Battery manufacturing is a highly complicated and multi-stage process with large number of parameters involved in each stage. Understanding the correlation of these factors and their impact on the performance of the cells is crucial and would provide opportunity to improve the quality of cells and reduce the production time and cost. While in traditional battery manufacturing, try and error approach is used to design experiments and reveal the correlations, a smart manufacturing requires advanced and systematic modelling techniques for representing the production line. Motivated by this urgent need, this study focuses on machine-learning models to correlate parameters of the manufacturing chain with the battery performance, i.e., to predict the cell areal capacity given electrode specifications. The study is dedicated to cathode manufacturing process, addresses capacities at various Crates and quantifies the predictability unlike the existing literature.The models here are supervised machine-learning models to classify the range of the cell capacity (mAh/cm2) given the characteristics of the cathode including thickness, and porosity. To prepare the data for training the models a set of experiments were run by altering the control parameters of cathode manufacturing, i.e., comma bar gap, coating ratio and coating line speed advised by an expert. Following the cathode coating process, cathode electrodes were obtained. The experiments led to cathode electrodes with various thickness and porosities all measured via high precision equipment. The cut electrodes were then used to build half coin-cells with lithium metal as the opposite electrode. To minimize the number of free factors, cathode formulation, calendaring control parameters, drying time and temperature were kept unchanged. Then the cells were cycled in controlled temperature of 25 C at various Crates, C/20, C/5, 1C and 5C.The classifier here is a support vector machine with radial basis function kernel. For training the model a set of 110 cell data are used after removing outliers and noisy measurements. The cell capacity data are then labelled by three labels of low, medium, and high for each Crate. The label ranges are selected to distinguish the cells with undesirably low, medium but acceptable, and high or desirable capacities. In order to validate the models single and double cross-validation (CV) approaches are utilised. In CV approach data are split in 5 folds, in each run 4 of those used for training and one for hyperparameter optimisation as well as testing the model, the process is repeated in a one or two loops for single and double validation. The accuracy of the models is considered to be the number of correctly classified items compared to all classified items.The modelling results and analysis reveal that the capacity range (A= low, B = medium and C = high) can be predicted with an accuracy of 96.68% for C/20 capacity, 97.3% for C/5 capacity, 96.67% for 1C capacity and 72.6% for 5C capacity in single CV. A summary of the results as well as the associated accuracies (confusion matrices) are given in Figure for C/20 and 5Crate capacities for single and double CV methods. The classification models confirm that capacities at all Crates are having stronger correlations with thickness compared to porosity (in range above 25%) therefore thickness is considered as a more fundamental feature to classify the batteries according to their performance index. For C/20, C/5 and 1Crate capacities, the values are directly related to cathode thickness, which means that thicker electrodes at almost all porosities are led to cells with higher capacities. However, for 5Crate, it’s the opposite and for thicker electrodes, the capacity drops off accordingly. This is matched with the fact that capacity at higher Crates is limited by electronic conduction in the solid components, ionic conduction in the pores of the electrode, or both. It is also evident that the classification accuracy is lower for higher Crate capacities and suggest that extra features of cathode electrodes should be taken into account for a more accurate classification.The model-based performance prediction discussed here quantifies the accuracy for the capacity prediction at various Crates given cathode characteristics. It is highly significant for performance prediction where physical cycling tests may take hours to be completed. Therefore, it can avoid waste of time and resource and reduce the production cost for the battery manufacturer. Figure 1
Read full abstract