Non-iterative structural topology optimization using deep learning

• Published in: Computer aided design Vol. 115; pp. 172 - 180
• Main Authors: Li, Baotong, Huang, Congjia, Li, Xin, Zheng, Shuai, Hong, Jun
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Ltd 01.10.2019; Elsevier BV
• Subjects: Artificial neural networks; Boundary conditions; Computer simulation; Computing time; Conductive heat transfer; Datasets; Deep learning; Generative adversarial network; Heat conduction; Hierarchical refinement; Iterative methods; Machine learning; Neural networks; Optimization; Pixels; Predictions; Topology optimization; Training; Deep learning; Hierarchical refinement; Heat conduction; Topology optimization; Generative adversarial network
• ISSN: 0010-4485; 1879-2685
• DOI: 10.1016/j.cad.2019.05.038

This paper presents a non-iterative topology optimizer for conductive heat transfer structures with the help of deep learning. An artificial neural network is trained to deal with the black-and-white pixel images and generate near-optimal structures. Our design is a two-stage hierarchical prediction–refinement pipeline consisting of two coupled neural networks: a generative adversarial network (GAN) for predicting a low resolution near-optimal structure and a super-resolution generative adversarial network (SRGAN) for predicting the refined structure in high resolution. Training datasets with given boundary conditions and the optimized pixel image structures are obtained after simulating a big amount of topology optimization procedures. For more effective training and inference, these datasets are generated with two different resolutions. Experiments demonstrated that our learning based optimizer can provide accurate estimation of the conductive heat transfer topology using negligible computational time. This effective incorporation of deep learning into topology optimization could enable promising applications in large-scale engineering structure design. •Non-iterative topology optimizer for conductive heat transfer structures with the help of deep learning.•Generative adversarial network with thermal boundary condition as input instead of simulated intermediate pixel images.•Two-stage hierarchical refinement pipeline for more effective training and prediction.