SOE·Crystal

Snow crystal growth from vapour

Growth ▸

Speed 4 steps/frame

These crystals take thousands of steps to form, but the GPU runs a step in under a tenth of a millisecond — fast enough that the whole thing is over in a second if you let it. This paces it back to something you can watch.

Stop at step 15,000

Growth halts here. The library presets carry the step count the paper grew them to — these crystals never truly finish, they just get bigger, so where you stop is a choice.

Gravner–Griffeath ▸

diffuse → freeze → attach → melt

ρ vapour density 0.58

β anisotropy 2.00

κ direct freezing 0.0000

α aftergrowth mass 0.000

θ aftergrowth vapour 0.0020

μ melting 0.000

γ sublimation 0.0000515

σ noise 0

At σ = 0 the model is fully deterministic: the same dials give you the identical crystal every single time, and total mass is exactly conserved — which is how the engine checks itself (see Readout).

Seed 1

Which particular crystal you get — but only once σ is above zero. With no noise there is nothing for a seed to vary: the model is deterministic and every run is the same flake. Real crystals differ because the air is noisy, and no two ever took the same path. That is what these two dials are, together.

ρ sweeps plate → sectored plate → dendrite. β sweeps fern → dendrite → sectored plate. μ promotes faceting. The authors note most random parameter choices give something dull — the interesting sets are narrow.

Reiter ▸

u′ = u + (a/12)(Σ₆ − 6u)

a diffusion 1.000

Centre weight is 1 − a/2, so a must stay below 2 or the centre cell takes negative weight and growth explodes.

b background vapour 0.40

c vapour addition 0.0001

Reiter's own notation is a, b, c — not α, β, γ. High c gives plates, sectors and broad stellar forms; low c with low b gives dendrites. Growth is very sensitive to a: 0.500 and 0.502 look quite different.

Lattice ▸

Grid 401 × 401

A triangular lattice held in a square array — six neighbours, not eight. Raising this on its own does not give a finer crystal: growth advances at the same rate in cells, so a bigger grid is simply a bigger room. For finer detail raise Stop at step too and let it grow on — more cells across the same flake is what buys the finer branches. Watch the radius against the grid: past about half, the crystal meets its own wrapped edge and what you are looking at stops being true. 1601² is 2.5 million cells and the GPU barely notices. Changing this reseeds.

Ice ▸

Render ice

Growth fade 1.00

Draws the quasi-liquid a cell gathers before it freezes, so ice arrives instead of appearing. This is the model's own boundary mass, not a timer — at the instant of attachment it becomes ice with no change in total mass. Turn it down to see the bare attachment lattice.

Relief 1.00

Refraction 0.55

Dispersion 0.35

Internal structure 0.60

Sparkle 1.00

Light off the arms' side walls and facets. A flat plate is a mirror, so it flares white at one exact angle and stays quiet elsewhere — that flash is real, not an effect.

Lighting ▸

Ice has no colour. Over a crystal this thin light is absorbed essentially not at all — blue ice needs metres, not microns — so every colour you see here is a lamp, seen through the flake. That is not a shortcut: it is how snow crystals are really photographed.

Backlight

Size 0.75

The lamp behind the crystal. The flake shows it to you, bent — this is where the colour lives.

Key

Angle 140° Elevation 38°

Low elevation rakes the side walls and makes the arms glitter; high hits the flat faces and gives the mirror flash. It is the single most useful control here.

Field

Deep field colour, and dark field on the right: blank the light on the camera's axis and leave it only off to the sides. Then the only light that reaches you is light the crystal bent, and it glows out of black.

Readout ▸

form—

fill fraction—

Ice counted against the disc it reaches — the one number that says which creature this is, and the instrument every preset in the library was found with. A solid hexagon reads 0.955 and a solid disc 1.155 (a lattice cell owns √3/2 of unit area, so this scale does not stop at 1). At or past 0.955 there is no structure left — that shapeless thing is, by the authors' own admission, much the most common outcome of picking dials at random. Down the other way: 0.85 plate, 0.5 sectored, 0.3 dendrite, 0.2 fern.

step0

attached cells1

radius0

room left—

steps / sec—

mass drift—

Room left is how far the crystal is from meeting its own wrapped edge — at 0% what you are looking at stops being true. Mass drift: with noise off, total mass (vapour + boundary + ice) cannot change, so any drift is the engine marking its own work.

Sweep ▸EXPERIMENT

Eight coupled dials, and the authors' own warning that most settings grow something dull. So don't guess — sweep. Pick a dial, grow a row of crystals across its range, and read the shape of the space directly. The paper's Figures 9–12 are exactly this experiment; here it takes about two seconds.

Dial

From

Crystals 7

Each one grows from its own seed to the current stop-at-step, on the current lattice. Click any result to keep it.

Library ▸RECIPES

SOE Crystal v1.0 · Gravner–Griffeath & Reiter

vapour · ice