Learning physics based models of Lithium-ion Batteries

Abstract

Abstract Lithium-ion (Li-ion) batteries are increasingly pervasive and important in daily life. We present a surrogate modeling approach that uses synthetic data generated by an electrochemical model to approximate Li-ion battery dynamics using a Deep Neural Network. Elechtrochemical models are needed to describe high current operation but are computationally costly. As an initial study, we prototype our approach for the Single Particle Model. We use a battery use-cycle and observations of concentrations and voltages to predict future battery behavior. Given only the use cycle and knowledge that the battery is fully charged, the surrogate model can accurately forecast observations of Li-ion concentrations and voltages for the entire use cycle, as well as give a window of time for battery depletion.