Physics-informed deep neural network framework for prediction of fatigue crack growth in LPBF-manufactured metallic alloys

A baseline Deep Neural Network (DNN) and two Physics-Informed Neural Networks (PINN-R and PINN-K<inf>max</inf>) were developed for predicting 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 only on driving force parameters such as ΔK and R , the network models integrate process parameters, mechanical properties, and fracture mechanics driving forces to capture complex interdependencies between manufacturing, material behavior, and crack growth response. The PINN models enforce monotonic constraints on the crack growth rate with respect to ΔK and either R (PINN-R) or the maximum stress-intensity factor K<inf>max</inf> (PINN-K<inf>max</inf>) by ensuring physical consistency while improving generalization. Models’ performance is assessed under two data splitting methods based on random K-fold cross-validation and grouped split by dataset IDs. A classical Walker model fitted on the same data provided a fracture-mechanics baseline. Under both data split methods, predictions from all neural models mainly fall within a ±3 × scatter band for all three alloys. The PINNs, particularly PINN-K <inf>max</inf> generally achieved better performance with lower RMSE and higher R² than the baseline DNN and Walker model, especially in the Paris and rapid-growth regimes. The results highlight the novelty of embedding physics into data-driven models by indicating a robust, physics-aware machine learning framework for fatigue crack growth prediction in LPBF alloys.