csml-rpi/ai-cfd-scientist/cfd-paper-writer

cfd-paper-writer

Adapted from Master-cai/Research-Paper-Writing-Skills (MIT) and Prof. Peng Sida's writing notes. Augmented with CFD-domain requirements (governing equations, turbulence closures, per-case contours, mesh independence, comparator definitions).

This skill is invoked from inside `cfd-paper` — by Phase D (writer / reviser) and by Phase F (reviewer). It never runs standalone. It receives a fully-built context bundle (analysis, manifest, literature, plan, figure paths, prior reviewer feedback) and emits either:

• a complete LaTeX manuscript, or
• a structured reviewer verdict.

Core philosophy (apply to every section)

1. One paragraph for one message. The first sentence states what the paragraph will do. Every subsequent sentence advances that one message (cause, contrast, consequence, refinement) — nothing else.
2. Claim → evidence → citation. Every non-trivial claim is followed by a number from a table or figure (evidence) and a \cite{} to the literature that supports the claim or that you are distinguishing from. No claim sits alone.
3. Visual quality is content, not decoration. Figures, captions, tables, and equation typography are part of the substantive contribution. Treat them as core writing — not afterthoughts.
4. Reverse-outline as you write. After each section, you should be able to write a 3–7 bullet outline of what it actually says. If you can't, the section is a token-soup draft.
5. Adversarial self-review before handing off. Read your own text as a hostile reviewer would. Flag every unsupported claim, every figure that doesn't carry its caption's promise, every term that drifts in meaning between sentences.

These rules eliminate the failure patterns that plagued earlier CFD paper drafts: n-gram word loops within paragraphs, "in introduction paragraph 1" meta-cross-references, citation clusters dumped at section ends, equations that are merely Re/Cf/RMSE definitions, missing geometry descriptions, and orphan figures.

Section-by-section writing guides

Each guide gives you: (a) pre-writing questions, (b) one or more concrete paragraph templates with sentence skeletons, (c) a quality checklist.

You pick the template that best fits the work and customize the sentence skeletons to your topic. Templates are guides, not prison cells — adapt them, but do not abandon their structural discipline.

1. Abstract

Pre-writing questions

1. What technical / physical problem do we solve, and why is there no well-established CFD solution?
2. What is our technical contribution? (a sensitivity study, a model modification, an open-ended discovery, a quantitative correlation, …)
3. Why does our method work in essence?
4. What new insight does the experiment produce? (e.g., "k-ω SST under-predicts reattachment by 20% on adverse-pressure-gradient flows", "a wall-distance-gated SA destruction reduction does not survive aggressive grids", …)

Templates

Version A (Challenge → Contribution): for studies with a single dominant contribution.

1. One sentence: the CFD task and physical setting.
2. One to two sentences: the technical challenge for previous closures / approaches.
3. One to two sentences: our technical contribution that addresses the challenge.
4. One sentence: the benefit (improvement vs baseline / reference, in concrete numbers).
5. One sentence: experiment summary (sweep / cases / metric outcomes).

Version B (Challenge → Insight → Contribution): for model-modification or OED papers where the insight is more interesting than the implementation.

1. Task + setting.
2. Technical challenge for previous methods.
3. One sentence: the insight (e.g., "near-wall production over-prediction on the recovery shoulder is the dominant error source for k-ω SST on periodic hill at Re=5600").
4. One to two sentences: the technical contribution implementing the insight.
5. One sentence: benefits + experiment summary.

Version C (Multi-contribution): for sweeps / sensitivities where each row of the design matrix is a contribution.

1. Task + setting.
2. If needed, one contrast sentence about prior approaches.
3. Contribution 1 + advantage.
4. Contribution 2 + advantage.
5. Contribution 3 + advantage.
6. Experiment summary.

Hard rules

• ≤ 250 words. No case IDs (case_001). No \cite{} calls. No equations beyond a single inline Reynolds-number formula if essential. No reference to section numbers or internal labels.
• Every claim in the Abstract must be re-derivable from a number in the Results table.

CFD specifics

• For sensitivity sweeps: state the parameter range, the QoI(s), and the strongest quantitative outcome (e.g., "the reattachment x_r/h shifted by 40% between k-ε and SA").
• For model-modification: state which closure was modified and the headline metric delta vs baseline.
• For OED: state how many families were tried, whether a winner was found, and either the winner's improvement or the no-winner family-coverage statement.

Quality checklist

1. Can a reader identify task, challenge, insight/contribution, and headline result in one pass?
2. Are all major claims supported by experimental numbers?
3. Are technical names self-contained (no jargon that requires reading another paper)?
4. Is there any sentence that mixes two or more messages?

2. Introduction

Logic map (think this through before writing)

What CFD task / flow problem
       ↓
What QoI(s) measure success
       ↓
SOTA methods fail to meet QoI target
       ↓
Root technical issue behind that failure
       ↓
Our technical solution + study design
       ↓
Why the solution / study works
       ↓
Additional contributions and impact

Backward reasoning (answer before drafting)

1. What CFD problem do we solve, and why is there no well-established solution? (most important)
2. What are the contributions of our study? (a new flow regime characterized? a new closure-modification family? a new correlation? a falsification of a published claim?)
3. What are the benefits of our contributions, why can they answer this challenge, and what new insight do they bring?
4. How do we use prior CFD work to lead readers to our challenge?

