CFD simulation to predict pressure and velocity fields
See original GitHub issue❓ Questions & Help
Thanks for sharing excellent work.
I would like to do computational fluid dynamics simulation. (Something like https://www.youtube.com/watch?v=U38cKk-sxyY&ab_channel=AutodeskResearch)
Paper Link: http://pub.ist.ac.at/~bbickel/downloads/2018_sigg_Learning3DAerodynamics.pdf
Dataset:
I have access to the dataset mentioned in the paper.
It contains
Inputs: shape S, incoming wind velocity V∞, mass density ρ, and viscosity μ.
Output: drag coefficient and velocity fields and pressure values on the surface of the object.
I have read individual geometry files (*.vtk format), I can get access to pos
, tetra
and output pressure values on the surface.
My understanding is that for predicting pressure values on the surface and drag coefficient from input geometry I can convert it to point cloud and do regression on each node.
I’m not sure about the prediction of velocity fields around the object?
Would it be possible to use Geodesic operator (paper: https://arxiv.org/pdf/1802.04016.pdf) with GMMConv, to learn geometric features of the object and do convolution on the object surface?
Could you please elaborate if it’s possible to work in this direction using GNNs and which one to use? And maybe give some pointers on how should I approach this problem.
Thanks!
Issue Analytics
- State:
- Created 3 years ago
- Comments:9 (6 by maintainers)
Top GitHub Comments
Yes, the current output is
n_nodes x n_nodes
. You can change it dependent on your use-case.The processor (Eq. (1)) can certainly be implemented in PyTorch Geometric, where you want to call
propagate
twice (for mesh and world edges, respectively). The Decoder then simply takes in latent node features to predict the next time state.