Redox flow batteries (RFBs) offer a reliable solution for long-term and grid-scale storage of renewable energy. However, the lack of effective approaches to designing electrodes and flow fields limits the performance of RFBs. The primary objective of this thesis is to fill the gap by exploiting the potential of machine learning in solving complex problems.

We begin with identifying the optimal structural parameters of RFB electrodes. A dataset containing 2,275 fibrous structures is generated by combining stochastic reconstruction method, morphological algorithm, and lattice Boltzmann method model. Combining machine learning regression models with genetic algorithm, we obtain 700 optimal solutions that exhibit up to 80% larger specific surface area and up to 50% higher hydraulic permea...[ Read more ]

Redox flow batteries (RFBs) offer a reliable solution for long-term and grid-scale storage of renewable energy. However, the lack of effective approaches to designing electrodes and flow fields limits the performance of RFBs. The primary objective of this thesis is to fill the gap by exploiting the potential of machine learning in solving complex problems.

We begin with identifying the optimal structural parameters of RFB electrodes. A dataset containing 2,275 fibrous structures is generated by combining stochastic reconstruction method, morphological algorithm, and lattice Boltzmann method model. Combining machine learning regression models with genetic algorithm, we obtain 700 optimal solutions that exhibit up to 80% larger specific surface area and up to 50% higher hydraulic permeability than commercial graphite felt electrodes. Moreover, the fiber diameter and electrode porosity of these optimal solutions show a triangle-like joint distribution, with a preference for fiber diameters of around 5 μm with aligned arrangements. When it comes to the design of electrode materials, an accurate and efficient microstructure reconstruction technique is highly required. To this end, we develop a generative adversarial network (GAN) to reconstruct the three-dimensional (3D) microstructure of RFB electrodes. The high quality of the images generated by GAN is shown through a statistical comparison between the real and generated datasets in terms of structural parameters (porosity, specific surface area, and tortuosity) and two-point correlation function. The strong similarity between the real and generated datasets is further proven using t-distributed stochastic neighbor embedding. With various spatial sizes of the latent input, the GAN demonstrates the ability to synthesize larger images while maintaining high quality. Furthermore, the computation time of reconstruction is linearly associated with the size of images, suggesting the high computational efficiency.

In addition to electrodes, flow fields play a critical role in determining the performance of RFBs. We develop a systematic and effective approach to designing flow fields, with a special focus on flow fields with one channel. A library of 11,564 flow fields is generated using a home-made path generation algorithm, in which flow fields are elegantly encoded by two-dimensional (2D) binary images. Through a collaborative screening process, eight novel designs are identified. Experimental validation shows that the battery with the new flow fields yields higher electrolyte utilization and exhibits about a 22% increase in limiting current density and up to 11% improvement in energy efficiency compared to a 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. To demonstrate the generalization of the above-mentioned approach, we extend the approach to designing flow fields with two channels through modifying the original path generation algorithm. From a library of 11,600 flow fields, eight promising candidates are identified. Simulation results show that the new flow fields exhibit higher electrochemical performance than a conventional interdigitated flow field.

Keywords: redox flow batteries; machine learning; electrode microstructure; flow field.