Navier–stokes Generative Adversarial Network: a physics-informed deep learning model for fluid flow generation

Numerical simulation in Computational Fluid Dynamics mainly relies on discretizing the governing equations in time or space to obtain numerical solutions, which is expensive and time-consuming. Deep learning significantly reduces the computational cost due to its great nonlinear curve fitting capability, however, the data-driven models is agnostic to latent relationships between input and output. In this paper, a novel deep learning named Navier–Stokes Generative Adversarial Network integrated with physical information is proposed. The Navier–Stokes Equation is added to the loss function of the generator in the form of residuals, which means physics loss in this paper. Then, the proposed model is trained in the framework of generative adversarial network. Experimental results show that proposed model significantly outperform similar models, mean absolute error are decreased by 62.29, 78.42 and 78.61% on pressure and velocity components. What’s more, effectiveness of introducing physics loss is also verified.