A generic physics-informed machine learning framework for battery remaining useful life prediction using small early-stage lifecycle data

Predicting the remaining useful life (RUL) of fast-charging lithium-ion batteries using early-stage lifecycle data is remains challenging due to limited run-to-failure data and lack of knowledge on battery degradation mechanisms. To address this issue, a generic Physics-Informed Machine Learning (PIML) framework is developed. The PIML framework consists of two parallel branches: a physics-informed (PI) branch and a data-driven branch. The PI branch is a neural network stacked by the linear projection layers with embedded physics knowledge, while the data-driven branch is a task-specific machine-learning model. In addition, a three-step training strategy is introduced, including (1) Training the data-driven branch, (2) Training the PI branch for aligning physical consistency without updating the hyperparameters in the data-driven branch, and (3) Fine-tuning both branches simultaneously to achieve optimal performance. To validate this framework, a physics-based model that represents the growth of solid electrolyte interphase (SEI) and a dilated convolutional neural network are implemented in the PI and data-driven branches, respectively. The solid electrolyte interphase-informed dilated convolutional neural network (SEI-DCN) model is demonstrated on the Stanford–MIT–Toyota-battery dataset. Using only four lifecycle data, the SEI-DCN model achieves very high prediction accuracy compared to standard dilated CNNs and other state-of-the-art models under various testing conditions and lifetime ranges. Moreover, the framework is generalizable to different physics-based battery degradation models. • Novel dual-branch parallel PIML framework to merge varied knowledge on battery degradation. • The new learning strategy ensures that PIML is lower bounded by the performance of data-driven models with imprecise knowledge. • Building SEI-informed DCNN validated on the fast-charging lithium-ion batteries and outperformed the SOTA model. • SEI-informed DCNN preserves SOTA prediction accuracy on completely new data with different operation conditions and life spans of the training set. • Investigation of the adaptability of the proposed framework for knowledge replacement across various SEI models.