Sharper starting points
Every forecast starts from a snapshot of the atmosphere. Ours learns that snapshot from every observation across time, and the physics fills in what no instrument saw.
The physics AI
of Earth.
IC Weather rebuilds the physics models that forecast the weather so they can learn from data the way AI does, while keeping every law of physics they were built on.
A storm, forecast by our model.
- Under contract with NOAALearning version of GFS
- Matches the originalDown to rounding error
- Exact gradientsThrough the whole model
- Built for GPUsFast and cheap enough to run thousands of times
How it works
Every equation, trainable.
Our models compute the exact gradient of a forecast with respect to everything that shaped it: the starting weather, the physics settings, the observations. With that gradient, a physics model trains on real data pass after pass, and every equation stays in place.
Pass 1Miss
- 01Forecast with the full physics.
- 02Measure the miss against observations.
- 03Send the gradient back through every equation.
- 04Update. The next pass lands closer.
Live demo
Watch physics learn.
A small twin of our method, running live in your browser. Our full models do the same thing for the whole planet.
Keyboard: arrow keys move, Enter acts, Shift moves faster.
Proof
As exact as the original.
Then it learns.
Closer to the original than the original is to itself.
We ran the same 15-day storm in our model and in the original model. Ours stays closer to the original than the original stays to itself when its starting point is nudged by a single rounding error.
A simulated storm on a global mesh. Lower means closer.
IC Weather vs. originalOriginal vs. itself, nudged by one rounding error
Half
the mismatch with real satellite observations, after the model learned from them in our first real-data test.
Exact
gradients, computed through every equation to the last digit a computer can hold.
Fast
Every component runs on the chips that power modern computing, so training a physics model becomes quick and affordable.
What it unlocks
When physics can learn, forecasting changes.
Models that tune themselves
Thousands of settings inside a weather model are still tuned by hand. Exact gradients let the model adjust them against real observations, all at once.
AI that obeys physics
Neural networks can train inside the full physics model. They learn what the equations miss, and the equations keep them consistent.
The value of every observation
Measure how much each satellite, balloon or ground station improves a forecast, and decide where the next one should go.
Who it's for
Built with the people who make forecasts.
Forecast centers and agencies
GPU speed and learning, without leaving the physics they trust.
Observation companies
A way to show, number by number, what each instrument adds to a forecast.
AI weather and climate teams
Physical gradients to train against, and physics that keeps learned models consistent.
Company
The learning layer for Earth-system physics.
IC Weather is a US company. We bring learning to the physics behind forecasts: the atmosphere, clouds and radiation, satellite observations, the ocean and space weather.
We are under contract with NOAA to build a learning version of the GFS that runs on GPUs with exact gradients, as a candidate path for the next generation of US weather forecasting.
IC Weather was founded by Xiaoxu Tian, Ph.D., a data assimilation scientist who has built adjoint and 4D-Var systems for MPAS.
Early access
Be among the first to train a physics model.
We are preparing our first release: Earth-system physics models that run on GPUs, compute their own gradients and train on your data. Join the early-access list, and tell us what you would build. Your answers shape what we release first.
- AtmosphereA global weather model with its physics, trainable end to end.
- Satellite viewSimulate what satellites see, with exact sensitivities to the atmosphere.
- Ocean and space weatherThe same approach, applied below the waves and above the sky.