Assessment of Na-Ion Battery Performance Using Machine Learning

In the last couple of decades, the smart design of battery electroactive materials and cells to satisfy the demand for efficient energy storage has attracted significant research interest. Li-ion batteries have been widely employed in hybrid cars, power plants, and electronic devices. However, sodium-ion batteries are getting increasing attention as sodium is much more earth-abundant and inexpensive than lithium. In this work, we studied the impact of critical materials and cell design factors, electrode preparation methods, and operational descriptors on the discharge capacity and the cycle life of the Na-ion batteries using machine learning. The dataset was created from 355 experimental papers published in 2015-2020 and contained 1227 different experiments: 747 and 380 cases had only the anode and the cathode studies, respectively, whereas 100 cases were for the full cells. In the analysis, 38 descriptors were used on the electrode materials, electrolyte, and electrode structure, in addition to material synthesis and electrode preparation methods. On the other hand, peak discharge capacity and cycle life (the highest cycle number at which 80% of the peak capacity is retained) were selected as the target (output) variables. Random forest, gradient boosting, support vector machines (for regression), and decision tree (for classification) methods were used for the analysis. In the analyses, the dataset was randomly divided into two subsets: 75% to build the models (training and validation) and 25% to test the performance of the models in predicting the unseen data. Five-fold cross-validation was applied for hyperparameter optimization. Root mean square error (RMSE) was calculated as the performance indicator. The pre-analysis of the dataset presented that the highest average discharge capacity is obtained with the alloy-based anodes, followed by metal sulfides. However, their average cycle life is relatively low. In contrast, carbon-based anodes present prolonged cycle life even though they have low discharge capacities. Moreover, full-cell applications that couple the alloy-based anodes with the metal oxide cathodes show the highest average discharge capacity. Lastly, for anode or cathode half-cell studies, using a single solvent leads to higher discharge capacities and cycle life. In contrast, mixed solvents perform better in full cells. Random forest models successfully predicted the discharge capacity and demonstrated the relative significance of descriptors (Figure 1). Boruta analysis, performed for feature importance, showed that the anode and cathode types are highly effective. At the same time, the synthesis method and crystal structure were also found to be influential. On the other hand, we used classification models, which can be considered as range prediction instead of point prediction, instead of regression models to analyze the cycle life data. Decision tree classification of cycle life was quite effective, leading to heuristic rules for selecting the anode and cathode materials or methods for synthesis and electrode preparation. Material synthesis conditions are essential for high cycle life for the anode, while solvent selection is also critical for the cathode studies. Figure 1. Random Forest Regression model for the prediction of the peak discharge capacity for a) anode training, b) anode testing, c) cathode training, and d) cathode testing sets References: Oral, B., Tekin, B., Eroglu, D., Yildirim, R. (2022). Performance analysis of Na-ion batteries by machine learning. Journal of Power Sources , 549 , 232126. Figure 1