fea

name: feadescription: Checks whether a part or an assembly of bonded parts holds under its loads, as a linear static stress study of a STEP, an xbf or a text-to-cad model, meshed by netgen and solved by CalculiX, giving peak von Mises stress, safety factor against yield and deflection drawn on the part. Use when the person asks whether a part will hold, break or is strong enough, how much it bends or sags, where it is most stressed, how much load it takes, how to make it stronger or lighter without failing, or for FEA, a stress analysis or a strength simulation. Static loads and steady g-loads only.license: MITmetadata:   repository: earthtojake/text-to-cad

FEA: will it hold, and how far does it move

A study holds some faces of a part, loads others, and has CalculiX solve the linear elastic response on a netgen mesh of quadratic tetrahedra: stress and displacement under forces, pressures, the part's own weight and a steady acceleration, with an isotropic material below yield, perfectly rigid fixtures and loads that do not move. An assembly's parts are bonded where they touch. It is a first-pass engineering check, not a certification; say so when you report. Vibration, buckling, heat, flow, fatigue, impact and parts that slide on each other are not in this version: say so plainly, and do not pass a static safety factor off as an answer to them. A drop or a braking load can be estimated as a steady g-load (Body loads).

Start with the task

This skill studies a part; it does not model one. It reads a saved .step or .xbf, or a model the engine holds by its document id (cad model solve <id>), and solves the whole document: to study one part of an assembly alone, export it as a document of its own.

Setup

Run the cad command (~/.local/bin/cad where it is not on your PATH; cad.exe on Windows). If it is missing, install text-to-cad, outside a sandbox or with its permission:

curl -fsSL https://texttocad.dev/install.sh | sh

On Windows, in PowerShell: irm https://www.texttocad.dev/install.ps1 | iex. If the CAD app was only just added, its server may still be starting: try again for up to a minute first. cad doctor --check explains any other failure, changing nothing; cad doctor repairs what it reports.

The study is the solve operation of the document's type: cad step solve, cad xbf solve, cad model solve. With no --study it lists the study's setup and writes nothing; with --study and --out DIR it solves and writes its results into DIR. Its two tools are integrations the cad command installs the first time a study needs them, nothing for you to install; cad run gives a script that calls cad.step.solve both, with the kernel that reads the part.

• netgen, the mesher: the netgen integration (cad[netgen], about 11 MB). netgen publishes it for macOS, Linux x86_64 and Windows x86_64 alone; elsewhere (Linux on ARM among them) a solve refuses, naming the platform.
• CalculiX (ccx), the solver: the calculix integration (cad[calculix]: the cad-ccx package, CalculiX 2.23 for each platform, about 13 MB). A person's own CalculiX is never run unless CAD_CCX names the file (2.20 or newer). It runs headless, on one thread, in a private folder, so the same study gives the same numbers every run, and every result names the version that solved it. cad run ccx ARGS runs that same ccx on a deck of your own.

Units are fixed: mm, N, MPa (pressure, modulus and stress), t/mm³ for density (steel 7.85e-9) and g (9.81 m/s²) for gravity and acceleration. Restate every load in them before you write it down.

Workflow

1. List the faces. A study names faces by the engine's refs, the ones the CAD Viewer copies; a face the user picked there arrives as /work/bracket.step#bracket/face 17 and needs no lookup.

   cad step solve part.step
   

   With no study this is the setup, and writes nothing: the parts, the pairs that touch, then each face of the part, one a line: ref, area, centre, surface type and a hint such as plane, normal -Z, largest. Match the user's words ("the mounting face") to a ref by the hint and the centre; a hole is a cylinder. A document of several parts lists no faces until one is named: cad step solve assembly.step --occurrence post. A face number is a position in the saved document, so a part edited and written again is listed again before its study is re-run.

2. Write the study (study-file.md):

   {
     "fixtures": [{"faces": ["bracket/face 17"], "type": "fixed"}],
     "loads": [{"faces": ["bracket/face 22"], "type": "force", "vector_N": [0, -500, 0]}]
   }
   

   fixed holds a face still; roller lets it slide in its plane but not off it (a frictionless support, or a symmetry cut). A force is the total over its faces; a pressure is in MPa, positive into the surface; gravity and acceleration take vector_g in g and no faces, and need the material's density. margin is the safety factor the part should keep (default 2, at least 1). checks adds a displacement check with the user's limit when they gave a sag or stiffness spec; left out, the stress check alone judges the part. Leave mesh out on the first run.

3. Solve into a folder.

   cad step solve part.step --study study.json --out /tmp/bracket-load --json
   

   --out is any folder, a temporary one included, made when missing; put it in the project when the user should keep the files. One folder per study: a solve replaces its own files there, touches nothing else, and a run that fails leaves the previous results as they were. It holds:

   File	What it is for
   result.json	the answer, what --json prints. Read this first.
   stress.png, displacement.png	the deformed part coloured by each field (blue low, red high), the deformation exaggerated by deformation_scale, with a colour bar giving the field, its units and its range. They show whether the load points the way you meant and which faces are held. Show the stress one to the user.
   result.glb	the same coloured surface, for any glTF viewer or cad viewer
   result.vtu	the volume mesh with displacement (mm) and von_mises (MPa) at its corner nodes (Your own questions)
   ccx.inp	with --deck: the input CalculiX solved (the finer solve's when there were two), for cad run ccx -i ccx in that folder

   Where the engine cannot draw (Linux with no display) the pictures are left out and a solver_note says so; everything else is written.

