Weather forecasting has been one of the great successes of computational science – decades of steady improvement from solving the physics of the atmosphere on ever-larger computers. Recently, models trained on historical data rather than physics have matched that performance, and the field is adjusting.
What changed
A machine learning model, given decades of past weather, learns to predict the next state from the current one. It does not simulate the physics; it recognises patterns.
That turns out to work remarkably well for medium-range forecasts – several days out – and runs in minutes on modest hardware instead of hours on a supercomputer.
What it is good at
Speed, which allows many forecasts to be run for the same period, giving better estimates of uncertainty. Skill at the ranges where most forecasting value lies. And tracking large-scale features like storm paths.
Where physics still matters
Extreme events outside the training data. A model that has learned from history handles what history contains; something unprecedented is exactly what it has not seen.
Fine local detail, where the physics of terrain and convection produce effects too small for the training data to capture.
And the observation network that provides the starting point, which both approaches depend on equally.
How forecasters are responding
Running both. The learned model for speed and ensemble breadth, the physical model for cases the learned one may miss, and human judgement to combine them.
Why it matters beyond forecasting
Because it demonstrates that learned models can replace simulation in a domain where simulation was considered essential. That has implications well beyond weather.