Trajectories
240

200 train + 40 holdout across both regimes

Regimes
2

Cylinder wakes and Kelvin–Helmholtz shear layers

Holdout design
Beyond

Higher Re and Mach than the training band

Field size
128²

Raw physical tensors, ready for operator models

What you get

Not a scraped benchmark. Each case is a high-order solver run, cropped to the scientifically useful window, and shipped as a tensor plus metadata so a training pipeline can ingest it without a CFD preprocessor.

Cylinder wake library

Incompressible (artificial-compressibility) flow past a circular cylinder, harvested after transients die out — the statistically stationary shedding window.

  • 150 train / 30 holdout
  • Re ≈ 62–179 train; holdout at Re ≈ 180–190
  • 41 frames at dt = 0.5, channels ρ, u, v, p
  • Solid body encoded in density (1 = fluid, 0 = solid)
Compressible mixing-layer library

Kelvin–Helmholtz double shear layer under compressible Navier–Stokes, density ratio 1:2, on a periodic domain.

  • 50 train / 10 holdout
  • Mach ≈ 0.20–0.32 train; holdout at Mach ≈ 0.32–0.35
  • 101 frames over 100 time units
  • Channels ρ, u, v, p, plus total energy E

Coverage at a glance

Holdout cases sit just outside the training band. If a model scores well here, it generalized — it did not interpolate a shuffled split.

Split Parameter range Cases Role
Cylinder train Re ≈ 62.4 – 179.3 150 Fit the wake
Cylinder holdout Re ≈ 180.2 – 189.8 30 Extrapolate in Reynolds number
Mixing-layer train Mach ≈ 0.201 – 0.319 50 Fit the shear layer
Mixing-layer holdout Mach ≈ 0.320 – 0.349 10 Extrapolate in Mach number

Shipped for training pipelines

Raw tensors

float32 NumPy arrays, shape (T, C, 128, 128). Physical units — not pre-normalized — so you control train-set statistics.

Case metadata

JSON per trajectory: Re / Mach, times, grid bounds, channel stats, harvest window. No hidden preprocessing.

Geometry in-channel

The cylinder is written into density. Models that already take (ρ, u, v, p) do not need an extra mask input.

PyFR provenance

Flux reconstruction at order 3. Fine scales are smoother than the 128×128 grid suggests — claims should stay at wake / shear-layer scale.

Who it is for

Teams training operator networks, PDE surrogates, or digital-twin rollouts who need solver-grade 2D wakes and mixing layers — not toy grids.

Paired with the model

This is the data behind Chaperone-Flow-1.0. Use the tensors alone, or start from the adapted Poseidon-B checkpoint.

Scope

2D only. Cylinder Re stays below 3D wake instability. Interpolated to 128×128 from the solver mesh. Complements a campaign on your geometry — it is not a substitute for it.

Need your geometry, your Re band, or 3D?

The public library proves the pipeline. We generate the campaign and adapt the model under a commercial license.

right-image