Discovery of Stable Sodium-Ion Battery Cathodes with Machine Learning

Despite their success in commercialization, lithium-ion batteries(LIBs) have a problem of cost increase due to limited Li resources. In this respect, sodium-ion batteries(SIBs) have similar chemical properties to analogous Li-based structures and high price competitiveness, making them a strong candidate to replace LIBs. The cathode mainly determines the stability and the performance of SIBs; hence, it is important to perform research on finding the next-generation materials to develop SIBs cathode materials with high voltage, capacity, and life cycle.<br/>Among the cathode candidates of SIBs, layered oxides, which have a high capacity and a simple structure, are mainly classified into O3 and P3, and especially, O3 has capable of improving electrochemical performance. However, an irreversible phase transition occurs from O3 to P3 during the (de)intercalation process. Since this phenomenon causes structural and performance degradation, it is imperative to develop a thermodynamically stable material for the commercialization of SIBs. The stability can be determined by obtaining the energy difference(ED) of the O3-P3 phase. In the process of developing new materials, selecting appropriate compounds is required, but testing all materials takes excessive time and cost. Density functional theory(DFT) calculations and machine learning(ML) are efficient for the initial new material screening. Therefore, in this study, we employed DFT and ML to explore cathode candidates which potentially resist irreversible O3-P3 phase changes during electrochemical cycling.<br/>In this study, two classification prediction models have been constructed for two states: pristine and desodiated. In the layered structure of Na_xTM_aTM^'_bO₂₄(6≤x≤12), 27 transition metals(TM) were combined at a ratio of a:b=5:1 and 3:3 to generate a total of 2,916 candidates. A total of 392 input features were generated using chemical descriptions(CDs), atomic features, and raw chemical information used in CDs calculations. The total energy of O3-P3 materials was calculated in each state using DFT calculation. ED is a value obtained by subtracting the energy of the P3 phase from the O3 phase in each of the predefined states. If pristine and desodiated ED is negative, it is classified as stable material, and this means that the phase would not be easily transformed.<br/>As a result of comparing several models using the Pycaret, the ExtraTrees Classifier showed the best performance and it was used as the base model. In the pristine state, unstable(194) and stable(1,257) structures were found, which is a serious data imbalance problem. However, in the desodiated state, it has a uniform distribution of unstable(697) and stable(754). To overcome such issues, data sampling algorithms(SMOTE, SMOTE+Tomek, SMOTE+ENN) and stratified sampling were conducted. As a result, in a pristine state, the model applying SMOTE+ENN to the base model performed the best, with significant performance improvement from an accuracy value of 0.878 to 0.976 and an F1 score of 0.404 to 0.979. In addition, the desodiated model also showed the highest performance with the SMOTE+ENN model and demonstrated that the accuracy and an F1 score increased from 0.650 to 0.945 and from 0.631 to 0.947, respectively.<br/>As a result of analyzing the calculated ED of the O3-P3 phase, 1,257 stable materials in the pristine state and 754 stable materials in the desodiated state were identified. The number of intersection data in the two states was 637, and a material with high stability that did not change the phase was found. Among them, 155 materials with a voltage of more than 3V and a theoretical capacity of more than 200mAh/g were selected. New cathode candidates were presented, showing values far beyond the performance of the existing Na cathode materials. We believe that the current study provides meaningful insights into the SIB research society by providing the database of O3-P3 phase stability and the potential cathode materials.