Approaches for Modelling the Mechanical Behaviour of Lithium-Ion Cells and the Application of Machine Learning Techniques within Numerical Simulation

  • Schmid, Alexander (Vehicle Safety Institut - TU Graz)
  • Pasquale, Angelo (ESI Group chair @ ENSAM)
  • Ellersdorfer, Christian (Vehicle Safety Institut - TU Graz)
  • Chinesta, Francisco (ESI Group chair @ ENSAM)
  • Ziane, Mustapha (ESI Group)
  • Raffler, Marco (Vehicle Safety Institut - TU Graz)
  • Feist, Florian (Vehicle Safety Institut - TU Graz)

Please login to view abstract download link

Lithium-ion cells are increasingly utilized in high-performance batteries for electric vehicles and other applications. These cells are constructed by stacking multiple layers of electrode (anode and cathode) and separator materials on top of each other to form a laminate structure. To ensure optimal performance and safety of these cells, simulation and experimentation are being used to analyse their behaviour under different conditions. Different modelling approaches are commonly used to simulate the mechanical behaviour. These differ in the level of detail. Detailed layer models represent the multitude of anodes, cathodes and separators individually. However, these models are very computationally intensive. With macroscopic approaches, a high level of detail is dispensed with in favour of computational efficiency. With these approaches, only the overall behaviour is depicted. The required parameters are calibrated against experiments of characteristic load cases. To minimize these issues, machine learning is used to improve existing modelling approaches and generate new ones. The Authors present a methodology where meta-modelling based on Proper Orthogonal Decomposition (POD) is used to streamline the calibration process of macroscopic models. To reduce the computational effort of detailed models the Proper Generalized Decomposition (PGD) method is used to split the behaviour of the structure into in-plane and out-of-plane characteristics. Due to that, the number of degree of freedom is reduced significantly, which has a positive influence on the computational effort. Additionally, a new modelling approach is presented which combines the advantages of both macroscopic and detailed models by using a neural network to reproduce the behaviour of a representative volume element in a computationally efficient manner. By improving existing modelling approaches and creating new ones, we can better understand and optimize the safety of lithium-ion cells.