From a problem description to checked CAD, fit-for-purpose meshes and one report

pinneapple_design.fitmesh turns a problem written in words (English or Portuguese) into reference geometries, checked CAD variants, meshes sized for what the result will be used for, and a single report that ties every decision to its evidence. The goal is the mesh and CAD quality the problem needs: not so fine that it wastes days, not so coarse that nothing in the result can be trusted.

from pinneapple_design.fitmesh import FitMeshSession

s = FitMeshSession("Erosion and pressure drop in a 3 inch pneumatic pipe bend, air at 20 m/s", "study/")
s.step()                         # the open questions and the geometry options, with references
s.answer(accuracy="design")      # screening / design / verification
s.choose_geometry("pipe_bend_90")
s.choose_parameters(n=4)         # the suggested parameters and ranges, or vary={"bend_ratio": [1, 3]}
s.generate_cad()                 # STEP + STL per variant, each with its checks
s.plan_meshes()                  # mesh plan per variant, with the rules and sources used
s.generate_meshes(procs=4)       # meshes + quality (and the coarser family for design / verification)
s.add_results("pressure drop", [f_fine, f_medium, f_coarse])   # optional: grid-convergence index
s.report()                       # study/report.html and study/report.json

FitMeshSession.run(text, out_dir, answers=..., geometry=..., vary=...) does it all in one call, and pinneapple_design.fitmesh.api.create_app() exposes the same steps over HTTP (one call per step, files and report downloadable).

The steps

Step	What happens	What is checked
Problem	physics, fluid, speed, size, Reynolds number, purpose, quantities, material and load are read from the text; anything missing becomes a question; every fact keeps its source (text, answer, default, model)	unanswered facts are listed in the report
Geometry	templates ranked by physics and keywords: NACA 4-digit wing, Ahmed body, cylinder, sphere, 90° pipe bend, backward-facing step, plate with a hole, cantilever beam, plate-fin heat sink	each template cites the published references it can be compared with
Parameters	high-influence parameters proposed for variation over their usual range (sizes within a factor 2 of the base), the rest kept; dependent sizes follow (development lengths in diameters); Latin hypercube, reference design first	validity limits of every parameter
CAD	STEP and STL per variant	valid solid (OpenCascade), volume against the exact formula, watertight STL, STL/solid volume within the tessellation tolerance for the purpose, dimensions against each reference object (same / inside the studied range / extrapolated)
Mesh plan	method, sizes, prism layers, domain, cell and cost estimate, and the mesh family	every rule is written into the rationale with its source
Meshes	Gmsh (solids: quadratic tetrahedra, .msh and CalculiX .inp; 2-D sections with a boundary layer), snappyHexMesh (3-D external with far field and ground, internal with inlet and outlet from the CAD faces)	pinneapple_data.cae quality (non-orthogonality, skewness, aspect ratio, scaled Jacobian, disconnected regions), checkMesh, prism-layer coverage
Report	one JSON and one HTML page, a verdict per stage	grid-convergence index when the results of the three meshes are given

The rules behind the mesh plan

Purpose	Surface cells	Wall treatment	Prism growth	Meshes
screening (~10 %)	L/40, 4 across the smallest feature	wall functions, y+ ≈ 50 (30-300)	1.3	one
design (a few %)	L/80, 8 across the smallest feature	y+ ≈ 1 where separation or wall quantities matter (k-ω SST without wall functions)	1.2	the mesh and one 1.5× coarser
verification	L/120, 12 across the smallest feature	y+ ≈ 1	1.15	three, ratio 1.5, for the GCI

• First-cell height from the skin friction at the reference length: Cf = 0.026 Re^-1/7 (White) turbulent, 0.664 Re^-1/2 laminar, and the pipe friction factor 0.316 Re^-1/4 for internal flows. y+ between 5 and 30 (the buffer layer) is never used.
• External domains: 5 L upstream, 10-20 L downstream, 5 body sizes to the sides; the blockage is reported. Extruded sections (wing, cylinder, step) are meshed as 2-D unless a 3-D wake is asked for.
• Solids: quadratic tetrahedra; 2-4 elements through the thinnest wall; 16-48 elements around holes, where the stress peaks; the closed-form check (Kt from Peterson's, beam formulas) is printed next to the plan.
• Grid convergence: observed order, Richardson extrapolation and GCI with a safety factor of 1.25 for three grids (Roache; Celik et al., J. Fluids Eng. 130:078001, 2008). Oscillatory convergence is reported as a range, not forced through the formula, and the asymptotic range is checked.

Language models, local or hosted

from pinneapple_design.fitmesh import FitMeshSession, LLMAssist
llm = LLMAssist(provider="ollama", model="qwen2.5:7b")          # or "anthropic" / "openai"
s = FitMeshSession("a bracket that holds a 2 kg motor on a wall", "study/", llm=llm)

The model can read the problem, rank the geometries, propose parameters and ranges, write a CadQuery part when no template fits, and review the mesh plan. It proposes; PINNeAPPle checks:

• unknown template or parameter names are dropped and ranges are clipped to the validity limits (the report says so);
• model-written CadQuery is parsed first (imports limited to cadquery, math and numpy; no file, system or dunder access), then runs in a separate isolated Python process with a timeout, and the part goes through the same CAD checks, marked "no published reference";
• the mesh review may change only the y+ target, the growth ratio and the layer count, inside the rule limits; refused changes are written into the rationale.

Calls can be logged with pinneapple_llm.ConversationStore. Without a model everything runs on the rules.

How this compares with the tools in use today

Commercial meshers (Fluent watertight and fault-tolerant workflows, STAR-CCM+ automated meshing, Fidelity Pointwise, ANSA batch meshing) and open ones (Gmsh, snappyHexMesh, cfMesh, Netgen) size meshes from curvature, proximity and gaps, but the first-layer height still comes from a y+ target the user supplies, and none checks the resolution against the quantity of interest before solving. Design-space tools (CAESES, optiSLang, HEEDS, Dakota, SALib) sample the ranges the engineer sets; none says where a range came from. Text-to-CAD (Zoo, Text2CAD, CAD-Coder, CADCodeVerify) now produces running CadQuery for simple parts, but dimensional accuracy is weak and failures can be silent. LLM simulation agents (MetaOpenFOAM, Foam-Agent, ChatCFD) report how often a case runs, not whether it is fit for purpose, and none runs a grid-convergence study.

This module's aim is the missing piece: references and ranges with their source, CAD checked against exact formulas and published objects, a mesh sized from the physics and the purpose, and one report with a verdict per stage. Template parts are not free-form CAD and the rules are not adjoint-based adaptation; when the stakes are high, the verification purpose and the GCI are what make a number quotable.

Limits

• Nine templates today; any other part goes through the language model (no published reference) or a STEP file.
• snappyHexMesh prism layers collapse at sharp edges and thin gaps; the coverage is measured and reported, and a low coverage is a warning, not hidden.
• Cell, cost and degree-of-freedom estimates are order-of-magnitude.
• Gmsh (GPL) and OpenFOAM (GPL) are optional external tools; install them separately (pip install gmsh).