PSI - Issue 57

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

74 2

Author name / Structural Integrity Procedia 00 (2019) 000 – 000

1. Introduction Through the combination of voids and micro-sized struts, metal lattice structures have significant advantages in weight reduction, thermal insulation, high specific strength, and energy damping, showing great potentials in the aerospace, biomedical, and transportation industries. The recent advancement of powder-bed based additive manufacturing (AM) technologies, such as electron beam melting (EBM) and selective laser melting (SLM), facilitates the manufacturing of these complex structures [1, 2]. During the AM process, various defects are produced because of the layer-based deposition characteristics. The surface defects existing in the micro-sized struts of lattice structures are easily becoming stress concentrators and initiating fatigue cracks under cyclic loads [2, 3]. To improve the fatigue strength of AM parts, various post-processing techniques are adopted, including machining, chemical etching and shot peening [4-6]. For parts with complex geometries such as lattice structures, it is difficult to adopt these post-processing techniques. Hence, the fatigue behaviors of the as-built AM parts are of great interest recently. Conventionally, the fatigue properties of AM parts are determined by experimental fatigue tests, which are expensive, time-consuming, and also rely on prior experiences. Therefore, considerable research has been conducted to study the fatigue properties of AM parts with theoretical models, such as Continuum Damage Mechanics (CDM) [7, 8] and Theory of Critical Distance (TCD) [9, 10]. Since the fatigue crack of as-built AM part is a local fatigue phenomenon, some researchers used Finite Element (FE) simulations to analyze the stress concentrations at the defects and predicted the fatigue lifetime [11, 12]. Since the fatigue scatter is an intrinsic fatigue behavior of as-built AM parts, which is resulted from the irregularities of surface defects. To reproduce the fatigue scatter using numerical approaches requires enough amount of FE models. In this study, a machine learning approach, namely Generative Adversarial Network (GAN), is applied to regenerate synthetic surface profiles of micro-sized as-built EBM parts using radial slices from X-ray tomography as input datasets. The synthetic surface profiles are then used to create FE models and investigate the stress concentrations around the surface defects. Eventually, a fatigue lifetime prediction model based on the combination of CDM and TCD is proposed to estimate and reproduce the scatter data observed experimentally. 2. Methodology 2.1. Surface regeneration with GAN The surface irregularities of as-built EBM parts increase the uncertainty of the fatigue behavior, producing a large fatigue scatter. To reproduce the fatigue scatter with numerical approaches, enough surface profile data is necessary. Previous studies showed the possibilities to generate randomly rough surface using filtering approach [13] and profile fitting function [14]. While for the rough surface of as-built EBM parts, it is difficult to adopt these methods because of the geometrical complexity of surface defects. In this work, a machine learning approach based on GAN [15] is applied to regenerate the synthetic surface profiles, using radial slices taken from literature [16] as input datasets. The training process of GAN to regenerate surface profiles is shown in Fig. 1.

Fig. 1. Schematic of the GAN training process for EBM surface profiles regeneration.