Most of us ship a model that returns one number and then bolt a confidence score onto it afterwards, usually because somebody asked. Operational meteorology went the other way decades ago: the distribution is the product, and the single run is just one sample from it. The pattern is worth stealing, so here is how it actually works.
The Physics Chain Underneath Every Forecast
Weather is not a bag of correlations, it is a solved-forward physics problem. The Sun heats Earth unevenly, temperature differences become pressure differences, pressure differences become wind, wind moves moisture, and moisture becomes clouds and precipitation. Earth's rotation deflects the moving air through the Coriolis effect, which is why cyclones spiral instead of collapsing straight into the low.
The energy accounting matters more than it looks. Evaporating water stores a large amount of latent heat in the vapor, and when that vapor condenses inside a cloud the heat dumps back into the surrounding air. That release is the fuel that turns an ordinary updraft into a thunderstorm and a tropical disturbance into a hurricane.
How A Numerical Model Actually Runs
Numerical weather prediction slices the atmosphere into a three dimensional grid and integrates the governing equations forward at every point: conservation of mass, momentum and energy, plus the ideal gas law. ECMWF's global system runs at roughly 9km grid spacing. Regional models at 3km or finer stop parameterizing convection and start resolving individual thunderstorms.
The inputs come from about 10,000 surface stations, roughly 900 radiosonde launches every 12 hours, plus radar, aircraft, buoys and satellites. Geostationary satellites give you the same hemisphere every few minutes, polar orbiters give you higher resolution on a slower revisit.
Why The Ensemble Is The Real Output
Every one of those observations carries error, and the system is chaotic, so two nearly identical initial states diverge. Lorenz found this inside a weather model in the first place, which means the field inherited the problem and the workaround at the same time.
The workaround is ensemble forecasting: run the same model many times from slightly perturbed initial conditions and read the spread across members. Tight clustering means high confidence. Wide scatter means several outcomes are genuinely open. A "40 percent chance of rain" is that scatter converted into a number, not a forecaster hedging. This is what pushed medium range forecasts, the 3 to 10 day window, from unusable to genuinely decision grade.
What To Take Back To Your Own Stack
If your model is deterministic and your users need to act on it, you are shipping the least informative part of what you computed. Perturb the inputs at the level of their real measurement error, run the thing N times, and report the spread. It costs N times the compute and it is usually worth it, because the honest answer to "how sure are you" is a distribution rather than a trained confidence head.
The full walkthrough of the physics, from dew point and cloud formation through fronts, supercells and hurricanes to how the forecast pipeline is assembled, is in our complete guide to how weather works.
Curious whether anyone here has actually shipped an ensemble instead of a point estimate, and what the compute bill looked like.