FieldCraft

An open-source, clean-room CAE application in pure Rust: from a parametric SolidCraft part to a verified finite element analysis, with tetrahedral and hexahedral meshing, its own implicit, explicit, thermal and lattice Boltzmann solvers, and decks for Abaqus, Nastran and LS-DYNA. The desktop app, the command line and AI agents (over MCP) drive the same commands.

Status: early but working end to end — open a part, mesh it, define the analysis, solve, see the results — with a nonlinear implicit solver (large deformation, plasticity, rubber) now in place. Hex meshing is the centre of gravity. Next: buckling and arc-length, contact, then the wider material and multiphysics library. See ROADMAP.md. Want to help? Pick an issue labelled good first issue or help wanted and read CONTRIBUTING.md.

Claude driving FieldCraft through its MCP server: one prompt, and it opens the part, meshes it, defines and solves the analysis, then checks the tip deflection against beam theory. Real session, waits shortened. Full video.

An AI agent drives FieldCraft over MCP, each stage captioned with the tool it calls: it opens an I-beam, hex meshes it, defines the material, the clamp and a 1000 N tip load, solves, and shows the von Mises stress and the deformed shape (tip deflection 0.176 mm; beam theory with shear 0.177 mm). MP4 · regenerated by tools/demo-video.

Download

• macOS: open the .dmg and drag FieldCraft into Applications (Apple silicon and Intel).
• Windows: unzip and run fieldcraft.exe.
• Linux: the .AppImage (allow it to run as a program), or the .tar.gz.
• Browser: try it at jyviko.github.io/fieldcraft, no install.
• Command line: fieldcraft-cli ships in every package; AI agents connect to it over MCP (docs/mcp.md).

Every release lists its checksums in SHA256SUMS.txt. Building from source is in CONTRIBUTING.md.

Hex meshing a part that does not sweep

A filleted L-bracket with a hole through each leg cannot be swept in one piece. Split into 30 cells in SolidCraft, FieldCraft meshes it all-hex as one conforming mesh — O-grids round the holes, graded through the fillet — and its hex20 tip deflection matches a fine tet10 solution to 0.1 %. The same part also meshes all-hex automatically with the octree mesher (mesh.hex.octree).

Cells	Conforming hex mesh	Von Mises (clamp and load drawn)

What it does today

• Geometry in: SolidCraft designs (.solidcraft) with persistent face, edge and body names, and face-tagged 3MF meshes; surface checks (weld, manifold, orientation, self-intersection).
• Meshing: constrained Delaunay tetrahedra (tet4, tet10) with exact predicates and a certified validity gate; Delaunay refinement, smoothing and sliver removal; structured hex by sweeping, revolving and mapping (hex8, hex20), and conforming multi-block hex across a part split into cells (mesh.hex.blocks); unstructured octree all-hex for any part (mesh.hex.octree); 2D quad and tri meshes; quality metrics per element and set.
• Elements: solids (tet4, tet10, wedge6, hex8 with incompatible modes, hex20 full and reduced, hybrid u–p for near-incompressible materials), plane stress, plane strain and axisymmetric, trusses, Timoshenko beams, MITC3/MITC4 shells, springs, dashpots and point masses.
• Model: sets bound to geometry names, so they survive remeshing; isotropic and orthotropic elasticity, J2 plasticity (isotropic and kinematic hardening, Johnson–Cook), Johnson–Cook damage, neo-Hookean, Mooney–Rivlin and Yeoh hyperelasticity, thermal properties; fixed, displacement, symmetry and temperature conditions; forces, moments, pressure, gravity, centrifugal and heat loads; reference points with kinematic and distributing couplings; undo, redo and a text-diffable .fieldcraft file.
• Solvers: linear static and frequency (supernodal Cholesky, block Lanczos); nonlinear static by Newton–Raphson with automatic increments — large deformation (total Lagrangian) on every solid element, follower pressure, finite- and small-strain plasticity sharing one return map with the explicit solver, rubber and thermal strain, with measured quadratic convergence; steady and transient heat transfer and coupled thermal stress; explicit dynamics with plasticity, damage, erosion and rigid walls; D3Q19 lattice Boltzmann flow with forces on the part's faces. Results never depend on the thread count.
• Decks: Abaqus, Nastran and LS-DYNA writers, each refusing by name what it cannot express.
• Results: contours of stress, strain and displacement with a legend, deformed shapes, mode shapes, explicit and nonlinear frames, probes and reaction sums; supports, loads, reference points and couplings drawn in the viewport.
• Verification: every capability is checked against closed forms and published benchmarks (NAFEMS, MacNeal–Harder, Scordelis–Lo, the elastica, Hill's plastic cylinder, Rivlin's tube) in tests/oracle, with results frozen so they cannot drift: 170 of 176 quantities pass in 50 cases, the 6 others documented known failures (docs/oracle.md); against Abaqus's published benchmark results in How FieldCraft compares with Abaqus.
• Driving it: a desktop app in the Craft family's layout, a CLI, a JSON control channel and an MCP server; every capability is a command (docs/features.md).

How FieldCraft compares with Abaqus

The Abaqus values are the ones Abaqus publishes in its Benchmarks Guide ¹ . We never run Abaqus. The FieldCraft values come from docs/oracle.md. Errors are measured against the benchmark's reference. For each element we list Abaqus's best published mesh. The Guide doesn't give element counts for the NAFEMS meshes.

