Multi-scale Topology Optimization for Manufacturable Graded Lattice Structure via Neural Network Prediction Model

Abstract This paper presents a multi-scale topology optimization method based on neural network (MSTO-NN) for manufacturable gradient lattice structures (MGLS), which takes into account the manufacturing limitation of fine feature size. The establishment of the optimization model is firstly based on mechanics-based homogenization method to obtain the equivalent properties of lattice structure with larger cells. To balance computational efficiency and accuracy, a prediction model of the equivalent properties of lattice structures is constructed using neural networks, which replaces micro-scale finite element analysis. The sensitivity analysis of neural networks is derived and integrated into multi-scale topology optimization. To ensure the manufacturability of lattice structure, the material interpolation for MGLS based on ordered SIMP is proposed to filter the fine feature structure during optimization process. Numerical examples under 2D and 3D boundary conditions are performed to validate the proposed method. The results show that the proposed multi-scale topology optimization method can effectively realize the utilization of materials, improve the stiffness of the structure and ensure the feasibility of the manufacturing process. Moreover, the comparison of numerical examples reveals that the size of cell and fine feature size limitation have a significant impact on the structural design and performance.