Estimating Effective Electrochemical Parameters for Li-ion Battery Models with Convolutional Neural Networks and Minkowski Tensors

Lithium-ion batteries (LIBs) have enabled a revolution in consumer electronics due to their exceptional energy and power density. Upscaling this technology from cell phone size (~10 Wh) to electric vehicle (~30-100 kWh) and grid storage size (>1 MWh) remains a challenge, however, partly due to the poor ability of understanding and predicting the microstructural behavior of such large systems. Since phenomena that occur at the microscale have a significant impact on system performance, it is crucial to accurately characterize these effects when modeling battery operation. This work thus focuses on incorporating microstructural electrode phenomena in macroscopic battery models to improve their capability to predict key battery performance metrics. This type of understanding improves the accuracy and computational efficiency of battery algorithms and can enable battery systems to be effectively designed to meet energy, calendar life, and safety requirements, and thus accelerate the adoption of electric vehicles and batteries as grid storage devices.

First, a multiscale framework for modeling the electrochemical behavior of LIBs is developed via homogenization. This framework establishes a closure problem, which is solved on a representative volume element (RVE) of the electrode microstructure, for determining effective properties such as diffusion and conductivity. The solution to this closure problem requires solving a coupled set of partial differential equations (PDEs), resulting in simplifying assumptions such as single spherical particles, which introduces inaccuracies because transport parameters are highly dependent on morphology. Thus, a Convolutional Neural Network (CNN) is developed for computational efficiency. The CNN, which is trained from solutions to the closure problem on computer-generated images representing various electrode morphologies, accurately predicts transport properties that can be used in macroscopic models. Results demonstrate that the CNN is significantly more accurate than the frequently-used Bruggeman relation and efficient enough for real-time quantification of evolving architectures. The CNN is further validated with sensitivity analysis and is applied to real SEM images and accurately predicts effective properties.

Next, the relationship between Minkowski tensors, which are powerful geometric descriptors for porous media calculated via integration, and electrochemical properties is studied. The Minkowski tensor values are then incorporated into the previous CNN by adding them in as nodes in a hidden layer to improve parameter estimation. This framework demonstrates how computer vision tools can be used to predict the solution of partial differential equations, which is a powerful model reduction technique that can be applied across many fields.