Visualization of Vortex Shredding Mesh Reduced Reconstruction Mesh Graph

[chairithinug]

Hi,

First of all, thank you for this amazing library that you provide!

I am currently following one of the examples from this GitHub repo. Specifically, I am looking at the Vortex Shredding Mesh Reduced from CFD. I have successfully trained the model, but I would like to visualize the reconstruction of the mesh graph, if possible.

Unlike Vortex Shredding dataset with the faces, this one does not seem to provide such faces for visualization and I could not find any example code that would visualize this reconstruction of mesh graph from the latent space against the ground truth.

So I wonder if this visualization is missing or if you could point me to the file, that would be great!

Thanks!

[mnabian]

Hi @chairithinug , thanks for the kind words and for exploring the Vortex Shedding Mesh-Reduced example — glad to hear you’ve been able to train the model successfully! You’re right that this particular example doesn’t include a built-in visualization script. The mesh-reduced implementation was externally contributed, and its scope focused on reproducing the encoder-decoder and temporal-attention training pipelines rather than visualization. To visualize the reconstructed mesh fields, you can reuse the vortex shedding (MeshGraphNet) visualization script (examples/cfd/vortex_shedding_mgn/inference.py) as a starting point. That code already handles:

• Extracting nodal quantities (u, v, p) from the graph
• Building a matplotlib.tri.Triangulation from node coordinates + faces
• Rendering predicted vs. ground-truth animations side by side

You can adapt it by replacing the MeshGraphNet model call with your Mesh_Reduced.decode() outputs. The main inputs you’ll need are:

• Position_mesh (node coordinates)
• Reconstructed node values from the decoder (x)
• The ground-truth graph.x from the dataset

Since the mesh-reduced dataset doesn’t include face connectivity, you can use Delaunay triangulation (via matplotlib.tri.Triangulation) or scipy.spatial.Delaunay to build approximate faces for plotting. Treat the MeshGraphNet visualization script as a general template and modify the data-loading and model-output parts to suit the mesh-reduced workflow.

Hope this helps you get started!