# A deep Convolutional Neural Network for topology optimization with strong generalization ability

**Authors:** Yiquan Zhang, Bo Peng, Xiaoyi Zhou, Cheng Xiang, Dalei Wang  
**arXiv:** [1901.07761](https://arxiv.org/abs/1901.07761)  
**Date:** 2019-01-23  
**Summary page:** https://www.summarizepaper.com/en/paper/1901.07761v3/

## Key points

- Deep Convolutional Neural Network (CNN) for structural topology optimization
- Focus on achieving strong generalization ability
- Neural network architecture with encoding and decoding parts
- Use of U-Net technique to enhance CNN performance
- Well-designed tensor as input, optimal structure layout as output
- Large dataset generated using conventional topology optimization approach (SIMP)
- Efficiency and accuracy evaluated by comparing with SIMP on various problems
- Significant reduction in computation costs without sacrificing optimality
- Ability to solve problems under unseen boundary conditions
- Strong generalization capabilities of the CNN
- Authors: Yiquan Zhang, Bo Peng, Xiaoyi Zhou, Cheng Xiang, Dalei Wang

## Summary

This paper presents a deep Convolutional Neural Network (CNN) for structural topology optimization, with a focus on achieving strong generalization ability. The neural network architecture consists of encoding and decoding parts, which perform down- and up-sampling operations. To enhance the performance of the CNN, the popular U-Net technique is employed. The input to the network is a well-designed tensor that includes different information channels for the problem, while the output is the layout of the optimal structure. To train the CNN, a large dataset is generated using a conventional topology optimization approach called SIMP. The proposed method's efficiency and accuracy are evaluated by comparing it with SIMP on various typical optimization problems. The results demonstrate that the proposed method significantly reduces computation costs without sacrificing design solution optimality. One notable advantage of this approach is its ability to intelligently solve problems under boundary conditions that were not included in the training dataset. This suggests that the CNN has strong generalization capabilities and can effectively handle unseen scenarios. The authors of this paper are Yiquan Zhang, Bo Peng, Xiaoyi Zhou, Cheng Xiang, and Dalei Wang. Their work contributes to advancing structural topology optimization through deep learning techniques, providing an efficient and accurate approach for solving complex optimization problems with generalization capabilities.