CAD thatsimulates.

A Formula 1 car coloured by surface pressure: blue suction on the wings and floor edges, red stagnation on the wing leading edges and tyre fronts

Geometry, meshing, CFD, FE and post-processing in one pipeline, running on the GPU. A design change goes from shape to a full aerodynamic report in seconds.

Open the Workbench demo See the benchmarks

33 swhole aero job, geometry to report9.6×OpenFOAM, meshing included0.4 sto mesh 0.64 M cells15 msper iteration on 2.5 M cells0.291Ahmed C_d, wind tunnel 0.285–0.2992.4 ¢per cloud run230+automated checks

// 01 Platform

Shape to answer on one device.

No file conversion, no hand meshing, no solver tuning. The geometry feeds every solver directly, and the whole chain runs where the data already is: on the GPU.

01 / 04

GPU + Rust

Geometry

A CAD kernel written from scratch: exact Booleans, smooth blends, sketches with fillets, extrude, revolve, pipe and patterns. STL and STEP parts join as nodes.

26–82×faster than NumPy

The car split into components: front wing, body, floor, wheels and rear wing in different colours

02 / 04

GPU

Mesh

Graded Cartesian grids with planar-exact cut cells follow from the geometry, refined where the part needs it. There is no meshing step to babysit.

2 sF1 car set-up

0.4 s0.64 M cells

The underside of the car coloured by surface pressure

03 / 04

CFD on GPUFE on CPU

Solve

RANS k-ω SST aerodynamics with moving ground and rotating wheels on the GPU. Flow, heat, structures and additive manufacturing run in the same runner.

15 msper iteration

64×a 16-thread CPU

Flow speed and streamlines on the centre plane of the car

04 / 04

GPU

Post-process

Forces with a per-component split, surface and floor pressure, wake planes, vorticity and VTK fields in one report, with every numerical risk flagged as a warning.

1command per study

Vorticity just above the ground: tyre wakes and floor-edge vortices

// 02 Capabilities

The datasheet.

What ships today and where it runs. Everything here is covered by automated checks.

Module	Method	Runs on
CAD kernel	Implicit geometry, exact Booleans and blends, sketches, extrude, revolve, pipe, patterns; STL and STEP import; watertight STL mesher	GPU + Rust
Meshing	Graded Cartesian grids with planar-exact cut cells, built from the geometry	GPU
External aerodynamics	Steady RANS k-ω SST with wall functions, moving ground, rotating wheels	GPU
Structures (FE)	Matrix-free voxel and finite-cell elasticity with multigrid: statics, modal (also prestressed), harmonic response, linear buckling	CPU · GPU next
Internal flow and heat	Laminar steady and time-accurate flow, conjugate heat transfer, coupled to stress	CPU
Additive manufacturing	Printability, build direction, supports, trapped powder, thin walls, LPBF distortion	CPU
Post-processing	Force split, surface and floor pressure, wake planes, vorticity, VTK; GPU rasterizer for surface renders	GPU
Agents	MCP tool server: an AI agent writes a spec, runs it and reads the KPIs, warnings and report	Any device

// 03 GPU

Any GPU. No lock-in.

One set of kernels runs through Vulkan, Metal and DX12: NVIDIA data-centre cards, AMD and Apple laptops, even the integrated GPU of a mini PC.

Whole aero job · F1 car · 2.5 M cells · seconds

L40S

H100

A10G

A100 80GB

Radeon 680M

165

CPU, 16 threads

702

The Radeon 680M is the integrated GPU of a mini PC. NVIDIA cloud runs cost 2.4 to 4.7 cents each.

Every stage on the deviceGeometry, grid, solver, forces, running means and surface renders. Post-processing stays on the GPU between iterations.
Bit-reproducibleRepeated runs are bit-identical, and every GPU kernel is checked against its CPU twin on T4, L4, A10G, L40S, A100, H100, Radeon 680M and a software driver.
Portable by designWGSL kernels through wgpu: Vulkan on NVIDIA and AMD, Metal on Apple, DX12 on Windows. No CUDA required.

