PDEBench

Broad time-dependent PDE benchmark with classical and realistic equations.

Registry ID: pdebench
Provider: scientific HDF5 repositories under the pdebench Hugging Face organization
Built-in configurations: burgers, advection, compressible_navier_stokes_1d

Inspect before downloading

navier datasets auth-status
navier datasets probe pdebench --configuration burgers

Small-compute loading

from navier_cfd import load_cfd_dataset

train = load_cfd_dataset(
    "pdebench",
    configuration="burgers",
    split="train",
    file_pattern="*.h5",
    trajectory_limit=32,
    max_windows=128,
    n_steps_input=4,
    n_steps_output=1,
)

The Hub repositories contain scientific trajectory HDF5 files rather than a conventional row-oriented datasets.load_dataset schema. NAVIER-CFD selectively downloads one file, opens it lazily, creates temporal windows, and returns canonical channel-last CFDSample objects.

Reproducibility

Pin the Hub revision and record the exact file, HDF5 tensor key, trajectory split seed, history length, forecast horizon, stride, resolution, and physical variables. Small-compute splits are deterministic by trajectory and must be labelled as subset splits unless they reproduce an upstream official partition exactly.