- Online only
- New
A baseline Deep Neural Network (DNN) and a Physics-Informed Neural Network (PINN) were developed to predict fatigue crack growth rates (da/dN) in Ti-6Al-4V, IN625, and 17-4PH alloys produced by laser powder bed fusion (LPBF). Unlike traditional analytical models that rely solely on the crack-driving parameters ΔK and R [1–2], the proposed frameworks in Figure 1 incorporate process parameters, mechanical properties, and fracture mechanics variables to capture the complex interdependencies between manufacturing, material behavior, and crack growth response. This integration is particularly important because LPBF process parameters such as laser power, scan speed, hatch spacing, layer thickness, and build orientation play a critical role in determining porosity, grain structure, and…
Datenschutzbedingungen (bearbeiten im Modul "Kundenvorteile")
Lieferbedingungen (bearbeiten im Modul "Kundenvorteile")
Rücksendebedingungen (bearbeiten im Modul "Kundenvorteile")
A baseline Deep Neural Network (DNN) and a Physics-Informed Neural Network (PINN) were developed to predict fatigue crack growth rates (da/dN) in Ti-6Al-4V, IN625, and 17-4PH alloys produced by laser powder bed fusion (LPBF). Unlike traditional analytical models that rely solely on the crack-driving parameters ΔK and R [1–2], the proposed frameworks in Figure 1 incorporate process parameters, mechanical properties, and fracture mechanics variables to capture the complex interdependencies between manufacturing, material behavior, and crack growth response. This integration is particularly important because LPBF process parameters such as laser power, scan speed, hatch spacing, layer thickness, and build orientation play a critical role in determining porosity, grain structure, and residual stresses, which in turn strongly affect both material strength and fatigue crack growth behavior [3-5].
The PINN model introduces a key innovation by enforcing monotonic fracture mechanics constraints on ΔK and R by ensuring physically consistent predictions while improving generalization and interpretability. Training was conducted using a hybrid 4-fold cross-validation scheme (80% training/validation, 20% testing), combined with dropout regularization and weight decay to avoid overfitting. Both models demonstrated strong predictive accuracy, with most predictions falling within the ±3× experimental scatter band. However, the PINN consistently outperformed the DNN by achieving lower RMSE, higher R², and more reliable predictions across all crack growth regimes, particularly in near-threshold and rapid-fracture regions where data scatter is most pronounced in Figure 2. Permutation importance analysis confirmed ΔK as the dominant variable across all alloys, while also revealing that the reduced apparent influence of process parameters was largely due to the limited variability of available datasets. Despite this constraint, the results show the essential role of process parameters for providing a critical influence on crack growth resistance.
Overall, these findings highlight the novelty and impact of embedding physics into machine learning frameworks. By combining fracture mechanics principles with data-driven learning, the PINN provides accurate, interpretable, and generalizable predictions of fatigue crack growth in LPBF alloys. This approach provides a robust pathway for extending predictive capabilities to underrepresented crack regimes, guiding process optimization, and ultimately improving the structural integrity and design reliability of additively manufactured components.