Development Roadmap

Authors: Anemoi Technical Subgroup (ATS)

Date: July 2025

This roadmap outlines the future development of the Anemoi project, emphasizing flexibility, adaptability, and high performance to meet the evolving needs of weather and climate modeling. We aim to build a modular, extensible ecosystem — much like a set of well-defined Lego blocks — where components such as model layers, training approaches, and data pipelines can be easily assembled, reused, and adapted to support diverse use cases, from regional to global forecasting. By promoting generalization, interoperability, and ease of integration, Anemoi enables users to experiment with new combinations or build on existing ones. The collaboration of diverse institutions remains a core strength, allowing the project to stay domain-agnostic and responsive to the specific needs of a wide range of stakeholders.

Key Development Themes

The technical direction for this project focuses on extending its usability, modularity, and flexibility across different domains and datasets. We have identified 4 key areas for improvement and expansion.

1. Flexible Model Architectures

   • Current Focus:

     • Focus on graph neural networks.

     • Enhance model architectures to be more modular.

     • Focus on implementing probabilistic and ensemble forecasting techniques (e.g., CRPS Loss, diffusion).

   • Long-Term Vision:

     • Develop a robust framework for model sharing across various ecosystems, supporting different types of models.

     • Ensure that the infrastructure for model development and training can easily scale from local to global forecasting tasks, while maintaining computational efficiency.

2. Data Handling

   • Current Focus:

     • Support for multiple datasets, particularly for high-resolution observational data (e.g., radar, satellite, SYNOP).

     • Implement multiple encoders and decoders.

   • Long-Term Vision:

     • Foster interoperability with other open-source tools, allowing users to easily integrate their datasets and models with external systems.

3. Operational Deployment and MLOps

   • Current Focus:

     • Develop CI/CD pipelines to automate testing, deployment, and operational inference.

     • Focus on optimizing performance through improved memory management and computational efficiency (e.g., Torch compilation, HPC optimizations).

     • Make inference as lightweight and efficient as possible to support real-time and large-scale operational deployment.

   • Long-Term Vision:

     • Strengthen MLOps practices to support full deployment cycles from development to production, ensuring that models can be updated and maintained effectively.

     • Integrate tools for model versioning, dataset tracking, and reproducibility across different use cases.

4. Collaboration and Interoperability

   • Current Focus:

     • Develop shared standards for graph representations, model architectures, datasets, and code contributions to ensure smooth collaboration across contributors.

     • Work towards better documentation and more educational resources to onboard new contributors.

   • Long-Term Vision:

     • Enable seamless collaboration across institutions and domains by developing common interfaces for datasets and models.

     • Encourage community-driven extensions and contributions to keep the ecosystem open and extensible.

For progress and updates on these topics, we refer to the relative Github issues and PRs across all Anemoi packages. An overview is available in the Anemoi Project Board.

Engaging More Users

In addition to the technical development, a key focus will be improving user engagement and onboarding, which includes:

• Developing comprehensive documentation and tutorials for users with various levels of expertise.

• Creating interactive guides and visualizations to explain complex concepts like model training, dataset preparation, and operational inference.

• Hosting community-driven workshops and outreach efforts to encourage broader participation and contributions.

Anemoi Community Meeting

The Anemoi Community Meeting is a biweekly online gathering for Anemoi developers and users.

When: Every other Tuesday from 11:00 to 12:00 CET, starting from July 1st, 2025.

Who: Anyone involved in or interested in the Anemoi ecosystem is welcome to join — including contributors, developers, researchers, and representatives from weather and climate modeling projects (e.g., NMS). Please feel free to share the invitation with colleagues who may benefit from attending.

Why: As the Anemoi framework grows in functionality and usage, the community meeting is a key touchpoint to maintain alignment, foster collaboration, and surface new ideas. It’s also a great opportunity to network and discuss shared challenges or feature needs.

Join: Click here to join the Anemoi Community Meeting on Microsoft Teams =======