🚗 How Chat CAD works

Describe or sketch a car in plain English, get a real 3D model you can edit part-by-part, see an AI photoreal concept, and test it like an engineer — aerodynamics and crash — from neural surrogates trained on thousands of real CFD cars.

1

Describe / sketch

Type “a sleek electric coupe, lower the roof, midnight blue”, click a body (Sedan / Coupe / SUV / Wagon), or upload a side sketch.

2

Real 3D model

The shape is built by morphing a real DrivAer CFD reference body to your dimensions — a clean, watertight 3D car, not a cartoon.

3

Edit parts & colour

Recolour the body, glass, wheels, lights and grille individually; tune the glass tint/clarity; stretch it longer / wider / taller / lower.

4

AI concept render

A diffusion model (SDXL) turns your description into a photoreal concept image — different prompt, genuinely different car.

5

Test it

Run engineering surrogates on the 3D model: drag, surface pressure, wall shear stress, and a 3D crash deformation.

The engineering under the hood

Aerodynamic drag (Cd)

A RegDGCNN graph network predicts drag from the car’s point cloud — no CFD run needed.

R² ≈ 0.81 on 825 held-out real cars

Surface pressure (Cp)

A per-point surface-field network paints the pressure distribution over the body.

DrivAerNet surface-field task

Wall shear stress (WSS)

A second surface-field network predicts skin-friction over the body.

trained on 5,350 real WSS cars

Crash safety

A frontal-impact crush animates on the car in 3D, with a moving stress band (Wierzbicki–Abramowicz mechanics).

3D, interactive

Where the intelligence comes from

Real data

Surrogates are trained on the DrivAerNet / DrivAerNet++ datasets — thousands of real CFD simulations of the DrivAer car.

Multi-agent pipeline

Styling → CAD → Meshing → Simulation → NVH agents collaborate to turn a brief into an analysed design.

Bring your own LLM

Paste an Anthropic or Google AI key for natural-language control, or use the built-in command parser — nothing is stored.