Article July 18, 2026 AeroFlow LLM · Procedural Geometry · In-Browser Simulation

AeroFlow NS: An LLM Generates Geometry & Runs a Fluid Solver in a Single HTML File

An AI model wrote the aircraft, the fluid solver, the particle system and the graphics, then folded the whole thing into one small HTML file that runs in any browser at 60 frames per second. No CAD, no installation, no plug-ins: open it and watch the physics solve itself.

Generated by an LLM Procedural geometry (no CAD) Real-time fluid solver ~2.2 MB standalone HTML 60 FPS in any browser Two aircraft models GPU tracers 500–10,000 Smoke & streamline modes Three.js · R3F
AeroFlow NS: Stealth UCAV generated by an LLM, visualized with GPU tracer particles in a real-time fluid solver
Cover. The stealth UCAV model, a blended flying-wing with lambda planform and flat-shaded skin, was generated entirely by an LLM from a natural-language prompt. The geometry, the solver, and the 9,000 GPU tracer particles all live inside a single ~2.2 MB HTML file.

AeroFlow NS is a real-time, three-dimensional wind tunnel that lives inside a single web page. What makes it remarkable is who built it. The aircraft, the fluid solver, the particles, the streamlines, the graphics and the controls were all generated by an LLM, a 2.8-trillion-parameter model, from a description written in plain language.

No CAD file was imported, no mesh loaded and no solver library linked. The model described the aircraft mathematically, wrote the solver from scratch, connected it to a Three.js scene and packed everything into a file small enough to email.

Disclaimer: This is a technology demonstration of what an AI model can generate, not a validated CFD tool. Treat it as a physically plausible real-time visualization, not a certified aerodynamic analysis.

Launch a simulation

Each simulation is a self-contained HTML file. Choose an aircraft below and it opens in a new tab, ready to run, with the same solver flowing around two very different shapes.

Once it’s open

Drag to orbit and scroll to zoom. The sliders set the airspeed, the angle of attack and the number of tracer particles and streamlines, and four modes switch between particles, streamlines, both, or smoke. Space freezes the flow, G shows the grid and O starts an automatic orbit. For a dramatic start, raise the angle of attack past 12° and watch the wingtip vortices strengthen, then switch to smoke to see the wake as one continuous stream.

What the model built

Everything you see was produced by a single AI model from a written brief, with no human-written code in the loop. It began with the aircraft: a refined jet airliner and a stealth flying wing, each described not as a mesh but as mathematical formulas for the fuselage, wings, winglets, nacelles and tail. It then wrote an incompressible fluid solver from scratch, the kind of code that normally takes a CFD engineer days to implement, which recalculates the airflow around the aircraft on every frame.

On top of the solver sit the visuals. Up to ten thousand tracer particles, drawn by the graphics card in a single pass and coloured by local speed. Up to two hundred streamlines seeded upstream and around the vortex cores. A smoke mode that turns the same particles into soft, diffuse puffs, like dye in a real tunnel. And around it all, the full rendering pipeline, the controls, the loading screen and a fallback for older browsers, packed by the model into a single file of about 2.2 MB.

Small enough to email

The delivery is the most striking part. A live 3D fluid solver with generated geometry, GPU particles and cinematic post-processing, and the whole thing is one file. It makes no requests to a server once open, runs in Chrome, Firefox, Safari or Edge on anything with WebGL, needs nothing installed, and holds 60 frames per second on a modest graphics card. The two aircraft share the identical solver; only the shapes and materials differ.

Why it matters

Sharing a CFD result used to mean shipping gigabytes of data, asking someone to install visualisation software, or settling for a video. Here the result is not a recording. It is a live calculation, running on the reader’s own graphics card the moment they open the link.

Two aircraft, one solver

The airliner has a slender fuselage, tapered wings with winglets, stabilisers and engine nacelles, and a wing aspect ratio of 5.14. The stealth flying wing is a blended lambda planform that tapers from an 8.1-metre centre chord to a one-metre tip, with an aspect ratio of 4.2. Run the same solver around both and the difference shows at once: the flying wing trails wider, more diffuse vortices and a broader wake, while the airliner’s vortices are tighter and persist further downstream. None of this is scripted. It emerges from the geometry meeting the physics, just as it would in a full CFD calculation.

What emerges on its own

Given nothing but the airspeed, the angle of attack and the shape, the solver produces the features any aerodynamicist would expect: a stagnation point at the nose, air accelerating over the top of the wing, a wake trailing behind, downwash beneath, and two counter-rotating wingtip vortices rolling up from the pressure difference between upper and lower surfaces. Change the angle of attack and the whole flow rearranges itself in real time.

The wingtip vortices you see trailing behind the aircraft are not injected from a formula. They emerge from the pressure difference between the upper and lower wing surfaces, exactly as they would in a real CFD calculation.

About the LLM

The simulation was produced with the LLM, a frontier-class open-weight model and, at the time of writing, the first open model in the three-trillion-parameter class. Its million-token context can hold an entire codebase at once; it can look at screenshots and live previews; and it is built for long, sustained coding tasks across many files with little supervision.

That combination suited this project unusually well. The aircraft demanded spatial reasoning: how a wing planform becomes a solid, how dihedral and thickness interact, how the body and the wind relate as the angle of attack changes. The solver demanded fluency in numerical methods, from the choice of advection scheme to the boundary conditions. And because the same agent also built the interface, packaged the file and checked the visuals by eye, refining particle counts, colours and camera framing from what it saw, the result looks right as well as computing right.

How good is it?

Having written and reviewed CFD solvers for years, I would call it a decent model, among the better AI-generated physics implementations I have seen. The architecture is right: a staggered grid, a stable advection scheme, a pressure projection and immersed-boundary geometry, giving numerically stable results with physically plausible behaviour. Its limits are those of any real-time solver: a coarse 0.5-metre grid, a loose pressure solve, first-order advection that smears fine detail, and no turbulence model. A production code would add multigrid pressure, higher-order schemes, boundary-layer resolution and turbulence modelling. As a live, browser-based demonstration of fluid physics, generated from a written description, it is remarkably competent.

The bigger picture

The distance between “describe what you want” and “a running simulation” is collapsing. The geometry, the solver, the graphics and the interface would each have taken days of expert work not long ago. Now the loop is simply this: describe the physics, review the result, refine. The question is no longer whether an AI can produce a simulation like this from a single request. It is how soon that request becomes the unremarkable part, and the interesting work moves to what we ask for next.


If you enjoyed this, you may also like the companion pieces on LLM-orchestrated CFD and why external aerodynamics is the easy part.