Deep-learning based prediction of chemo-mechanics and damage in battery active materials

Abstract

Layer-structured cathode active materials of Li-ion batteries such as

(NMC) provide benefits including high specific capacity and energy density. However, NMC materials (secondary particles) consist of randomly oriented grains (primary particles), which features anisotropic lattice chemical strain inside each grain and weak intergranular bonding. During

insertion into and extraction from the active material, high stresses arise at the interfaces between primary particles and particle disconnection occurs. Therefore, material microstructure characteristics such as grain orientation and morphology play a critical role in determining cycling performance of the active material. However, resolving particle microstructures with different characteristics remains challenging due to high computational costs and limited statistical generalizability. In this work, ConvLSTM is employed to predict the dynamic evolution of critical physical fields — including

concentration, stresses and damage — inside secondary particles with diverse microstructures. First, the microstructure of active particles are generated with a certain number of primary particles, whose sizes and orientations can strictly follow given statistical distributions with binning method, even with limited particle numbers. Second, images carrying essential characteristics of microstructure evolution are incorporated into the model. A hybrid loss combining Mean Squared Error (MSE) and Structural Similarity Index (SSIM) is employed, along with a scheduled sampling training strategy, to enhance prediction accuracy. The model’s out-of-sample predictive performance has also been evaluated. Additionally, a microcrack density-based damage model is also used to assess microstructure damage evolution. This work reveals that the proposed approach achieves highly accurate predictions, providing valuable insights into microstructure behavior.