Forget the weather forecast for 14 days: AI should calculate it for 30 days

Contents

In a new study, researchers have shown for the first time that an accurate weather forecast over a month could be possible – with the help of AI.

Deepmind's AI weather model Graphcast could calculate forecasts for a period of a month or more. This is according to a paper that Trent Vonich, PhD student at the University of Washington (UW) and his team recently published as a preprint on arXiv. Conventional weather models can predict the weather for a maximum of about 14 days.

Who invented modern weather forecasting?

The British mathematician Fry Lewis Richardson came up with the idea at the beginning of the twentieth century that it should be possible to predict the weather using the laws of physics. Until then, meteorologists had largely relied on experience and observation and developed empirical models based on the development of similar, known weather patterns.

Richardson, on the other hand, wanted to rely entirely on physics. However, his methods were not yet practicable at the time. His first hand-calculated predictions for changes in air pressure were far removed from the actual measurements – but are now considered classics of meteorology.

How do weather forecasts work?

Weather models describe the temporal – i.e. future – development of pressure, temperature and humidity based on specific initial values. However, the physical equations only indicate the extent to which variables change – are usually coupled differential equations. To calculate what the weather will be like, these equations have to be solved numerically. This is done in a grid in fixed time steps.

How precise a weather model is depends on how small the grid is. The Cosmo-DE weather model of the German Weather Service, for example, uses a grid of 2.8 km × 2.8 km with 50 height layers. Clouds, which are usually smaller than such cells, must therefore be "parameterized" – They only appear in the adjustment of individual model parameters. This is why forecasts are always particularly inaccurate when there are small-scale disturbances in a larger air flow.

Why is their forecast time limited?

The rule that weather forecasts are only useful for about 14 days goes back to the American mathematician and meteorologist Edward Lorenz. In the early 1960s, Lorenz investigated the possibilities of numerical weather forecasting in a highly simplified system – known as a convection cell. This is a volume of air that is heated evenly from below. The warm air rises, cools down and then flows back down again. He discovered that even small inaccuracies in the initial values led to large inaccuracies in the forecast after a while. Lorenz therefore assumed that the time horizon of the prediction would be limited even if the initial values were measured with arbitrary accuracy. Chaos theory later provided the theoretical underpinning for this so-called "butterfly effect". In short: weather models are chaotic systems.

So how do the AI models work?

The AI models do not attempt to calculate a physical model of the weather. They use training data to learn how meteorological variables change over time.

The idea of using data-driven predictions instead of physical models is an obvious one. This is because the computational effort for such models is much smaller and they produce results more quickly. Research groups around the world have been working in this field since the 1990s – not only for weather – but also for climate models. However, the accuracy of the results has so far left much to be desired.

In 2019, Google researchers presented their version of a nowcast. The short-term weather forecast for a few hours uses images from a rain radar as training data for a neural network. Many countries regularly publish radar measurements throughout the day that show how clouds form and move throughout the day. In the UK, a new reading is published every five minutes. Putting these snapshots together results in an actual stop-motion video showing how the rain patterns move across the country, similar to the forecast images on TV. The neural network is then trained to predict the next rain pattern. This has worked well for short periods of time, but the AI approaches have a problem: they do not automatically take into account natural laws such as the conservation of energy and mass. AI models tend to behave unphysically.

There are various approaches to solving this problem. In 2023, for example, Huawei's research department published a paper on a deep neural network that generates global weather forecasts for up to seven days. With comparable accuracy, but 10,000 times faster than current weather models. As the resolution of numerical weather models is still severely limited by the amount of computing power required, such neural networks could also be used to calculate much finer-resolution forecasts.

Pangu-Weather, Huawei's model, uses two new techniques to improve the results: Firstly, Kaifeng Bi's team works with a transformer model that processes three-dimensional data. The software calculates meteorological variables such as temperature, pressure, wind speed and humidity at different altitudes. By keeping the calculations for the different heights within the model consistent with each other, the researchers elegantly prevent the model from becoming unphysical – and thus getting out of hand.

Why can Graphcast now make even better predictions?

To make the AI predictions even more long-term, Vonich and his team used an elegant trick. They wanted to see how well Graphcast would work if they were able to radically improve the accuracy of the initial conditions, i.e. the starting data.

To do this, they compared the final state of the atmosphere from so-called reanalysis data – i.e. weather data from the past – with the predictions from Graphcast. Shortcomings in a prediction could then be used to adjust the initial conditions of the reanalysis data that the model had used to start its prediction. With the improved initial conditions, the accuracy of Graphcast's 10-day forecasts improved by an average of 86 percent. The model was also able to deliver good results when forecasting the weather for more than 33 days in the future.

The open question now is how well this trick would also work for real forecasts. In general, AI models only tend to deliver meaningful predictions if the conditions are not too far removed from those of their training examples. Nevertheless, the work has caused quite a stir – The possibilities of AI weather forecasting are clearly far from exhausted.

This article first appeared on t3n.de.

(emw)