by using optimization algorithms to guide the neural network in calculating the optimal thermoelectric conversion efficiency. First, a dataset of 3000 three-stage segmented thermoelectric generator configurations is generated using a COMSOL simulation model. A backpropagation neural network is then constructed and trained on this dataset, using particle swarm optimization to adjust the network weights and computational parameters. The trained neural network accurately predicted the conversion efficiency of three-stage segmented thermoelectric generators over various temperature ranges. When compared to COMSOL calculations, the values are generally within the 95% confidence interval of the backpropagation neural network predictions. Finally, genetic algorithms and artificial fish swarm algorithms are used to drive the neural network to calculate the optimal thermoelectric conversion efficiency. Compared to COMSOL, the thermoelectric conversion efficiency increased by 2.48% and 3.61%, respectively. Despite the initial dataset construction time, the artificial neural network can determine the optimal geometric structure and conversion efficiency of a three-stage segmented thermoelectric generator within 5 min. This method is expected to significantly improve the design and production efficiency of three-stage segmented thermoelectric generators.