AeroPhon is a browser-native design-to-simulation loop: natural language, sketches or hand gestures drive real parametric CAD geometry, and machine-learned CFD surrogates trained on thousands of high-fidelity simulations return drag, surface pressure, wall shear and crash response — instantly, while you design.
Five stages run behind every request — from a sentence to an engineering readout.
Plain-English commands, an uploaded side-profile sketch, or webcam hand gestures (rotate, zoom, explode, select).
A real, part-labelled DrivAer CFD reference body is morphed to your length, width, height, roofline, nose and rake.
Graph neural networks predict drag and full surface fields on the new shape in under a second.
One click escalates to a genuine OpenFOAM RANS solve (k-ω SST) or a frontal-crash estimate on the same geometry.
PBR viewport with clearcoat paint, Blender Cycles photoreal renders, diffusion concept styling, teardown animation.
Every capability below is live in the app — none of it is mocked.
A RegDGCNN graph network evaluates the morphed body as a 3-D point cloud and returns the drag coefficient without meshing or solving.
R² ≈ 0.81 on held-out real cars · <1 sPer-vertex Cp and wall-shear-stress fields painted directly onto the body — the same views a CFD post-processor gives, at interactive speed.
Full-body field payloads · ~1 s responseWhen the surrogate is not enough: automated blockMesh → snappyHexMesh → simpleFoam (k-ω SST) with force-coefficient extraction on your exact geometry.
Measured Cd vs. surrogate, side by sideFrontal-impact crush with staged fold kinematics and a von-Mises stress proxy, animated frame by frame on the car body.
Wierzbicki–Abramowicz fold mechanicsStable Diffusion + ControlNet renders photoreal concepts that follow your model's true silhouette; Blender Cycles renders the actual geometry.
27–44 s per render on a laptop GPU21-landmark hand tracking runs in your browser: open palm rotates, pinch zooms, two hands explode the assembly, dwell selects a part.
One-Euro-filtered · no server round-tripThe car splits into its 18 real components — each wheel, rim, lamp and glass pane — and disassembles piece by piece to a service layout, then rebuilds in reverse.
Component split via mesh-graph analysisRetrieve the most similar real cars, interpolate between designs, morph along learned shape modes, and see where your car sits on a t-SNE map of 4,165 CFD-validated designs.
DeepSDF · latent interpolation · t-SNEUpload any car image and a ResNet-18, fine-tuned on 7,967 CFD-labelled renderings, estimates its drag coefficient.
Validation R² ≈ 0.70The learned components, their training data, and honestly reported accuracy.
| Component | Architecture | Training data | Reported accuracy |
|---|---|---|---|
| Drag surrogate point cloud → Cd | RegDGCNN (graph CNN) | DrivAerNet++ point clouds, real RANS Cd labels | R² 0.81 (held-out) |
| Surface pressure per-vertex Cp | DGCNN segmentation head | DrivAerNet++ surface fields | R² 0.69 (per-point) |
| Wall shear stress per-vertex |τw| | DGCNN segmentation head | 5,350 cars, VTK surface data | R² 0.24 (v1 — improving) |
| Image → Cd | ResNet-18 (fine-tuned) | 7,967 rendered DrivAerNet++ views | R² 0.70 (val) |
| Verification solver | OpenFOAM simpleFoam, k-ω SST | — (physics, not learned) | measured, not predicted |
No installation. Runs in the browser; heavy lifting happens server-side.
| Step | What to do |
|---|---|
| 1 | Type "draw me a car" — a real morphable car appears with its predicted drag. |
| 2 | Say "make it longer", "lower the roof", "paint it midnight blue" — geometry and physics update. |
| 3 | Click Pressure (Cp) or Wall shear to paint predicted flow fields on the body. |
| 4 | Click Disassemble for the workshop teardown, or Gesture control and drive the viewer with your hands. |
| 5 | Finish with Run real CFD to verify the surrogate against OpenFOAM on your design. |