Real-time topology optimization based on convolutional neural network by using retrain skill

Yan, Jun, Geng, Dongling, Xu, Qi and Li, Haijiang ORCID: https://orcid.org/0000-0001-6326-8133 2023. Real-time topology optimization based on convolutional neural network by using retrain skill. Engineering with Computers 39 , pp. 4045-4059. 10.1007/s00366-023-01846-3

Preview

PDF - Accepted Post-Print Version Download (1MB) | Preview

Abstract

To realize a real-time structural topology optimization (TO), it is essential to use the information during the TO process. A step-to-step training method is proposed to improve the deep learning model prediction accuracy based on the solid isotropic material with penalization (SIMP) TO method. By increasing the use of optimization history information (such as the structure density matrix), the step-to-step method improves the model utilization efficiency for each sample data. This training method can effectively improve the deep learning model prediction accuracy without increasing the sample set size. The step-to-step training method combines several independent deep learning models (sub-models). The sub-models could have the same model layers and hyperparameters. It can be trained in parallel to speed up the training process. During the deep learning model training process, these features reduce the difficulties in adjusting sub-model parameters and the model training time cost. Meanwhile, this method is achieved by the local end-to-end training process. During the deep learning model predicting process, the increase in total prediction time cost can be ignored. The trained deep learning models can predict the optimized structures in real time. Maximization of first eigenfrequency topology optimization problem with three constraint conditions is used to verify the effectiveness of the proposed training method. The method proposed in this study provides an implementation technology for the real-time TO of structures. The authors also provide the deep learning model code and the dataset in this manuscript (git-hub).

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Engineering
Publisher:	Springer Verlag
ISSN:	0177-0667
Date of First Compliant Deposit:	14 May 2023
Date of Acceptance:	13 May 2023
Last Modified:	09 Nov 2024 11:45
URI:	https://orca.cardiff.ac.uk/id/eprint/159470

Actions (repository staff only)

Edit Item