Data-driven three-dimensional topology optimization using the ResUNet-based generative adversarial network

Topology optimization is a layout design method that determines material distribution in the design domain to obtain the optimal structural performance under specific constraints. However, the high design freedom in three-dimensional structures enormously increases computational complexity, which limits the engineering applicability of conventional topology optimization methods. In this paper, a novel data-driven topology optimization method is presented to enhance the computational efficiency of three-dimensional structural design. First, a ResUNet-based generative adversarial network is proposed to establish the nonlinear mapping from the design domain to the corresponding three-dimensional topological configuration. Subsequently, a universal three-dimensional parametric encoding strategy is developed to extend the three-dimensional topology optimization dataset for training the data-driven model. Finally, the well-trained ResUNet-based generative adversarial network is deployed to realize the end-to-end design for three-dimensional structures. Numerical simulations on three representative cases—a cantilever beam, an MBB (Messerschmitt-B o ·· lkow-Blohm) beam, and a practical wheel design—validate the effectiveness, robustness, and generalization capability of ResUNet-based generative adversarial network. This research significantly enhanced the design efficiency of three-dimensional topology optimization in engineering, providing a new paradigm for data-driven topology optimization methods.