This paper proposes a physics-guided machine learning (PGML) framework for predicting the aging trajectory of lithium-ion (Li-ion) batteries, with a focus on accurately identifying the knee point (KP) in the capacity retention curves. The KP marks the transition from linear to nonlinear aging behavior, and its identification is crucial for predicting battery lifetime. Traditional data-driven machine learning (ML) methods often require large datasets and struggle with small samples and weak feature correlations, leading to inaccurate predictions. The PGML framework integrates physical insights into the ML model training process, enabling accurate KP prediction with limited data. The proposed method first incorporates electrode-level physical information into the model to predict the KP. It then explores the relationship between the KP and accelerated aging behavior, developing an algorithm for aging trajectory prediction. By leveraging physical knowledge, the PGML framework can transfer valuable insights to achieve accurate KP predictions with small datasets. The method was validated using a dataset of Li[NiCoMn]O₂ (NCM) cells, demonstrating that only 14 cells are needed to train a PGML model to achieve a lifetime prediction error of 2.02% using the data of the first 50 cycles, compared to 100 cells without physical insights. The PGML framework includes two main steps: analyzing aging mechanisms by reconstructing OCV curves to obtain physical information and extracting mechanism features from discharge curves, incremental capacity (IC) curves, and differential voltage (DV) curves; and employing physics-guided feature relationship recognition to minimize physical inconsistency and enhance the prediction performance of ML models on small datasets. The method was tested on NCM and LiFePO₄ (LFP) datasets, showing high accuracy in predicting the KP and aging trajectory. The results demonstrate that the PGML framework significantly improves the accuracy of KP prediction and aging trajectory prediction, especially for small datasets with limited physical information. The method outperforms conventional ML approaches in terms of prediction accuracy and is particularly effective when the features do not strongly correlate with the prediction target. The PGML framework also shows strong adaptability and accuracy for both NCM and LFP cells, highlighting its potential for battery health management and prognostics.This paper proposes a physics-guided machine learning (PGML) framework for predicting the aging trajectory of lithium-ion (Li-ion) batteries, with a focus on accurately identifying the knee point (KP) in the capacity retention curves. The KP marks the transition from linear to nonlinear aging behavior, and its identification is crucial for predicting battery lifetime. Traditional data-driven machine learning (ML) methods often require large datasets and struggle with small samples and weak feature correlations, leading to inaccurate predictions. The PGML framework integrates physical insights into the ML model training process, enabling accurate KP prediction with limited data. The proposed method first incorporates electrode-level physical information into the model to predict the KP. It then explores the relationship between the KP and accelerated aging behavior, developing an algorithm for aging trajectory prediction. By leveraging physical knowledge, the PGML framework can transfer valuable insights to achieve accurate KP predictions with small datasets. The method was validated using a dataset of Li[NiCoMn]O₂ (NCM) cells, demonstrating that only 14 cells are needed to train a PGML model to achieve a lifetime prediction error of 2.02% using the data of the first 50 cycles, compared to 100 cells without physical insights. The PGML framework includes two main steps: analyzing aging mechanisms by reconstructing OCV curves to obtain physical information and extracting mechanism features from discharge curves, incremental capacity (IC) curves, and differential voltage (DV) curves; and employing physics-guided feature relationship recognition to minimize physical inconsistency and enhance the prediction performance of ML models on small datasets. The method was tested on NCM and LiFePO₄ (LFP) datasets, showing high accuracy in predicting the KP and aging trajectory. The results demonstrate that the PGML framework significantly improves the accuracy of KP prediction and aging trajectory prediction, especially for small datasets with limited physical information. The method outperforms conventional ML approaches in terms of prediction accuracy and is particularly effective when the features do not strongly correlate with the prediction target. The PGML framework also shows strong adaptability and accuracy for both NCM and LFP cells, highlighting its potential for battery health management and prognostics.