| name | physicsnemo-cfd-create-dataset-adapter |
| description | Create a new dataset adapter for the PhysicsNeMo CFD benchmarking workflow. Use when the user wants to add a new CFD dataset, write a DatasetAdapter, integrate a new mesh format, or benchmark models on custom data. |
| license | Apache-2.0 |
Create a Dataset Adapter
Guide the user through adding a new CFD dataset to the benchmarking
workflow by writing a DatasetAdapter subclass.
Reference files to read first
Before starting, read these files for context:
physicsnemo/cfd/evaluation/datasets/adapter_registry.py — base class and registry
physicsnemo/cfd/evaluation/datasets/schema.py — CanonicalCase and build_predictions_dict
physicsnemo/cfd/evaluation/datasets/adapters/drivaerml.py —
reference adapter implementation
workflows/benchmarking/notebooks/adding_a_new_dataset.ipynb —
end-to-end tutorial (writes a DrivAerStar adapter: format conversion,
field renaming, WSS sign flip, STL creation)
Step 1: Explore the new dataset
Ask the user for the dataset path, then inspect one file. Report not just
array names but their component count, dtype, and value range, plus
mesh.bounds and any geometry arrays — the decision table below needs
all of these:
import numpy as np
import pyvista as pv
mesh = pv.read("<path_to_one_file>")
print(f"Type: {type(mesh).__name__}, Points: {mesh.n_points}, Cells: {mesh.n_cells}")
print(f"Bounds (xmin,xmax,ymin,ymax,zmin,zmax): {mesh.bounds}")
for loc, data in [("cell", mesh.cell_data), ("point", mesh.point_data)]:
for name in data.keys():
arr = np.asarray(data[name])
comps = arr.shape[1] if arr.ndim > 1 else 1
print(f" [{loc}] {name}: comps={comps}, dtype={arr.dtype}, "
f"range=({arr.min():.3g}, {arr.max():.3g})")
print("Has Normals:", "Normals" in mesh.cell_data or "Normals" in mesh.point_data)
print("Has Area:", "Area" in mesh.cell_data or "Area" in mesh.point_data)
Identify these differences from the canonical schema:
| Question | What to look for |
|---|
| File format | .vtp, .vtu, .vtk, or a non-VTK format (CGNS, OpenFOAM, HDF5, CSV, ...)? Model wrappers ultimately read .vtp (surface) or .vtu (volume) XML — see "Reading non-PyVista source formats". |
| Directory layout | Flat directory? Nested run_<id>/ dirs? How are case IDs derived from filenames? |
| Pressure field name | The canonical key is pressure. What is the VTK array name? |
| WSS field name | The canonical key is shear_stress (N, 3). Is it a single vector or separate scalar components? |
| Sign conventions | Compare field ranges with DrivAerML. Are normals, WSS, or pressure flipped? |
| Extra arrays | Are there explicit Normals or Area arrays? DrivAerML has none — remove them if present. |
| STL files | Are separate STL geometry files available? If not, the surface mesh itself is the geometry. |
| Coordinate frame & scale | Compare mesh.bounds and units against the training dataset. Matters only for geometry-referenced checkpoints (e.g. DrivAerML-trained). See "Match geometry orientation and scale". |
| Inference domain | Surface (.vtp) or volume (.vtu)? |
Step 2: Write the adapter class
Subclass DatasetAdapter with these methods:
from pathlib import Path
from physicsnemo.cfd.evaluation.datasets.adapter_registry import DatasetAdapter, register_adapter
from physicsnemo.cfd.evaluation.datasets.schema import CanonicalCase
class MyDatasetAdapter(DatasetAdapter):
def __init__(self, root: str, **kwargs):
self._root = Path(root)
@classmethod
def inference_domain_from_kwargs(cls, kwargs=None):
return "surface"
def list_cases(self):
...
def load_case(self, case_id: str) -> CanonicalCase:
...
Map source arrays to canonical keys
ground_truth must use the framework's canonical keys, but source files
rarely use those names. The canonical vocabulary (see schema.py /
build_predictions_dict) is:
| Canonical key | Shape | Domain |
|---|
pressure | (N,) | surface, volume |
shear_stress | (N, 3) | surface |
velocity | (N, 3) | volume |
turbulent_viscosity | (N,) | volume |
Build an explicit rename map from the source names you found in Step 1:
RENAME = {"pMean": "pressure", "wallShearStress": "shear_stress"}
ground_truth = {
canon: np.asarray(mesh.cell_data[src], dtype=np.float32)
for src, canon in RENAME.items()
}
When names are ambiguous, disambiguate by: component count (a 3-comp
field is velocity or shear_stress), dtype/value range, and —
decisively — what the model's training data called each field (see
"Why conventions must match the training data"). Do not confuse this
source→canonical map with the separate canonical→VTK-name map in
output.mesh_field_names (Step 4), which controls the written arrays.
Common transformations in load_case
Reading non-PyVista source formats: pv.read handles VTK-family
files, but CFD ground truth often ships as CGNS, OpenFOAM cases, Ensight,
Tecplot, HDF5/.npz, or CSV point clouds. Only reading changes — the
target is still a canonical .vtp/.vtu mesh plus a ground_truth
dict:
import meshio, pyvista as pv
mesh = pv.wrap(meshio.read(src_path))
mesh = pv.OpenFOAMReader(case_foam_file).read()
cloud = pv.PolyData(points_xyz)
cloud["pressure"] = p_values
Format conversion (legacy .vtk → .vtp):
mesh = pv.read(vtk_path).extract_surface()
mesh.save(vtp_path)
Combining separate WSS scalars into a vector:
wss = np.stack([mesh.cell_data["WSSx"], mesh.cell_data["WSSy"], mesh.cell_data["WSSz"]], axis=1)
Removing explicit Normals/Area (DrivAerML convention):
for key in ["Normals", "Area"]:
if key in mesh.cell_data:
del mesh.cell_data[key]
Creating STL from surface mesh (when no STL is shipped):
mesh.extract_surface().triangulate().save(stl_path)
The STL must be named drivaer_{int(case_id)}.stl in the same directory
as the VTP for the model wrappers to find it.
