AI CONTEXTS with Dr. Jan Keller (Images: DWD/Private)

“Fortunately, weather models are occasionally inaccurate” – How the German Weather Service uses artificial intelligence

16 October 2025, 10:37 a.m. | Reading time: 9 min

By the KIWIT editorial team

“The weather forecast is never right!” – Not true, says Jan Keller from the German Weather Service (DWD). The meteorologist heads the Data Assimilation division at Germany’s renowned weather authority. He works on AI applications that can optimize weather forecasting. We spoke with Jan Keller about how these are combined with classical forecasting models, the opportunities artificial intelligence offers for meteorology – and why natural intelligence remains important.

This is Episode 1 of our new special series AI CONTEXTS.

KIWIT: Dr. Keller, let’s start with a suspicion that you will probably confirm: you become a meteorologist if you were already fascinated by the weather as a child. How did you get into weather?

Keller: In fact, I too was fascinated by the weather as a child – especially thunderstorms. I wanted to understand them better. So around the 10th grade I decided to become a meteorologist. By the way, not all scientists at the DWD are meteorologists. We also have a large number of physicists and mathematicians on our team.

A common, general everyday observation: many people think that weather forecasts are often not accurate at all.

We continuously evaluate the quality of our forecasts for various parameters. And these evaluations show that weather forecasts are constantly improving – although not equally for all parameters. My assumption is that people get used to better forecasts, especially since the improvements are rather gradual and steady, not sudden. For example, a forecast for the next day about 15 years ago was about as good as a forecast three days ahead is today. But of course, forecasts are sometimes wrong. People may notice that more readily now that forecasts have generally become better.

National Weather Universalist – internationally respected and in demand: the DWD in Offenbach. According to the DWD Act, the DWD is a scientific institution providing comprehensive meteorological services for the state and the public. Image: DWD

Also, give it a listen: our new podcast KInote

Follow and subscribe to us!

Share this article

As in numerous other domains, the significance of artificial intelligence (AI) is increasingly being acknowledged within meteorology. What role does AI currently play in weather research and forecasting?

When considering AI in its original sense as statistical learning methods, it is evident that such approaches have been employed in meteorological research for several decades. What is novel, however, is the utilization of statistical models based on neural networks, trained on extensive datasets. Today, this is generally what is understood by the term “AI.”

Given the vast repositories of observational data that meteorologists have accumulated over time, these methods can be employed with considerable efficacy, as neural networks improve in performance proportionally to the quantity of data available for training. Consequently, AI now occupies a prominent role in weather research. Nonetheless, limitations are already encountered, particularly when training comprehensive AI-based weather models, as the requisite hardware—primarily graphics processing units (GPUs)—is not inherently optimized for applications with substantial memory demands.

In the current year, the DWD established a dedicated AI Center. What is the purpose of this initiative, and what objectives does it pursue?

The application of AI at the DWD extends beyond numerical weather prediction to encompass a wide range of meteorological and climatological applications. Furthermore, these AI initiatives are not restricted to meteorological data alone; efforts are also underway to develop proprietary language models. To advance these initiatives in a structured manner and to maximize potential synergies, the DWD AI Center was established. Within this framework, AI developers collaborate with personnel from specialized divisions—including weather forecasting, climate monitoring, and administrative units—to develop and refine AI methodologies.

Image Gallery | Insights into the diverse work conducted at the DWD (navigate by clicking or sliding).

Alle weltweiten Messungen und Beobachtungen sowie die Ergebnisse der Computervorhersagen laufen in der DWD-Zentrale in Offenbach zusammen. Bild: DWD

Wettermeldungen, Computervorhersagen für die nächsten zehn Tage und viel persönliche Erfahrung sind Grundlagen amtlicher Wetterwarnungen des DWD. Bild: DWD

Wind und Wetter unserer Welt: Eine Satellitenaufnahme bildet die Pressewand des DWD. Bild: DWD

Alle weltweiten Messungen und Beobachtungen sowie die Ergebnisse der Computervorhersagen laufen in der DWD-Zentrale in Offenbach zusammen. Bild: DWD

1/12

On weather websites, you increasingly read about “AI weather models.” What does that mean – and how do they differ from classical numerical models?

First, let’s look at the classical – or more precisely, physics-based – models: the fundamental physical laws and the temporal evolution of the atmosphere are implemented in computer programs as partial differential equations. These equations are then solved for each time step – essentially computing a “mini” forecast. The spatial resolution (that is, how finely you want to predict the processes) determines the length of each time step – the finer the resolution, the shorter the step. As a result, a 24-hour forecast requires hundreds or even thousands of individual steps.

AI-based models, on the other hand, are statistical models – or more precisely, trained neural networks. During training, these models learn to reproduce a certain final state of the atmosphere as accurately as possible from a given initial state. In this sense, an AI model “knows” the true target state of the atmosphere during training. However, there are no restrictions regarding the time step; current global AI weather models require fewer than ten steps for a 24-hour forecast. In addition to this significant speed gain, these models also show slightly better forecast accuracy.

AI weather models learn from past weather developments. How do classical models like DWD’s ICON, the European ECMWF, or the US GFS use past data? What is the difference?

Physics-based models only take past weather into account in the sense that all improvements are tested against historical periods and then evaluated against the “truth.” In this way, the past is considered indirectly – only by the developers themselves.

Worth Knowing

Weather models are complex computational programs based on physical laws that simulate the behavior of the atmosphere. They process vast quantities of observational data—such as temperature, air pressure, and wind—and use these inputs to calculate future weather developments. Artificial intelligence (AI) is increasingly being employed to enhance these models: it assists in filling data gaps, refining forecasts, and more rapidly identifying patterns within large climate datasets.

