Machine learning-assisted design of flow fields for redox flow batteries

Abstract

Flow fields are a crucial component of redox flow batteries (RFBs). Conventional flow fields, designed by trial-and-error approaches and limited human intuition, are difficult to optimize, thus limiting the performance of RFBs. Here, we develop an end-to-end approach to the design of flow fields by combining machine learning and experimental methods. A library of 11 564 flow fields is generated using a custom-made path generation algorithm, in which flow fields are elegantly encoded by two-dimensional binary images. To accelerate the discovery process, we train convolutional neural networks with low test errors for predicting the uniformity factor and pressure drop of flow fields (0.59% and 1.37%, respectively). Through a collaborative screening process, eight promising candidates are successfully identified. Experimental validation shows that the battery with the flow fields designed with this approach yields higher electrolyte utilization and exhibits about a 22% increase in limiting current density and up to 11% improvement in energy efficiency compared to the conventional serpentine flow field. Furthermore, statistical analysis suggests that the promising candidates have a saved channel length of 1490 ± 100 and a torque integral of 20.1 ± 1.8, revealing the quantitative design rules of flow fields for the first time.