Title:Data-driven Topology Optimization of Channel Flow Problems

Abstract:Typical topology optimization methods require complex iterative calculations, which cannot meet the requirements of fast computing applications. The neural network is studied to reduce the time of computing the optimization result, however, the data-driven method for fluid topology optimization is less of discussion. This paper intends to introduce a neural network architecture that avoids time-consuming iterative processes and has a strong generalization ability for topology optimization for Stokes flow. Different neural network methods including Convolution Neural Networks (CNN), conditional Generative Adversarial Networks (cGAN), and Denoising Diffusion Implicit Models (DDIM) which have been already successfully used in solid structure optimization problems are mutated and examined for fluid topology optimization cases. The presented neural network method is tested on the channel flow topology optimization problems for Stokes flow. The results have shown that our presented method has high pixel accuracy and gains a 663 times decrease on average in execution time compared with the conventional method.