Optimizing interleaved distributed streamline obstacle flow channels through genetic algorithm coupled with backpropagation neural network regression predictive surrogate model

Abstract

The design of the proton exchange membrane fuel cell (PEMFC) channel has a significant impact on mass transfer and chemical reactions. In this study, an innovative channel design with streamlined obstacles arranged in an interleaved manner is proposed to achieve optimal performance for PEMFC. The design utilizes genetic algorithm optimization (GA) combined with backpropagation (BP) neural network as a predictive regression proxy model. The input variables include the radius, semi-axis, width, and average interleaved distribution interval of the streamlined obstacle, while the output variables are current density and cathode pressure drop. This approach aims to achieve the optimal design prediction of streamlined obstacle structures. The dataset is obtained from a three-dimensional two-phase model of PEMFC established in the COMSOL software. The results indicate that the GA-BP neural network regression predictive proxy model aligns well with the physical model and demonstrates high prediction speed and accuracy. The training time is approximately 15 min, and the prediction time is less than 3 s. The determination coefficients for current density and cathode pressure drop are 0.99977 and 0.99999, respectively. The physical model and proxy model are validated using an orthogonal table design of input variables. Additionally, using net output power as a comprehensive evaluation index, 9,720,000 operating conditions are predicted. The optimal geometric parameters for the streamlined obstacles are determined as follows: a radius (R) of 0.41 mm, a semi-axis (L) of 0.2 mm, a width (W) of 0.94 mm, and an average interleaved distribution interval (D) of 0 mm. These dimensions are validated in the physical model at an operating voltage of 0.4 V, resulting in a current density of 1.5901 A/cm2 and a pressure drop of 502 Pa, with relative errors of 0.18 % and 4.38 %, respectively, compared to predictions. The optimized channel, when compared to the basic straight channel, shows an increase of 3.25 % in current density and a 3.16 % increase in net power.