Surrogate Modeling of Aerodynamic Flows using Implicit Neural Representations

A framework for surrogate modeling of parametric problems built upon implicit neural representations is applied to the prediction of aerodynamic flow fields. Implicit neural representations parameterize an unknown function, implicitly defined by some set of relations, using neural networks and recent advances in computer graphics have shown the viability of such an approach for geometric shape representations and modeling of boundary value problems. The solution to partial differential equations may be viewed through this lens and the non-linear independent dual system (NIDS) framework leverages these concepts. The method allows for continuous prediction of field variables, independent of discretization and topology. This is in contrast to other methods using convolutional neural networks or snapshot based decomposition, which are mesh dependent, provide predictions only at discrete locations, and may require interpolation. The NIDS framework is used to construct surrogate models for two dimensional RANS flows around bodies of varying geometry. Such models ability to accurately predict flow fields and compute aerodynamic quantities of interest from predictions on seen and unseen geometries and mesh topologies is investigated.