PSI - Issue 57

Yixuan Hou et al. / Procedia Structural Integrity 57 (2024) 73–78 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

75 3

The GAN architecture includes two neural networks, a generator and a discriminator. The generator produces a random synthetic 3D volume, where 2D surface profiles are radially cropped from the 3D volume. The slices sampled from the real image and the synthetic 3D volume is judged by a discriminator, and the generator will be punished if the generated surface profiles are not qualified and trained to generate new surface profiles until the discriminator cannot identify the generated surface profiles. With this, the surface profiles of EBM-manufactured micro-sized parts are virtually manufactured, and the irregularity of the surface defects can be reproduced. 2.2. FE modelling To investigate the stress concentrations at the rough surface of as-built EBM parts, an axisymmetric FE model is created with the synthetic surface profiles regenerated using GAN approach. The FE modelling process using an experimental surface profile extracted from [16] is shown in

Fig . 2 . Radial slice is cropped from 3D rendering of X-ray tomography, then it is converted into cad sketches and imported into Abaqus software to build the FE model.

Fig. 2. FE modelling process using regenerated surface profile.

In this way, a 2D axisymmetric finite element model is created with the material properties listed in Table 1. The boundary conditions are set to simulate the uniaxial tensile fatigue tests, where the applied loads are set between 150 and 370 MPa, the load ratio is 0.1. On the rough surface of the model, the mesh size is set as 0.1µm with curvature control to localize the critical notch, which results in the maximal value of Maximum Principal Stress (MPS) and is considered as the crack initiation site. A 2D, 4-node bilinear, axisymmetric stress, reduced integration with hourglass control element (CAX4R) is used to mesh the model. Besides, mesh convergence studies are performed to make sure the calculated stresses at the notch tip and critical distance are converged.

Table 1. Material properties of as-built EBM Ti6Al4V [16] Material E (Gpa) ν YS 0.2 (MPa)

UTS (MPa)

Ti6Al4V

108

0.34

753

824

2.3. Lifetime prediction model According to CDM theory, the material degradation under fatigue load can be represented by a damage variable, , which is used to address the micro and macro scale damages on the material mechanical response. The Representative Volume Element (RVE) is introduced as a scaler damage parameter, and the mechanical properties inside RVE are assumed as homogeneous [17]. The relationship between fatigue lifetime and fatigue load can be described with the following equation: