Computer More accurate weather forecasts thanks to AI
More accurate, faster and more reliable weather forecasts: according to a study, a new AI model from Google subsidiary DeepMind should make this possible.
10.12.2024, 05:30
SDA
An expert from the German Weather Service (DWD) says that in certain aspects, AI forecasting systems cannot yet compete with classic, physically-based models. AI models should therefore be seen as a supplement, not a replacement.
Roland Potthast, head of Numerical Weather Prediction at the DWD, describes the study published in the journal Nature as an "important step": such models have a lot of potential that now needs to be tapped into. The approaches of Google and other tech companies could "complement, inspire and advance" weather services. This could provide the general public with ever better forecasts and warnings.
The machine-learning weather forecasting method called "GenCast" was developed by a team led by Ilan Price from the London-based company DeepMind. The study was carried out exclusively by DeepMind employees, but was then evaluated by independent experts for the journal. The team came to the conclusion that "GenCast" outperforms the best conventional medium-term weather forecasts. The model is also able to better predict extreme weather conditions, the path of tropical cyclones and the development of wind speeds.
The AI was trained on the basis of analysis data from 40 years of weather events (1979 to 2018). The research group then tested how well "GenCast" could forecast the weather for 2019.
Global 15-day forecasts
It is generally accepted that weather forecasts become less accurate the further into the future they look. "GenCast" is able to produce global 15-day forecasts within eight minutes, according to the report. For such medium-term forecasts, the European Center for Medium-Range Weather Forecasts (ECMWF) was previously considered the most accurate in the world. According to the developers, "GenCast" has now achieved over 97 percent better results in forecasting 1320 wind speeds, temperatures and other atmospheric characteristics.
GenCast does not calculate its forecasts once, but a total of 50 times per forecast. The probability of predictive reliability increases accordingly. According to the team, the system promises greater accuracy, efficiency and accessibility in a wide range of situations.
DWD tests its own AI model
The DWD is currently testing its own AI model, with others in the pipeline that will be used to supplement existing methods, said Potthast. "Physically based models and AI models are combined in the DWD's forecasting chain in order to provide the best possible forecasts on each time scale and for the targeted forecast variables - such as precipitation, temperature, winds, pressure, humidity, gusts, ice saturation and much more," said Potthast.
The DWD expert explained that AI as a new tool does not make humans superfluous. In fact, more work is currently required in order to continue to reliably provide the current quality of the physically-based systems. "The AI models are not yet able to provide this quality, breadth, variety and reliability, but are only faster or better in selected variables or scores," says Potthast. However, there is a very steep learning curve in this area.
AI does not respect the laws of nature
Weather is created by many interconnected processes, explained the DWD expert. "Physical models, such as those used in weather forecasting, adhere to the laws of nature." This makes their predictions consistent and comprehensible. "Machine learning models work differently. They concentrate on predicting individual values as accurately as possible without directly observing the laws of nature."
Such models often do not distribute the energy that drives the weather - large movements such as winds and small details such as turbulence - in the way that happens in nature. This can result in forecasts that look good at first glance but are not quite right in reality, especially when the weather becomes more complicated. "Physical models do this better because they are designed from the outset to adhere to these relationships."
