The further into the future you look, the less accurate the weather forecast becomes - the reliability of the forecast decreases after just a few days. Now researchers have used machine learning to train an AI model that creates reliable and fast 10-day weather forecasts with little computational effort. The model, called GraphCast, was trained on historical weather data and performed better in tests than the best traditional systems. It was even able to predict extreme weather events, although it was not explicitly trained for this. From the researchers' perspective, GraphCast can complement conventional systems and overcome new challenges.
The dynamics of weather is a complex physical phenomenon that depends on numerous factors, including air pressure, humidity, temperature and wind speed. To predict the weather, complex models have been used that describe these factors using differential equations and convert them into algorithms. These so-called numerical weather forecast models (NWPs) have so far been considered the gold standard for weather forecasts that are as precise as possible. They are used to create forecasts based on current weather data. However, they require powerful mainframe computers, and each refinement requires additional computing capacity.
Training with historical data
A team led by Remi Lam from Google DeepMind in London has now developed a model with a different approach. Their model GraphCast is based on coupled neural networks trained on historical weather data from 1979 to 2017. The system learned how certain combinations of meteorological parameters typically develop and what weather conditions result from them. This is done without the complex physical calculations on which traditional models are based. As a result, the AI system requires significantly less computing capacity: instead of several mainframe computers, a computing chip developed for deep learning that is barely larger than a postage stamp is sufficient.
In less than a minute, GraphCast can generate 10-day weather forecasts with a resolution of about 0.25 degrees of latitude and longitude, or about 28 by 28 kilometers. The current weather and the weather from six hours ago serve as input. Based on this, the model predicts what the weather will be like in six hours. This result is then fed back into the model and assumed as the new current state, on the basis of which the weather is again predicted in another six hours - and so on.
More accurate than the previous gold standard
To determine the reliability of the AI system, the researchers pitted it against the most accurate conventional medium-range weather model to date. The result: “In 90 percent of cases, GraphCast was significantly superior to the best deterministic model to date,” report the researchers. “It also better predicted extreme weather events, including the tracks of tropical cyclones and extreme heat and cold waves, even though it was not specifically trained for this.” The predictions in the test runs referred to weather events in 2018 that were not included in the training data set.
When making predictions for 2021, GraphCast was slightly less accurate, but still about as good as the previous best model. The researchers conclude that it is helpful for accuracy if the training data is as up-to-date as possible. In fact, when they additionally trained GraphCast with weather data through 2020, their 2021 forecast performance improved. “Regular retraining based on current weather data makes it possible to capture weather phenomena that change over time, for example due to climate change,” writes the team.
Supplement for conventional models
However, a weak point of GraphCast so far is its handling of uncertainty, as Lam and his colleagues admit. As with normal weather forecasts, the forecasts become less reliable the further into the future they go, but the model is not yet able to indicate the degree of uncertainty. “This is an important next step in further development,” say the researchers. They also want to further refine the spatial resolution of the model in the future.
From the team's perspective, GraphCast represents a turning point in weather forecasting. However, it is not intended to replace traditional methods. “Previous prediction methods have been developed over decades, tested in many real-world scenarios, and offer numerous possibilities, many of which we have not yet considered,” they write. “Our work aims to be proof that machine learning is capable of meeting the challenges of real-world prediction problems and has the potential to complement and improve current best methods.”
Source: Remi Lam (Google DeepMind, London, UK) et al., Science, doi: 10.1126/science.adi2336