AI for the questions
the internet can’t answer

We are a geometric deep learning lab building foundation models that inherit the symmetries of physical law, trained on high-fidelity simulation data we create in house. Our models turn hours of simulation into seconds. We are currently focused on aerodynamics and defense.

Ratio-Aero modelVisualization tool
Angle of attack6.0°
Lift coefficient0.34

A NACA 0012 wing at Reynolds number 10,000, in the averaged flow from our own simulations. Move the slider to pitch the wing. The flows that matter have no closed form. We learn them.

156,000×

Faster. A 130-hour simulation of NASA’s Orion capsule at Mach 5 now takes under 3 seconds, compared with the standard method (CFD). Run on two H100 nodes.

Read the paper

440,000+

ILES & RANS simulations of wings across 10,000+ geometries, 11 angles of attack and 12 Reynolds numbers.

Explore the dataset

5,892

Downloads of our open Downforce Airfoil dataset on Hugging Face.

Get it on Hugging Face

Physics has no internet.
We are building it.

  1. Data

    Language models learned from text that already existed online. The data physical AI needs, how fluids flow and how structures behave, was never put online at scale. We generate it ourselves with high-fidelity simulation.

  2. Symmetry

    Physics has structure: it holds the same when a system moves or rotates. Our models are built to respect it, so they need less data and generalise further.

  3. Speed

    Classical solvers take hours or days per design. Our models predict the flow in seconds, so engineers can test hundreds of designs instead of a handful.

  4. Open

    We publish as we go: papers, datasets, and model weights as each release is ready.

What we build

Simulation data, models, evals.
Built to order.

Reach out
  1. Data
  2. Models
  3. Evals
  4. Deployment
  5. Design search

  • Data

    Every case comes with its setup.

    CFD simulation data made to order, at the scale a model needs and from fast RANS runs up to high-fidelity simulation. Every case records its solver, mesh, boundary conditions and convergence, so you know exactly how it was made.

  • Models

    Answers in seconds.

    Physics models trained on your designs and operating conditions, or adapted from ours. They respect the structure of physics, so they learn more from less data and hold up on designs they haven’t seen.

  • Evals

    A solver is the judge.

    We test any physics model, ours or anyone else’s, against fresh solver runs it has never seen, and show exactly where it holds and where it breaks. Public benchmarks and private tests.

  • Deployment

    Runs inside your walls.

    In your cloud, on your own servers, or fully offline (air-gapped). Outside the range it was tested on, the model hands the question to a solver, so every answer stays within what was tested.

  • Design search

    Try thousands of designs.

    Our models run inside your optimization process. Test thousands of shapes in the time one standard simulation takes, then confirm the shortlist with a solver.

Working today in aerospace and defense.

Now

Aero

Now

Defense

Our research

Models, data and papers

Follow ratiolabs on Hugging Face ↗ to get new weights and data the day they ship.

London · San Francisco

Team

Founders

Founding team

We run joint research with universities, labs and Enterprises.

We’re hiring

Work with us.

We’re hiring researchers and engineers in London and San Francisco. Pick the area closest to your work and send us something you have built, with the area in the subject line.

Machine learning and engineering 8

  • Geometric deep learning Build models that understand irregular geometries, physical symmetries, and complex spatial relationships and make them useful for real-world engineering. Reach out
  • Physics foundation models Build scalable foundation models that learn from high-fidelity physics simulations and generalize across geometries, flow regimes, and operating conditions. Reach out
  • World models and robotics Connect physics foundation models to robotics and world modeling, with a focus on learning physically consistent representations of dynamic systems. Reach out
  • Evals and environments Build scientific benchmarks and simulation-backed environments that rigorously establish where physics AI models work and where they fail. Reach out
  • Optimisation Turn fast physics predictions into design optimization tools that explore engineering tradeoffs and verify promising candidate geometries. Reach out
  • Data generation Build and operate the compute and data workflows that generate trustworthy simulation datasets for training physics foundation models. Reach out
  • Platform Build secure, high-performance systems that move physics models from research experiments into dependable services and customer environments. Reach out
  • Product Transform physics foundation models and datasets into accessible, interactive products for scientists and engineering teams. Reach out

Physics and simulation 2

  • Aerospace Define and validate meaningful aerodynamic problems for physics AI models, connecting model development with real aerospace engineering needs. Reach out
  • CFD Generate and validate the high-fidelity CFD simulations that support model training and evaluation, ensuring datasets are physically and numerically defensible. Reach out

Work with us

Tell us the question.

We will tell you if we can answer it, and what it takes.

Careers

Come build it.

We’re hiring researchers and engineers in London and San Francisco. Send us what you’ve built.