AI-CFD-Scientist

Per-case + cross-case TEXT/JSON analysis — collect all PROCEED-flagged cases, compute QoI tables, identify trends and correlations, produce analysis.json that downstream skills consume. Self-contained — embeds ResultsInterpreterAgent.system_prompt + user_prompt verbatim. Scope boundary — this skill produces text + tables only. Cross-case figure generation (Cf overlays, RMSE bars, Δ-fields, parametric trends) is owned by cfd-cross-analyze. Per-case PyVista renders are owned by cfd-viz mode=interpret. Paper figure planning + LaTeX writing is owned by cfd-paper.

Implement a custom OpenFOAM model (viscosity / turbulence / source term / fvOption) inside a case-local customModels directory. Self-contained — embeds the full openfoam_literature_change_agent_prompt_v2 protocol AND the compile_fix prompts verbatim. Builds payload, generates C++, compiles with wmake, smoke-tests. NEVER edits OpenFOAM installation directories.

Cross-experiment quantitative analysis. Author Python scripts (numpy + matplotlib + PyVista + pandas) that load every PROCEED case, compute aggregated metrics, and generate the comparison figures the paper needs (Cf overlays vs reference, RMSE bar charts, parametric trends, Δ-field contours). Then run a vision+text interpretation pass over the generated plots. This is the cross-case Python authoring layer cfd-paper consumes — without it, papers end up with only per-case contours.

Run CFD cases. Two modes — sweep (ARIS-style, agent-driven, no FoamAgent handoff) for parameter variants on a known-good baseline; single-case (FoamAgent-driven via scripts/foam_run.py) for fresh case generation that needs RAG. Embeds RunValidityAgent prompts (Allrun preflight + runtime investigation) verbatim. Writes <case_dir>/run_result.json.

Generate one OpenFOAM case from a natural-language requirement using a markdown-driven FoamAgent loop. Self-contained — embeds parser/planner/writer/reviewer prompts verbatim from Foam-Agent/src/services/. The agent runtime drives the loop natively via Read/Write/Bash. FAISS RAG retrieval uses scripts/rag_query.py; pdflatex, xvfb-run, inline Python, jq, and all other tools are also available — the only forbidden script across the entire system is scripts/orchestrator_run.py.

Convert literature + topic into a list of testable, FoamAgent-executable CFD hypotheses. Self-contained — embeds the HypothesisAgent system + user prompts verbatim from prompts/prompts.yaml. Outputs hypotheses.json. Use after cfd-literature in the general pipeline.

Per-case decision agent — PROCEED, REVISE, or RERUN. Reads run_result.json + diagnostic figures, judges physics plausibility through a vision-LLM pass, returns decision.json. Self-contained — embeds ResultsInterpreterAgent.interpretation_* and vision_* prompts verbatim.

Retrieve CFD literature for a topic via Semantic Scholar (with optional OpenAlex / arXiv / web supplements) and write a normalized lit.json. Use when the user wants a literature survey for a CFD study, or when cfd-pipeline / cfd-hypothesis / cfd-paper needs lit.json upstream. Self-contained — no Python helper required, but a one-shot script fast-path is documented at the bottom.

Mandatory mesh-independence study before any experiments. Per physics group, run baseline + refined meshes (10% near-wall, 5% away-from-wall), compare QoIs against the 5% threshold, escalate to GCI if needed, lock the selected mesh. Self-contained protocol embedded in this file (merged from skills/cfd-mesh-independence). Outputs selected_mesh_spec.json and mesh_independence_context.json.

Open-ended discovery loop — propose novel CFD model modifications (e.g. SA / k-omega variants, viscosity laws), implement via cfd-code-modify, run via cfd-experiment, score against baseline + reference, repeat until budget. Self-contained — embeds MetricProposer, ComparatorAuthor (text + PyVista), ComparatorVerifier, and the LLM-judge prompt verbatim. Use for topics with "open-ended", "discover", "find a novel", "beat baseline", "best model".

Full paper writing in skill mode. ONE skill that walks the agent through a strict numbered protocol: Stage 0 experiment planner, then a planner-viz-writer-compile-reviewer-revise cycle that iterates until the reviewer returns pass=true or the cap is hit. Each phase is one concrete agent action with the LLM prompt embedded verbatim and an immediate schema-validation gate. No script delegation. The agent itself is the planner, the viz author, the vision QA, the writer, the compile-fixer, the reviewer, and the reviser — same as every other skill-based AI scientist.

General-purpose academic paper writer skill for CFD research, adapted from Master-cai/Research-Paper-Writing-Skills with CFD-domain overlays. Embeds section-by-section structural guidance (Abstract / Introduction / Related Work / Methods / Experiments / Discussion / Conclusion), Master-cai's reviewer rubric, and a claim-evidence-citation discipline. Used by cfd-paper Phase D (writer / reviser) and Phase F (reviewer). Replaces the previous ad-hoc HARD-rule list with a structural writer skill — the audit then enforces only objectively-checkable properties.

Top-level CFD Scientist research pipeline. Chains the cfd-* sub-skills end-to-end — literature → hypothesis → mesh-gate → experiments → analysis → paper. Routes to code-mod / open-ended / general subflows based on intent. Self-contained — embeds MetricSetupAgent + MetricSetupVerifier prompts verbatim for the metric-setup stage.

Convert hypotheses (or a direct simulation request) into FoamAgent-ready user-requirement strings, one per case. Self-contained — embeds the IdeationAgent + HypothesisAgent expert prompts verbatim. Use after cfd-hypothesis (general flow) or directly when the user gave a simulation-only request. Outputs requirements.json.

Numerical analysis patterns for OpenFOAM run outputs — log parsing, D²-law fits, slope/RMSE/Richardson extrapolation, residual-plateau checks, drift checks. Self-contained — the agent writes 30-line Python snippets via Bash using these patterns. No Python wrapper.

Generate PyVista figures from a CFD case. Two modes — interpret (per-case diagnostics for cfd-interpret) and full (publication-quality cross-case figures for cfd-paper). Self-contained — embeds viz_quality and vision QA prompts verbatim. Agent writes and runs the PyVista script directly; the batch helper is optional.

LangGraph-mode runtime contract for scripts/foam_run.py — preserves Foam-Agent RAG and prompt behavior. For skill-mode case generation, use cfd-skills/cfd-foamagent/SKILL.md instead (fully self-sufficient, agent-driven, no foam_run.py wrapper).

Top-level CFD scientist orchestrator. Reads a user CFD topic, picks one route, initializes a run directory, and invokes the first skill of a fixed chain. From that point on, every per-stage skill's `## Next` block names the next skill, and the agent walks the chain to a finished paper. No user shell commands between stages.

Custom OpenFOAM model implementation entry — alias for cfd-skills/cfd-code-modify. Use when the user wants to implement, modify, or add a custom viscosity / turbulence / source term / fvOption inside a case-local customModels directory. The full self-contained recipe (with the OPENFOAM 10 LITERATURE CHANGE AGENT v2 protocol embedded verbatim) lives in cfd-skills/cfd-code-modify/SKILL.md.

Mesh independence study using rapid mesh-sensitivity assessment with Richardson extrapolation and GCI. Follows the mesh sensitivity protocol exactly.

Top-level CFD research entry — alias for cfd-skills/cfd-pipeline (general path). Use when the user gives a CFD research topic and wants the full literature → hypothesis → mesh-gate → experiments → analysis → paper flow. The full self-contained recipe lives in cfd-skills/cfd-pipeline/SKILL.md.