Weather models built with artificial intelligence now match traditional forecasts for wind and temperature. Rain is harder. This project tests how well such models forecast rain, from the next day to several weeks ahead.
At a glance
- Theme: Rain forecasts from days to weeks
- Observations: Satellite rainfall (GPM), rain gauges
- Approach: Artificial-intelligence weather models, compared with forecasts from weather centres
- Scale: Global

The basics
A weather forecast starts from the state of the atmosphere today, as measured by satellites, weather stations and balloons. For decades there was one way to get from today to tomorrow: a computer model that solves the equations of physics for the air, step by step, on a supercomputer. In the last few years a second way has appeared. A model built with artificial intelligence learns from decades of past weather how the atmosphere changes from one moment to the next, and then makes a forecast in minutes.

For wind, pressure and temperature the new models have become as accurate as the traditional ones. Rain is another matter. It does not vary smoothly as temperature does: it falls in bursts, over small areas and short times, and a few very wet days bring a large part of the year's total.
The challenge
A model that learns by keeping its average error small is tempted to play safe. Light rain over a wide area is rarely far off, while a sharp storm forecast slightly out of place counts as wrong twice. Such a forecast scores well and still misses what matters most, and the further ahead it looks, the smoother it becomes. The yardstick is imperfect too: satellites see rain everywhere but not exactly, and gauges measure it exactly but only at single points.

Our focus
We let an artificial-intelligence weather model learn rain from satellite observations and compare its forecasts with those of established weather centres, around the globe and from one day to several weeks ahead. We look beyond the average error: at heavy rain, at where the rain falls, and at how much of its variability a forecast keeps.
Why it matters
Flood warnings, reservoirs and farming depend on rain forecasts. Knowing what an artificial-intelligence forecast gets right, what it misses and how far ahead it can be trusted shows how to use it, on its own or together with traditional forecasts.












