“Li-ion Cell modeling & machine learning based diagnosis"

“Li-ion Cell modeling & machine learning based diagnosis" With the current energy transition towards renewable sources and the rise of electric vehicles, optimizing battery use and improving our understanding of their degradation is becoming increasingly important. This thesis focuses on the modeling and diagnosis of Li-ion cells, specifically on designing and training a machine learning (ML) model capable of quantifying and characterizing degradation modes within the cell. These modes consist of the loss of active material (within both anode and cathode) and the loss of cyclable lithium throughout the battery's life. The model uses the cell's voltage response during low C-rate charging as the only input, or more specifically, using the incremental capacity method. This measurement can be carried out in electric vehicles utilizing the already installed and standard voltage and current sensors. One of the main challenges in using ML is the need for large datasets to train the models effectively. In this case, synthetic data is generated based on digital twins of the cell, which replicate the cell's behavior under certain degradation conditions. This battery model is based on the mechanistic approach, in which both electrodes are characterized independently, allowing estimation of the individual contribution of the active material of each one on the voltage response of the full cell. The synthetic data used for the ML model training comes from this approach and from the available data sets generated with the ‘Alawa Toolbox software on previous work on the topic conducted by Matthieu Dubarry and his team. The work includes the design of a new machine learning architecture. This original approach outperformed all similar models and architectures identified in the available literature, achieving a better generalization (lower error and higher accuracy) of degradation mode diagnosis for different cell configurations. Specifically, it was tested on different cells configuration not used during its training. The model was independently trained and tested for 3 chemistries, based on the synthetic databases previously mentioned. All the chemistries had graphite as negative electrode and NMC811, NCA, and LFP as positive electrode respectively. In summary, the developed model was able to quantify the extent of degradation modes in new cells (not used in the training set) with significantly higher accuracy compared to state-of-the-art models, even with 20 times fewer trainable parameters than the previous best performing model found in the literature for this task. Additionally, this work also leaves the path ready for further model verification and improvements based on the testing on aged cells. It includes the selection and programming of specific duty cycles, designed to expose each electrode of a specific commercial cell (LG MJ1 18650) to aggressive ageing and degradation conditions based on literature review and previous tests on the same chemistry. This was done with the aim of further verifying the application of the model and to track the degradation mode evolution of each of the implemented duty cycles. Finally, it introduces the possibility of using higher C-rates and Intermittent current interruption as a method to reduce the diagnosis