Enhancing Battery Consistency Through Physics-Machine Learning Integration: A Calendering Process-Oriented Optimization Strategy

Manufacturing tolerances inevitably induce cell-to-cell inconsistencies. These inconsistent cells are connected in series and parallel to form battery packs, which will affect the safety and reliability of the battery system. This study presents a novel optimization framework integrating the multi-level physical model with machine learning to improve battery consistency from the manufacturing perspective. The multi-level physical modeling approach is applied to establish the link between the parameter deviations of the calendering process and the battery inconsistency performance. Based on the multi-level physical model, the Monte Carlo method is used to describe parameter deviations and generate datasets of electrochemical properties. The coefficients of variations in battery capacity and resistance are calculated as the consistency evaluation index based on these datasets. The proposed optimization approach applies machine learning to reduce the computational cost of the multi-level physical simulations due to lots of Monte Carlo simulations. Combined with the multi-level physical model and neural network model, the multi-objective particle swarm optimization algorithm is adopted to provide the optimal calendering process parameter deviations by achieving the trade-off between battery consistency performance and manufacturing cost. Results indicate that the battery consistency performance is improved by controlling the precision of the calendering process and manufacturing cost. This approach can effectively give feedback and guidance to the inverse design of the manufacturing process.