Match geometry orientation and scale
Geometry-referenced models (e.g. DoMINO) normalize the mesh/STL
coordinates against a fixed bounding box baked into the checkpoint
from its training dataset: DoMINO reads
cfg.data.bounding_box_surface.min/max (and bounding_box.min/max for
volume) and maps every coordinate into that box. If the new dataset's
geometry sits in a different frame, origin, or unit scale, it lands in
the wrong normalized space — predictions are wrong even when field names
and signs are correct.
This only matters when the checkpoint was trained on a specific
geometry-referenced dataset (e.g. DrivAerML). For
scale/translation-invariant models, or when the model was trained on
this same dataset, skip it.
Match three things to the training dataset (DrivAerML reference bounds
below, in meters, from the DoMINO config):
| Box | min (x, y, z) | max (x, y, z) |
|---|
| Surface | -1.5, -1.4, -0.32 | 5.0, 1.4, 1.4 |
| Volume | -3.5, -2.25, -0.32 | 8.5, 2.25, 3.00 |
- Orientation / axes: same convention — x streamwise (length), y
width, z up. Permute or rotate if the new data uses a different
up-axis or flipped sign.
- Origin / position: the bounding box should start near the same
(x, y, z) minimum, so the geometry falls inside the model's domain
box.
- Scale / units: extents must be the same order of magnitude.
Millimetre data must be scaled to meters (×0.001).
Check mesh.bounds and transform in load_case before saving the prepared VTP/STL:
b = mesh.bounds
mesh.points *= 0.001
mesh.points += np.array([x_off, y_off, z_off], dtype=np.float32)
Caching pattern
Do expensive conversions lazily and cache:
def _prepare_case(self, case_id):
prepared_path = self._root / "_prepared" / f"{case_id}.vtp"
if not prepared_path.exists():
return str(prepared_path)
Step 3: Register and test
register_adapter("my_dataset", MyDatasetAdapter)
adapter = MyDatasetAdapter(root="/path/to/data")
cases = adapter.list_cases()
case = adapter.load_case(cases[0])
assert case.ground_truth is not None
assert "pressure" in case.ground_truth
Step 4: Run inference and benchmark
Build a config and run:
from physicsnemo.cfd.evaluation.config import Config
from physicsnemo.cfd.evaluation.benchmarks.engine import run_benchmark
config = Config.from_dict({
"run": {"device": "cuda:0", "output_dir": "results"},
"model": {"name": "<model_name>", "inference_domain": "<surface|volume>", ...},
"dataset": {"name": "my_dataset", "root": "/path/to/data", "case_ids": cases[:2]},
"output": {
"ground_truth_mesh_field_names": {"pressure": "<vtk_gt_name>", "shear_stress": "<vtk_gt_name>"},
"mesh_field_names": {"pressure": "<vtk_pred_name>", "shear_stress": "<vtk_pred_name>"},
},
"metrics": ["l2_pressure", "l2_shear_stress", "drag", "lift"],
"reports": {"enabled": False},
})
results = run_benchmark(config)
Step 5: Make permanent (optional)
Save the adapter to
physicsnemo/cfd/evaluation/datasets/adapters/<name>.py and register in
adapters/__init__.py:
from physicsnemo.cfd.evaluation.datasets.adapters.<name> import MyDatasetAdapter
register_adapter("my_dataset", MyDatasetAdapter)
Why conventions must match the training data
The field name mappings, sign conventions, and format conversions in the
adapter exist because the model checkpoint was trained on a specific
dataset (e.g., DrivAerML) with specific conventions. The adapter bridges
the gap between the new dataset's conventions and the training data's
conventions — not some abstract standard. If a model is retrained
directly on the new dataset, the adapter would not need these
transformations. When writing an adapter, always ask: "What conventions
did the model's training data use?" and map to those.
Gotchas
- DistributedManager: Model wrappers call
DistributedManager.initialize(). In notebooks without torchrun,
set env vars first: WORLD_SIZE=1, RANK=0, LOCAL_RANK=0,
MASTER_ADDR=localhost, MASTER_PORT=12355.
- STL naming: DoMINO looks for
drivaer_{tag}.stl, GeoTransolver
looks for drivaer_{tag}_single_solid.stl then *.stl. Both now fall
back to any *.stl in the directory.
- VTP vs VTK: Model wrappers use VTK XML readers internally. Legacy
.vtk files must be converted to .vtp/.vtu.
- Checkpoint loading: Some wrappers need
trusted_torch_load_context() for PyTorch 2.6+ checkpoint
compatibility.
- Domain-scoped metrics:
l2_pressure resolves to different
implementations for surface vs volume based on inference_domain. Use
the same metric name for both.
- Geometry frame: geometry-referenced checkpoints (DoMINO) assume
the training dataset's coordinate frame and scale. A mm-vs-m or
flipped-axis mismatch produces wrong predictions with no error raised.
See "Match geometry orientation and scale".