RealPDEBench

Paired real measurements and matched simulations for sim-to-real forecasting.

Registry ID: realpdebench
Scenarios: cylinder, controlled_cylinder, fsi, foil, combustion
Provider repository: AI4Science-WestlakeU/RealPDEBench

Inspect the provider

navier datasets auth-status
navier datasets probe realpdebench --configuration cylinder

Small-compute loading

from navier_cfd import load_cfd_dataset

train = load_cfd_dataset(
    "realpdebench",
    configuration="cylinder",
    data_type="real",
    split="train",
    max_arrow_files=1,
    trajectory_limit=16,
    max_windows=64,
    n_steps_input=20,
    n_steps_output=20,
)

The Hub release stores complete trajectories in scenario-level Arrow shards and associated metadata/index files. NAVIER-CFD selectively downloads a limited number of Arrow shards, decodes the physical fields from the stored shape metadata, creates time windows, and returns canonical structured CFDSample objects.

Use data_type="real" or data_type="numerical". For combustion data, numerical channels can be selected by index.

Split and interpretation

A partial Arrow-shard run is explicitly marked subset_mode=True and official_split=False. It is suitable for small-machine portability and pipeline benchmarks, but it must not be presented as the full RealPDEBench official benchmark protocol.

Preserve numerical, real-only, fine-tuning, in-distribution, and OOD paradigms without leakage. Record the resolved revision, selected Arrow files, scenario, data type, trajectory identities, temporal window, field mapping, and split seed.