The long summer of AI
The artificial intelligence boom has only just started. With its unique computer infrastructure and technical expertise, Jülich can make an important contribution.
Intelligent weather forecast
Artificial intelligence has found its way into weather and climate research as well, says Priv.-Doz. Martin Schultz from the Institute for Advanced Simulation (IAS). Just five years ago, meteorologists were sceptical about AI models, but the algorithms have now proven their efficiency.
What is special about AI models for weather forecasting?
Classic forecasting models are based on systems of equations that reflect the physics of weather events. They are so complex that they can only be solved with great computational effort, as they consist of a large number of individual equations that have to be calculated as closely as possible in spatial terms.
AI models, on the other hand, learn from the past. They are trained using historical weather data that goes back four decades. The AI models “memorize” what happened during this period and can use this knowledge to make predictions for the current weather conditions.
What are the advantages of AI algorithms?
Firstly, they in effect provide better predictions in certain areas than the classic models. Secondly, they are particularly efficient. Once an AI model has been trained, the costs for computing power go down significantly – they are around a hundred to a thousand times lower. This would allow forecasts to be updated much more frequently, for example every hour.
What restrictions apply to the new models?
The training is very time-consuming. The models only really pay off if you can manage with just a few training runs and can use the models for a long time. There are also areas in which AI algorithms are inferior to traditional models. A typhoon may be forecast earlier using conventional methods, while a cold snap may be predicted better by artificial intelligence. We are currently still gaining experience as to which model is more suitable in which case.
Why did many meteorologists have reservations about AI models for a long time?
AI predictions were not convincing until five years ago, as the state of the art at that time was simply not sufficient to generate reliable forecasts. The models were much smaller. They were not able to depict the complexity of the atmosphere. But this has now changed and the AI models have become considerably larger. The latest generation of AI models delivers results that are close to or even better than conventional predictions.
Cover story AI
How can the predictive power of a model be assessed?
There are extensive methods for the evaluation of weather models that can also be used to assess AI models. This makes it possible to quantify how well the prediction patterns correspond to reality. The robustness of the models is also important, however, that is, how well they can predict situations for which there is no training data. There are first indications that large AI models have actually developed something like an understanding of the physics behind the weather based on the training data – without explicitly knowing these laws of nature. That gives us hope.
Are such models also suitable for long-term climate forecasts?
We are still at the beginning of research in this area. This is because simulations in climate research depend heavily on the boundary conditions, such as on CO2 concentrations and ice sheets, for example. In other words, on variables that change over very long time scales. However, there is not enough data available on these variables. Hybrid models are therefore a current approach: at their core, they are classic physical models that work on a rather coarse scale. Only the small-scale processes, such as cloud formation, are calculated by an AI.
Interview by Arndt Reuning
Illustration (created with the help of artificial intelligence): SeitenPlan with Stable Diffusion and Adobe Firefly
contact
Further Information
Last Modified: 29.02.2024