   The result is a check's: ok, findings (each with severity, code, message, location and hint), written, and measurements: the study as given, the material, the faces it named (named_faces), the summary's figures (max_von_mises_MPa and max_von_mises_at_mm, max_von_mises_gauss_MPa, yield_MPa, safety_factor, max_displacement_mm and max_displacement_at_mm, applied_force_N, reaction_force_N, deformation_scale, checks; an assembly's weakest_part and parts), each fixture's reaction, mesh (size, elements, degrees of freedom, straightened, solver), timings, and refined, the first solve's size and peak when the run solved twice. Without --json the same prints as lines.

   • When the safety factor comes out under 3 the run meshes again finer (half the element size, less on a big model) and solves again by itself; the finer result is the one written (Mesh).
   • Progress ([fea] meshing ...) goes to stderr, never into the result; --verbose adds mesh and solve detail there.
   • A malformed study (an unknown key, a face not on the part, a zero load, a load on a held face, no fixture) is refused before meshing, naming the field. A part the fixtures leave free to move, or that no load reaches, is refused by the solve; so is a solve whose reactions do not balance the applied load.
   • The command exits 1 ("ok": false) when a finding is an error: the part yields or misses a check, with the results written all the same; or the study stopped before the solve (not_connected), with nothing written. Warnings leave it at 0. Read the result either way.

4. Report with units: the peak von Mises and which feature it is near, the safety factor and the yield it is against (the low end of the model material's range), the max displacement and where, the applied load against the reaction, the mesh size and whether a finer re-solve ran; then the findings, what the model assumes, and what would change the answer most. Show stress.png, saying the deformation is exaggerated by deformation_scale.

Your own questions

For anything the summary does not answer (the stress at the mounting holes, the sag along an edge, how much of the part is above some stress), read result.vtu rather than solving again. meshio or PyVista read it as it is (uv run --with meshio):

import meshio, numpy as np

mesh = meshio.read("/tmp/bracket-load/result.vtu")
xyz, stress = mesh.points, mesh.point_data["von_mises"]          # mm, MPa
moved = np.linalg.norm(mesh.point_data["displacement"], axis=1)  # mm
near = np.linalg.norm(xyz - [42, 12, 3], axis=1) < 5             # within 5 mm of a point
print(stress[near].max(), moved[xyz[:, 2] > 49].max())

It holds the corner nodes of the tetrahedra, inside the part as well as on its skin; the reported peak also counts the mid-edge nodes, so it can be a little higher than the VTU's own maximum. An assembly's cells carry their part (part, the index into the summary's parts).

Assemblies

A document of several parts is solved as one assembly: the parts that touch are bonded into one conforming mesh, each keeps its own material, and each part's stress is judged against its own yield. Every part must reach a fixed part through bonded neighbours. The keys are in study-file.md.

The setup (cad step solve assembly.step) prints each part (ref, name, volume), then each pair: bonded, with its contact area and gap; not connected for a near miss (a gap up to 1 mm, beyond contact_tolerance_mm); overlapping · N mm³ where the solids share volume, which is never bonded. An overlap no deeper than the tolerance is an interference (a press fit, or a modelling slip): it is listed bonded with interference N mm and closed like a gap. Read the listing against the design: is each detected joint a real one, and is a joint you expect missing? It judges by the 0.1 mm default; a study that should bond a wider gap sets contact_tolerance_mm from the gap printed. Names can repeat (two parts called bar): use the ref (frame/bar.2) when one does.

Each part is solved with its model's own material; parts gives a part the numbers the user gives, and a part with no material in its model and none in the study is refused, naming it (Materials). connections only overrides a detection (free for a pair that must not be glued); touching pairs are bonded without it, and a pair set free inside a group glued through other parts is refused. Fixtures and loads name faces on any part (post/face 23), but not a face wholly covered by a bonded joint (refused, naming the face); a face another part only partly covers is held or loaded on its uncovered area. Connections are bonded or free: parts that press and slide on each other, and bolts, are not in this version.

The finer re-solve runs when the weakest part's safety factor is under 3. The summary's yield_MPa and safety_factor are the weakest part's (weakest_part, weakest_part_peak_MPa), while max_von_mises_MPa and max_displacement_mm span the whole assembly: the highest stress can sit in a stronger part, so do not divide one by the other; each part's own numbers are in parts. Each finding names its part ("'post' yields: ..."): fix that part (a thicker section, a fillet, its material in parts, a larger joint, whose area_mm2 is in the result's connections), not the assembly in general.

Compare variants

Each version of a design is its own document; run the same study on each into its own folder and compare the result.jsons. Face numbers are not stable between versions: an edit renumbers them, and a ref reused unchecked lands on a different face with no error, so list each variant's faces and map the baseline's by surface type, normal, centre and area, asking when two fit. Give every study the baseline's mesh.size_mm, so a difference is the design's and not the mesh's, and give the mapping (baseline bracket/face 17 → bracket/face 21, the mounting face) with the table so the user can check it.

Findings

Two von Mises peaks are reported. The nodal peak (max_von_mises_MPa, the one the safety factor uses) is each element's stress at its integration points, extrapolated to its nodes and averaged over the elements at a node, per part in an assembly; it is usually the higher of the two. The integration-point peak (max_von_mises_gauss_MPa) is the raw element value; more than 50 % above the nodal one, it means a concentration the mesh has not resolved. Safety factor = yield / nodal peak of the solve that is written. The edge of a perfectly rigid fixture or bonded joint and a sharp inside corner are singularities: the peak there grows with every refinement and never converges, so the findings below say when to read the stress a little away from it instead.

Findings list errors first: an error makes the part unfit to use, a warning is a judgment to weigh and act on or explain. Most carry the face (bracket/face 12, the original part's ref) when the point is on one, and the point in mm. peak_at_fixture and peak_concentration appear only when the safety factor is under the study's margin and a finer solve did not settle the peak (moved it under 10 %); above the margin they would not change the answer.