Standard PhysicsSend a design problem

Standard Physics

AI agents that do the engineering physics.

We build AI agents that do the engineering physics. Reynolds runs the simulation you would have asked an engineer for. Archimedes goes further: give it a goal and it designs, tests and improves thousands of candidates, then hands you the best one with the evidence.

Early access for Archimedes. Prefer to write? kabir@inviscidai.com

01 Reynolds

Ask for a simulation. Get the answer.

Running CFD takes a specialist and a week of setup. Reynolds is an agent that does that part. It is open source, and free for individual engineers.

A free-surface wave pattern past a Wigley hull, computed by Reynolds.
  • Describe the flow in plain words. It writes the case, builds the mesh and runs it.
  • The solver underneath is OpenFOAM, so the answer is a real computed field, not a guess.
  • You get the figures, the numbers, and an ordinary case an engineer can open and rerun.
pip install openreynolds

More about Reynolds → How the solver is validated →

02 Archimedes

Set a goal. Get a better design.

One simulation answers a question. Archimedes asks thousands of them on your behalf, writing each design as code and testing it, until it has something better than what you started with.

  • Send the goal, the limits that cannot change, and how you would score a design.
  • A swarm of agents writes candidate designs as code and simulates every one of them.
  • The best is re-checked on a stricter simulation. When it does not hold up, the report says so.
Writing geometryDesign 238 of 1,251
Shape
End-plate tips
Blades
3
Across
240 mm
Best score
0.4150

Plates on the tips to stop flow spilling round them.

The brief it was given

One propeller for a 2 kg drone that hovers in air and cruises underwater, on the same motor. Score it on how much battery power turns into thrust.

  1. An agent writes the blade as code: chord, pitch, sweep and thickness at every station along the span. Each round is a different idea, not a nudge to the last one.

  2. A mesh is built around the part and a real flow solver runs it, in air and underwater. No model guesses the answer; the equations are solved.

  3. Only the score decides. Designs that lose are dropped, the ones that win get built on, and the leaders are re-run on a finer mesh before anyone sees them.

    Best score0.4150 after 238 designs

A real run: 1,251 designs, 20 hours, about $160 of compute. Underwater, the stock propeller drew 47.5 W; the one it kept draws 3.91 W.

  • 7case studies published in full, including the ones that went wrong
  • 1,500designs a run, each one simulated rather than predicted
  • 3 to 30 hfrom brief to a checked answer, for a few hundred dollars of compute

The one it kept.

A stock drone propeller stalls its motor underwater and wastes 97% of the battery. Archimedes redesigned the blade and the gearbox together. Underwater draw fell from 47.5 W to 3.91 W.

Read the propeller case →
What it was givenStock propeller, 47.5 W underwater
What it keptDesign 1,255, 3.91 W underwater

What you get back.

When a result doesn't hold up on the stricter check, the report says so. Three of the published cases say exactly that.

Example report (PDF) →
  1. A report, written for a reviewer.

    What it tried, what won, what it rejected and why. Not a score with no working shown.

  2. The design itself.

    The geometry and the numbers behind it, so your team can check it, build it or take it into CAD.

  3. The run behind it.

    How many designs were tested, what each one scored, and the finer simulation the winner was re-checked on.

Have something that should perform better?

Send the goal and the limits, and it goes straight to our team. Or start on your own with Reynolds, free, today.

Every result comes from flow simulation, the same kind engineers already use. See the accuracy tests →

Send a design problem