| name | scientific-journal-writing |
| description | Use when writing, revising, outlining, or critiquing CFD-AI and scientific ML journal papers; focuses on argument structure, claim-evidence alignment, novelty, and reviewer defense. |
| version | 0.4.0 |
| author | CFD-AI Paper Skills maintainers |
| metadata | {"short-description":"Structure and sharpen CFD-AI/SciML journal manuscripts","gold-standard-authors":["Kai Fukami","Steven L. Brunton","Romit Maulik","Sangseung Lee","Ricardo Vinuesa"]} |
Scientific Journal Writing for CFD-AI / SciML
Trigger
Use when the user asks for paper outline, abstract, introduction, methods, results, discussion, conclusion, novelty framing, contribution bullets, or manuscript critique for CFD-AI/SciML work.
Core rule
Do not merely improve English. First make the manuscript defensible.
A paper is defensible only if it has:
- a physics/engineering problem,
- a specific gap,
- a calibrated contribution,
- evidence for every major claim,
- limitations before reviewers weaponize them.
Gold-standard writing target
Model the discipline of papers by Fukami, Brunton, Maulik, Lee, and Vinuesa:
- physics-first setup,
- ML role clearly scoped,
- reproducible numerical setup,
- fair baselines,
- physical diagnostics,
- generalization tests,
- figure sequence that proves claims,
- cautious wording.
The target is not “more academic English.” It is normal CFD-AI/SciML paper texture: physical problem first, narrow gap second, method role third, evidence sequence fourth, limitation boundary last. If the prose could fit any AI-for-science paper after replacing two nouns, it is not close enough to the gold-paper target.
Use progressive disclosure:
| Task | Read first |
|---|
| Super-resolution/reconstruction abstract, intro, or results | references/gold-papers/fukami-2019-super-resolution-jfm.md |
| Spatio-temporal super-resolution | references/gold-papers/fukami-2021-spatiotemporal-super-resolution-jfm.md |
| Field positioning or related-work taxonomy | references/gold-papers/brunton-2020-machine-learning-fluid-mechanics.md and references/gold-papers/vinuesa-2022-enhancing-cfd-ml.md |
| ML workflow staging / physics insertion points | references/gold-papers/brunton-2021-applying-ml-fluid-mechanics.md |
| Uncertainty/trustworthy surrogate claims | references/gold-papers/maulik-2020-probabilistic-neural-networks-prf.md |
| Cylinder-wake prediction/generalization | references/gold-papers/lee-2019-cylinder-wake-jfm.md |
| Experimental-fluid-mechanics opportunity framing | references/gold-papers/vinuesa-2023-transformative-ml-experiments-nrp.md |
| Claim strength decisions | rubrics/claim-evidence-rubric.md |
| AI-ish / field-inappropriate vocabulary | references/field-terminology-style-guide.md, rubrics/vocabulary-style-rubric.md, examples/ai-ish-to-field-native-prose.md |
| Gold-paper closeness / normal paper texture | references/gold-paper-style-patterns.md, rubrics/gold-paper-closeness-rubric.md, examples/generic-ai-to-gold-paper-prose.md |
Workflow
1. Classify manuscript state
Classify as:
- idea only
- outline
- partial draft
- full draft
- revision
- response letter
- LaTeX production
If missing, infer. Ask only when ambiguity blocks work.
2. Extract one-sentence claim
Use:
We propose X for Y under Z physical/numerical conditions, using W, and demonstrate A/B/C against baselines D/E.
If this cannot be written, stop polishing and diagnose the missing piece.
3. Contribution taxonomy
Classify each contribution:
- method/architecture
- physical modeling
- numerical algorithm
- surrogate/ROM/closure
- dataset/benchmark
- empirical physical insight
- uncertainty/verifiability
- software/workflow
Reject vague contribution bullets such as “we propose a novel framework.” Require mechanism + evidence.
