Neural Network Enhanced Physicsbased Surrogate Model Framework for Multiphysics Battery Simulations

Sammanfattning

Physics-based battery models are important for understanding and predicting the behaviorof batteries during operation. To fully describe their performance under differentconditions high-fidelity multiphysics models are used. While these offer high accuracy,their computational cost is high. Simplified models, such as equivalent circuit models,provide a faster alternative. However, these models are less flexible and need to be recalibratedfor different circumstances, often involving extensive testing.This thesis presents a framework for using neural networks to enhance simplified physicsbasedmodels based on high-fidelity simulation data. Battery voltage and swelling forceare simulated with a high-fidelity electrochemical-mechanical battery model for differentapplied currents and external pre-compression loads, while accounting for electrochemicallyinduced swelling. The voltage response is then modeled with a first-order non-linearequivalent circuit model, and the mechanical response with a non-linear spring along witha swelling rate. Based on the high-fidelity data, neural networks are used to approximatethe circuit elements, spring stiffness and swelling rate. Finally, symbolic regression isused to discover closed-form expressions for each neural network, revealing insights independencies.The result suggests that highly non-linear circuit elements are required to capture thedynamics of battery voltage. Compared to a fully data-driven baseline, embedding neuralnetworks within an equivalent circuit model preserved physical relations when extrapolatingbeyond the training domain. This work is limited to fully reversible swelling, andthe voltage data considers only discharges. Although the thesis project demonstrates thisframework on mechanically coupled battery simulations, we believe that the methodologyand the overall workflow can be extended to include other processes, such as thermaleffects or degradations in future work.