The Role of Data Science in Numerical Modeling of Lithium Ion Batteries

Abstract

Batteries are complex electrochemical devices which are nearly ubiquitous in today's society. As the energy demands of mobile devices increase, the performance of batteries must also improve to keep pace. One of the key elements in iterative battery design is the application of numerical models which can predict the properties of potential batteries at significantly reduced cost compared to cell development, enabling high-throughput screening of potential materials and geometries. These models can vary in complexity from simple empirical fits, through continuum-scale models, up to molecular dynamics simulations, which offer increased fidelity, but at an extremely high computational cost. In addition to first-principles models, data-driven models have become popular as the available computational resources and amount of available data have grown astronomically. These models use self-tuning algorithms which form highly accurate nonlinear mappings from inputs to outputs, providing excellent accuracy for relatively low computational cost at runtime. While traditional data-driven models can achieve impressive results given a large amount of data, the acquisition of data at the proper scale is expensive, does not generalize well to other use conditions or battery chemistries, and offers little guidance in the form of physical interpretability. In this work, combinations of physical models and data-driven models are utilized in order to provide highly accurate, flexible applications of the information contained within the first-principles models, while also significantly reducing computational cost at runtime. While design applications of equivalent techniques are conceivable, this work focuses on applications for the calibration of first-principles models for the purposes of improved control of existing electrochemical cells.