Bringing together users, researchers and coders in ECMWF’s machine learning efforts

Jesper Dramsch

ECMWF’s endeavour to create first-class machine learning weather forecasts relies on the expertise of its staff, including Jesper Dramsch, who joined ECMWF in 2021.

Jesper identifies as non-binary and is referred to with ‘they/them’ pronouns.

As early as during secondary school in Hamburg, Germany, they took a rather unusual course in geology. They became interested in geophysics and did an elective A-level in the subject. They then went on to study geophysics and oceanography at the University of Hamburg.

Jesper followed up their studies with a PhD in geophysics at the Technical University of Denmark, which they obtained in 2021. The main project was related to imaging the geology of oil fields in the North Sea off Denmark.

Jesper Dramsch field work

Jesper (left) during a field work trip as part of their PhD in the Norwegian archipelago of Svalbard in the Arctic Ocean.

But this was also a time of Jesper’s emerging interest in machine learning. As early as 2016, they participated in a machine learning competition. In 2019, they gave courses in Python and machine learning to geoscientists, and from 2020 to 2021 they worked as a machine learning engineer in Oxford, UK.

“After my studies, I decided not to develop a career in oil and gas,” Jesper says. “Instead, I was keen to pursue my activities in machine learning in an area of work which I considered a positive influence in society.”

In 2021, they joined ECMWF to work on new machine learning initiatives in weather prediction.

Hybrid machine learning

At first, Jesper worked on various schemes to introduce machine learning techniques into some of ECMWF’s traditional weather forecasting activities.

“At the time most people thought that you can’t do predictions exclusively using machine learning tools,” Jesper says. “So, I was mostly working on hybrid models and post-processing.”

In particular, they helped the team working on sub-seasonal to seasonal (S2S) predictions, who organised an AI challenge with the World Meteorological Organization (WMO) to improve forecasts up to several months using artificial intelligence (AI).

They also looked into observational operators, which mapped relatively sparse soil moisture measurements to a consistent grid, which is the format required for numerical weather prediction systems.

“It took me a while to get to grips with weather forecasting terminology, such as ‘S2S’, ‘MJO’ (Madden–Julian Oscillation), and ‘observational operators’,” Jesper says.

Ensemble forecast post-processing diagram

One of Jesper’s early activities was their involvement in a project to improve ensemble forecast post-processing methods. This system improves each member of the forecast individually to improve the skill of the overall forecast. N stands for the number of ensemble members, C for the number of feature channels (e.g. atmospheric variables), and H and W for height and width of the grid. (See

ECMWF’s machine learning forecasting system

Since early 2023, Jesper has been involved in building ECMWF’s first machine learning forecasting system, called the Artificial Intelligence/Integrated Forecasting System (AIFS).

This relies on traditional methods to establish the initial conditions of forecasts, but it uses machine learning to predict future conditions of the weather.

“The AIFS uses machine learning techniques on 40 years of weather data, conveniently brought together in ECMWF’s ERA5 reanalysis, to learn how weather systems evolve,” Jesper explains.

“It then predicts the weather six hours from now, and this can be done repeatedly so that longer forecasts are achieved.”

Training of the AIFS

This diagram shows the principle of the training of the AIFS. The weather as recorded in ERA5 is sampled, and ‘input’ and ‘output’ separated by six hours is batched together. These batches are used to train the AIFS so that it can make predictions six hours ahead.

The system already achieves better evaluation results than traditional forecasts for many variables, but it has several limitations, including its horizontal resolution, its range of variables, and its current lack of ensemble forecasts. For example, the resolution is currently about 28 km compared to 9 km for ECMWF’s physics-based forecasts, and it does not yet predict the same range of variables as ECMWF’s Integrated Forecasting System (IFS). Nevertheless, these are current areas of development.

Jesper worked on a few additional areas, starting with making the AIFS highly configurable for other researchers and developers. They have also worked on a custom data routing solution that defines which weather variables are used as forcing, prognostic, and diagnostic variables. More recently, Jesper has worked on improving the build system, dependencies, and quality of the code.

“My task was to make the code more robust, and that’s what I’ve done,” Jesper says. “For example, I’ve identified and rewritten all the pieces of code that can exist independently and that can be used in different sections, while keeping researchers, coders and operational use in mind.”

As a result, Jesper is familiar with, and understands, a large part of the code: “About 90%, I’d say, but I’m still learning!”

Screenshot of machine learning forecasts on the ECMWF website

This is a screenshot of AIFS and other machine learning forecasts on the ECMWF website. Jesper has written the plugins used to run most of these models daily.

They are also involved in Anemoi, the framework for developing the AIFS, which is being designed to support training models for many organisations. This is about enabling the weather community to come together to work on data-driven weather forecasts. It includes ECMWF’s work on the AIFS, distributed training on supercomputers, and expanding towards limited-area machine learning forecasts at a higher resolution in ECMWF’s Member and Co-operating States.

“Because of the code adaptation we’ve done, we can easily expand parts of the AIFS and Anemoi to regional systems,” Jesper explains.

Jesper’s work is thus not just very future-oriented, it also goes beyond ECMWF’s own forecasting system to benefit the wider community.

This view beyond is also exemplified by Jesper organising and teaching in the 2023 ECMWF MOOC on machine learning for weather and climate prediction; in-person machine learning training at ECMWF; co-chairing the working group on modelling of the ITU/WMO/UNEP Focus Group on AI for natural disaster management; and public presentations, such as representing ECMWF’s work at the science night in Bonn on 17 May 2024.