Machine Learning-Assisted Multi-Objective Optimization of Battery Manufacturing from Synthetic Data Generated by Physics-Based Simulations

Abstract

Recent efforts in the ongoing energy transition showed that Lithium-Ion Batteries (LIBs) represent a suitable technology for portable devices and electric vehicles due to their high performances and relatively good cell durability. However, the optimization of the electrode manufacturing process establishes a critical step to ensure high-quality LIBs. Nowadays, such optimization is currently based on trial-and-error and costly approaches. Indeed, LIB electrode manufacturing is a complex process involving multiple interlinked steps, including numerous process parameters that unravel sparse conditions to explore. In that sense, the digitalization of the manufacturing process highlights deep insights into its impact on the electrode and cell properties to accelerate and support autonomous productions and the prototyping activities at the experimental level. [1] We have shown in previous works that 3D-resolved physics-based models constitute meaningful tools to simulate the manufacturing process chain and bring to the community a series of unique computational 3D-resolved models for unlocking the relationships between fabrication parameters and electrode properties. [2, 3] Nevertheless, the physics-based models can still have a huge computational cost for high-throughput screening and their integration in optimization loops. In this work, we reported an innovative computational approach for digital-based manufacturing process optimization. Supported by a deterministic Machine Learning (ML)-based pipeline, we raised a multi-objective optimization of LIB electrode properties and inverse design of its fabrication process using synthetic data, as Figure 1A illustrates. In the end, this gave us the possibility to manufacture the real electrode to validate its physical relevance. To do so, we first focused on the setting of a smart design of experiments based on low-discrepancy sequences, exploring our manufacturing parameters space, to then evaluate various synthetic electrode properties related to the kinetic, ionic, and electronic transport properties of 3D-generated microstructures. The latter are generated using the ARTISTIC physics-based models' chain. [2, 3] Those synthetic data constituted a relevant database, to train and validate an accurate deterministic regression approach for fitting the electrode properties as a function of the manufacturing parameters. These so-derived regression models represented a meaningful tool to bypass the manufacturing physical modeling chain for focusing on unexplored manufacturing conditions. [4] Indeed, we took the benefits of combining these models to raise a multi-objective optimization loop using the Bayesian optimization (BO) framework to pinpoint the optimal manufacturing conditions to produce a high-performance 3D microstructure (Figure 1B). [5] This objective function granted equal weights for each electrode property when optimizing to guarantee a good balance of the optimal properties with minimal tortuosity factor, maximal effective electronic conductivity, maximal active surface area between active material and pores, and maximal density. The good results were carried out by combining BO and regression functions to considerably accelerate the search for the best manufacturing parameters to adopt. In the end, we have manufactured the electrode cell using the aforementioned optimal condition predicted by the optimization loop, to experimentally validate the physical relevance of this BO prediction. As a perspective in our future research, we aim to extend this study to more manufacturing steps (e.g. electrolyte infiltration, formation, and electrochemical performance) and additional fabrication parameters (e.g. calendering speed, particle size) and show that our approach can be transferred to accelerate the manufacturing of battery technologies and the manufacturing of composite materials in general.[1] Meyer, C, Kosfeld, M., Haselrieder, W., Kwade, A. Process modeling of the electrode calendering of lithium-ion batteries regarding the variation of cathode active materials and mass loadings. Journal of Energy Storage, 18, 2018[2] Lombardo, T., Caro, F., Ngandjong, A. C., Hoock, J-B., Duquesnoy, M., Delepine, J-C., Ponchelet, A., Doison, S., Franco, A. A. The ARTISTIC online calculator: exploring the impact of lithium-ion battery electrode manufacturing parameters interactively through your browser. Batteries & Supercaps, 5(3), 2022.[3] Duquesnoy, M., Lombardo, T., Caro, F., Haudiquez, F., Ngandjong, A. C., Xu, J., Oularbi, H., Franco, A. A. Functional data-driven framework for fast forecasting of electrode slurry rheology simulated by molecular dynamics. npj Computational Materials, 8(161), 2022.[4] ARTISTIC Project. https://www.erc-artistic.eu/. Accessed on November 2022.[5] Gunantara, N. A review of multi-objective optimization: methods and its applications. Cogent Engineering, 5(1), 2018. Figure 1