Abstract：Ti-6Al-4V alloy fabricated by laser powder bed fusion (L-PBF) holds significant promise in aerospace applications due to its superior specific strength and good corrosion resistance. However, the unique manufacturing process inevitably introduces complex internal defects, such as pores and lack of fusion, alongside metastable microstructures. These factors cause significant scatter in fatigue performance, making it challenging for traditional physical models to accurately map the nonlinear relationships along the "process—structure—property" chain. Consequently, Machine Learning (ML) has emerged as a robust data-driven approach to address these challenges. This paper provides a comprehensive review of the research progress in ML-based fatigue property prediction for L-PBF Ti-6Al-4V. First, critical influencing factors—including defect characteristics (size, location, and morphology), microstructure, surface roughness, and residual stresses—are systematically analyzed to establish a physical basis. Then, for deterministic fatigue life prediction, this paper introduces three main streams: feature engineering-based data-driven methods, which rely on manually crafted descriptors; microscopy image-based deep learning methods, which automatically extract features from raw image data; and physics-informed ML models, which integrate domain knowledge to enhance interpretability and extrapolation capability, especially with small datasets. Beyond deterministic prediction, the main ML approaches for quantification of fatigue performance uncertainty are explored. Techniques such as Bayesian Neural Networks (BNN), Dropout NN, and Gaussian Process Regression (GPR) are evaluated, specifically analyzing their strengths and limitations in decomposing and quantifying both aleatoric uncertainty (from material intrinsic randomness) and epistemic uncertainty (from model/knowledge limitations). Through this analysis, several persistent challenges are identified, including data scarcity, the lack of interpretability in complex models, and the difficulty of modeling multi-factor coupling. Finally, promising future directions are outlined. These include fostering open data platforms to alleviate data bottlenecks, advancing hybrid modeling that deeply integrates physical mechanisms, and promoting the transition of ML models from lab research to integrated digital twin platforms for in-situ quality monitoring and component-level life prediction. This paper aims to provide a methodological reference and highlight a clear path toward more reliable and industrially applicable fatigue life prediction for additively manufactured titanium alloys.
2026, Vol. 47 
