(Invited) Accelerating Battery Manufacturing Optimization by Combining Process Modeling with Machine Learning

Abstract

The manufacturing process of lithium ion battery (LIB) electrodes impact their architecture and practical properties, such as their energy and power densities, their durability and safety. Therefore, it is very important to optimize this manufacturing process in a proper manner. Such an optimization is highly complex because of the very significant amount of parameters and numerous interdependencies between the process steps, such as the slurry mixing, the casting, the drying, the calendering, the assembly, the electrolyte filling and the solid electrolyte interphase formation. As a consequence, traditional trial and error approaches can lead to high scrap rates in LIB cell manufacturing (as high as 30 % in prototyping activities).1 In this lecture I present a computational infrastructure able to optimize the LIB manufacturing process. Such infrastructure, called ARTISTIC,2 is supported on multiscale physics-based modeling and machine learning. Mixing, casting and drying stages are simulated by using a 3D-resolved Coarse Grain Molecular Dynamics approach, while the calendering is simulated by using a 3D-resolved Discrete Element Method. The electrolyte filling is simulated by using a 3D-resolved Lattice Boltzmann Method and the formation stage is simulated by using a 3D-resolved Finite Element Method (FEM)-based model. These physics-based models are sequentially linked in a way that the outputs of one simulated process step goes as inputs for the following one. In such a way, the interdependencies between process steps can be captured. At the end of this workflow, the 3D-FEM-based model is used to simulate the electrochemical and transport processes in the electrodes and capture the influence of manufacturing parameters on the spatiotemporal heterogeneities of lithiation/delithiation. Machine learning models, particle swarm optimization and Bayesian optimization are used to speed up the parameterization of these different models. Machine learning is also used to derive surrogate models implemented in a Bayesian optimizer predicting which manufacturing parameters to adopt in order to maximize/minimize different electrode properties.All this modeling work is supported in a 1-to-1 comparison with experimental data acquired in our battery manufacturing pilot line. The optimization capabilities of our computational infrastructure is demonstrated for LIBs with electrodes with different active material chemistries (NMC111, NMC622, LFP, graphite, silicon-graphite blends), for Sodium Ion and Solid State Batteries. Illustrative examples are provided during my lecture.3 I also discuss the ARTISTIC Online Calculator,4 a free service offered by the project to perform battery manufacturing simulations from an internet browser. Finally, I demonstrate our Virtual Reality-based digital twin of our battery manufacturing pilot line5 and how it is being used by us to train the next generation researchers and pilot line operators.