~/aixsci
weekly updates

fields/climate-weather/what the AI was for

Structure determination in weather, climate and oceans

How neural networks are used in weather, climate and ocean science to recover a hidden field, such as surface currents or a full atmospheric state, from measurements that only hint at it.

How it is used

Weather, climate and ocean science tries to describe the state of the air and the sea: how fast water moves, how warm it is, where the winds blow. Almost none of this is measured directly everywhere at once. Satellites see the surface, ships and floats sample single points, and much of the ocean and atmosphere is never touched. Structure determination means working backwards from these indirect traces to the field that produced them. In the examples here, networks are trained on pairings of the two: a sequence of satellite-like temperature gradient images against the surface velocities of an ocean simulation, or a mixture of gridded satellite data and scattered in situ readings against a complete gridded state of the atmosphere.

The models are of two familiar kinds. One is the U-Net, an image-to-image network that shrinks a picture down through layers to catch large patterns and then builds it back up, with links across the two halves so fine detail is not lost. It turns one map into another map of the same place. The other is a transformer, a network that weighs all parts of its input against each other, used here to fuse observations of different kinds into a single gridded state. One example adds a layer with a learnable length scale so readings taken at arbitrary points, not on a grid, can be placed onto one.

Why researchers use it

Researchers reach for this because the quantity they want is often not the quantity they can see. Ocean currents at fine scale are hard to observe, while surface temperature is imaged often and widely; if temperature patterns carry a reliable signature of the flow, a network trained on a simulation can read the one from the other. In the other example, the network stands in for the assimilation step that normally fits observations to a physical model to produce a starting state for a forecast. Both replace a hand-built chain of equations and corrections with a mapping learned from paired data.

What is being substituted for, in both cases, is an expensive inverse calculation. Traditionally one writes down the physics, then searches for the state that best explains the measurements, a costly process repeated at every forecast cycle. A trained network does the search once, during training, and afterwards gives an answer in a single pass. That also lets different kinds of measurement, already on a grid or scattered irregularly, be handled inside one system rather than stitched together beforehand.

In the record

Limits

Both examples learn from paired data, and the pairing usually comes from a simulation or a reanalysis, so the network inherits whatever that source gets wrong and may not hold where the real ocean or atmosphere behaves differently. A mapping from surface traces to interior motion is not unique: different flows can leave similar temperature patterns, and the network returns one plausible answer without saying how uncertain it is. Features deliberately filtered out of the training targets, such as tides, will not be recovered, and performance outside the regions and seasons trained on is unknown.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

Every article