Deep-learning accelerating topology optimization of three-dimensional coolant channels for flow and heat transfer in a proton exchange membrane fuel cell

Abstract

To quickly optimize the cooling performance of three-dimensional coolant channels in a proton exchange membrane fuel (PEMFC) cell, a generative adversarial network (GAN) model that adopts small-scale sample data to predict the coolant channel topology structure is constructed. The coordinate information of the topological pseudo-density field is directionally input into the neural network to improve prediction accuracy. The results show that more microtopologies in the coolant channels can be predicted with the GAN model than with traditional thermal–fluid–structural topology optimization. The topological configuration with a maximum temperature of 0.15 K, lower than the training dataset, is achieved. The well-trained GAN model can generate topological structures for the optimization objective, that is, thermal performance superior to the training set. The proposed model can also generate topological structures with better fluid and structural performance, which is 20.44% and 9.25% superior to that of the training dataset, respectively. The computation time required by the GAN model is <1 s, while the computation time for the same case is 30.2 h using the traditional thermal–fluid–structural topology optimization. These findings can aid in designing coolant channels in PEMFCs.