Mesh Based Neural Networks for Estimating High Fidelity CFD from Low Fidelity Input

Abstract

In this paper, we propose the design of "mesh-based deep neural network" architectures that explicitly model the spatial dependencies between the nodes of a computational fluid dynamics (CFD) mesh. Our goal is to solve the entrenched partial differential equations for the problem of dynamic high fidelity state space prediction at specific freestream conditions. Building high fidelity CFD models is computationally intensive and requires accurate modeling of the dependencies of the flow field around the aerodynamic system. We build the mesh based neural network for the nodes of the CFD mesh, on and around the aerodynamic geometry and use it to predict the high fidelity models from a low fidelity model. We call these networks mesh based neural networks as they encode the connectivity of the CFD mesh. We conduct experiments using a simulated CFD with pressure data from fluid flow fields, for the task of predicting high fidelity pressure using data from a low fidelity mesh. Our results demonstrate the feasibility of this approach and opens up the possibility of using such systems for boot strapping high fidelity computations and their use in the real world.