csml-rpi/ai-cfd-scientist/cfd-viz

cfd-viz

PyVista-based figure generation. The agent authors the script, runs it off-screen, and validates the output with a vision-LLM QA pass before declaring done.

Modes

mode=interpret — per-case diagnostics

Quick figures for one case sufficient for cfd-interpret to judge physics plausibility:

• Mesh overview (full domain + zoomed near-wall)
• Velocity-magnitude contour at the latest time
• Pressure contour at the latest time
• Streamlines on a midplane
• y+ map on no-slip walls
• Residual history (matplotlib for this one)

mode=full — publication figures

Driven by a paper plan (<out-dir>/paper_unified_plan.json):

• Comparison panels across cases (same field, same colormap, same clim)
• Profile plots at probe locations (overlay across cases)
• Streamlines / vortex iso-surfaces with consistent styling
• Error-vs-reference plots (when reference data is available)

Inputs

• out-dir (required)
• mode (required) — interpret or full
• For mode=interpret:
• case_dir (required) — single case
• For mode=full:
• paper_plan (required) — path to paper_unified_plan.json (figure jobs list)
• case_dirs (required) — list of case dirs to draw from

Outputs

• mode=interpret: <case_dir>/figs/*.png (5–8 PNGs)
• mode=full: <out-dir>/paper_figs/*.png plus <out-dir>/paper_figs/cases_config.json (manifest of which figure used which case + field + time)

Recipe (primary, agent-driven)

Step 1 — Inspect available data

For each case_dir:

1. Verify <case_dir>/case.foam exists; create an empty file if missing (touch case.foam) — PyVista's OpenFOAMReader needs this entry-point file.
2. List time directories (<case_dir>/<float>/); pick the latest non-zero (the gate is: refuse to plot at t=0 unless this is a steady-state case where t=0 is the only available time, in which case warn).
3. List internal fields (U, p, etc.) and patch fields (<time>/<field> boundary file, plus boundary patch list from constant/polyMesh/boundary — only for patch identification, never to read mesh geometry; PyVista handles geometry via OpenFOAMReader).
4. For mode=full, read paper_plan to get the list of figure jobs.

Step 2 — Author PyVista script(s)

Embed viz-quality awareness using the prompts below. The agent (you) writes a self-contained Python script under <out-dir>/_viz_scripts/ and runs it via python <script>.py.

System prompt — figure-quality contract (from prompts/prompts.yaml: ResultsInterpreterAgent.viz_quality_system_prompt)

You are a CFD visualization reviewer. You receive the user requirement and images generated from an OpenFOAM simulation (mesh, field contours, different angles). Decide if these visualizations are sufficient to judge whether the run succeeded and met the user requirement. These images may be embedded in a manuscript, so treat them as publication candidates.
Camera and zoom must be appropriate for paper-quality figures. For localized, feature-focused requests (recirculation bubble, reattachment point/length, shear layer, step lip/corner, jets, wall quantities like Cp/Cf/y+), require zoomed-in framing so the feature is large and inspectable when printed. It is acceptable if the full computational domain is not visible in every view, as long as the requested feature is readable. Treat extremely zoomed-out views where features are too small/unreadable as unacceptable.
Layout and typography: reject if a colorbar, scalar bar, or legend overlaps or obscures contours, streamlines, mesh, or plotted data (a common PyVista failure). Reject if titles, axis labels, tick labels, legend text, or colorbar labels are too small to read at typical manuscript figure size. If a vertical colorbar crowds the plot, the generator should prefer a horizontal bar or reposition margins—flag that in your reason when rejecting.
Geometry/aspect: reject primary 2D full-domain plots where the domain looks like a thin line or sliver (one in-plane dimension visually collapsed), so channel/duct geometry cannot be recognized—common when Lx/Ly is large and the render window is wrong.
If images are missing, blurry, empty, framed/zoomed so poorly that features cannot be inspected, do not show relevant variables/times, or fail layout/legibility checks above, set viz_acceptable to false.
Return ONLY valid JSON: {"viz_acceptable": true or false, "reason": "short string"}.

User prompt — QA per-figure (from prompts/prompts.yaml: ResultsInterpreterAgent.viz_quality_user_prompt)

User requirement:
{user_requirement}
 
The following images are from the simulation (mesh and field visualizations). Are they acceptable to judge the run—and suitable as paper figures (no colorbar/legend overlapping data; fonts large enough)? Return JSON only: {"viz_acceptable": bool, "reason": "string"}.

Step 3 — Generate per-mode

mode=interpret — script template

The agent writes a script equivalent to (concrete fields/clims chosen based on Step 1 inspection):

