Deep learning driven real time topology optimization based on improved convolutional block attention (Cba-U-Net) model

Abstract

Topology optimization design provides innovative structures with excellent thermal, mechanical and acoustic performance for modern engineering. Moving Morphable Component (MMC), as an emerging explicit topology optimization method, can effectively avoid many optimization problems such as the checkerboard phenomenon, however, its optimization iteration process still consumes considerable time, which makes real-time structural topology optimization impossible. Therefore, a lightweight and high-efficiency convolutional neural network, the improved convolutional block attention U-Net (Cba-U-Net) model, is proposed for topology-optimized configuration prediction, which avoids its own tedious iterative computation process and acquires the topology configuration in real-time. It is demonstrated that the proposed network not only obtains accurate topology-optimized configurations in negligible time but also has an accuracy rate of 91.42% compared to other deep learning models. The improved Cba-U-Net model is suitable not only for Moving Morphable Components but also for other optimization algorithms, such as Solid Isotropic Material with Penalization (SIMP) and Evolutionary Structural Optimization Method (ESO). By combining deep learning with topological optimization algorithms, this form of optimization is highly generalizable for practical large-scale projects.