Benchmark	Reference	Abaqus (element, result, error)	FieldCraft (element, result, error)
NAFEMS LE1 elliptic membrane, σ_yy at D	92.7 MPa	CPS8 fine: 92.81, +0.12 %; CPS8R fine: 90.07, −2.8 % 2	quad8 24 × 48: 92.648, −0.06 %; hex20 layer 24 × 48: 92.869, +0.18 %
NAFEMS LE10 thick plate, σ_yy at D	−5.38 MPa	C3D20R fine: −5.53, +2.8 %; C3D20 fine: −5.64, +4.8 %; C3D10HS fine: −5.51, +2.4 % 3	hex20, 2 × (8 × 16 × 3): −5.334, −0.86 %
NAFEMS FV32 tapered membrane, f₁…f₆	44.623 … 391.44 Hz	CPS8R: worst mode 5, 381.32, +0.37 %; CPS6: worst +0.02 % 4	hex20 layer 16 × 8: every mode within 0.004 %
NAFEMS LE11 thermal cylinder, σ_zz at A	−105 MPa	C3D20 fine: −103.26, −1.7 %; C3D20R fine: −99.60, −5.1 % 5	pending (thermal loads)
Scordelis–Lo roof, free-edge midpoint u_z	0.3024 (MacNeal–Harder); Abaqus uses 91.2 mm	S4 18 × 18: 91.89 mm, +0.77 %; S8R 9 × 9: 91.90 mm, +0.8 % 6	hex20R one layer, quarter 16 × 16: −0.3014, −0.33 %
Pinched cylinder, radial u under the load	1.825 × 10⁻⁵	S8R5, 2646 dof: 1.833 × 10⁻⁵, +0.44 %; S4R, 2646 dof: 1.779 × 10⁻⁵, −2.5 % 7	pending (shells)
MacNeal–Harder twisted beam, thick, in-plane tip u	5.424 × 10⁻³	S4R 8 × 48: 5.430 × 10⁻³, +0.11 % 8	pending (shells)
MacNeal–Harder pinched hemisphere	0.0940	not found (Abaqus publishes NAFEMS LE3, a different load and material)	pending (shells)
MacNeal–Harder patch, straight and curved cantilever, thick cylinder	see docs/oracle.md	not found	hex8 incompatible: patch exact; straight cantilever −1.2 % to −10.8 %; curved −12 % / −20 %; thick cylinder −1.4 %
Cook's membrane, Euler and plate buckling	—	not found	see docs/oracle.md

Not like for like:

• LE1 and FV32: our hex20 rows model plane stress with a thin solid layer. Only the quad8 row matches Abaqus's CPS8 directly.
• LE10: our mesh is probably finer than Abaqus's "fine" mesh, but the Guide gives no element count, so this isn't certain. The Guide notes that only one element touches D in its fine mesh.
• Scordelis–Lo: Abaqus publishes shell elements only. Our row uses a single layer of solid elements. The two references differ slightly: 91.2 mm against MacNeal–Harder's normalised 0.3024.
• Pinched cylinder: Abaqus reports mesh size as degrees of freedom.

Build

cargo xtask ci                                  # fmt, assets, licences, layers, clippy, tests, wasm
cargo xtask test --heavy                        # the heavy (ignored) tests and the doc tests
cargo run -p fieldcraft-cli -- commands         # every command and its parameters
cargo run -p fieldcraft-cli -- run script.json  # run a command script
cargo run -p fieldcraft-cli -- mcp              # MCP server on stdio for AI agents (docs/mcp.md)

Rust 1.95 or later, with the wasm32-unknown-unknown target for cargo xtask ci.

Documentation

• Architecture: crates, layers and the meshing pipeline
• Element library and verification and validation
• Capability checklist and roadmap
• Libraries and licences
• AI agents (MCP): claude mcp add fieldcraft -- /path/to/fieldcraft-cli mcp
• Control protocol: driving the desktop app
• Releasing: how a release is built, verified and published
• Contributing and agent instructions

License

Dual-licensed under MIT or Apache-2.0, at your option. Notices are in NOTICE; third-party licences in THIRD-PARTY-LICENSES.txt.

FieldCraft is an independent project, not affiliated with, sponsored or endorsed by Dassault Systèmes, MSC Software, Siemens or Ansys. Abaqus, Nastran, MSC Nastran, Simcenter Nastran and LS-DYNA are trademarks of their owners and are named only to describe file compatibility.

---

1. Abaqus 2017 Benchmarks Guide, hosted at https://abaqus-docs.mit.edu/2017/English/SIMACAEBMKRefMap/. Each footnote names the page, and the values are in its "Results and discussion". ↩︎

2. "LE1: Plane stress elements—elliptic membrane", Table 1, Abaqus/Standard: simabmk-c-le1.htm ↩︎

3. "LE10: Thick plate under pressure", Table 1, Abaqus/Standard: simabmk-c-le10.htm ↩︎

4. "FV32: Cantilevered tapered membrane", frequency table: simabmk-c-fv32.htm ↩︎

5. "LE11: Solid cylinder/taper/sphere—temperature loading": simabmk-c-le11.htm ↩︎

6. "Barrel vault roof problem", Table 1, regular meshes: simabmk-c-barrelvault.htm ↩︎

7. "Pinched cylinder", regular-mesh table: simabmk-c-pinchcyl.htm ↩︎

8. "Twisted beam" results table: simabmk-c-twistedbeam.htm ↩︎