Physics-informed neural networks for battery response prediction

This work introduces methods of physics-informed neural networks (PINNs) lying on a spectrum between these two approaches, enabling prediction of lithium-ion cell response with low computational cost while utilising available data and known behaviours.<br><br>For electrical behaviour, the equivalent-circuit model (ECM) was replaced with one of two approaches: a Gradient-Informed Neural Network (GINN) incorporating gradient constraints into its loss function, and a Model-Informed Neural Network (MINN) that embeds an ECM within the training loop.

Both approaches reduced computation time by up to two magnitudes compared to the ECM (2.34% of runtime). Memory requirements were also lower when simulating more than five cells simultaneously.<br><br>For thermal predictions, a cloud-point PINN was developed which learns the solution of the heat equation for given boundary conditions. By enforcing the governing PDE and boundary conditions, this model learns to predict cell temperature at any resolution for static and dynamic load profiles.