~/aixsci
weekly updates

fields/climate-weather/what the AI was for

Attribution in weather, climate and oceans

How researchers in weather, climate and ocean science use AI to work out what caused something: which conditions, emissions or instrument settings account for a measurement or an observed change, and by how much.

How it is used

Attribution here means apportionment rather than prediction. The question is not what the air quality will be tomorrow but how much of yesterday's clean air was the weather and how much was policy; not what a laboratory instrument will read but which conditions shaped the reading it gave. The usual approach is to train a statistical model (a set of rules fitted to past data) that links an outcome to many candidate drivers at once, and then interrogate the fitted model: hold the drivers fixed, vary them one at a time, or read off which ones the model leaned on. Tree ensembles, which average many simple branching rules, are common because they tolerate messy, correlated inputs.

The inputs are typically whatever can be measured alongside the outcome: wind, temperature, humidity and pressure from weather stations or reanalysis (a reconstruction of past weather on a grid), plus instrument settings, and plain clock variables such as hour of day and day of week that stand in for human activity. The model is fitted separately for each site or each quantity, then used to estimate what the outcome would have been under some other set of conditions. The difference between that estimate and what was actually observed is the share attributed to the remaining cause.

Why researchers use it

The difficulty in this field is that almost nothing happens for one reason. Pollution falls because a policy worked, or because it was windier. An instrument signal shifts because the chemistry changed, or because the sample was heavier. Weather varies enormously on its own, and that variation is large enough to hide or mimic the effect anyone is trying to measure. The controlled experiment is unavailable: there is only one atmosphere and one ocean, and no second copy to leave alone as a comparison.

What the fitted model stands in for is that missing comparison. It is used to construct a counterfactual: an estimate of what would have been observed had the weather behaved differently, or had one driver taken another value. Simpler alternatives exist, such as adjusting for a few variables by regression or comparing similar days, but they assume relationships are straightforward and handle only a handful of drivers. Flexible models absorb curved and interacting relationships between many inputs without those relationships having to be written down in advance, which is why they are reached for when the drivers number in the dozens.

In the record

Limits

A fitted model shows association, not mechanism. When drivers move together, as weather and emissions often do, the model can credit the wrong one, and clock variables that stand in for human activity will also absorb anything else that follows a daily or weekly rhythm. The counterfactual is only as good as the range of conditions in the training data: push beyond what was observed and the estimate becomes guesswork. Results depend on which drivers were offered, so a cause left out of the inputs cannot be attributed, and will instead be quietly shared among those included.

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

Related

Every article