import pyvista as pv
import numpy as np
from pathlib import Path
case = Path("<case_dir>")
figs = case / "figs"
figs.mkdir(exist_ok=True)
pv.OFF_SCREEN = True
reader = pv.OpenFOAMReader(str(case / "case.foam"))
reader.set_active_time_value(reader.time_values[-1])  # latest non-zero
data = reader.read()
internal = data["internalMesh"]
# 1. Mesh overview
p = pv.Plotter(off_screen=True, window_size=(1600, 900))
p.add_mesh(internal, style="wireframe", color="black", line_width=0.5)
p.view_xy()
p.add_text("Mesh", font_size=14)
p.screenshot(str(figs / "01_mesh_overview.png"))
# 2. U magnitude contour
internal["Umag"] = np.linalg.norm(internal["U"], axis=1)
p = pv.Plotter(off_screen=True, window_size=(1600, 900))
p.add_mesh(internal, scalars="Umag", cmap="viridis",
           scalar_bar_args={"title": "|U| [m/s]", "position_x": 0.25,
                            "position_y": 0.05, "width": 0.5, "height": 0.04,
                            "vertical": False, "title_font_size": 16,
                            "label_font_size": 14})
p.view_xy()
p.screenshot(str(figs / "02_U_contour.png"))
# 3. Pressure contour
p = pv.Plotter(off_screen=True, window_size=(1600, 900))
p.add_mesh(internal, scalars="p", cmap="RdBu_r",
           scalar_bar_args={"title": "p [Pa]", "position_x": 0.25,
                            "position_y": 0.05, "width": 0.5, "height": 0.04,
                            "vertical": False})
p.view_xy()
p.screenshot(str(figs / "03_p_contour.png"))
# 4. Streamlines (midplane)
streamlines = internal.streamlines("U", n_points=50, source_radius=0.1,
                                    integration_direction="both", max_steps=2000)
p = pv.Plotter(off_screen=True, window_size=(1600, 900))
p.add_mesh(internal, scalars="Umag", cmap="viridis", opacity=0.4)
p.add_mesh(streamlines.tube(radius=0.005), color="black")
p.view_xy()
p.screenshot(str(figs / "04_streamlines.png"))
# 5. y+ map (if wallShearStress + boundary patches available)
# (omitted for brevity; agent fills in when these are present)
# 6. Residual history — matplotlib, parsed from log.<solver>
# (use re to extract lines like "smoothSolver: ... Initial residual = X")

Key styling rules (the embedded viz_quality_system_prompt enforces these):

• Always horizontal colorbar at position_y ≈ 0.05, height ≈ 0.04, width ≈ 0.5 — never vertical, never default position.
• title_font_size >= 16, label_font_size >= 14, font_size >= 14 for plotter text — readable at print size.
• window_size=(1600, 900) for thin/wide channel-like geometries; (1600, 1200) for tall/square; never fall back to 800×600 default.
• For zoomed-feature plots (BFS step, jet exit, etc.), call p.camera.zoom(1.5–2.5) after setting view.

mode=full — paper-figure script

For each entry in paper_plan.figure_jobs:

• Same styling rules.
• Identical clim across cases compared in the same panel — compute clim from the union of all case data first, then apply to every panel.
• Same colormap across the whole manuscript — pick once (e.g. viridis for U, RdBu_r for signed quantities, cividis for monotone scalars).
• Use pv.Plotter(off_screen=True, shape=(1, n_cases)) for side-by-side panels.

Write all figures to <out-dir>/paper_figs/. Write a manifest:

{
  "figures": [
    {"path": "fig_01_U_contour_panel.png", "cases": ["case_001", "case_003"], "field": "U", "time": 5000.0, "spec_id": "U_panel"}
  ]
}

Step 4 — Run the script(s)

python <out-dir>/_viz_scripts/<script>.py

If the script crashes (PyVista off-screen issues, missing fields), inspect the error, regenerate the script with the fix (often: missing OFF_SCREEN, wrong field name capitalization, a field that doesn't exist at the chosen time), and retry up to 3 times.

Step 5 — Vision QA pass

For each generated PNG, send it to a vision-LLM with the embedded viz_quality_* prompts above. Render {user_requirement} as the case requirement (or, in mode=full, the figure's spec from the paper plan).

If the response is {"viz_acceptable": false, "reason": "..."}, parse the reason:

• "colorbar overlaps" → adjust scalar_bar_args.position_x/position_y to clear data, regenerate, re-QA.
• "fonts too small" → bump font sizes by 4 pt and regenerate.
• "domain looks like a sliver" → adjust window_size aspect ratio or camera.parallel_scale.
• "feature not inspectable / too zoomed-out" → add camera.zoom(2.0) or use set_focus() near the feature.
• "wrong field shown" → fix the field/time selection in the script.

Max 10 inner attempts per figure. After that, keep the latest version and emit a warning in the timeline.

Step 6 — Write outputs and timeline

mode=interpret: figures land in <case_dir>/figs/. No manifest needed — cfd-interpret reads the directory.

mode=full: figures + manifest in <out-dir>/paper_figs/. Append timeline event per figure plus an aggregate:

{"stage": "viz", "event": "complete", "ts": "<iso>", "mode": "full", "n_figures": 12, "n_qa_retries": 3}

Skip logic

• interpret: regenerate if any source field file is newer than existing figs, or if figs/ has fewer than the expected number of PNGs.
• full: skip a figure if its PNG exists AND the plan entry hasn't changed (compare paper_plan mtime vs PNG mtime); regenerate if plan is newer.

Anti-hallucination rules

• Never plot a field that doesn't exist in the time directory. If <time>/wallShearStress is missing for a y+-style figure, write the script to detect that and emit a stub that plots only the mesh near the wall.
• Never invent reference data. If a comparison-vs-DNS plot is requested but no reference CSV exists in reference_data/, skip the comparison series (don't draw a fake DNS line) and add a caption note.
• Never report a figure as "accepted" without an actual QA response from the vision LLM.

LangGraph mode only — do NOT call from skill mode

Two valid execution paths

This skill has two equally valid paths — pick whichever fits the situation. Both produce the same artifacts.

1. Inline agent recipe (above). The agent authors the PyVista script, runs it via Bash, validates each PNG with the embedded vision-QA prompt, and retries on rejection. Use this when the visualization is per-case and easy to spec.

1. Script fast-path. scripts/viz.py (interpret mode) and src/cfd_langgraph/paper_unified/batch_paper_viz.py (paper-batch mode) implement the same loop in Python. They use the same prompts as the inline recipe and write the same outputs. Invoke them via Bash with the relevant CLI args. The scripts/orchestrator_run.py LangGraph end-to-end driver is the only Python file skill-mode should NOT call directly — it bypasses skill mode entirely. Everything else under scripts/ is fair game.