Skip to content
tensorhaze

Menu

Motion

Reduced motion shows still images instead of animation. Everything else stays the same.

Study 02 · Physics data

Field Study

Three physics simulations from The Well, each shown as one frame redrawn as vector contours. The equations behind each sit next to it.

  • Stills · real data

01Acoustic scatteringSound scattering off inclusions

02Rayleigh–Bénard convectionHeat carried by a rolling fluid

03Radiative mixing layerHot and cold gas mixing as it cools

01

The three fields

Each is one time step of one recorded run, drawn as contour lines, with the equations the run solved.

Panel 01 · Acoustic scattering

Sound scattering off inclusions

Acoustic scattering, pressure contours and inclusion outlines Meandering lines of constant pressure crossing the frame, a few of them closed loops, with dotted outlines of the inclusions.
  1. Level −0.3Pressure, line weight light.
  2. Level 0.3Pressure, line weight medium.
  3. Level 0.6Pressure, line weight strong.
  4. InclusionsOutlines of the speed-of-sound field at 0.5 and 1.7.
Figure 1Pressure contours at −0.3, 0.3 and 0.6. Dotted outlines are the inclusions, from the speed-of-sound field at 0.5 and 1.7. Rows 145 to 255 of the 256 × 256 grid, all inside the region the wave has reached at this step.1

A pressure wave crosses a square domain that holds randomly placed inclusions, small regions where the density differs strongly from the background. The speed of sound changes sharply at each inclusion, so each one scatters part of the wave.1

The runs were made with Clawpack.2 They are in the collection because they are hard for machine learning: the medium is discontinuous, where most physics datasets are smooth.

∂p/∂t + K(x, y) (∂u/∂x + ∂v/∂y) = 0
∂u/∂t + ρ(x, y)⁻¹ ∂p/∂x = 0
∂v/∂t + ρ(x, y)⁻¹ ∂p/∂y = 0

p pressure, u and v velocity, ρ density, K bulk modulus. K is fixed at 4.0 in this dataset.

File
data/train/acoustic_scattering_inclusions_chunk_0.hdf5
Trajectory and step
Trajectory 4, step 10 of 102 stored
Grid
256 × 256
Fields traced
Pressure, and speed of sound for the outlines

Panel 02 · Rayleigh–Bénard convection

Heat carried by a rolling fluid

Rayleigh–Bénard convection, buoyancy contours Contour lines of the buoyancy field: a smooth line along the bottom edge, a narrow column rising from it on the right and a rounded lobe on the left.
  1. Level 0.35Buoyancy, line weight light.
  2. Level 0.55Buoyancy, line weight medium.
  3. Level 0.75Buoyancy, line weight strong.
Figure 2Buoyancy contours at 0.35, 0.55 and 0.75. A window of 295 of the 512 columns, full height. Buoyancy is 1 at the bottom plate and 0 at the top.1

A fluid layer is heated from below and cooled from above, and is periodic from side to side. Warm fluid rises, cool fluid sinks, and the layer breaks into convection cells with boundary layers at the plates.1

This run has Rayleigh number 1e6 and Prandtl number 1, so the thermal diffusivity κ and the viscosity ν are both 1e-3. The runs were made with software based on Dedalus.3

∂b/∂t − κΔb = −u·∇b
∂u/∂t − νΔu + ∇p − b ê_z = −u·∇u
κ = (Ra · Pr)^−½    ν = (Ra / Pr)^−½

b buoyancy, u velocity, p pressure, ê_z the vertical unit vector, Δ the Laplacian. Ra is the Rayleigh number and Pr the Prandtl number.

File
data/train/rayleigh_benard_Rayleigh_1e6_Prandtl_1.hdf5
Trajectory and step
Trajectory 7, step 150 of 200, which is t = 37.5 of 50 at 0.25 per step
Grid
512 × 128, horizontal by vertical
Field traced
Buoyancy

Panel 03 · Radiative mixing layer

Hot and cold gas mixing as it cools

Turbulent radiative layer, log density contours Contour lines of the logarithm of gas density: doubled, ragged outlines of curling filaments in a band near the middle, with empty space on either side.
  1. Level 1.5Log density, line weight light.
  2. Level 2.3Log density, line weight medium.
  3. Level 3.8Log density, line weight strong.
Figure 3Contours of the natural log of density at 1.5, 2.3 and 3.8, with the dense gas on the left. A window of 295 of the 384 cells along the long axis, full width.1

Cold dense gas and hot dilute gas move past each other at a speed well below the speed of sound. The Kelvin–Helmholtz instability, seeded with small noise, mixes them. Each gas on its own is in thermal balance; the mixed gas at in-between temperatures is not, and cooling beats heating there. The result is a net flow of mass from the hot gas to the cold.1

This run has cooling time t_cool = 0.32, in the simulation's own units, not seconds. The step shown is the last one stored. The runs were made with Athena++; the setup is described by Fielding et al.4

∂ρ/∂t + ∇·(ρv) = 0
∂(ρv)/∂t + ∇·(ρvv + P) = 0
∂E/∂t + ∇·((E + P)v) = −E / t_cool
E = P / (γ − 1),  γ = 5/3

ρ density, v the 2D velocity, P pressure, E total energy, t_cool the cooling time.

File
data/train/turbulent_radiative_layer_tcool_0.32.hdf5
Trajectory and step
Trajectory 6, step 100 of 101, the last, at t = 159.7
Grid
128 × 384
Field traced
Density, in log space

02

How the stills are made

One slice of one run is read, traced and reduced to plain SVG paths.

