...
To support this, the pilot has worked with technologies developed by ECMWF, including FDB, Polytope, Earthkitearthkit and Aviso, as well as resources provided through the European Weather Cloud (EWC). Rather than focusing only on the technologies themselves, the project develops practical workflows and open-source blueprints to demonstrate how they can be combined and adapted for real applications. The aim is to provide experience, examples, and feedback that can benefit both ECMWF and the wider Member State community.
...
For the extension year, the work is organised around three technical work streams defined in the Statement of Work for the project extension:
- WS1– Earthkit earthkit–Anemoi alignment: assess and reduce duplicated data-handling functionality and support the progressive use of Earthkitearthkit as a common data layer in Anemoi.
...
- WS3– Native unstructured-grid support:validate and extend Earthkitearthkit and Polytope capabilities for loading, retrieving, regridding and processing native ICON data, supported by reusable example notebooks.
...
- Polytope/FDB: Tiago Quintino, James Hawkes, Emanuele Danovaro, Peter Tsrunchev, Mathilde Leuridan and Christopher Bradley
- Earthkitearthkit: Sandor Kertesz, Pedro Maciel, Iain Russel and James Varndell
...
Extension work stream and planned output | Objectives addressed | Progress and achievements to date (August 2026) |
WS1 – Integration of Earthkit earthkit in the Anemoi framework(D1.1 Migration to Earthkitearthkit; D1.2 Missing or improved Earthkit earthkit functionality)
| O3 · O4
| The project established an exchange with the Anemoi machine-learning community, with the goal to assess overlaps between Anemoi and Earthkitearthkit in the handling of meteorological data. Some functionality currently implemented or wrapped within Anemoi can increasingly be provided by Earthkit earthkit as a common data-handling layer, allowing Anemoi to focus more clearly on its machine-learning workflows. Following the stable Earthkitearthkit 1.0 release at the end of June 2026, existing migration work within Anemoi is updating the packages to the new interfaces and reducing duplicated data-handling code. Testing of the migrated interfaces and identification of remaining gaps continue during Q3. |
WS2 – Publishing the FLEXPART workflow as an EWC blueprint (D2.1 Deployable blueprint; D2.2 Documentation and user guidance) | O1 · O2 · O4 | The FLEXPARTatmospheric dispersion modelhas been integrated into a documented, deployable blueprint, and published on the EWC Community Hub, where EWC users can discover and reuse applications. The blueprint automates the full FLEXPART chain, from retrieval and preprocessing of ECMWF forecast data to model execution and visualisation, and can be deployed using EWCCLI, the EWC command-line tool. Beyond the FLEXPART use case, it provides a practical example of how Member States can turn technical workflows into reproducible, shareable EWC resources. |
WS3 – Support for native unstructured-grid data and meteorological processing in the Earthkit earthkit ecosystem(D3.1 Testing, gaps and feedback; D3.2 Demonstration notebooks) | O1 · O2 · O3 | The project has advanced and demonstrated support for unstructured-grid data across ECMWF’s data-access and processing tools. Using the operational ICON model configurations at MeteoSwiss as practical test cases, notebooks demonstrate how Earthkitearthkit can load, regrid and visualise these data, while Polytope enables selective retrieval of areas, time series, trajectories and profiles. The resulting reusable workflows and examples demonstrate how ECMWF tools can be applied to native unstructured-grid data and provide a basis for evaluating similar approaches for other Member State datasets.Further work is ongoing during Q3 and Q4 to consolidate these workflows and extend meteorological processing capabilities in earthkit-meteo. |
Cross-cutting – Community dissemination and reuse | O3 · O4 | The extension year has increased the emphasis on making project outputs openly accessible and reusable beyond their original implementation. Public notebooksdemonstrate complementary ways of working with NWP data: open-data access and processing with Earthkitearthkit, and data access and processing with FDB, Polytope and Earthkitearthkit. Together, the examples show how NWP data can be retrieved, processed and visualised using the technologies explored by the project. The EWC blueprint is likewise accompanied by source code, deploymentdocumentation and usage guidance.The experience of moving from a technical prototype to a reusable Community Hub resource will also be presented at the EWC UserWorkshop on 22 September 2026, providing a practical example for other users interested in publishing and sharing their own EWC workflows. |
...
