SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsscientific-figures

scientific-figures

Writing & research

View repository

Updated 10-15 09:49

About this skill

Scientific Figures & Visualization — Craft Guide

Figures carry a paper's evidence. Make them accurate, clear, accessible and reproducible. A misleading or unreadable figure undermines an otherwise sound study.

1. Choose the chart for the message + data

  • Distribution: histogram, box plot, violin/strip/beeswarm (show the raw points for small n).
  • Comparison across groups: bar with error bars/CIs, or dot/forest plots (cleaner than many bars).
  • Relationship: scatter (+ regression line & CI), bubble (3rd var by size), hexbin for dense data.
  • Trend / time / energy series: line with uncertainty band; small multiples for many series.
  • Composition: stacked/100% bar; avoid pie charts (hard to compare) except for 2–3 large slices.
  • Part-to-whole over categories: grouped bars or heatmap.
  • Multidimensional: small multiples (faceting), parallel coordinates, Pareto front scatter for multi-objective optimization.
  • Networks: node-link (bibliometric maps, BIM/data graphs) with meaningful layout — see bibliometric-analysis.

2. Honest encoding (non-negotiable)

  • Bars start at zero; don't truncate axes to exaggerate differences.
  • Always show uncertainty where it exists (error bars/CIs/bands) and state what they represent (SD? SE? 95% CI?).
  • Label every axis with quantity + unit; use consistent significant figures.
  • One message per figure; don't overload. No 3D effects, no chart-junk, no dual y-axes (they mislead).
  • Don't imply more precision than the data supports; don't connect categorical points with lines.

3. Colour & accessibility

  • Use colour-blind-safe palettes: Okabe–Ito (qualitative), viridis/cividis (sequential), diverging maps (e.g. RdBu) only for true midpoints. ~8% of men have colour-vision deficiency.
  • Add redundant encoding — vary shape/linestyle/hatching as well as colour so figures survive greyscale printing.
  • Ensure sufficient contrast and legible font sizes at final print scale (check at single-column width).

4. Format & resolution

  • Vector (PDF/SVG/EPS) for plots, diagrams and anything with text/lines — scales without pixelation.
  • High-DPI raster (≥300 dpi TIFF/PNG) only for photographs/point-cloud renders.
  • Embed fonts; match the manuscript font where the venue requires it; respect column-width limits.

5. Captions & references

  • Captions go below figures (above tables). Make them self-contained: what the figure shows, key abbreviations, sample size, and what error bars denote.
  • Every figure must be referenced in the text and numbered in order. Use \cref in LaTeX (see latex-academic).

6. Reproducible figure code (prefer this over manual drawing)

  • Python: matplotlib/seaborn — set rcParams (font size, dpi, savefig format='pdf'); seaborn for statistical plots; export with plt.savefig('fig.pdf', bbox_inches='tight').
  • R: ggplot2 — theme_minimal()/theme_classic(), scale_color_viridis_*, ggsave('fig.pdf', ...).
  • Interactive/exploration: plotly/bokeh (export static for print).
  • Keep the script + data in-repo so the figure regenerates exactly (reproducibility — see open-science-reproducibility).

Example (matplotlib, accessible defaults):

import matplotlib.pyplot as plt
plt.rcParams.update({"figure.dpi": 150, "savefig.format": "pdf",
                     "font.size": 9, "axes.spines.top": False, "axes.spines.right": False})
okabe_ito = ["#E69F00","#56B4E9","#009E73","#F0E442","#0072B2","#D55E00","#CC79A7","#000000"]
# ... plot with these colors + distinct markers/linestyles, labelled axes (units), error bars ...
plt.savefig("fig_results.pdf", bbox_inches="tight")

7. Diagrams (workflows, systems, concept maps)

  • Mermaid — quick flowcharts/graphs that render in markdown (use for research-workflow and PRISMA diagrams in this toolkit).
  • TikZ — publication-grade vector diagrams inside LaTeX (full control, consistent fonts).
  • Graphviz (DOT) — automatic layout for graphs/ontologies/BIM data structures.
  • draw.io/Inkscape — manual vector diagrams (export PDF/SVG).
  • Keep hierarchy clear, label relationships, and avoid crossing edges where possible.

8. Visual / graphical abstracts

  • One clear figure conveying the study's core idea/finding for the table of contents (many AEC journals offer this). Self-explanatory, minimal text, on-brand with the paper's figures.

9. Integrity

  • Never fabricate or "beautify" data, splice images, or selectively crop to mislead (image manipulation is misconduct — see research-ethics-integrity). Disclose any legitimate image processing.
  • If the user hasn't supplied data, deliver the design + reproducible code with clearly marked placeholder data, never invented results.