MEMFNet: Toward a knowledge-guided paradigm for interpretable electrochemical performance prediction

Understanding and predicting the electrochemical behavior of high-nickel cathode materials remains a central challenge in developing advanced lithium-ion energy storage systems. Although recent machine learning methods have achieved remarkable predictive performance, their generic architectures seldom embody the underlying physical and chemical mechanisms governing electrochemical processes, which limits both interpretability and generalization. We present MEMFNet, a deep learning framework specifically designed to reflect materials science knowledge through a dual-pathway architecture that mirrors the distinction between static material properties and dynamic electrochemical processes. Trained on 158,200 voltage-capacity data points from 791 discharge profiles of high-nickel cathode materials, MEMFNet reduces prediction error by 48.64 % compared to state-of-the-art methods. More importantly, the knowledge-guided architecture transforms the model from a black box into an interpretable system whose learned representations align with established electrochemical principles. By integrating domain knowledge, MEMFNet enables interpretable and scientifically meaningful learning in materials informatics.