The third Workshop on Machine Learning for the Earth System, organised by the Center for Earth System Observation and Computational analysis (CESOC), brought together leading scientists from around the world, who are advancing the role of ML in weather and climate prediction. ECMWF and its partners in the EU Destination Earth initiative presented cutting-edge work showcasing how data-driven methods are moving beyond the atmosphere to encompass the ocean, sea ice, waves, land surface, and hydrology, and beyond medium-range forecasting to encompass sub-seasonal, to multi-annual timescales.
Kicking Off with the AIFS
Mario Santa Cruz opened the ECMWF contributions with an overview of the Artificial Intelligence Forecasting System (AIFS). His presentation introduced the current state of AIFS development, including the first operational version, AIFS Single 1.0, released earlier this year, and the probabilistic AIFS ENS 1.0, implemented on 1 July 2025. The talk helped set the stage for the workshop by highlighting how ML-based weather forecasting has moved from research to operations at ECMWF, and how this serves as a foundation for expanding data-driven methods to other Earth system components in the framework of DestinE.
Extending ML Success Beyond the Atmosphere
Scientists Rachel Furner and Nina Raoult opened the session on “Earth system components” with their extended talk “Beyond the atmosphere: Building a data-driven Earth System Model at ECMWF”. They showcased the achievements of the DestinE Earth System ML-based components work at ECMWF, reviewing how ML models are being developed to extend data-driven forecasting across the entire Earth system. The AI activities will enhance the capabilities of DestinE’s Digital Twins, complementing the high-resolution physics-based models.
While ML has already demonstrated impressive skill in atmospheric prediction through the AIFS, applying these methods to the ocean and land surface presents unique scientific and technical challenges. These include differences in data quality and quantity, the need to represent processes across different temporal and spatial scales, and —in the case of the land surface — less well-understood physics. This talk highlighted some of the tailored approaches implemented at ECMWF to deal with these technical challenges, such as tendency scaling for slowly evolving variables, handling missing values, and applying bound constraints to ensure physically consistent predictions.
One of the key takeaways from the talk was that promising ML-based prototypes now exist for ocean, sea ice, waves, land and hydrology. In fact, the workshop coincided with the release of two blog posts about the advances in ML-driven wave and hydrology modelling.
Towards Coupled ML Earth System Models
The talk also addressed how to bring these components together into a unified framework. Strategies range from standard physical coupling—where atmosphere, ocean, and land models exchange information at well-defined classical boundaries—to fully end-to-end ML systems trained with a shared loss function. Each approach has implications for physical consistency and forecast skill, and research is ongoing to compare the two approaches.
AIFL: an ML-driven hydrology model
ECMWF hydrology expert Maria Luisa Taccari further presented the hydrology component in her poster “Towards a data-driven flood forecasting system”. The system, known as AIFL (AI for Flood Forecasting), is a global deep learning model developed at ECMWF to predict daily streamflow at the catchment scale quickly and reliably. This work demonstrates how ML is being applied not only to atmospheric prediction but also to forecasting the potentially devastating impacts of floods. Find more information in our blog post about AI for hydrology.
Moving to sub-seasonal timescales
Machine learning specialist at ECMWF Jakob Schloer presented the talk “A data-driven model for sub-seasonal forecasts”, which adapts the AIFS-CRPS framework to cover the critical weeks between medium-range and seasonal prediction. By using longer time steps, the model shows improved skill and stability at sub-seasonal scales. He also introduced the AI Weather Quest competition, designed to encourage community-wide comparison of data-driven models for sub-seasonal forecasting in a real-time forecasting setting. ECMWF is participating with three different versions of the AIFS in the challenge with their forecasts being now publicly available.
Contributions from other DestinE partners
The wider DestinE efforts were also strongly represented at the workshop. Sophie Buurman (KNMI) highlighted “Multi-domain dynamical graph training of Bris for high-resolution on-demand forecasts”. This work demonstrated how graph-based ML architectures such as those used in MetNorway’s Bris model can deliver scalable, flexible forecasts across multiple domains.
