Cross-material battery capacity estimation using hybrid-model fusion transfer learning

Evaluating battery health involves navigating the intricate interplay of physical, chemical, and electrochemical processes across multiple scales-a task that becomes even more complex with the introduction of new battery materials. This necessitates substantial development, modeling, and recalibration of boundary conditions. Our study introduces a hybrid fusion model that combines convolutional neural networks (CNNs) with self-attention mechanisms to enhance battery health assessments. In total, three datasets are involved-covering 77 LFP, 20 NMC, and 18 NCA batteries-encompassing over 170,000 cycles across a broad spectrum of battery materials and operational conditions for pre-training the base model and for transfer learning. Our findings reveal that, when transferring aging knowledge from LFP to ternary batteries (NMC and NCA) under diverse chemistries, temperatures, and operational strategies, the model achieved root mean square errors (RMSEs) of 7.47 mAh and 12.4 mAh, mean absolute percentage errors (MAPEs) of 0.67 % and 1.14 %, and coefficients of determination (R2) of 0.922 and 0.918, respectively. These results demonstrate the effectiveness of our hybrid fusion model, which uses deep transfer learning and combines CNNs with self-attention mechanisms to accurately diagnose battery capacity across various types by analyzing short cycle sequences and integrating insights throughout the cell operational history.