The paper presents a novel method for predicting the aging trajectory of lithium-ion (Li-ion) batteries using physics-guided machine learning (PGML). The key challenge addressed is the accurate prediction of the knee point (KP), which marks the onset of accelerated aging, to improve the accuracy of battery lifetime predictions. The proposed method integrates physical information with machine learning to enhance the prediction accuracy, especially for small datasets. The KP is identified using a polynomial fit to the capacity retention curves, and its relationship with the end-of-life (EOL) is analyzed. The PGML framework includes three main components: feature extraction and physical information acquisition, physics-guided feature relationship separation, and binary machine learning training and prediction. The effectiveness of the method is demonstrated on two datasets of Li[NiCoMn]O₂ (NCM) and LiFePO₄ (LFP) cells, showing that the KP can be accurately predicted with only 14 cells for NCM and 123 cells for LFP, achieving EOL prediction errors of 2.02% and 7.48%, respectively, using the first 50 cycles of data. The results highlight the importance of incorporating physical insights to improve the accuracy of machine learning models, especially in the context of accelerated aging conditions.The paper presents a novel method for predicting the aging trajectory of lithium-ion (Li-ion) batteries using physics-guided machine learning (PGML). The key challenge addressed is the accurate prediction of the knee point (KP), which marks the onset of accelerated aging, to improve the accuracy of battery lifetime predictions. The proposed method integrates physical information with machine learning to enhance the prediction accuracy, especially for small datasets. The KP is identified using a polynomial fit to the capacity retention curves, and its relationship with the end-of-life (EOL) is analyzed. The PGML framework includes three main components: feature extraction and physical information acquisition, physics-guided feature relationship separation, and binary machine learning training and prediction. The effectiveness of the method is demonstrated on two datasets of Li[NiCoMn]O₂ (NCM) and LiFePO₄ (LFP) cells, showing that the KP can be accurately predicted with only 14 cells for NCM and 123 cells for LFP, achieving EOL prediction errors of 2.02% and 7.48%, respectively, using the first 50 cycles of data. The results highlight the importance of incorporating physical insights to improve the accuracy of machine learning models, especially in the context of accelerated aging conditions.