Introducing random defects is a type of the dominant causes of fatigue scatter of additive manufacturing (AM) materials. The fracture mechanics-based models oversimplify the fatigue damage process and ignore the key influencing factors. Given data-driven nature of machine learning (ML), sufficient samples are required to train, which generally cannot be met in practice. In view of this, this work develops two physics-guided machine learning frameworks which combine physics-based models and ML algorithms, for improving the prediction ability. In the proposed framework, ML models can consider those factors ignored by physics-based models and physics-based models can ensure the consistency with physical results. Finally, fatigue data of three AM materials are used for model evaluation and comparison. The results show that, physics-based models lack sufficient explanatory power for scatter of fatigue life. Moreover, compared with the purely data-driven methods, the proposed framework can maintain high accuracy and alleviate the over fitting phenomenon under insufficient fatigue data sources.