关键词: 增材制造, 疲劳寿命预测, 小样本, 物理引导, 多保真建模, 迁移学习, 不确定性量化

Abstract: Additive manufacturing (AM) shows great promise for lightweight aerospace structures and the integrated forming of complex components; however, the fatigue life of AM metallic parts exhibits significant scatter and uncertainty due to the coupled influences of processing parameters, microstructures, defects, and residual stresses, making accurate prediction highly challenging. While machine learning (ML) provides a powerful tool for capturing the nonlinear processing–microstructure–defect–life relationships, the high cost of fatigue experiments, long labeling cycles, and cross-platform data heterogeneity lead to persistent small-sample, high-noise, and poor-generalization issues. This review systematically summarizes recent progress in small-sample ML-based fatigue life prediction for AM metals, covering physics-guided ML, mechanism-driven data augmentation, multi-fidelity modeling, hybrid physics–data frameworks, transfer learning and meta-learning, uncertainty quantification, and probabilistic life prediction. We also discuss critical problems including missing metadata in databases, a lack of unified benchmarks, and the gap between experimental design and algorithm evaluation. Looking forward, considering the stringent reliability demands of aerospace applications, we identify key directions such as establishing standardized databases, achieving physically consistent data fusion, and developing unified evaluation frameworks geared toward engineering qualification.

Key words: Additively manufactured, Fatigue life prediction, Small sample, Physics-guided, Multi-fidelity modeling, Transfer learning, Uncertainty quantification