2D Fluid Flows Prediction Based on U-Net Architecture

Computational fluid dynamics (CFD) solvers provide helpful components and amenities for industrial development and the advancement of fluid flow simulations. Nevertheless, CFD solvers are not advantageous since iterative simulations require high computational resources and enormous memory for complex calculations. The deep neural network-based CFD data-driven learning method eliminates these constraints by lowering the compensation between model complicatedness and precision. In this paper, we present the feasibility of predicting fluid flow velocity fields based on CFD using U-Net architecture, a subdomain of deep learning. The experimental results show that the U-Net architecture can predict fluid flow with a total loss of 0.2223, a validation loss of 0.2728, and an accuracy of 86% from our private dataset. Our U-Net model can be used to predict fluid flows, which has been proven.