TCD–Statistics–Machine Learning Framework for Defect-Controlled Fatigue in Recycled Powder LPBF Inconel 718

Additive manufacturing (AM) represents a promising route towards more resource-efficient production of high-performance metallic components. However, the reliable assessment of fatigue behaviour in the presence of process-induced defects remains a major challenge, particularly for defect-sensitive alloys such as laser powder bed fusion (LPBF) Inconel 718. This work proposes a fatigue modelling framework integrating the Theory of Critical Distances (TCD), extreme value statistics (EVS), finite element (FE) simulations, and computational surrogate models (CSMs) to quantify the combined influence of surface roughness, internal porosity, macro-notches and size effect on fatigue strength. The methodology was validated using plain and V-notched LPBF Inconel 718 specimens manufactured from two partially recycled powder feedstocks. Surface topography was characterized through profilometry, while metallographic analyses were employed to determine pore size and location distributions. FE simulations were used to evaluate fatigue stress concentration factors (FSCFs) of surface roughness, and a surrogate model, based on FE simulation results, was used to calculate FSCFs of pores. The proposed computational framework enabled us to obtain high correlation factor R2≈0.90 and low symmetric mean absolute percentage error SMAPE≈17.8%, through FSCFs at 99th percentile, while FSCFs at 95th percentile provided R2≈0.65 and SMAPE≈34.2%. The numerical results also confirmed the fractographic evidence that surface roughness was the primary crack initiation site, with roughness-related FSCFs about 25% higher than pore-related ones. The proposed approach offers a computationally efficient engineering methodology for defect-driven fatigue assessment of additively manufactured metallic components under LPBF conditions dominated by near-spherical gas porosity.