fields/climate-weather/what the AI was for
Forecasting in weather, climate and oceans
How researchers in weather, climate and ocean science use artificial intelligence to predict what the atmosphere and the sea will do next, from the next hour of rain to the next two years of the tropical Pacific.
How it is used
Weather and climate science tries to say what the atmosphere and ocean will do before it happens. The usual tool is a physics model: equations for air and water, solved step by step on a grid of boxes covering the planet. AI enters this work in several places. Some models learn the whole forecast directly from past observations, taking recent maps of wind, temperature and pressure and producing the maps that follow. Others learn only a part, such as the small-scale processes that happen inside a single grid box, and hand the rest back to the physics. Still others learn to turn the last few radar pictures of rain into the next few.
The models used follow the shape of the data. Where the data is a single sequence of numbers over time, such as a storm's track and strength, networks built to carry information forward through a series are used. Where the data is a stack of maps, models that compare every part of a map with every other part, or that slide small filters across it, are more common. Some produce not one forecast but a spread of plausible ones, by drawing random numbers and letting the model turn each draw into a different future. Longer forecasts are often made by feeding the model's own output back in as its next input.
Why researchers use it
Physics models are expensive. A fine grid over the whole atmosphere means an enormous number of calculations, and running the same forecast many times over to gauge uncertainty multiplies the cost. AI models, once trained, are cheap to run, so a spread of forecasts becomes affordable. They also stand in for parts of the physics that nobody can write down exactly: cloud formation, rainfall and turbulence happen on scales smaller than any practical grid box, and are usually represented by rough approximate formulas. A learned replacement, fitted to reanalysis data that blends observations with model estimates, can take that place.
There is a second reason: some things are predictable in the data without being tractable in the equations. The slow swing of warm and cool water across the tropical Pacific is one. Forecasts of it have long struggled through the northern spring, when the signal is weak and small errors grow. Models fitted to long records of wind stress and ocean temperature at many depths can pick up relationships between ocean basins that help carry a forecast through that weak patch. Rainfall in the coming hour is another: too fast for a global model, but visible in radar.
In the record
Limits
These models learn from the past, so they describe the climate of their training record rather than a changing one, and extreme events are rare in that record and poorly constrained. They can be smooth where the real atmosphere is sharp, understating heavy rain or deep pressure. Scores on held-out data flatter them: nearby times are similar, so a split can leak. Recursive forecasts drift as errors feed on themselves. And a learned component inside a physics model can be accurate while remaining hard to inspect, which makes failures difficult to anticipate.
Written by AIxSci: the background from general knowledge, the examples from checked records.
Related
Every article
aix-00283
Deep learning forecast of El Niño suggests Indian and Atlantic oceans extend warning time
aix-00285
Neural network forecasts freak waves minutes ahead from a single buoy's record
aix-00286
AI fills gaps in river temperature records to track overlapping heatwaves
aix-00282
Neural network trained on a decade of records forecasts typhoon strength and position
aix-00291
Statistical model estimates what Makkah's air would have held without lockdown
aix-00292
Neural network forecasts daily sea surface temperature up to 90 days ahead
aix-00294
Neural network trained on simple equations spots tipping points in real data
aix-00295
Diffusion model trained on past weather produces 15-day global ensemble forecasts
aix-00296
Neural networks trained on 39 years of weather data produce week-ahead global forecasts
aix-00297
Weather and climate model built from physics equations plus a trained neural network
aix-00298
Generative AI model forecasts UK rainfall 90 minutes ahead from radar images
aix-00299