2D wind tunnel
A fluid dynamics simulation running on your device's graphics card, in real time: pick a case or draw an obstacle with the mouse, set the Reynolds number and watch wakes, vortices and separation develop. Drag, lift and vortex shedding frequency are measured as the flow evolves.
Flow
Aerodynamic coefficients
Results
How it works
- Method: D2Q9 Lattice Boltzmann. Instead of solving the Navier–Stokes equations directly, the solver evolves particle distributions on a regular grid with a local collision step and a streaming step; the macroscopic limit is incompressible flow. Each cell is a pixel computed in parallel by the GPU, which is how it reaches hundreds of millions of updates per second.
- Collision: regularized model, which keeps only the hydrodynamic part of the distributions and stays stable at very low viscosity, with a high bulk viscosity to damp sound waves and a Smagorinsky turbulence model (LES) phased in above Re ≈ 200, where the grid no longer resolves the smallest eddies.
- Boundaries: prescribed velocity at the inlet, prescribed pressure at the outlet with an absorbing layer, free-slip tunnel walls (or a moving ground for the car), no-slip on the obstacles with interpolated bounce-back on the exact shape.
- Forces: drag and lift come from the momentum exchanged at the wall links; the Strouhal number St = f·L/U from the oscillation frequency of the lift.
- Verification: the "Cylinder in a channel" case is the Schäfer and Turek (1996) benchmark at Re = 100 (reference: Cd max 3.23, Cl max 1.0, St 0.30). With 20–30 cells across the diameter the Strouhal number is within 1% and Cd and Cl within 4–5%, mostly because of the method's slight compressibility; Poiseuille flow in a channel is reproduced within 0.3%.
Limitations
Two-dimensional, incompressible (low Mach) flow: in 2D vortices do not break down as they do in reality, so above a few thousand Reynolds the result is qualitative. The grid is uniform, so thin boundary layers are under-resolved, and the tunnel has finite blockage. Reliable numbers for a design need 3D analysis, wall-adapted meshes and calibrated turbulence models.
From simulation to design
On real projects we couple 3D CFD with neural surrogate models trained on the simulations: once trained they answer in milliseconds and make it possible to optimize the shape across thousands of variants. The result is CAD-ready geometry.