Efficient multi-material topology optimization design with minimum compliance based on ResUNet involved generative adversarial network

Topology optimization is a common approach for material distribution in continuous structure due to its rigorous mathematical theory. However, with the increase of material types in design domain, the computational efficiency of traditional topology optimization for multiple materials problem is greatly decreased. In this paper, a novel deep learning-based topology optimization method is proposed to achieve multi-material structural design for improving computational efficiency. A large number of multi-material topological configurations are simulated by solid isotropic material with penalization (SIMP), to construct multi-material topology optimization dataset. Subsequently, ResUNet involved generative adversarial network (ResUNet-GAN) is developed for high-dimensional mapping from design parameters to the corresponding multi-material topological configuration. Finally, the ResUNet-GAN, trained by the multi-material dataset, is utilized to design multi-material topological configuration. Numerical simulations verify that the well-trained ResUNet-GAN is successfully applied to three types of cases: the cantilever beam with double materials, the cantilever beam with triple materials, and the half-MBB with triple materials. The deep learning-based topology optimization approach is superior to the conventional methods in terms of higher computational efficiency, performing the potential of such a data-driven method to accelerate the calculation of structural optimization design.