// 04 Benchmarks

Faster than OpenFOAM. True to the wind tunnel.

Ahmed body with a 25° slant, k-ω SST with wall functions, same Apple M3 Pro laptop. Time from geometry to a settled drag, meshing included.

Run	Cells	Time to settled drag	Speed-up	C_d	C_l
LeonSim, GPU (Metal)	636 k	3.6 s	9.6×	0.289	0.381
LeonSim, native Rust CPU, 11 threads	636 k	23.3 s	1.5×	0.276	0.390
OpenFOAM v2412 (snappyHexMesh + simpleFoam)	327 k	34.8 s	1.0×	0.313	0.381
Wind tunnel (Ahmed 1984 / Meile 2011)				0.285 / 0.299	0.345

OpenFOAM is the best of 2, 5 and 8 ranks. The Ahmed wake oscillates rather than reaching a fixed point, so coefficients move by a few percent with the solver path; with the stricter settle-then-average protocol all LeonSim back-ends agree (C_d 0.291 to 0.298) and the GPU is about 4× faster than OpenFOAM.

230+automated checks against analytic solutions, published data, OpenFOAM and a real browser

±2 %drag of a single Formula 1 run; downforce ±10 %, reported with every result

8 / 8GPU kernel checks on eight devices, and on a software GPU in CI on every change

// 05 Showcase

From an STL to a full aero report.

Formula 1 is the hardest test we run: thin wings, ground effect, rotating wheels and a mesh full of display-model defects. One command does all of it.

2026 concept car

From a public STL, repaired automatically.

Holes, loose patches, thin plates and open rims are fixed on import, then the car is solved with a moving ground and rotating wheels.

1.53 m²downforce

1.09 m²drag

29 %front balance

8.5 M cells, 10 minutes on an AMD Radeon 680M. Drag ±2 %, downforce ±10 % (why).

Read the full study

Surface pressure on the 2026 concept car, three-quarter rear view

Flow speed and streamlines on the centre plane of the car, showing the slow wake behind it — Speed on the centre plane with streamlines: accelerated flow over the nose and engine cover, the slow wake behind the rear wing and diffuser.

Surface pressure on the underside of the car — From below: suction under the floor and in the diffuser carries most of the downforce.

Cold plateFlow, conjugate heat and coolant-pressure stress in one run, about 5 s.

Bolted bracketA CAD part held by its bolt holes and loaded through its pin hole.

Stiffened panelStress and linear buckling of a thin-walled structure.

Tuning forkNatural frequencies from the geometry: it rings near 440 Hz.

LeonSim Workbench: spec editor, 3D temperature view with coolant streamlines and KPIs for a liquid cold plate

// 06 Workbench and agents

Drive it from a browser, a script or an AI agent.

Write a spec, run it and explore the result in 3D: colour by temperature, stress or pressure, follow streamlines, cut sections, animate mode shapes, and click the part to place supports, loads and heat sources.

# a full F1 aero report on the GPU
$ leonsim aero f1 --level medium --backend gpu
# flow, heat and stress of a cold plate
$ leonsim run examples/cases/coldplate.yaml
# the Workbench, and the MCP server for agents
$ leonsim web --open
$ leonsim mcp

// 07 Roadmap

What comes next.

Not shipped yet, in the order we are building it.

NextStronger GPU pressure solveThe pressure correction is about 80 % of each GPU iteration: the next big speed lever.

NextStructures on the GPUBring the FE solvers onto the same device pipeline as the CFD.

PlannedUnsteady RANS and DDESTime-accurate wakes, for credible wheel and diffuser flows and tighter downforce.

PlannedAdaptive octree gridsFewer cells in empty air, more in wing slots and gaps.

PlannedExact NURBS B-repA full STEP round trip alongside the implicit kernel.

PlannedValidation at scaleHundreds of Ahmed and Windsor body variants against published data sets.

See a full simulation in your browser.

The Workbench demo holds stored results of 18 example cases, from a cold plate to a tuning fork.

Open the Workbench demo F1 2026 study