February 1, 2020
Journal Article

Machine learning coupled multi-scale modeling for redox flow batteries


Reaction distributions at macro or device scales have been studied for redox flow batteries by many researchers using both experimental and modeling approaches. The reaction distribution on the pore-scale structure of electrodes, however, is not well understood. Especially lacking is knowledge about how the reaction distribution at the pore scale may impact overall flow-battery performance. Recent advances from the rapidly growing field of machine learning have helped researchers in various areas to advance their understanding of complex systems. This study introduces the framework of a multi-scale model that couples a deep neural network (DNN), a widely used machine learning approach, with a partial differential equation solver and provides understanding of the relationship between the pore-scale electrode structure reaction and device-scale electrochemical reaction uniformity within a flow battery. We trained and validated a DNN using 128 pore-scale simulations that provided a quantitative relationship between battery operating conditions (e.g., inlet velocity, current density, and inlet concentration) and uniformity of the surface reaction for the pore-scale sample. Using the framework and DNN, we were able to upscale information about surface reaction uniformity at the pore level to combined uniformity at the device level. We also validated the information obtained using the framework and DNN against the experimental measurements. Based on the multi-scale model results, we established a time-varying optimization of electrolyte inlet velocity, which led to a significant reduction in pump power consumption for targeted surface reaction uniformity but little reduction in electric power output for discharging. The DNN provides predictive results that are much more accurate than the traditional regression/reduced order model, and it can efficiently estimate the full cumulative distribution function of the electrode surface reaction, which provides researchers a more complete view of reaction uniformity. The multi-scale model coupled with the DNN approach establishes the critical link between the micro-structure of a flow-battery component and its performance at the device scale, thereby providing rationale for further operational or material optimization.

Revised: February 11, 2021 | Published: February 1, 2020


Bao J., V. Murugesan, C.J. Kamp, Y. Shao, L. Yan, and W. Wang. 2020. "Machine learning coupled multi-scale modeling for redox flow batteries." Advanced Theory and Simulations 3, no. 2:Article No. 1900167. PNNL-SA-148857. doi:10.1002/adts.201900167

Research topics