Feature-driven machine learning screening of sodium-ion battery electrodes to support large-scale energy storage

Abstract

This study combines the Materials Genome Initiative (MGI) and machine learning to explore the screening and performance prediction of sodium-ion battery cathode materials, contributing to the development of large-scale energy storage. By performing feature importance analysis, three models (model #1, model #2, and model #3) are constructed using 21, 5, and 1 input features, respectively, to represent different stages of the electrode material screening process. Using PCA and weighting methods, the output feature ACE (average voltage, specific capacity, specific energy) is constructed to describe the overall performance of energy storage batteries. This method enables an approximate evaluation of the overall battery performance (as represented by the ACE feature) from a single, easily accessible input, making it suitable for early-stage screening, significantly enhancing machine learning efficiency while maintaining high accuracy and robustness. Furthermore, this paper discusses selecting ACE weights for different grid energy storage applications and designing specific output feature weight ratios for grid energy storage and frequency regulation. In the forward application, four promising electrode materials are identified, while in the reverse application, the model helps define the input feature range for designing new materials.