Fatigue strength prediction of additively manufactured Ti-6Al-4V: A defect-based physics-informed neural network

Laser powder bed fusion (LPBF)-produced Ti-6Al-4V components show large scatter in fatigue strength due to defects such as surface roughness, subsurface and internal pores, residual stress and a metastable microstructure. While machine learning (ML) can capture these trends, purely data-driven approaches often lack physical grounding and generalisability. To address this limitation, this study introduces a novel modular, physics-informed framework that combines (i) feature-sensitive inputs, (ii) a unique simulation-based pretraining strategy on a NASGRO-derived synthetic database parameterised by XCT statistics, and (iii) a physics-informed neural network (PINN) with fracture-mechanics constraints, namely a Kitagawa-Takahashi (KT) envelope, monotonicity in threshold stress intensity factor ( ∆ K th ) and load-ratio ordering. The KT-constrained PINN is applied to fatigue-strength prediction in LPBF-produced Ti-6Al-4V. On a fixed stratified split of 143 experimental specimens, validation R² improves from 0.70 (baseline ANN) to 0.77 (pretrained ANN) and 0.82 (PINN), while MAE reduces from 40 to 29.9 MPa. Parity plots, residuals and KT-diagram checks show that predictions reflect defect-size trends and heat-treatment effects. Results demonstrate that this framework uniquely bridges the gap between data-driven flexibility and fracture-mechanics rigour, offering a transferable, physically consistent approach for AM fatigue prediction even when experimental defect data is sparse.