How to make a good figure and how to give a talk
After this lecture you can
Of the people who decide they like the title of your paper:
| Share | … will |
|---|---|
| 1% | actually read the paper |
| 5% | read the abstract |
| 10% | read the figure captions |
| 100% | look at the figures |
Your story should be in the figures.
Condensed from Rougier, Droettboom, Bourne, Ten simple rules for better figures, PLOS Comput. Biol. 10, e1003833 (2014). They are mandatory for the final colloquium.
Every axis has a label; every label is legible at the back of the lecture hall.
If the reader has to squint, the figure is broken.
Multi-panel figures get (a), (b), (c) … in the top left of each panel.
Caption and talk refer to panels by letter – don’t make the reader hunt.
No wasted white space, oversized markers, or decoration – and no four ideas crammed into one figure.
Give the important content the space it deserves.
Panels meant to be compared get the same axis ranges and tick spacing.
Different scales hide the very effect you want to show.
Points = measurements; lines = a model, a fit, or linear interpolation.
Don’t draw a smooth curve through measurements unless it is a fit you can defend.
The caption explains the figure; the legend only identifies the curves.
A skimming reader must understand the figure from the caption alone.
Same data – the truncated axis turns a 2% difference into a dramatic visual gap.
Same data – fix fonts, line widths, ticks, and label curves directly.
Jet is not perceptually uniform and dies in greyscale – use viridis/cividis (or a diverging map around a meaningful zero).
Backgrounds, heavy grids, shadows, legend frames: ink without information.
A paper figure is dense; a slide figure is sparse.
Re-make figures for talks: bigger fonts, fewer curves, larger markers.
A polished but unclear figure is worse than a rough but clear one.
Decide on the message first; let the visual decisions follow.
Specific units in parentheses, ideally with the quantity name:
σ (GPa) / ✓ Shear stress σ (GPa)σ / GPa, σ in GPa (DIN style)σ [GPa] – neverTell a (simple) story.
This means you have to know what your story is before you start making slides.
Corollary: anything that does not help the story has to be cut – less is more. Slides need to flow logically.
Your story flows through the slide messages. A message is a sentence, e.g. “Kokkos is slower than CUDA”.
While someone reads your slide, they are not listening to you.
Practice
Learn by heart
Polish your slides
| Section | Content |
|---|---|
| Introduction | What did you do? Which programming model? |
| Methods | How does your model work? How does it differ from what we did in the lecture? Advantages and pitfalls. |
| Results | Show the code works (a video of the lid-driven cavity is great). Performance: in million lattice updates per second. |
| Discussion | What did you learn? What would you do differently next time? |
10–15 minutes talk + 15–20 minutes discussion. All group members present.
Figure 1: Strong / weak scaling plot from Greiner & Pastewka, Proceedings of the bwHPC Symposium (2019). The \(x\)-axis is number of MPI processes; the \(y\)-axis is million lattice updates per second; the dashed lines indicate ideal scaling.
The three rules people break most:
And for the talk: one simple story, one message per slide, as little text as possible.
Figures (free): Jupyter Lab, matplotlib, Inkscape, Gimp, IPE, Blender, ColorBrewer
Figures (commercial): Origin, DataGraph, Adobe Illustrator
Slides (open source): LibreOffice Impress, Quarto, Beamer (LaTeX), Jupyter + RISE, Marimo
Slides (commercial/free): PowerPoint, Keynote, SoftMaker Office, OnlyOffice, WPS Office
HPC: Parallelization on GPUs and Accelerators