dataset_card.md

Dataset Card: SparseWake

Dataset Name

SparseWake

Intended Use

Evaluation of sparse temporal hydrodynamic state estimation from body-fixed induced-flow histories.

Not Intended Use

SparseWake is not a full biological lateral-line pressure model, a multi-neighbor schooling model, or a replacement for direct CFD solvers.

Data Source

The processed data were generated from WakeSchool, an external fish-schooling simulator with DNS-parameterized wake and body-flow components. The artifact starts from processed HDF5 benchmark datasets and does not redistribute simulator source code.

Released Fields

• X_raw: wake-plus-potential induced velocity features, sample shape [N, 6, 3] after loading.
• X_wake_raw: wake-only induced velocity features where available.
• X_potential_raw: potential-only induced velocity features where available.
• X_external_raw, X_self_raw, X_total_raw: self-signal controls where available.
• y: labels with columns delta_x, delta_y, theta_rel, sin_phi, cos_phi, phi.
• groups: wake phase index.
• pose_id: pose index for pose-holdout splitting in the sample dataset; full datasets can derive it from phase blocks.
• region_id: region code for close wake, near side, and mid wake.
• sensor_world_positions: sensor coordinates for each sample.

Units and Conventions

Lengths are nondimensionalized by fish body length. Angles are stored in radians. Reported orientation metrics use degrees. The fish body is represented with semi-axes a = 0.5 L and b = 0.075 L.

Sensor Layout

Six body-fixed sensors are arranged as anterior, midbody, and posterior left/right pairs. Sensor names are stored in documentation and sample HDF5 attributes.

Train/Validation/Test Protocol

Main results use a held-out-pose test protocol. For each seed, 20% of follower pose identities are assigned to the test set, and all saved wake phases for those poses are withheld from training. The remaining non-test samples are split into training and validation with a 15% validation fraction. Thus the test set is pose-disjoint from training, while training and validation may contain different phases of the same non-test pose.

Known Limitations

The current benchmark uses induced velocity rather than pressure, pressure gradients, or wall shear. It is single-neighbor unless explicitly stated. Multi-neighbor assignment, closed-loop control, and biological self-filtering are outside the released benchmark.

Ethical Considerations

The dataset is synthetic fluid-dynamics data. It contains no human subjects, no animal measurements, and no personal data.