A coupled machine learning (ML) and multi-objective genetic algorithm (MOGA) approach to the design of porous electrode structures for redox flow batteries (RFBs) is developed in this thesis, and over 700 new electrodes that have up to 80% enhancement of specific surface area and 50% enhancement of hydraulic permeability compared with commercial graphite felt electrodes are successfully designed.

In order to fasten the commercialization of RFBs, further optimization of the batteries to reduce the cost, increase the power density and electrolyte utilization are desired. An important route is to optimize the porous electrode structures. For RFB electrodes, large specific surface area and high hydraulic permeability are very desirable properties. Considering that traditional trial-and-erro...[ Read more ]

A coupled machine learning (ML) and multi-objective genetic algorithm (MOGA) approach to the design of porous electrode structures for redox flow batteries (RFBs) is developed in this thesis, and over 700 new electrodes that have up to 80% enhancement of specific surface area and 50% enhancement of hydraulic permeability compared with commercial graphite felt electrodes are successfully designed.

In order to fasten the commercialization of RFBs, further optimization of the batteries to reduce the cost, increase the power density and electrolyte utilization are desired. An important route is to optimize the porous electrode structures. For RFB electrodes, large specific surface area and high hydraulic permeability are very desirable properties. Considering that traditional trial-and-error approaches are hindered by limited human intuition, here in this thesis, a coupled ML and MOGA strategy is developed for the design of porous RFB electrodes. First, a dataset containing over 2,000 porous electrode structures is generated by a stochastic reconstruction program and the two properties are computed through numerical simulations using morphological algorithm and lattice Boltzmann method (LBM), respectively. Based on the dataset, ML models are trained to learn the underlying relationship between electrode structures and the two properties. The best models, which are trained with artificial neural network (ANN) algorithm, achieve the lowest fitting errors. Based on the ANN models, the MOGA non-dominated sorting genetic algorithm II (NSGA-II) is adopted to design new electrode structures for RFBs. As a result, more than 700 promising candidates are successfully identified. The methodology proposed here is also applicable to other material structure design problems.