Fernando Iglesias-Suarez (Predictia Intelligent Data Solutions) presented “Harnessing Machine Learning for Climate Modelling: Towards the Development of a DestinE Climate Emulator”. His poster showcased progress in building ML-based emulators that can accelerate long-term climate simulations.
Together, these contributions underline how the DestinE initiative unites expertise from across Europe to develop innovative ML tools spanning both weather and climate scales.
Showcasing ECMWF’s Broader ML Efforts
In addition to the DestinE funded ML activities, the workshop featured a broader range of ML initiatives underway at ECMWF. These included the Horizon Europe Weather Generator project that develops a foundation model for weather and climate closely linked to the DestinE efforts (Christian Lessig), Graph-DOP (end-to-end data-driven system learnt purely on observations developed by ECMWF, Peter Lean and Ewan Pinnington), verification of machine learning models (Massimo Bonavita), and data-driven sub-seasonal forecasts with a representation of the stratosphere (Steffen Tietsche). These talks illustrated the diversity of ML approaches being explored at ECMWF, highlighting how ML is becoming embedded across ECMWF’s modelling chain.
A great success
The workshop organised by CESOC with contributions of ECMWF scientists proved to be a great success, not only for the scientific advances presented but also the spirit of collaboration it fostered. The event showed the dynamism of machine learning research across the world (from Canada to South Korea), bringing together diverse expertise from operational centres, research institutes, and industry. The breadth of topics, from numerical weather prediction to the role of observations and climate, underscores how innovation and the exchange of knowledge is accelerating advance in data-driven Earth system modelling.
Hackathon on ML for the Earth System
The workshop was followed by a hands-on hackathon, where participants worked intensively to tackle real-world challenges in data-driven Earth system modelling. ECMWF scientists played a central role in shaping the event by setting up challenges and sharing expertise. Topics included investigating long-term stability of ML-based forecasts (co-led by Lorenzo Zampieri), fine-tuning models on observations (co-led by Mario Santa Cruz), and verification of ML models (co-led by Joffrey Dumond le Brazidec). The hackathon provided a fast-paced environment for collaboration and innovation, fostering exchange across the community and driving forward practical solutions. During the event, Harrison Cook presented “forecast in a box”, a modular AI forecasting concept being developed at ECMWF in the framework of DestinE.
Shaping the Future of Digital Twins
The advances presented at the workshop mark an important step towards fully data-driven Earth system models, capable of strengthening both short-term forecasts and long-term climate projections. By combining expertise across domains and driving innovation in model design and training, machine learning is becoming central to ECMWF and well as to its partners working on the Digital Twins of the Earth system within Destination Earth.
The event underlined how technical progress and scientific creativity are coming together to expand the role of ML in weather and climate prediction—laying the groundwork for more flexible, scalable, and impactful forecasting tools.
Destination Earth is a European Union funded initiative launched in 2022, with the aim to build a digital replica of the Earth system by 2030. The initiative is being jointly implemented by three entrusted entities: the European Centre for Medium-Range Weather Forecasts (ECMWF) responsible for the creation of the first two ‘digital twins’ and the ‘Digital Twin Engine’, the European Space Agency (ESA) responsible for building the ‘Core Service Platform’, and the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT), responsible for the creation of the ‘Data Lake’.
We acknowledge the EuroHPC Joint Undertaking for awarding this project strategic access to the EuroHPC supercomputers LUMI, hosted by CSC (Finland) and the LUMI consortium, Marenostrum5, hosted by BSC (Spain) Leonardo, hosted by Cineca (Italy) and MeluXina, hosted by LuxProvide (Luxembourg) through a EuroHPC Special Access call.
More information about Destination Earth is on the Destination Earth website and the EU Commission website.
For more information about ECMWF’s role visit ecmwf.int/DestinE
For any questions related to the role of ECMWF in Destination Earth, please use the following email links:
