Physics-Informed Neural Networks for Simulation of Fluid Flow

Abstract

Physics-informed neural networks have been attracted to perform computational physics. We employ the physics-informed neural networks (PINNs) to solve fluid mechanics problems. Starting with transforming boundary value problems for Navier-Stokes equations into unconstrained optimization problems, we use feed-forward neural networks for minimizing the loss function which models physics. In this thesis, we show the capability of neural networks to simulate fluid flow by solving Navier-Stokes equations based on PINN. We present the correlation between the model structures and the performance, which are computational speed and accuracy, and observe the trainability of PINN with respect to the variations of Navier-Stokes equations.