Combined Machine Learning and Electrochemical Impedance Spectroscopy to Diagnose and Predict the State-of-Health of Commercial SMD Solid-State Batteries

Abstract

Solid-state batteries (SSBs) are considered as one alternative to conventional lithium-ion batteries as they enable safer operation. A detailed look into the reaction kinetics and the possible aging mechanisms of the solid electrolytes (SEs) and the interface of SE and active material at the composite electrodes was of major interest in recent years12. Cell scale modeling and prediction of the cell performance and the degradation however are lacking, although they would offer fast evaluation beyond material-intensive and time-consuming experiments. Besides, diagnosis with non-destructive tests is crucial for the development, design, production and application of SSBs. For example, cell with degradation need to be identified by battery management systems and balanced for the safety of the entire energy system.Herein we characterize a group of commercial solid-state batteries3 using electrochemical impedance spectroscopy (EIS) and train a machine learning model to diagnose their state-of-health (SoH) throughout cycling. The SSBs used have multilayered electrodes with a ceramic SE and need to be formed by customers. To shorten the measurements period, the experiments were conducted with distinct formation and cycling protocols, which leads to faster and various degradation modes during their lifetime. The SoH of the tested SSBs ranges from 40% to 140% SoH for the customer-defined rated capacity of 200 µAh. The EIS data was collected intermittently between the 5th cycle to the 100th cycle and was organized into 108 datasets for training and validation. Fig.1a) shows parts of the EIS experiment data. The machine learning (ML) model was then trained by using the EIS data and operating condition parameters as input values and the SoH during cycling as output property. The EIS data was preprocessed and selected with a feature selection model. Significant features, e. g., impedance data from a certain frequency range, were selected and analyzed qualitatively. The applied feature engineering transfers data from non-gaussian into gaussian distribution, which enables the usage of Bayesian ridge regression algorithms to avoid overfitting.Despite the manufacturing tolerance and inherent production deviation among the cells, the model achieves a root mean squared error of 2.5% for the SoH diagnose during validation. To predict the future SoH after 10 more cycles, it achieves a root mean squared error of 2.7%. Fig.1b) and 1c) present the diagnosed/predicted SoH vs. the observed SoH. Different from semi-empirical models using EIS data to fit equivalent circuit models (ECM), we have preprocessed EIS data in the frequency domain and applied ML directly for the SoH prediction. Our method avoids importing additional inaccuracies during the ECM fitting process. Furthermore, the feature analysis reflects the dominating changes in certain frequency ranges. The tested SSBs have stronger correlations between the imaginary part of the impedance in the high frequency range and their SoH prediction. We presume that the mechanical failures in the materials applied in the SSB might dominate its degradation in this phase, as the capacitance in bulk impedance has the strongest correlation with the SoH fading.All in all, the herein presented work highlights the promise of combining data-driven modeling with EIS characterization to predict the performance of complex electrochemical systems, and can be expanded to other cell geometries and further battery technologies.