Employing Neural Network Regression Models to Screen the Electrode Design Space for Redox Flow Batteries

Abstract

Redox flow batteries (RFBs) are a promising electrochemical technology for integrating renewable energies into the current grid infrastructure, but further cost reductions are needed for widespread deployment.1 Porous electrodes are an integral component of the RFB reactor providing heterogeneous active sites for electrochemical reactions, a solid-void matrix that facilitates electrolyte dispersion, and a conductive solid for transporting electrons throughout the reaction zones. Many electrodes employed in RFBs use gas diffusion layers (GDLs)—such as carbon papers, cloths, and felts—that are repurposed from fuel cell applications. While GDLs possess some favorable properties, they are not broadly tailored to support electrochemical reactions and forced liquid-phase flow. As such, these materials are suboptimal for RFB applications, and, consequently, research has focused on post-process modifications to improve electrochemical and fluid dynamic performance.2,3 In contrast, the ideal electrode would be a bottom-up design with a property profile specific to the particular redox chemistry and operating conditions. Unfortunately, manufacturing limitations and time-intensive experimentations stymie broad screening of electrode designs. This motivates the use of computational tools.The advances of computational fluid dynamics and multi-physics solvers enable a platform to construct novel electrode designs and investigate resulting structure-performance relationships. By incorporating physical properties and operating conditions, the electrode can be modeled at a variety of length scales.4 Simulation domains can be used to span a wide variety of electrode and system parameter values allowing for the generation of large data sets for statistical and sensitivity analysis. Further, the current epoch of machine learning algorithms engender opportunities to build regressive models for predicting RFB electrochemical and fluid dynamic performance that can be subsequently used for optimization protocols.In this presentation, we use an experimentally-validated RFB model to generate >8,000 electrochemical simulations across a range of electrode properties. We then use feature engineering to reduce data complexity for building a robust and accurate neural network to predict output current density. Subsequently, we use this trained and validated neural network in conjunction with metaheuristics to search the electrode design space and report optimal morphological properties. This approach—as shown in the figure below—represents a first step to exploring the electrode design space for advanced RFBs and may ultimately support performance optimization of existing battery installation, materials selection for prototypes, or new approaches to electrode design.AcknowledgmentsThis work was funded by the Joint Center for Energy Storage Research, an Energy Innovation Hub of the U.S. Department of Energy, Office of Science, Basic Energy (De-AC02-06CH11357). K.M.T. recognizes additional support from the U.S. NSF Graduate Research Fellowship (1122374).References A. Z. Weber et al., J Appl Electrochem, 41, 1137 (2011).K. J. Kim, Y.-J. Kim, J.-H. Kim, and M.-S. Park, Materials Chemistry and Physics, 131, 547–553 (2011).K. V. Greco, A. Forner-Cuenca, A. Mularczyk, J. Eller, and F. R. Brushett, ACS Appl. Mater. Interfaces, 10, 44430–44442 (2018).A. Gayon Lombardo, B. A. Simon, O. Taiwo, S. J. Neethling, and N. P. Brandon, Journal of Energy Storage, 24, 100736 (2019). Figure 1