Scientific visualisation, from the pipeline up

What it is

Not a chart,and not an illustration

Scientific visualisation draws data that already has a shape in space. The geometry is given; your job is to decide what to show of it.

The geometry is given

A mesh, a grid, a set of particles. You are not choosing axes — the data already lives somewhere.

Exploration, then explanation

Two different jobs. One is for you and can be ugly; the other is for a reader and cannot.

Every choice is a claim

An isosurface threshold, a colour range, a camera angle. Each one asserts what is worth seeing.

The pipeline

Four stages,in every tool

ParaView, VisIt and everything else implement the same sequence. Knowing it makes an unfamiliar tool legible.

ReadGet the data into memory, or arrange not to — large datasets are read in pieces, on demand.

FilterSlice, threshold, contour, resample. Reduce it to the part that answers the question.

MapTurn numbers into geometry and colour. This is where the interpretive decisions live.

RenderDraw it, to a window or straight to a file. The only stage that needs a GPU, and often not even then.

Choosing a technique

Match the drawingto the data

Scalar fields, vector fields and tensors each have a small set of things that work, and a long list that mislead.

Scalars: slice, contour, volume

A plane through it, a surface at one value, or the whole field with opacity. In roughly that order of cost.

Vectors: glyphs and streamlines

Arrows show direction locally; streamlines show where a massless particle would go. Both beat a magnitude map.

Derived quantities first

Vorticity, Q-criterion, divergence. Compute the thing you mean, then draw that, rather than hinting at it.

Time: animate, or small multiples

An animation shows evolution; a row of stills lets a reader compare. Papers usually want the second.

The commonest mistake is drawing a vector field as a colour map of its magnitude. It looks clean and it throws away direction, which is usually the thing that mattered. The second commonest is an isosurface at a value chosen because it looked good, which is an aesthetic decision presented as a result.

Colour, and scale

The rainbow mapis a bug

The one part of visualisation with a right answer, and the reason these tools are built the way they are.

viridisSequentiallow to high, perceptually even

Cool-warmDivergingwhen zero is meaningful

RainbowNeverinvents edges that are not there

A rainbow map is not perceptually uniform: the yellow-green band compresses a wide range of values into a narrow visual step, and the cyan-to-blue band does the opposite. Readers see boundaries where the data is smooth and miss changes where it is not. Use a perceptually uniform sequential map — viridis, or ParaView's Black-Body — for data that runs from low to high, and a diverging map centred on zero when the sign is what matters.

Fix the colour range across a time series before rendering frames. A per-frame automatic range makes an animation where the colours change because the scale moved, not because the physics did.

Client and server splitThe interface runs where you are; the data stays where it was produced. Only pixels cross the network.

Parallel renderingThe dataset is divided across processes, each renders its part, and the images are composited.

Batch, with no windowpvbatch and pvpython render offscreen. For a hundred frames this is the only sensible route.

On a machine

ParaView isalready built

Nothing here needs compiling. Pick a machine, and these two commands put it on there.

hub://paraview

$ dxflow workflow create --identity paraview hub://paraview

$ dxflow workflow start paraview

Start a machine What ParaView needs

Read it, then run itPrepaid and by the hour, on a machine that is yours about a minute after you ask.

Start a machine All notes