3.2 Results and activities
WS1– Alignment of Earthkitearthkit and Anemoi data-processing components
Anemoi, ECMWF’s machine-learning framework for developing and running data-driven weather models, and Earthkit earthkit evolved partly in parallel, leading to some overlapping and duplicated meteorological data-handling functionality. Maintaining similar capabilities in both ecosystems increases complexity and maintenance effort. The aim of this workstream is therefore to use Earthkit earthkit as the common data-handling layer, allowing Anemoi to focus on its machine-learning-specific functionality.
The project reviewed the relevant Anemoi components and identified the clearest duplication in anemoi-datasets, where custom xarray handling covered conversion to field objects, coordinate and metadata handling, time handling and variable selection. These functions can now increasingly rely on the common EarthkitearthkitField/FieldList abstractions instead of being maintained separately.
Following the stable Earthkit earthkit 1.0 release, migration work is now updating anemoi-datasets, anemoi-transform and anemoi-inference to the new interfaces and reducing duplicated data-handling code.
The next step is totest the migrated interfaces with representative Anemoi workflows, gather community feedback and identify any remaining gaps that should be addressed in Earthkitearthkit.
WS2 – Reusable FLEXPART blueprint on the European Weather Cloud
...
As illustrated in Figure 1, the blueprint reduces deployment of the complete workflow to only a few user commands. From a local computer, a user with access to an EWC tenancy can deploy a pre-configured virtual machine and then start a dispersion simulation by selecting, for example, the forecast cycle, simulation length and release location. The workflow then automatically retrieves the required IFS forecast data through ECMWF MARS, preprocesses it usingflexprep, developed within the project as an Earthkitearthkit-based reimplementation of the original FLEXPART preprocessing, runs FLEXPART-IFS in a container and generates ready-to-view PDF plots.
...
During the extension year, the project has enabled native unstructured-grid data to be used across key parts of the Earthkitearthkit workflow. Using operational ICON data from MeteoSwiss as a practical test case, Earthkitearthkit can now load native-grid fields together with the latitude and longitude of each grid point, convert them into the widely used xarray format, regrid them to regular latitude–longitude grids and visualise the results. This closes an important gap that previously limited the use of unstructured-grid data with common Earthkitearthkit processing tools.
The work was developed jointly with ECMWF through biweekly technical exchanges, using operational MeteoSwiss data for testing and feedback. The resulting workflow has been consolidated intopublic example notebooks, demonstrating the sequence from native-grid loading and xarray conversion to regridding and plotting.
...
On the data-access side, the project also worked with Polytope, ECMWF’s service for extracting selected subsets of forecast data. Polytope’s feature-extraction algorithms can now operate directly on the native ICON unstructured grid, which was not previously possible. This allows users to select only the data required for geographical areas, time series, trajectories or vertical profiles, without first regridding the data. Together with the Earthkit earthkit developments, this enables selective access to and processing of data while retaining the native unstructured grid.
...
Public example notebooks
Earthkit earthkit workflow for unstructured-grid data
View the Earthkit earthkit notebook →
Polytope feature extraction on native unstructured-grid data
View the Polytope notebooks →
...
Further work will focus on consolidating these workflows and extending meteorological processing capabilities in earthkitearthkit-meteo.
3.3 Challenges
A key challenge is involving more Member States in concrete use cases.The project continues to seek wider engagement through community events and has also been raised in discussions at ECMWF TAC/PAC meetings, but additional Member State use cases remain limited. Broader participation would help assess how the technologies explored in the project can be applied to different Member State use cases and operational environments.
The later-than-planned release of Earthkit earthkit 1.0, originally expected by the end of Q1 2026, has also shifted parts of the Earthkit–Anemoi earthkit–Anemoi alignment and testing into Q3. This reduces the time available for testing the migrated interfaces and consolidating the remaining work during the extension year.
...