An interpretable machine learning model for Mn-based cathode development: mapping synthesis parameters to cycling stability
Abstract
Manganese-based layered cathode materials for sodium-ion batteries have faced the challenge of capacity degradation, although numerous mechanism insights and modified strategies have been probed. Herein, a machine learning-based framework is constructed to predict manganese-based layered material systems, to overcome the deficiencies from conventional experimental approaches. Based on the chosen XGBoost model, fourteen key features are screened out by integrating improved SMOTE data enhancement and recursive feature elimination techniques. Moreover, critical parameter thresholds of the features are identified by SHAP and ALE interpretability analyses. The model is ultimately validated by experimental verification, with 83% of the designed materials for cycling stability falling within reasonable predictions. The proposed methodology bridges predictive modeling, material design, and performance improvement, through which a machine learning model is well-developed for predicting high-performance cathode materials for sodium-ion batteries.