Data‐Efficient Cycle‐Level Capacity Prediction Using 1D Deep Convolutional Network

Abstract

Fundamental battery research is constrained by the time‐consuming process of long‐term charge–discharge cycling stability testing. Here, we introduce DeepBat, a deep learning framework designed to shorten cycling tests by achieving accurate, long‐horizon prediction of cycle‐level capacities under limited data conditions. Unlike conventional lifetime forecasting models requiring extensive datasets, DeepBat leverages cross‐formulation learning to deliver high‐accuracy, cycle‐by‐cycle specific capacity predictions over extended cycling horizons using only sparse early‐cycle input data. Validated on a custom‐built library of microstructurally diverse organic electrodes encompassing 0D, 1D, and 2D carbon additives, the model achieved 98.8% mean prediction accuracy across 500 unseen cycles, demonstrating the potential to shorten experimental duration by 50% while maintaining high predictive fidelity. DeepBat outperforms traditional machine learning baselines and general‐purpose large language models, exhibiting superior robustness against data noise. SHapley Additive exPlanations‐based interpretability analysis confirms that DeepBat captures specific capacity evolution trends by assigning physically consistent importance to carbon additive dimensionality, loading, and cycle number. This work establishes DeepBat as a robust, interpretable tool for high‐accuracy, cycle‐level specific capacity predictions over extended cycling intervals, enabling model predictions to replace 50%–75% of long‐term cycling data required for material screening while maintaining reliable degradation assessment, thereby accelerating fundamental research on battery‐material technologies.