4. Claim-evidence map
For every major claim, build:
| Claim | Evidence needed | Current evidence | Risk |
|---|
| accurate | fair metrics + baselines | ? | high if no baseline |
| physically consistent | conservation/residual/BC/spectrum | ? | high |
| generalizes | unseen Re/geometry/BC/time | ? | high |
| efficient | wall-clock/memory/params | ? | medium |
| robust | noise/sparsity/OOD tests | ? | high |
Unsupported claims become TODOs, not prose.
5. Section logic
Abstract
Five moves:
- physical/engineering context,
- unresolved limitation,
- proposed method,
- concrete evidence/result,
- implication/limited scope.
No “promising,” “robust,” “SOTA,” or “physically consistent” without measurable evidence.
Gold-paper closeness gate: the first sentence must name a physical/numerical object or bottleneck, not AI capability. Accept “high-resolution vorticity reconstruction from low-resolution snapshots”; reject “AI has revolutionized CFD.”
Introduction
Flow:
- Why the fluid/process problem matters.
- Why conventional CFD/experiments/current ML are limited.
- Narrow gap.
- Proposed approach.
- Contribution bullets with evidence hooks.
Related work
Use taxonomy, not chronology:
- fluid-mechanics workflow role: discovery, modeling, sensing/reconstruction, experiments, closure, ROM, control, optimization, acceleration,
- ML insertion point: problem, data, architecture, loss, optimization, deployment/control,
- then method family: PINN, operator, CNN, GNN, ROM, sparse identification, RL, generative, or domain-specific model.
End with precise contrast.
Methods
Must include governing equations/assumptions, solver/data generation, BC/IC, nondimensional groups where relevant, architecture, loss, training, implementation details.
Results
Recommended sequence:
- validation/sanity case,
- main quantitative comparison,
- field visualization GT/pred/error,
- physics diagnostic,
- ablation,
- generalization,
- efficiency,
- failure/limitations.
Discussion
Explain why results matter physically. Do not restate tables.
Output template
Return:
- One-sentence paper claim
- Contribution bullets
- Claim-evidence map
- Section-level structure
- Reviewer attack points
- Required evidence/TODOs
- Gold-paper closeness score, if writing/reconstruction is requested
- Exact rewrite, if requested
For abstracts, use examples/bad-to-good-abstract.md as the style and evidence standard. For prose that sounds generic or AI-written, use references/gold-paper-style-patterns.md, rubrics/gold-paper-closeness-rubric.md, and examples/generic-ai-to-gold-paper-prose.md. When the user wants reviewer criticism to drive the rewrite, route through skills/paper-revision-loop-manager/SKILL.md and use templates/reviewer-editor-loop-report.md. For experiments, route to templates/experiment-plan.md and rubrics/sciml-experiment-rubric.md.
v0.4 eval-backed checks
These checks were added after simulated baseline-agent failures in evaluation/runs/2026-06-30-v04-*.
| Trigger | Extra requirement |
|---|
| Spatio-temporal super-resolution | Separate spatial downsampling, temporal frame interval, high-resolution reference, and time-axis diagnostics. |
| Brunton-style taxonomy | Classify by workflow role before model family; gap paragraph must name CFD bottleneck, closest prior role, unresolved boundary, mechanism, and decisive evidence. |
| Cylinder/wake prediction | State Re/regime, split type, rollout horizon, and physical-consistency evidence before polishing prose. |
Anti-patterns
- “novel” repeated as a substitute for novelty.
- “physics-informed” without saying where physics enters.
- “generalizable” from interpolation-only tests.
- “real-time” without hardware/timing.
- “state-of-the-art” without a fair baseline matrix.
- Generic opening context that could belong to any AI-for-science paper.
- Method-first paragraphs that name the network before the flow case, variable, and reference data.
Verification
Before finalizing:
- Every claim is supported or marked TODO.
- Unknown numerical details are not invented.
- Reviewer objections are explicit.
- Suggested text preserves scientific tone and does not overclaim.
- The final answer can be scored by
rubrics/claim-evidence-rubric.md.
- If manuscript prose is generated, the final answer can also be scored by
rubrics/gold-paper-closeness-rubric.md.
