Converting to DGL graphs

Again, start from a PyVista dataset:

from cooldata.pyvista_flow_field_dataset import PyvistaFlowFieldDataset

ds_pv = PyvistaFlowFieldDataset.load_from_huggingface(
    num_samples=2, data_dir="datasets/pyvista"
)
ds_pv[0]

Loaded 2 samples from '/nfs/homedirs/peo/flow_field_dataset/docs/usage/datasets/pyvista'.
Loaded 2 samples from 'datasets/pyvista'.

2025-06-08 21:24:03.278 ( 148.697s) [    7479E03EC740]      vtkCGNSReader.cxx:4267  WARN| vtkCGNSReader (0x5857e36b72c0): Skipping BC_t node: BC_t type 'BCInflow' not supported yet.
2025-06-08 21:24:03.279 ( 148.697s) [    7479E03EC740]      vtkCGNSReader.cxx:4267  WARN| vtkCGNSReader (0x5857e36b72c0): Skipping BC_t node: BC_t type 'BCSymmetryPlane' not supported yet.
2025-06-08 21:24:03.279 ( 148.697s) [    7479E03EC740]      vtkCGNSReader.cxx:4267  WARN| vtkCGNSReader (0x5857e36b72c0): Skipping BC_t node: BC_t type 'BCTunnelOutflow' not supported yet.

Design ID: 125002

Quader 1\nTemp: 24.8\nPos: (0.02, 0.07)\nSize: (0.15, 0.05) Quader 2\nTemp: 72.1\nPos: (0.08, 0.09)\nSize: (0.14, 0.06) Quader 3\nTemp: 40.7\nPos: (0.20, 1.00)\nSize: (0.05, 0.01) Quader 4\nTemp: 29.5\nPos: (0.12, 0.03)\nSize: (0.08, 0.01) Cylinder 1\nTemp: 70.8\nPos: (0.07, 1.00)\nRadius: 0.00 Cylinder 2\nTemp: 79.8\nPos: (0.11, 1.00)\nRadius: 0.01

Bounding Box (xmin, xmax, ymin, ymax): (-0.00, 0.50, 0.00, 0.10)

Now, let us convert the PyVista dataset to a DGL graph. This requires some computation, and is implemented in a parallelized way. Also, you need to install the dgl library (see here).

from cooldata.dgl_flow_field_dataset import DGLSurfaceFlowFieldDataset

ds_dgl_surface = DGLSurfaceFlowFieldDataset(
    "datasets/dgl_surface", ds_pv, normalize=False
    
)

Let us now explore what the graphs look like:

ds_dgl_surface[0]

Graph(num_nodes=35350, num_edges=282774,
      ndata_schemes={'HeatTransferCoefficient': Scheme(shape=(), dtype=torch.float32), 'BodyID': Scheme(shape=(), dtype=torch.int32), 'SurfaceType': Scheme(shape=(), dtype=torch.int32), 'CellArea': Scheme(shape=(), dtype=torch.float32), 'Normal': Scheme(shape=(3,), dtype=torch.float32), 'ShearStress': Scheme(shape=(3,), dtype=torch.float32), 'Position': Scheme(shape=(3,), dtype=torch.float32), 'Temperature': Scheme(shape=(), dtype=torch.float32), 'Pressure': Scheme(shape=(), dtype=torch.float32)}
      edata_schemes={'dx': Scheme(shape=(3,), dtype=torch.float32)})

Again, we can visualize the graph:

Hide code cell content

 
import pyvista as pv
pv.set_jupyter_backend('static')

ds_dgl_surface.plot_surface(ds_dgl_surface[0],"Pressure")
../_images/ed36d598a41da314d99bebc4e2cf80648db0b76302059fb7fb42e7006a97fcb6.png

Now, let us normalize the data:

ds_dgl_surface[0].ndata["Pressure"].mean()

tensor(-1.7514)

ds_dgl_surface.do_normalization = True
ds_dgl_surface[0].ndata["Pressure"].mean()

tensor(-0.3887)

You can find a full training script here.

It is possible to use the volumetric data as a graph as well:

from cooldata.dgl_flow_field_dataset import DGLVolumeFlowFieldDataset


ds_dgl_volume = DGLVolumeFlowFieldDataset("datasets/dgl_volume", ds_pv)

The graphs will have a significantly larger number of nodes and edges:

ds_dgl_volume[0]

Graph(num_nodes=192498, num_edges=4690894,
      ndata_schemes={'Volume': Scheme(shape=(), dtype=torch.float32), 'TurbulentKineticEnergy': Scheme(shape=(), dtype=torch.float32), 'TurbulentDissipationRate': Scheme(shape=(), dtype=torch.float32), 'Velocity': Scheme(shape=(3,), dtype=torch.float32), 'Position': Scheme(shape=(3,), dtype=torch.float32), 'Temperature': Scheme(shape=(), dtype=torch.float32), 'Pressure': Scheme(shape=(), dtype=torch.float32)}
      edata_schemes={'dx': Scheme(shape=(3,), dtype=torch.float32)})

Again, we can normalize the graph features:

ds_dgl_volume[0].ndata["Velocity"].mean()

tensor(1.1449)

ds_dgl_volume.normalize()
ds_dgl_volume[0].ndata["Velocity"].mean()

tensor(0.1931)