~/aixsci
weekly updates

fields/climate-weather/model families

Recurrent neural networks in weather, climate and oceans

How recurrent neural networks, which read data in order and carry a memory of what came before, are used to forecast storms, waves, rainfall and river conditions, and why researchers reach for them.

How it is used

Weather, climate and ocean science works mostly with sequences: readings taken hour after hour, day after day, at a station, a buoy, a river basin or across a radar image. A recurrent network is built for that shape of data. It steps through a series one observation at a time and keeps an internal state, a running summary of what it has seen, which it updates at each step. Common versions add small gates that decide what to keep and what to discard, so useful information can survive many steps. In the research here the networks are given past measurements and asked for the next ones, or for a label describing what is about to happen.

The uses in these examples fall into a few shapes. One is straightforward forecasting: feed in several consecutive readings and predict the following one, separately for each quantity of interest. Another is filling gaps, where a network learns the relationship between widely measured drivers, such as air temperature, rainfall and river flow, and a quantity measured only patchily, then reconstructs the missing record. A third treats the question as classification, asking whether a particular event is coming rather than predicting the numbers themselves. Recurrent layers also appear as one component inside larger models, combined with layers that handle spatial structure in images.

Why researchers use it

Researchers reach for these networks because the physics-based models they normally use are expensive to run and demand that every relevant process be written down in advance. A recurrent network instead learns the pattern of change from the record itself, including relationships nobody specified. It stands in for a mechanism: where a quantity depends on its own recent history in ways that are hard to state as equations, the network's internal state carries that history for you.

The appeal is also practical. Once fitted, such a model runs quickly enough for use when time is short, as in emergency forecasting or short-range rainfall prediction. It can be fitted jointly across many sites so that one model learns from all of them rather than each in isolation, which helps where individual records are thin. It can also be trained on simulated series from simple mathematical models and then applied to real observations, which matters when real examples of the event of interest are rare.

In the record

Limits

These networks learn from the record they are given, so they tend to reproduce conditions seen in training and may fail as the climate shifts beyond that range. Rare events are the hardest cases and the least represented, which is why the examples resort to rebalancing data, simulating it or pairing events with quiet periods. Long dependencies still fade despite the gates. The internal state is not interpretable, so a forecast comes without a physical explanation, and nothing guarantees that outputs respect conservation of mass or energy.

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

Related

Every article