Open Access

PROCEEDINGS

A Machine Learning Framework for Isogeometric Topology Optimization

by

1 The State Key Lab of Digital Manufacturing Equipment and Technology, School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, 430074, China
2 School of Mathematics & Computer Science, Wuhan Polytechnic University, Wuhan, 430023, China

* Corresponding Author: Zhaohui Xia. Email:

The International Conference on Computational & Experimental Engineering and Sciences 2023, 27(3), 1-2. https://doi.org/10.32604/icces.2023.09091

Abstract

Topology optimization (TO) is an important and powerful tool to obtain efficient and lightweight structures in conceptional design stage and a series of representative methods are implemented [1-5]. TO are mainly based on the classical finite element analysis (FEA), resulting in an inconsistency between geometric model and analytical model. Besides, there are some drawbacks of low analysis accuracy, poor continuity between adjacent elements, and high computational cost for high-order meshes. Thus, isogeometric analysis (IGA) is proposed [6] to replace FEA in TO. Using the Non-Uniform Rational B-Splines (NURBS), IGA successfully eliminates the defects of the conventional FEA and forms a new research direction, isogeometric topology optimization (ITO) [7-12]. Although these methods have high accuracy and high solution efficiency, they are accompanied with significant time costs. With the rapid development of machine learning (ML) and deep learning (DL), data-driven TO methods are generated to reduce computational costs and speed up optimization process [13-18]. In this paper, an ITO method based on deep neural networks is proposed. The computational time of optimization can be effectively reduced while ensuring high accuracy. With an IGAFEA two-resolution SIMP method, the ML dataset can be obtained during early iterations of the topology optimization process. Thus, there is no need to pre-process in the whole framework. Unlike existing datadriven methods, generating datasets online both significantly saves data collection time and enhances relevance to the design problem. In the optimization process, the ML model can be updated online by continuously collecting new data to ensure that the optimized topology structures are closer to the standard results. Through a series of 2D and 3D design examples, the generality and reliability of the method are effectively demonstrated by the relative differences and speedups and its time-saving advantage becomes more obvious as the design scale increases. Furthermore, the impacts of framework parameters and neural network hyper-parameters on the results are studied through several controlled experiments.

---

Keywords

---