PSI - Issue 79

Qinghui Huang et al. / Procedia Structural Integrity 79 (2026) 291–297

293

framework to predict both the micromechanical responses and macroscopic mechanical behavior of polycrystalline metals under complex loading conditions. The core premise is that the macroscopic nonlinear mechanical behavior of materials originates from the collective manifestation of microscopic physical processes within their constituent grains, including crystal orientation, morphological evolution, and dislocation slip (Hutchinson, (1976)). This study focuses on the fabrication of GH4169 (Inconel 718) Ni-based superalloy materials using a commercial EP-M250 LPBF system. To accurately capture the micromechanical behavior of LPBF GH4169, this study developed a CPFEM simulation model based on the representative volume element (RVE) of material. The Voronoi method was employed to construct an equivalent RVE model with an equiaxed structure and random grain orientation. Three distinct strain amplitudes (εₐ) were selected: 0.15%, 0.195% and 0.25%, with a load ratio (R) of -1. According to the experimental data of LPBF GH4169, these model parameters were calibrated, as summarized in Table 1. The accumulated plastic strain was recorded during the first 50 cycles of simulation (Li et al. (2015)). The CPFEM simulation data were subsequently used for training and validation of the neural network model. Table 1. Elasticity and crystal plasticity model parameters of the LPBF GH4169 alloy. C 11 [MPa] C 12 [MPa] C 44 [MPa] 0 γ  n κ [MPa] c 1 , c 2 b Q [MPa] 311145.511 153250.774 78947.3684 0.00025 12 130.5 1,0.5 10 12 2.2 Incremental neural network The INN designed in this paper aims to capture the evolution behavior of plastic strain from incremental value under cyclic load.

Figure 1 Incremental neural network architectures for Δ prediction. The INN adopts the concept of incremental learning to capture the evolution of material response from stepwise data. Instead of directly mapping the total value, the INN predicts the incremental quantity Δ between successive time steps, which effectively embeds the derivative information and augments the dataset (Ma et al. (2021)). The investigated network architectures include: 1) a simple network, 2) a single-hidden-layer network, and 3) a double hidden-layer network, as illustrated in Fig. 1. The model employs strain amplitude ( a ) and the current timestep accumulated plastic strain as input to predict the incremental plastic strain ( ∆ ). The architecture consists of two input layer neurons and one output layer neuron, with all three networks utilizing ReLU activation function. To