Physics-Informed Machine Learning for Fatigue and Fracture Analysis and Prediction: A Scoping Review

Physics-Informed Machine Learning has emerged as a powerful paradigm for fatigue and fracture analysis, combining the data-driven flexibility of neural networks with the consistency of physical laws. This scoping review maps the extent, range, and nature of research on PIML applied to fatigue life prediction, crack growth and propagation, multiaxial fatigue, damage mechanics, prognostics and health management, and structural health monitoring, covering the literature published from 2018 up to August 2026. A comprehensive search strategy integrated the core terminology of the field with broader synonym terms, complemented by forward and backward citation searching and manual screening of the most productive venues. The combined strategy retrieved 124 records, of which 76 primary research articles satisfied the inclusion criteria and were charted and characterized in this review. The included studies are characterized through the lens of a three-bias taxonomy—observational, learning, and inductive—which organizes how physical knowledge enters the learning pipeline: through data-level constraints, loss-function modifications, and structural architectural changes, respectively. The review maps the architectural landscape of the field, from standard multi-layer perceptron-based PINNs to specialized variants including sequential attention models, physics-informed Kolmogorov–Arnold networks, geometry-aware finite encodings, graph neural networks, and probabilistic Bayesian formulations. The mapped literature indicates that PIML methods are reported to improve generalization from sparse and noisy experimental data, achieve computational gains over conventional simulation, and quantify prediction uncertainty. Persistent challenges nonetheless emerge across the mapped literature, including high training costs, sensitivity to hyperparameter and loss-weighting choices, limited transferability across problem configurations, and the black-box nature of deep models. The review identifies knowledge gaps and concludes with an agenda for future research, emphasizing generalizable and adaptive architectures, multi-scale and multi-physics formulations, and standardized benchmarks.