fields/climate-weather/what the AI was for
Simulation surrogate in weather, climate and oceans
How researchers in weather, climate and ocean science train neural networks to stand in for expensive physics-based simulations, producing forecasts and model components at a fraction of the computing cost.
How it is used
Weather and climate science rests on numerical simulation: computer models that divide the atmosphere and ocean into a grid of boxes and step the equations of fluid motion forward in time. These runs are costly, so researchers train neural networks, programs that learn patterns from past data rather than following written rules, to reproduce what the simulation would have produced. Sometimes the network replaces the whole forecast. A transformer model, GL-Geoformer, was trained on fields of wind stress and ocean temperature across the tropics and predicts two years of the same fields ahead, feeding its own output back in to extend the forecast step by step.
Elsewhere the network replaces only one part of the simulation while the physics solver keeps running. Processes smaller than a grid box, cloud formation, rainfall, radiation, cannot be resolved directly and have traditionally been approximated by hand-written formulas. A network can learn these effects instead, taking the state of a single column of atmosphere and returning how that column should change. The same approach supplies corrections where data on pressure levels must be converted to the coordinates the solver uses.
Why researchers use it
The appeal is speed. A simulation that resolves the relevant physics honestly can take hours or days on a supercomputer, which limits how many scenarios can be explored and how many repeated runs can be made to gauge uncertainty. A trained network, once built, runs in a small fraction of that time, so it stands in for computing capacity the field does not have.
It also stands in for formulas nobody knows how to write. The unresolved processes in a climate model are approximated by expressions chosen partly on judgement, and they are a long-standing source of disagreement between models. Learning those effects from data shifts the problem from writing a formula to assembling a training set. In forecasting, a learned model may also capture relationships the equations as implemented do not, as with the interactions between tropical ocean basins that affect how far ahead El Niño can be predicted.
In the record
Limits
A surrogate is only as good as what it was trained on. These models learn from past states, so behaviour outside that range, a warmer climate, an unusual event, may be reproduced poorly or not at all, which matters most for the climate projections the field cares about. Errors can also compound when a model is fed its own output over long runs. And a network gives no account of why it predicts what it does, so a wrong answer is hard to diagnose, and physical conservation laws hold only approximately unless enforced.
Written by AIxSci: the background from general knowledge, the examples from checked records.
Related
Every article
aix-00306
Benchmarking the cost of ensemble Kalman methods for calibrating chaotic models
aix-00303
Neural network stands in for a weather model inside a forecast-correction scheme
aix-00283
Deep learning forecast of El Niño suggests Indian and Atlantic oceans extend warning time
aix-00297