AI-based optimization of PEM fuel cell catalyst layers for maximum power density via data-driven surrogate modeling

作     者：Wang, Bowen Xie, Biao Xuan, Jin Jiao, Kui 

作者机构：Tianjin Univ State Key Lab Engines 135 Yaguan Rd Tianjin 300350 Peoples R China Loughborough Univ Dept Chem Engn Loughborough Leics England 

出 版 物：《ENERGY CONVERSION AND MANAGEMENT》 (能量转换与管理)

年 卷 期：2020年第205卷第0期

页      面：112460-000页

核心收录：

学科分类：0820[工学-石油与天然气工程] 08[工学] 0807[工学-动力工程及工程热物理] 0801[工学-力学（可授工学、理学学位）] 

基　　金：China-UK International Cooperation and Exchange Project (Newton Advanced Fellowship) - National Natural Science Foundation of China China-UK International Cooperation and Exchange Project (Newton Advanced Fellowship) - UK Royal Society [NAF\R1\180146] National Natural Science Foundation of Tianjin (China) [18JCJQJC46700] 

主　　题：Proton exchange membrane fuel cell Catalyst layer composition Agglomerate model Data-driven surrogate model Stochastic optimization algorithm 

摘      要：Catalyst layer (CL) is the core electrochemical reaction region of proton exchange membrane fuel cells (PEMFCs). Its composition directly determines PEMFC output performance. Existing experimental or modeling methods are still insufficient on the deep optimization of CL composition. This work develops a novel artificial intelligence (AI) framework combining a data-driven surrogate model and a stochastic optimization algorithm to achieve multi-variables global optimization for improving the maximum power density of PEMFCs. Simulation results of a three-dimensional computational fluid dynamics (CFD) PEMFC model coupled with the CL agglomerate model constitutes the database, which is then used to train the data-driven surrogate model based on Support Vector Machine (SVM), a typical AI algorithm. Prediction performance shows that the squared correlation coefficient (R-square) and mean percentage error in the test set are 0.9908 and 3.3375%, respectively. The surrogate model has demonstrated comparable accuracy to the physical model, but with much greater computation-resource efficiency: the calculation of one polarization curve will be within one second by the surrogate model, while it may cost hundreds of processor-hours by the physical CFD model. The surrogate model is then fed into a Genetic Algorithm (GA) to obtain the optimal solution of CL composition. For verification, the optimal CL composition is returned to the physical model, and the percentage error between the surrogate model predicted and physical model simulated maximum power densities under the optimal CL composition is only 1.3950%. The results indicate that the proposed framework can guide the multi-variables optimization of complex systems.