Foil

Flow around a tapered NACA0025 hydrofoil: planar PIV measurements on a fixed cross-section, paired with matched numerical simulations.

Visualizations

Real-world
Simulated

Key stats

Item Value
n_traj 99 × 2 (paired real + numerical)
n_frame 3990
\(\Delta t\) \(2.5\times 10^{-3}\) s
Resolution (real) 128×256
Resolution (sim) 128×256
Modalities (real) \(u,v\)
Modalities (sim) \(u,v,p\)
Memory 335.64 GB

Note

We use n_traj = X × 2 to indicate paired trajectories: X real-world and X numerical trajectories for the same scenario.

Physical parameters

  • Airfoil: NACA0025 tapered hydrofoil (root chord 100 mm, tip chord 20 mm)
  • Measurement plane: cross-section at \(z=50\) mm
  • Angles of attack: {0°, 5°, 10°, 15°, 20°}
  • Reynolds numbers: 19 discrete values spanning 2968–17031

HF Datasets format

This scenario is distributed as Hugging Face Datasets (Arrow) under foil/hf_dataset/ using a lazy-slicing architecture.

Data organization

  • real/ — Arrow dataset containing complete real-world trajectories
  • numerical/ — Arrow dataset containing complete numerical trajectories
  • {train|val|test}_index_{real|numerical}.json — Index files defining splits

Schema (high level)

Each Arrow row stores one complete trajectory (all 3990 frames):

  • sim_id (string): trajectory identifier (e.g., 0_2968.h5)
  • u, v (bytes): float32 arrays of shape (3990, H, W) — complete time series
  • p (bytes; numerical only): float32 array (3990, H, W)
  • vo (bytes): float32 array (3990, H, W) — vorticity
  • x (bytes): float32 array (H, W) — spatial x-coordinate grid (time-invariant)
  • y (bytes): float32 array (H, W) — spatial y-coordinate grid (time-invariant)
  • t (bytes): float32 array (3990,) — time stamps
  • shape_t (int): complete trajectory length (3990)
  • shape_h, shape_w (int): spatial dimensions

Train/val/test splits are defined by the index JSON files, which map sample indices to (sim_id, time_id) pairs.

Eval splits & subsets

We provide two layers of splitting:

  • Dataset split (train/val/test): defined by {split}_index_{type}.json files.
  • Eval subset (test_mode): an optional filter inside val/test to select trajectories by parameter regime.

The subset membership is defined by JSON mapping files (downloaded as "metadata"):

  • foil/in_dist_test_params_real.json
  • foil/out_dist_test_params_real.json
  • foil/remain_params_real.json
  • foil/in_dist_test_params_numerical.json
  • foil/out_dist_test_params_numerical.json
  • foil/remain_params_numerical.json

How to interpret these files and test_mode:

  • in_dist: in-distribution parameter settings (held out for evaluation).
  • out_dist: out-of-distribution / boundary parameter settings (OOD generalization).
  • seen: parameter settings used for training (defined by remain_params_*).
  • unseen: parameter settings not used for training (union of in_dist + out_dist).

Download

See Getting Started for full setup. Quick commands:

# Evaluation metadata (small; includes the JSON mapping files)
realpdebench download --dataset-root <DATASET_ROOT> --scenario foil --what metadata

# HF dataset shards (large)
realpdebench download --dataset-root <DATASET_ROOT> --scenario foil --what hf_dataset