Data-Driven and Hybrid Modeling for Metal Fatigue: A Review of Classical Methods, Machine Learning, and Physics-Informed Neural Networks

Yuzhou Shi, Arko Suryadip Dey, Yazhou Qin

The prediction of metal fatigue life has evolved from classical empirical approaches to advanced, data-driven computational models. However, traditional methods struggle with large data scatter, complex variable-amplitude loading, and the cost of experimental testing. These limitations are particularly pronounced in additively manufactured (AM) components, which exhibit random porosity and are highly sensitive to process parameters. This review integrates classical fatigue mechanics with modern data-driven methodologies. It evaluates fatigue-life prediction for metallic alloys, welded assemblies, and AM materials. We review classical prediction tools, machine learning (ML) algorithms, deep learning architectures, and physics-informed neural networks (PINNs). ML models capture nonlinear degradation patterns but suffer from limited interpretability (“black-box” behavior) and are unable to extrapolate from small datasets. Embedding governing physical laws into PINNs helps mitigate these limitations. This approach enhances physical consistency, reduces training-data requirements, and strengthens extrapolation capability. In additively manufactured metals, defect location is often a more critical predictor of fatigue failure than defect size or morphology. To address data scarcity, we highlight the use of generative adversarial networks and transfer learning. Integrated models, combined with real-time structural health monitoring data, enable accurate, dynamic digital twins and preemptive fatigue prognosis.