ECMWF Newsletter #174

ECMWF–ESA Workshop on Machine Learning for Earth Observation and Prediction

Massimo Bonavita (ECMWF)
Rochelle Schneider (ESA ESRIN Φ-lab)


The third edition of the ECMWF–ESA Workshop on Machine Learning for Earth Observation and Prediction took place from 14 to 17 November 2022 at ECMWF’s headquarters in Reading, UK ( While the first two editions of the workshop were held online due to Covid restrictions, this one ran in a hybrid format, with an in-person component (about 120 people) and a large and active online participation (about 700 registered participants). These attendance numbers, together with a record number of 121 abstract submissions, confirm both the large interest in machine learning (ML) in the Earth system sciences and the fact that the ECMWF–ESA workshops have established themselves as a reference meeting and discussion venue in this area.

One of the aims of the event was to provide an up-to-date snapshot of the state of the art in this rapidly evolving field. The two invited talks by Prof. Stephen Penny (Sofar Ocean Technologies) and Prof. Damien Borth (University of St. Gallen) set the stage, with overview presentations of the state of the art, current challenges, and opportunities for adopting AI/ML solutions in data assimilation and Earth observation. From these talks, it was apparent that increasingly sophisticated ML techniques have further spread into research and operational practice in the Earth sciences and, more importantly, they are being tailored to this specific domain with compelling results.

On-site participants
On-site participants. The event was a hybrid on-site and online event, with about 120 people attending in person.

Thematic focus

The workshop was structured according to separate themes designed to cover the main application areas of ML in Earth observation, numerical weather prediction (NWP), and climate prediction:

  • Machine learning for Earth observations
  • Hybrid machine learning in data assimilation
  • Machine learning for model emulation and model discovery
  • Machine learning for user-oriented Earth science applications
  • Machine learning at the network edge and high-performance computing

In this edition, the last theme was added to the first four traditional areas. The aim was to encourage discussion of emerging topical areas of ML applications, such as observation processing on board of satellites (see the talk by Vit Ruzicka and poster presentations by Giacomo Acciarini and Andrea Spichtinger) and futuristic applications of AI/ML in quantum computing. In this latter area, the talks by Lisa Wörner and Bertrand Le Saux provided a fascinating snapshot of current and planned developments for the application of these techniques to Earth observation.

Another topic at the heart of numerous presentations and discussions is the possibility that in the not-too-distant future ML tools will completely supersede foundational NWP activities, such as data assimilation, model development and model emulation. Unsurprisingly, views were varied. While most participants felt that a more likely development path would involve ML components being introduced in specific parts of the NWP value chain, others advocated a bolder approach. This would involve substituting entire NWP activities with potentially faster and cheaper ML counterparts (see for example the talks by Sid Boukabara and Stephen Rasp). What is not controversial is the fact that things appear destined to move fast as big commercial players, such as NVIDIA and Google, enter the field of ML model emulators.

Working groups

Working group discussions for different thematic areas were organised to help participants explore the main ideas emerging from the in-person and poster presentations. They also served to report on the main current and predicted trends in each of the areas. The working groups were held in a hybrid format, with both in‑person and online attendees, which allowed a broad and very diverse participation. Preliminary findings from the working groups are available on the workshop website (, while a more detailed report is in preparation.


We are very encouraged by the large and unabated interest in this series of workshops, the excellent level of the presentations and discussions, and the very positive feedback received from participants. This confirms that the field of ML in the Earth sciences is still on an upward trajectory. The ECMWF–ESA workshops have a significant role to play in its development and in building a strong community of developers and practitioners.

Stay tuned for the next edition!