Recurrent neural network modeling of the large deformation of lithium-ion battery cells

Abstract

As the automotive industry transitions from combustion to electric motors, there is a growing demand for efficient computational models that can describe the homogenized large deformation response of Li-ion batteries. Here, a detailed three-dimensional unit cell model with periodic boundary conditions is developed to describe the large deformation response of a typical anode-separator-cathode lay-up of a pouch cell. The model makes use of a Deshpande-Fleck foam model for the porous polymer separator and Drucker-Prager cap models of the granular cathode and anode coatings. Using the unit cell model, the stress-strain response of a battery cell is computed for 20’000 random loading paths in the six-dimensional strain space. Based on this data, a recurrent neural network (RNN) model is trained, validated and tested. It is found that an RNN model composed of two gated recurrent units in series with a deep fully connected network is capable to describe the large deformation response with a high level of accuracy. As a byproduct, it is shown that advanced conventional constitutive models such as the anisotropic Deshpande-Fleck model cannot provide any predictions of satisfactory accuracy.