Predicting the future of battery lifetime and knee point ultra-early at the formation stage

Abstract

Early lifetime prediction can accelerate lithium-ion battery development, support factory-side quality control, and reduce testing-related energy consumption. However, most existing methods rely on post-production cycling data and long test windows, limiting manufacturing scalability. Here, we report an uncertainty-aware framework that jointly predicts battery end of life and the cycle to the knee point directly from formation-stage data. The framework combines physics-informed semantic encoding, a multi-scale convolutional neural network, and temporal attention to extract electrochemically meaningful features, while Monte Carlo Dropout quantifies predictive uncertainty. It achieves a mean absolute error of 46.57 cycles and a mean absolute percentage error of 6.39% for end of life, and 60.80 cycles and 8.57%, respectively, for knee-point prediction. By moving prognostics to the formation stage, this approach reduces dependence on additional cycling tests and enables faster manufacturing feedback for cell screening and formation-protocol evaluation.