摘要
近些年来随着深度学习技术的发展,利用数据驱动技术加速拓扑优化成为可能。文中基于图像处理的思想,利用卷积自编码器学习有限元网格与固体各向同性材料密度场之间的特征表示,将拓扑优化的过程表示为一个端到端的深度学习模型。随后利用自编码器内部的反向传播和传统的有限元求解器,在拓扑优化的迭代过程中建立在线学习机制对自编码器模型的参数进行更新。最后通过数值研究和设计实例,证明了基于卷积自编码器的拓扑优化框架具有良好的优化效果和可扩展性,能够有效处理不同载荷和边界条件的设计问题。
With the development of deep learning in recent years, it is possible to accelerate topology optimization with the help of data-driven techniques. In this article, based on the idea of image processing, topology optimization is represented as an end-to-end deep learning model by means of the convolutional autoencoder network, through which efforts are made to work out the feature representation between the finite-element mesh and the popular Solid Isotropic Material with Penalization(SIMP) density field. Then, with the help of back propagation inside the autoencoder and the traditional finite-element solver, an online learning mechanism is set up to update the parameters of the autoencoder model during the iterative process of topology optimization. Finally, through numerical research and design examples, it is proved that the framework of topology optimization based on the convolutional autoencoder has good optimized effect and scalability, and can effectively deal with the design problems with different loads and boundary conditions.
作者
陈延展
陈勇
刘文涛
张承霖
郭虎
陈佶思
CHEN Yan-zhan;CHEN Yong;LIU Wen-tao;ZHANG Cheng-lin;GUO Hu;CHEN Ji-si(State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body,Hunan University,Changsha 410082;Huda Aisheng Automotive Technology Development Co.,Ltd.,Changsha 410205)
出处
《机械设计》
CSCD
北大核心
2022年第4期19-24,共6页
Journal of Machine Design
基金
国家重点研发计划项目资助(2019YFB1706504)。
关键词
卷积自编码器
深度学习
拓扑优化
有限元分析
convolutional autoencoder
deep learning
topology optimization
finite-element analysis
