What is it about?
This paper describes a new method to teach incompressible fluid dynamics to convolutional neural networks (CNNs) on a 3D grid. Our method does not rely on any precomputed training data and therefore avoids the need to generate expensive training sets. The resulting CNNs yield fast and stable simulations of incompressible fluids for a wide range of Reynolds numbers, feature the Magnus effect and generalize to new fluid domains that were no covered during training.
Featured Image
Photo by Jong Marshes on Unsplash
Why is it important?
Neural surrogate models promise to drastically speed up fluid simulations. Our method allows to train such models without ground truth data. Trained models work for a wide range of Reynolds numbers, feature the Magnus effect and generalize to new fluid domains.
Read the Original
This page is a summary of: Teaching the incompressible Navier–Stokes equations to fast neural surrogate models in three dimensions, Physics of Fluids, April 2021, American Institute of Physics,
DOI: 10.1063/5.0047428.
You can read the full text:
Resources
Contributors
The following have contributed to this page