figure-and-result-storytelling

name	figure-and-result-storytelling
description	Use when designing figures, result flow, captions, and visual evidence for CFD-AI/SciML journal manuscripts.
version	0.4.0
author	CFD-AI Paper Skills maintainers
metadata	{"short-description":"Evidence-first figure planning for CFD-AI/SciML papers"}

Figure and Result Storytelling

Trigger

Use when planning or critiquing paper figures, result sections, captions, visual comparisons, or graphical abstracts.

Core rule

A figure is not decoration. It is evidence for a claim.

Progressive disclosure

• Read rubrics/figure-evidence-rubric.md before scoring a figure plan.
• For super-resolution/reconstruction figure grammar, read references/gold-papers/fukami-2019-super-resolution-jfm.md.
• For spatio-temporal super-resolution, read references/gold-papers/fukami-2021-spatiotemporal-super-resolution-jfm.md.
• For cylinder-wake prediction/rollout figures, read references/gold-papers/lee-2019-cylinder-wake-jfm.md.
• For CNN mechanism/interpretability figures, read references/gold-papers/lee-2021-cnn-wake-analysis-pof.md.
• For uncertainty figures, read references/gold-papers/maulik-2020-probabilistic-neural-networks-prf.md.
• For captions that sound decorative or AI-written, read references/gold-paper-style-patterns.md, rubrics/gold-paper-closeness-rubric.md, and examples/generic-ai-to-gold-paper-prose.md.

Figure archetypes

1. Problem schematic — domain, BC/IC, inputs/outputs, data pipeline.
2. Method schematic — architecture or workflow, but only if it clarifies mechanism.
3. Dataset/regime map — Re/geometry/parameter split; train vs test vs OOD.
4. Main quantitative comparison — metrics table/bar with fair baselines.
5. GT/pred/error fields — same color scale, units, colorbar.
6. Physical diagnostic — spectrum, conservation residual, drag/lift, heat flux, etc.
7. Ablation — mechanism evidence.
8. Generalization/OOD — unseen regime proof.
9. Efficiency — speed/memory/params.
10. Failure case — honest boundary of method.

v0.4 figure add-ons

Claim type	Required visual evidence
Spatio-temporal SR	GT/coarse-input/pred/error fields plus time-axis or frame-interval evidence.
Wake rollout	Time horizon, not only isolated snapshots; include force/frequency/residual where relevant.
Interpretability	The interpreted artifact itself: feature maps, spectra, latent masks, graph nodes, sparse terms, or policy behavior.
Control	Controlled vs uncontrolled vs classical baseline time histories and actuation constraints.

Workflow

1. Extract claims.
2. Assign at least one figure/table to each major claim.
3. Check whether every figure has a quantitative takeaway.
4. Ensure visual comparisons use same scales and units.
5. Draft captions as mini-arguments.

Caption template

Use:

Figure X. [What is shown.] [Experimental condition.] [Key quantitative/physical takeaway.] [Why it supports the claim.]

Bad caption:

Velocity field comparison.

Good caption:

Predicted and reference vorticity fields for the unseen $Re=2000$ cylinder wake. The proposed model preserves shear-layer roll-up and reduces relative $L_2$ error by 31% compared with FNO, while the error map shows remaining bias near vortex shedding extrema.

Gold-paper-like captions should behave as evidence, not decoration. They should name the condition, the compared panels, the metric or diagnostic, and the claim boundary. Avoid captions that say only “comparison of results” or “effectiveness of the proposed method.”

Output template

1. Figure sequence
2. Claim supported by each figure
3. Required panels
4. Caption draft
5. Missing plots/data
6. Visual risk checklist
7. Gold-paper closeness issues in captions, if any

Also include:

Figure/table	Claim	Rubric score 0-3	Missing evidence	Required fix

For gold-paper reconstruction tasks, add a confidence column: confirmed, likely, or TODO.

Anti-patterns

• Field snapshots without GT/pred/error or metric.
• Different color scales for compared fields.
• Caption states only what is shown, not what is learned.
• Turbulence claim with no spectrum/statistical diagnostic.
• Generalization claim with no unseen-regime panel.
• Caption uses effective, robust, complex dynamics, or physically consistent without metric/diagnostic/scope.

Verification

• Same colorbar/scale for comparisons.
• Units and nondimensional quantities are labeled.
• Every figure is referenced in text.
• Captions state takeaway, not just content.
• Pretty fields are paired with quantitative/physical diagnostics.
• Each major claim has at least one figure/table with rubric score 2 or 3.
• Captions can be scored at least 2 on the figure/caption axis of rubrics/gold-paper-closeness-rubric.md.