Accurate and real-time structural topology prediction driven by deep learning under moving morphable component-based framework

In the present work, we intended to discuss how to achieve real-time structural topology optimization with a significantly higher accuracy. Ideally, with an adequate computation time cost requirement, the topology optimization design problem can be formulated and solved using a direct topology optimization process, such as moving morphable component (MMC). However, the direct optimization approaches are estimated over hundreds and even thousands of design iterations, costing an innegligible computational time. There is, therefore, a need for a different approach that will be able to optimize the topologies accurately and in real-time. In this study, a topology optimization mathematical model based on a convolutional neural network is developed to replace the iterative calculations in direct topology optimization methods. The network is constructed by introducing residual learning and attention schemes into the U-Net framework. The network is trained through a dataset generated from direct MMC method. By carefully tuning the parameters during the training stage of the neural network, the network can generate topologies in real-time without any further need of the direct MMC method. Compared with state-of-the-art machine learning driven topology optimization approaches, our model achieves a significantly higher accuracy.