Xavier Abellan, Senior Analyst, Forecasts and Services Department
Machine learning (ML) and artificial intelligence (AI) are transforming numerical weather prediction (NWP).
At ECMWF, this shift is embodied by the Artificial Intelligence Forecasting System (AIFS), which became operational in early 2025. The AIFS is based on graph neural networks and transformers to deliver global forecasts. Our ensemble version, AIFS ENS, which was introduced in mid-2025, offers probabilistic forecasts with remarkable speed and efficiency, using up to 1,000 times less energy than traditional physics-based models.
To support the development and integration of such models, ECMWF has also introduced the Anemoi framework – a modular, open-source toolkit designed to streamline the training, evaluation, and deployment of machine learning models for Earth system science. Anemoi enables reproducible workflows, supports hybrid modelling, and is already being co-developed in collaboration with our Member States and external partners.
Such rapid progress in this new field would not be possible without appropriate computing resources. At ECMWF, we make use of our high-performance computing facility (HPCF) and cloud infrastructure not only to develop and run new applications, but also to support collaboration and innovation across our community.
The European Weather Cloud (EWC): a hub for the meteorological community
The European Weather Cloud (EWC), jointly operated by ECMWF and EUMETSAT, is a distributed cloud platform designed for the meteorological community in Europe. It allows users to run workloads very close to ECMWF and EUMETSAT data services in a flexible way, with virtual CPUs, storage and networking, as well as specialised hardware such as GPUs.
Hundreds of users across the national meteorological and hydrological services (NMHSs), as well as researchers in academia within Member and Co-operating States, already benefit from this platform to carry out their official duties, accelerate their research, and collaborate with other partners in the community.
The European Weather Cloud service is available to different groups of users.
AI/ML templates: lowering the barrier
In the past year, we have seen a growing interest and demand for workloads related to AI and ML, especially on the EWC. The landscape is evolving rapidly, and the technical setup required to use many of the popular tools in the field can be daunting even for expert users. Organisations and individuals with limited resources may find it particularly difficult to keep pace.
To democratise the access to these new technologies, at ECMWF we have created a set of ready-made templates that contain pre-configured environments and automation scripts that allow users to quickly set up their working environment for the desired use case. The templates are designed in a way that allows them to be installed on both new and existing computing instances, and can be easily interchanged or combined, helping users focus on science rather than infrastructure.
The templates currently part of this collection are:
- ML Basic: provides a software environment with the basic AI/ML tools in python such as torch, tensorflow, keras, scikit-learn, and others.
- AI Models: sets up a software environment with the AI-models package, which allows users to easily run some third-party data-driven weather forecasting models, such as Pangu-Weather or GraphCast.
- Anemoi: installs a software environment featuring all the Anemoi components. It includes the basic packages such as datasets, training, graphs, models and inference.
- AIFS Single mean squared error (MSE): installs the ECMWF AIFS Single MSE data-driven forecasting system and supporting dependencies.
- AIFS ENS continuous ranked probability score (CRPS): installs the ECMWF AIFS ENS CRPS data-driven forecasting system and supporting dependencies.
These templates use well established technologies, such as Conda, Python environments and Ansible, for automation and reproducibility. Each come with their own instructions and customisation options for advanced users. As new tools and applications emerge, they will be added to the collection.
All of these resources are publicly available in GitHub, and also featured in the EWC Community Hub, encouraging community development and enhancements.
EWC Community Hub
The EWC Community Hub, launched in September 2025, is a centralised platform where EWC users can discover, evaluate, select, and deploy items and services designed for the European Meteorological Community and built to run on the EWC.
The Community Hub brings together a wide range of open-source contributions, designed to accelerate collaboration and make cloud technologies easier to use in meteorological applications.
The platform is made up of three main parts:
- Community Hub items: a growing collection of open-source resources, including Infrastructure as Code (IaC), configuration blueprints, containerised applications, algorithms, datasets, and documentation – all tailored to the EWC.
- Community Hub dashboard: a single interface where users can browse, filter, and select the items that best fit their needs.
- Community Hub tooling and documentation: step-by-step support to use the items effectively, regardless of the technology.
In addition, users can join the EWC discussion platform, hosted on Rocket.Chat to connect with others to collaborate, troubleshoot and share developments.
A screenshot of the EWC Community Hub.
The initial set of resources, including the AI/ML templates, were primarily developed and released by the EWC teams from ECMWF and EUMETSAT, but the goal is a community-driven ecosystem where users can share their own items in the Hub. There are already some examples of Community Items and we look forward to accepting further contributions from our users.
The EWC Command Line Interface (CLI) has also been released, allowing users to interact with EWC services and consume any items in the EWC Community Hub with ease. This tool is also open source, publicly available in GitHub, and easily installable as a Python package with pip.
AI innovation on the Cloud
Beyond the core templates and operational models, the EWC is home to a growing range of AI innovation. One of the most promising developments is Forecast-in-a-Box (FiaB) – a portable, containerised solution that packages AI models and their inference environments into a single deployable unit. Designed to be platform-agnostic, FiaB can run on local machines, clusters, or directly within the EWC, making it ideal for rapid prototyping and operational use.
Screenshot of Forecast-in-a-Box (FiaB).
FiaB dramatically reduces the time and cost of generating forecasts. In recent demonstrations, AI-based forecasts using FiaB were produced in under three minutes, compared to over 30 minutes for traditional physics-based models. This efficiency opens the door for use in resource-constrained environments, such as early warning systems in developing countries. For example, a World Meteorological Organization (WMO)-supported pilot project in Malawi is exploring FiaB to strengthen local forecasting capabilities with minimal infrastructure requirements.
AI is no longer experimental
With AIFS and other ML models now complementing traditional NWP, and with the EWC providing infrastructure, templates, and community support, AI in weather prediction has entered a new era. By lowering technical barriers, and encouraging collaboration, ECMWF is enabling users to explore, develop and deploy AI-driven solutions for better forecasting.