Datenbasierte Parameteroptimierung der Zellfinalisierung durch maschinelles Lernen im Produktionshochlauf in der Batterieproduktion

RWTH Publications (RWTH Aachen)

Abstract

With the growing demand for cost-efficient and high-quality battery cells driven by the global transition to sustainable mobility, the optimization of production processes has become a key focus in both industrial and academic research. Among these processes, formation – an essential step at the end of the battery production process – represents a critical factor due to its significant impact on final cell quality and its high contribution to overall energy and investment costs. In early ramp-up phases of production lines, formation is often associated with high reject rates and extended ramp-up times. This is largely due to the complex interplay of material properties, upstream process parameters, and formation conditions, all of which make optimal parameterization challenging. This thesis presents the development of a data-driven optimization approach for systematic parameter setting in the formation of lithium-ion cells. The aim was to design a model using machine learning techniques that identifies optimal formation conditions while accounting for cell and quality parameters – ultimately enabling a substantial reduction in process time and cost without compromising cell quality. To support this goal, key influencing factors – particularly current rate, pressure, and formation temperature – were identified and analyzed within a experimental based framework. The analysis revealed strong interdependencies among these parameters, showing that their combined effect significantly influences formation time and electro-chemical performance. While pressure and temperature alone have a limited direct impact on formation time, they enable the use of higher current rates, thereby indirectly reducing cycle time. Additionally, the results confirmed a material-specific dependency of optimal process parameters, highlighting the need for adaptive, data-driven modeling approaches. Building on these findings, a two-stage model was developed, methodologically based on a combination of the CRISP-DM framework and DoE elements. In the first stage, the model predicts the minimally required formation time based on defined quality criteria. In the second stage, it derives the corresponding optimal process parameters. Real-world experimental data were used for training, and multiple machine learning algorithms – such as Random Forest, Gradient Boosting, and Neural Networks – were implemented and evaluated in terms of accuracy, robustness, and scalability. The best performance was achieved by combining a neural network for time prediction with a Gradient Boosting model for parameter selection. The results demonstrate the potential of the developed approach to significantly ac-celerate the parameterization process in formation and reliably predict the minimum required formation time – while maintaining cell quality and requiring only limited data input.

Authors 1

1. Sarah Wennemar corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen