基于深度学习的多材料结构拓扑优化方法

Topology Optimization of Multi-material Structures Based on Deep Learning

摘要:

提出了一种基于深度卷积神经网络（CNN）的多材料结构拓扑优化方法，实现在不需要任何迭代分析的情况下，在极短的时间内预测出多材料优化结构。研究中，采用了流行的U-Net网络结构，以提高神经网络的边界提取能力。通过有序多材料SIMP（各向同性实材料惩罚密度法）插值方法（Ordered SIMP）生成随机加载条件、质量分数及成本分数下的多材料优化结构数据集，训练得到深度学习神经网络。将所提出方法的效率和精度与传统算法进行比较，对该方法的性能进行评价，结果表明，该方法在几乎不牺牲设计方案性能的前提下，显著降低计算成本。该方法对于拓扑优化在未来多材料结构设计实践中具有巨大潜力和广阔应用前景。

Abstract:

A topology optimization method of multi-material structure based on deep convolution neural network （CNN） is proposed， which can predict the optimized structure of multi-material in a very short time without any iteration. The popular U-Net network structure is adopted to improve the edge extraction ability of neural network. To train the network， the ordered multi-material SIMP （isotropic real material penalty density method） interpolation method （Ordered SIMP） is used to generate multi-material optimal structure data sets under random loading conditions， mass fraction and cost fraction. The efficiency and accuracy of the proposed method are compared with traditional algorithms， and the performance of the proposed method is evaluated. The results show that the proposed method can significantly reduce the computational cost with little sacrifice on the performance of the design scheme. The proposed method has great potential and broad application prospects for topology optimization in the practice of multi-material structure design in the future.