Bi-directional evolutionary 3D topology optimization with a deep neural network

The FEM-based topology optimization repeats usually finite element analyses many times to converge to the stopping criteria. If the near-optimal topology data are available in advance at the beginning of an optimization process, the iterative computation could be greatly reduced. In an effort to obtain swiftly optimum topology solutions, the deep learning and neural networks with a special segmentation scheme of digital images are combined with the BESO (bi-directional evolutionary structural optimization) topology method in this study. The pre-trained digital images of 3200 optimum topologies construct the design domain for the main topology optimization. Additionally, a new post-processor is developed in order to reconstruct the relative locations among finite elements in the raw outputs generated by the neural network. The proposed method has been demonstrated to be efficient in lowering the iterations with several 2D and 3D optimization examples. The iteration counts can be reduced 63% for a 2D example and by 72.5% for a 3D one, compared to BESO results alone.

Automated summary

This study combines deep learning and neural networks with the BESO topology method, utilizing pre-trained digital images to quickly obtain optimal topology solutions. A new post-processor is developed to reconstruct the relative locations among finite elements in the raw outputs, resulting in improved optimization efficiency.