Supervisor:

Daniel-Ioan Stroe

Co-Supervisor:
Søren Byg Vilsen

Assessment Committee:
Huai Wang(Chair)
Professor Daniel Brandell, Uppsala University, Sweden
Professor Andreas Jossen, Technical University of Munich, Germany

Moderator:
Tamas Kerekes

Abstract:

In recent years, there’s been a big push in the transportation and energy sectors towards the use of lithium-ion batteries (LiBs) for their high energy density, efficient charging, long lifetime, and low self-discharge. But as these batteries are used, they degrade, leading to shorter lifetime and potential safety issues. To address this, it’s crucial to predict the State of Health (SOH) of LiBs accurately.

However, current methods rely heavily on specific data and lack a deeper understanding of how and why batteries degrade. This PhD project aims to improve the accuracy and applicability of predicting LiBs’ degradation, especially for electric vehicle (EV) applications. Instead of relying solely on data like voltage, current, and temperature, a model that combines machine learning with the underlying physics of battery degradation will be developed. This approach enables more reliable predictions of LiB performance over time, improving both efficiency and safety.

Accelerated aging experiments are being conducted on Nickel-Manganese-Cobalt-Oxide (NMC) battery cells to investigate their degradation under varied conditions including fast charging, temperature fluctuations, and dynamic discharging profiles. The aim is to stimulate different dominant mechanisms and generate battery aging dataset. To strike a balance between efficiency and ensuring the consistency of aging mechanisms, careful selection of stress factors is imperative. Extensive calendar and cyclic aging tests have been performed to identify stress rankings and operational intervals for commercial LFP/C batteries using nonlinear mixed effects models. This process aids in the development of testing protocols that enable more accurate prediction of battery lifetime. Subsequently, a test matrix has been devised based on the identified stress factors and aging mechanisms, including SEI layer growth, anode cracking propagation, lithium plating, and electrolyte consumption. These fundamental insights serve as the basis for constructing digital twins and developing physics-informed machine learning algorithms, facilitating a deeper understanding and more precise prediction of battery performance.

Traditional LiB models struggle to accurately predict battery performance under real dynamic conditions, especially considering various aging modes and mechanisms. To address this limitation, a LiB digital twin is proposed. The digital twin is capable of capturing real measurement data and integrating the intricate coupling between SEI layer growth, anode crack propagation, and lithium plating. The dominant mechanism for the tested NMC532 cells from BOL to EOL is identified as anode particle cracking. This digital twin offers several advantages: it can estimate aging behavior from a macroscopic full-cell level down to a microscopic particle level, including voltage-current profiles in dynamic aging conditions. It enables the prediction of degradation behavior in NMC-based LiBs and supports electrochemical analysis. Moreover, an enhanced digital twin facilitates the quantification of aging effects and identification of aging modes by combining electrochemical techniques with post-mortem analysis to assess chemical and structural degradation. The effectiveness of employing an electrochemical-based digital twin to quantify the impacts of each aging mode and mechanism has been demonstrated, providing a robust physical foundation for physics-informed machine learning in predicting LiB aging behavior.

To address the limitations of black-box models and computationally intensive multiphysics models in predicting LiB degradation, a promising approach is to develop hybrid models that combine physics insights from LiBs’ digital twin model with machine learning (ML). A pure machine learning method, called mixed-input LSTM, was initially proposed to create a unified model for SOH estimation. To further enhance prediction performance, a strategy called Physics-Informed Neural Network (PINN) was introduced. In this approach, a partial differential equation governing anode particle cracking is used to constrain the neural network (NN) in predicting capacity loss. Compared to baseline NN models, PINN demonstrates improved generalization and accuracy. Specifically, the PINN achieved an average MAE, MAPE, and RMSE of 1.6%, 0.11%, and 1.9%, respectively, compared to 6.1%, 0.42%, and 8.3% for the NN model when using 50% of historical data for retraining.

By exploring digital twin model and physics-informed machine learning rooted in digital twin knowledge, the PhD thesis improves the accuracy and adaptability of battery degradation prediction while minimizing extensive data needs. The outcomes of this Ph.D. project will advance intelligent battery management, charging protocol optimization, and offer valuable insights for the design efforts of next-generation batteries.