Forward story (write in this order)

1. Introduce the CFD task / flow problem and applications (heat transfer, drag prediction, mixing efficiency, separation control, …).
2. Use prior CFD work to lead to the technical challenge we solve.
3. Present our contribution(s).
4. Explain the technical advantage and the new insight explicitly.

Section skeleton

\section{Introduction}
% Task and application
% Technical challenge for previous methods (limitation + technical reason)
% Our pipeline / study design
% Experiment summary
% Contributions

Opening templates

• Version 1 (task is niche → define task, then applications).
• Version 2 (task is familiar → open with applications).
• Version 3 (general → specific: state general task, then narrow to this paper's setting).
• Version 4 (familiar task → open with applications, then expose the unresolved failure in the very first paragraph via a prior method).

Technical-challenge templates

• Version 1 (existing task, existing methods): general challenge → traditional methods + limitation → recent methods + limitation → exact challenge we solve.
• Version 2 (existing task with insight historically grounded): mainstream limitation → classical line with related insight → why that classical line is still insufficient → modern methods unresolved → bridge to our method.
• Version 3 (novel task / OED): state the goal, decompose into N independent challenge points using First/Second/Finally.

Pipeline / study-design templates

• Version 1 (one contribution, one teaser figure): framework + key novelty + teaser pointer + implementation + multi-advantage list.
• Version 2 (two contributions, one teaser figure): framework + contribution 1 + advantage 1 + remaining challenge + contribution 2.
• Version 3 (new module on existing pipeline): prior pipeline setup + new module + observation that motivates it + mechanism + comparison to alternatives.
• Version 4 (observation-driven): innovation first + intuitive observation as motivation + implementation + technical advantage + empirical gain.

Hard rules

• ≥ 4 paragraphs, each with a single message.
• ≥ 3 \cite{} calls in the Introduction (claim → reference, not citation cluster at end).
• The first paragraph must already gesture at what problem the paper solves — even if the technical-challenge paragraph comes later.
• No hiding concrete contribution in abstract "novelty" prose. Be concrete.

CFD specifics

• Cite prior CFD studies on the same geometry / closure / flow regime — not just generic turbulence references.
• If using DNS/experimental reference data, cite the canonical source (Driver-Seegmiller for BFS, Frohlich et al. for periodic hill, etc.).
• For sensitivity studies, state the range of Re / parameter explored.
• For OED / code-modification, name the closure being modified in the second paragraph at the latest.

Quality checklist

1. Does the first sentence of each paragraph state its message?
2. Does each paragraph carry one message only?
3. Are the technical challenge, the technical reason, and the solution mechanism all explicit?
4. Are claims in the Introduction aligned with the Results table?
5. Is terminology stable across all sections (e.g., "skin friction coefficient" not switching between "Cf", "C_f", "wall shear coefficient")?

3. Related Work

Goal

Position your work against the most relevant lines of CFD research. Make your novelty / specific contribution easy for a reviewer to verify.

Workflow

1. List directly competing and recent papers first (last 3–5 years preferred).
2. Group literature by technical topic, not by publication year.
3. For each topic: summarize the common paradigm in 1–2 sentences, then the key limitation tied to your study's QoI.
4. End each topic with a transition sentence that distinguishes your work.

Suggested topic groupings for CFD

For a turbulence-model sensitivity study: (a) RANS closures for separated flows; (b) DNS / LES reference data; (c) prior sensitivity studies on similar geometries.

For a viscosity / rheology study: (a) constitutive law variants; (b) prior solver implementations; (c) experimental validation studies.

For an OED / model-modification study: (a) original closure and canonical extensions; (b) prior model-modification attempts and their failure modes; (c) discovery-loop methodology (if relevant).

For a sweep / correlation study: (a) prior empirical correlations on this geometry; (b) prior simulation studies bracketing the parameter range; (c) the experimental literature that grounds the QoI.

Paragraph template (per topic)

1. Topic sentence: define the scope of this topic in one sentence.
2. Representative methods: one compact summary of the dominant paradigm, with citations.
3. Limitation: tied to your target technical challenge / QoI / parameter range.
4. Transition sentence: leads to your method / study.

Hard rules

• 2–4 focused topics. Not more (becomes a citation dump). Not fewer (insufficient coverage).
• ≥ 1 \cite{} call per paragraph, woven into the sentence — not appended at the end.
• Compare mechanisms, assumptions, and failure modes — not just method names.
• Do not hide the strongest baseline. If a recent paper directly competes with yours, cite it in the first or second topic.
• Never write meta-references ("in this section we…", "as paragraph 3 noted…", "in introduction paragraph 1…"). Every sentence stands on its own.

Quality checklist