Boosting the optimization of membrane electrode assembly in proton exchange membrane fuel cells guided by explainable artificial intelligence

•Machine learning could greatly reduce the computing cost of multiphysics models.•Explainable AI intelligently revealed impact of each parameter in the complex system.•Precise AI prediction costs time orders of magnitude shorter than experiments.•AI-guided optimization achieved 3.2 times the Pt util...

Full description

Saved in:
Bibliographic Details
Published inEnergy and AI Vol. 5; p. 100098
Main Authors Ding, Rui, Yin, Wenjuan, Cheng, Gang, Chen, Yawen, Wang, Jiankang, Wang, Ran, Rui, Zhiyan, Li, Jia, Liu, Jianguo
Format Journal Article
LanguageEnglish
Published Elsevier Ltd 01.09.2021
Elsevier
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:•Machine learning could greatly reduce the computing cost of multiphysics models.•Explainable AI intelligently revealed impact of each parameter in the complex system.•Precise AI prediction costs time orders of magnitude shorter than experiments.•AI-guided optimization achieved 3.2 times the Pt utilization of commercial products. The utilization of environmentally friendly hydrogen energy requires proton exchange membrane fuel cell devices that offer high power output while remaining affordable. However, the current optimization of their key component, i.e., the membrane electrode assembly, is still based on intuition-guided, inefficient trial-and-error cycles due to its complexity. Hence, we introduce an innovative, explainable artificial intelligence (AI) tool trained as a reliable assistant for a variable analysis and optimum-value prediction. Among the 8 algorithms considered, the surrogate model built with an artificial neural network achieves high replaceability in the experimentally validated multiphysics simulation (R2= 0.99845) and a much lower computational cost. For interpretation, partial dependence plots and the Shapley value method are applied to black-box models to intelligently simulate the impact of each parameter on performance. These methods show that a tradeoff existed in the catalyst layer thickness. The AI-guided optimization suggestions regarding catalyst loading and the ionomer content are fully supported by the experimental results, and the final product achieves 3.2 times the Pt utilization of commercial products with a time cost orders of magnitude smaller. [Display omitted]
ISSN:2666-5468
2666-5468
DOI:10.1016/j.egyai.2021.100098