Intuitively, one might assume that conventional and AI-based models complement each other, thereby improving forecast accuracy. Is this the case, and what are the limitations?

There are several approaches to combining classical and AI-based methods. One approach is the use of AI models for ensemble forecasting. In this method, numerous model simulations are conducted with slightly different initial conditions to estimate forecast uncertainties. In classical models, the number of realizations is limited to fewer than 50 due to computational constraints. AI models, however, could enable much larger ensembles, comprising hundreds or even thousands of realizations.

Another approach is spectral nudging, which is based on the observation that AI models generally provide more accurate forecasts at larger scales. To leverage this accuracy within physics-based models, the physical model is continuously nudged toward the AI forecast primarily in the middle and upper troposphere. Near the surface, where smaller-scale processes tend to dominate, the accuracy of the physics-based model is preserved.

Furthermore, AI methods can be directly integrated into physics-based models. In this approach, specific processes that are poorly represented—or only representable at very high computational cost—within the physical model are replaced by trained AI components. However, this requires substantial development effort.

Ensemble forecasting: Does the generation of such large ensembles make it more difficult to approximate the most probable outcome?

Both yes and no. Large ensembles—comprising hundreds or thousands of realizations—allow for a more precise estimation of the uncertainty distribution of forecasts. However, it is true that such extensive datasets are more challenging to manage and analyze, though this represents a technical rather than a conceptual limitation.

Image Gallery | AI in Autumn (navigate by clicking or sliding)

Das mögen viele: Altweibersommer und Goldenen Oktober. Den gibt es dieses Jahr verbreitet eher sparsam. Mit KI-Anwendungen lässt sich inzwischen der Verlauf der Laubfärbung prognostizieren. Bild: KIWIT

Die interaktive Karte von Explore Fall prognostiziert mit KI, wie sich die Laubfärbung in Nordamerika für die nächsten Tage entfaltet. Bild: https://www.explorefall.com

Garantiert KI: Eine Lösung von ChatGPT 5 für den Promt: „Hamburg im Goldenen Oktober“. Gut zu erkennen ist die Übergeneralisierung. Bild: Gemeinfrei

1/4

What role will meteorologists themselves play in the future? Will there still be room for “human judgment”—for experience, intuition, and the final refinement of forecasts, for example in relation to the specific characteristics of mountains and valleys, coasts, and inland areas?

At the DWD, numerical weather forecasts—that is, computer models—are calculated and made available on a regular basis. However, the forecast products themselves are the result of a process chain that includes meteorologists as an essential pillar. The forecasts and warnings published by the DWD are therefore created or verified by humans. The term “human judgment” aptly describes this role.

A particularly challenging aspect remains cloud cover: in certain situations, short-term forecasts often struggle. Sometimes the model predicts “10 hours of sunshine,” yet the sky remains overcast; at other times, it predicts “completely cloudy,” and by midday gray turns to blue. Why is sunshine sometimes so difficult to predict, and could AI help optimize this?

The underlying processes are relatively complex and, therefore, not always accurately predicted even by modern weather models. In general, machine learning can only provide limited assistance, but there are several proposed AI-based methods that would also incorporate observational data—satellite data, in this case.

Can AI also extend the forecast horizon for weather? Classical models reach physical limits after a few days. Can AI push these limits further?

This remains an open question in the scientific community. Some experts believe that the current forecast horizon of approximately two weeks will remain largely unchanged even with AI models. Others suggest that it might be extended, as these models operate differently. Nevertheless, realistic forecasts for several months in advance remain unattainable based on current knowledge, even with AI.

Video | How is a weather forecast created? (DWD)

The DWD communicates in a factual and restrained manner, whereas other providers often use sensational headlines such as “heatwave horror” or “snow avalanche.” Is sober communication part of the DWD’s self-conception, particularly as a scientific institution?

At the DWD, we do not perceive ourselves as a media organization aiming to attract attention, but rather as a governmental specialized authority with a scientific research mandate. Our responsibility is to provide objective and professional information on weather, climate, and their potential impacts, to the best of our knowledge and judgment.

Finally, a brief look at the current weather is in order: the so-called “Golden October” remains, in many regions, literally overcast—high-pressure systems are unfavorably positioned, and “North Sea clouds” are drifting inland. Are there still opportunities for a few bright autumn days outside the higher elevations?

Regrettably, the forecast for the next two weeks suggests a continuation of a gray October with only a few brief periods of sunlight—but fortunately, weather models are sometimes inaccurate.

We thank you for the insightful perspectives on the work of the DWD and the use of AI in meteorological research.

Weather Researcher and AI Developer

Dr. Jan Keller holds a diploma and a doctorate in meteorology. At the German Weather Service (DWD), he leads the Data Assimilation division. Previously, he headed a research group and was also a visiting researcher at the National Center for Atmospheric Research in Colorado, USA. He studied meteorology at the University of Cologne and completed his doctoral thesis at the University of Bonn.

In the field of data assimilation, Keller focuses on integrating new observational data into weather and climate models. Together with his DWD colleague Roland Potthast, Keller published an important study in 2024: for the first time, the researchers succeeded in integrating weather observation data into forecasting models and analysis processes solely through the use of AI.

The series AI CONTEXTS presents applications and developments of artificial intelligence across various scientific disciplines. This bimonthly series covers interdisciplinary research and applied fields related to public services and infrastructure—from the national weather service to energy, mobility, and education.

Responsible for AI CONTEXTS: Prof. Dr. Marcel Schütz (Northern Business School Hamburg)