A coupled machine learning and genetic algorithm approach to the design of porous electrodes for redox flow batteries

Abstract

The design of porous electrodes with large specific surface area and high hydraulic permeability is a longstanding target for the development of redox flow batteries (RFBs), but traditional trial-and-error strategies are hindered by the heavy cost of collecting large amounts of data and the limitation of human intuition when multiple trade-offs are at play. In this work, a novel framework coupling machine learning and genetic algorithm is developed to identify the optimal electrode structures for RFBs. A custom-made dataset containing 2275 fibrous structures is first generated by adopting a combination of stochastic reconstruction method, morphological algorithm, and lattice Boltzmann method. Based on the dataset, our best machine learning models allow to achieve test errors of 1.91% and 11.48% for predicting specific surface area and hydraulic permeability, respectively. Combined with well-trained prediction models, the genetic algorithm is developed to screen more than 700 promising candidates with up to 80% larger specific surface area and up to 50% higher hydraulic permeability than the commercial graphite felt electrodes. Results show that the fiber diameter and electrode porosity of these promising candidates exhibit a triangle-like joint distribution, with a preference for fiber diameters of around 5 μm with aligned arrangements.