Rapid prediction of particle-scale state-of-lithiation in Li-ion battery microstructures using convolutional neural networks

A machine learning (ML) model was developed to study the discharge behaviour of a LixNi0.33Mn0.33Co0.33O2 half-cell with particle-scale resolution. The ML model could predict the state-of-lithiation of the particles as a function of time and C-rate. Although direct numerical simulation has been well established in this area as the prevalent method of modeling batteries, computational expense increases going from 1D-homogenized model to particle-resolved 3D models. The present ML model was trained on a total of sixty different electrodes with various lengths for a total of 4 different C-rates: 0.25, 1, 2, and 3C. The ML model used convolutional layers, resulting in an image-to-image regression network. To evaluate model performance, the root mean squared error was compared between the state of lithiation (SoL) predicted by the ML model and ground truth results from pore-scale direct numerical simulation (DNS) on unseen electrode configurations. It was shown that the ML model can predict the SoL at better than 99% accuracy in terms of relative error, but almost an order of magnitude faster than the DNS approach. The present work was limited to 2D cases but demonstrates that ML is a viable path forward for studying real 3D microstructures.