A machine-learning framework for isogeometric topology optimization

Isogeometric analysis has been widely applied in topology optimization in recent years, and various methods have been derived. However, most methods are accompanied by significant computational costs, which make it difficult to deal with complex models and large-scale design problems. In this paper, an isogeometric topology optimization method based on deep neural networks is proposed. The computational time of optimization can be effectively reduced while ensuring high accuracy. With the IGA-FEA two-resolution SIMP method, the machine-learning dataset can be obtained during early iterations. Unlike existing data-driven methods, online dataset generation both significantly reduces data collection time and enhances relevance to the design problem. As the iterations process, the machine learning model can be updated online by continuously collecting new data to ensure that the optimized topology structures approach the standard results. Through a series of 2D and 3D design examples, the generality and reliability of the proposed model have been verified and its time-saving advantage becomes more pronounced as the design scale increases. Furthermore, the impacts of neural network parameters on the results are studied through several controlled experiments.