PSI - Issue 57

Yixuan Hou et al. / Procedia Structural Integrity 57 (2024) 73–78

76

4

Author name / Structural Integrity Procedia 00 (2019) 000 – 000 = 1 ( 1 + ) [ ] （ 1 ） Where N f is the cycles to rupture, is the stress amplitude, is the resistance stress, which is the material parameter that resists against the fatigue damage accumulation [18], m is an experimentally obtained material constant. In this way, the fatigue lifetime of as-built EBM parts can be estimated by the local stress at the surface defects. However, since the surface defects in EBM part are very sharp, producing severe stress concentration around the notch. Therefore, using stress at the notch tip to predict fatigue lifetime usually gets very conservative estimation [19]. Therefore, the Theory of Critical Distance (TCD) is integrated into the lifetime prediction model. Based on linear elastic fracture mechanics, TCD assumes that the notched components reach their fatigue limits when the effective stress ∆ eff at the critical distance from the notch root tip equals the plain fatigue limit, ∆ 0 [20]: ∆ eff =∆ 0 （ 2 ） According to the work by Peterson [21] and Neuber [22], the TCD can be formalized as point method (PM) and line method (LM), respectively. Specially, based on the linear-elastic TCD, the critical distance can be obtained by [20]: = 1 (∆ ℎ ∆ 0 ) 2 （ 3 ） where ∆ ℎ is the range of the threshold stress intensity factor and ∆ 0 is the range of the plain fatigue limit, which are determined under the same load ratio. In this way, the stress at the critical distance from the critical notch tip is treated as effective stress, which can be integrated in the damage model to estimate the lifetime. The relationship between estimated fatigue lifetime and effective stress is expressed as follows: = 1 ( 1 + ) [ ∆ eff ] （ 4 ） Using the experimental S-N data from Reference [16], the material parameters can be calibrated, and the numbers of cycles to failure can be written as = 0 . 2146 ×[ 8592 . 69 ∆ eff ] 3 . 66 （ 5 ） 3. Results and discussion During the GAN training process, two real surface profiles extracted from [16] are fed as input datasets, after the training is completed, 80 synthetic surface profiles are cropped from the regenerated 3D volume. The surface roughness of the regenerated surface profiles (i.e. R a = 43.81 ± 3.34) shows similar values with the two input surface profiles (i.e. R a = 42.26 ± 3.23), and the R a values of all the radial slices extracted from literature are 44.10 ± 4.36. In Fig. 3, the examples of real (a) and regenerated synthetic surface (b) profiles are shown.

(a)

(b)

Fig. 3. (a) Radial slice from X-ray tomography [16], used as input dataset to train the GAN; (b) Regenerated surface profile using GAN.

To reproduce the fatigue scatter, Python programming language embedded in Abaqus is used in FE modelling. The regenerated surface profiles are randomly selected and imported into Abaqus software to analyze the stress fields under a random axial fatigue load. Next, the MPS at the critical distance is used as effective stress to estimate the