models/Aardvark Weather
Aardvark Weather
Aardvark Weather is a weather forecasting system that learns to go straight from raw satellite and weather-station readings to a forecast, replacing every stage of the usual forecasting chain at once.
What it is
Conventional weather forecasting runs in stages. First, scattered readings from satellites, balloons, ships and ground stations are combined into a tidy picture of the atmosphere on a grid of points, a step called data assimilation. Then a physics model steps that picture forward in time. Finally the output is adjusted for particular places and particular quantities, such as the temperature at a named airport. Each stage is a separate piece of software, and each is expensive to run.
An end-to-end system replaces the whole chain with one trained model. A neural network, a program that adjusts millions of internal numbers until its outputs match examples it has been shown, is given the raw observations and the later weather that followed them. In Aardvark Weather the chain is divided into three learned parts: one that turns the mixed observations into a gridded starting state, one that steps that state forward, and one that reads off forecasts for chosen places.
Why scientists use it
The appeal is cost and speed. The physics stage of a traditional forecast needs a supercomputer, and the assimilation stage needs one too. A trained network does its arithmetic once at training time, so making a forecast afterwards is comparatively cheap. That matters for regions and organisations without large computing budgets, and for anyone wanting many forecasts rather than one.
Replacing the whole chain, rather than one stage, also removes the handovers. In a staged system each stage is tuned on its own and passes a fixed, summarised product to the next. Trained end to end, the early parts can keep whatever information turns out to help the final forecast, including detail that a human-designed intermediate format would have discarded. The same argument is made in speech recognition, translation and image analysis.
In the record
Limits
A learned system can only imitate the weather it was trained on. Unusual conditions, a storm of a kind rarely recorded, are where it is least reliable and where its confident-looking output is most misleading. It carries no physical laws, so it can produce states that conserve nothing and could not occur. It also inherits the gaps and biases of the observing network, which is sparse over oceans and poorer regions. And it cannot explain its reasoning, so a forecaster cannot check why it said what it said.
Written by AIxSci: the background from general knowledge, the examples from checked records.