  1. Read one sliceOne field, one trajectory and one step are read from the dataset's HDF5 file on Hugging Face by byte-range requests, without downloading the file. All three stills took 2,477,236 bytes.
  2. TraceMarching squares (scikit-image) finds the lines where the field crosses fixed levels. Lines shorter than 25 points are dropped.
  3. SimplifyEach line is reduced with Douglas–Peucker at a tolerance of 0.6 on a 300 × 130 drawing and rounded to 0.1. The three drawings are about 12 KB of path data.

Nothing else is added. The lines are the recorded field's own level sets, so they show where the field crosses a value and say nothing about values between the lines. The only motion on this page is each drawing tracing itself once as it comes into view.

Every choice, the dataset, file, field, trajectory, step, crop and levels, is a constant at the top of the tracing script, so the stills can be regenerated.

03

Other recorded runs

Each dataset holds many more runs than the one traced here.

DatasetWhat varies between runsRunsTotal size
acoustic_scattering_inclusions1 to 4 source rings, 1 to 15 random inclusions, and the density background4,000283.8 GB
rayleigh_benardRayleigh number 1e6 to 1e10 (5 values), Prandtl number 0.1 to 10 (7 values), and 50 initial conditions1,750358.4 GB
turbulent_radiative_layer_2Dt_cool from 0.03 to 3.16 (9 values), with 10 noise seeds each906.9 GB

No solver runs on this page. It shows recorded data only: three drawings, each from one time step of one run. Sizes are the dataset cards' estimates for all the runs together.

04

Checks and limits

What the three drawings show, and what they do not claim. Each point is stated elsewhere on this page.

One step of one run
Each drawing is a single time step of a single recorded run. It is not an average and not a sequence.
Level sets only
The lines are where the recorded field crosses a few fixed values. They say nothing about values between the lines.
A crop, not the whole field
Each figure caption gives the rows or columns drawn out of the full grid.
Simplified paths
Lines shorter than 25 points are dropped. Each remaining line is reduced at a tolerance of 0.6 on a 300 × 130 drawing and rounded to 0.1.
No solver
No solver runs on this page. It shows recorded data only.
Regenerating
Every choice, the dataset, file, field, trajectory, step, crop and levels, is a constant at the top of the tracing script, so the stills can be regenerated.

05

Sources

  1. R. Ohana et al., "The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning", NeurIPS 2024, arXiv:2412.00568. The collection is by Polymathic AI; the three datasets are released under CC BY 4.0. arxiv.org/abs/2412.00568
  2. K. T. Mandli et al., "Clawpack: building an open source ecosystem for solving hyperbolic PDEs", PeerJ Computer Science 2, e68 (2016): the solver behind the acoustic runs.
  3. K. J. Burns, G. M. Vasil, J. S. Oishi, D. Lecoanet and B. P. Brown, "Dedalus: a flexible framework for numerical simulations with spectral methods", Physical Review Research 2, 023068 (2020): the framework behind the Rayleigh–Bénard runs.
  4. D. B. Fielding, E. C. Ostriker, G. L. Bryan and A. S. Jermyn, "Multiphase gas and the fractal nature of radiative turbulent mixing layers", The Astrophysical Journal Letters 894, L24 (2020): the setup of the radiative-layer runs. ui.adsabs.harvard.edu/abs/2020ApJ...894L..24F/abstract
  5. The Well, Polymathic AI: the code and dataset index. github.com/PolymathicAI/the_well
  6. Lab project notes, 2026: the datasets, trajectories, tracing steps and citations used on this page.
  7. Lab tracing script: the output and byte count were checked by running it again on 2 Oct 2026.

Credits

Field Study simulations

The Field Study drawings simplify vector contours traced from one time step in each of three datasets in The Well: acoustic_scattering_inclusions, rayleigh_benard and turbulent_radiative_layer_2D. Polymathic AI released these datasets under CC BY 4.0.

Licence: CC BY 4.0

  • R. Ohana et al., The Well, NeurIPS 2024 Datasets and Benchmarks, arXiv:2412.00568.
  • Mandli et al., 2016 (Clawpack): acoustic scattering.
  • Burns et al., 2020 (Dedalus): Rayleigh–Bénard convection.
  • Fielding et al., 2020, The Astrophysical Journal Letters 894, L24: turbulent radiative mixing layer.

github.com/PolymathicAI/the_well

Field Study tracing software

The tracing script reads HDF5 slices with h5py and fsspec, processes the field arrays with NumPy, and extracts contours with scikit-image.

Licence: BSD 3-Clause (project code; see upstream notices for individual components)

h5py · fsspec · NumPy · scikit-image

Fonts

Tensorhaze Sans (derived from Inter 4.1, the Inter Project Authors), Geist and Geist Mono (the Geist Project Authors), Outfit (the Outfit Project Authors) and Literata (the Literata Project Authors), served from this site under the SIL Open Font License 1.1. The licence notices ship with the font files.

Licence: SIL Open Font License 1.1

  • Tensorhaze Sans: derived from Inter, Copyright 2016 The Inter Project Authors (github.com/rsms/inter).
  • Geist: Copyright 2024 The Geist Project Authors (github.com/vercel/geist-font).
  • Outfit: Copyright 2021 The Outfit Project Authors (github.com/Outfitio/Outfit-Fonts).
  • Literata: Copyright 2017 The Literata Project Authors (github.com/googlefonts/literata).

openfontlicense.org · OFL-TensorhazeSans.txt · OFL-Geist.txt · OFL-Outfit.txt · OFL-Literata.txt

Facts as of 2 Oct 2026.