CDT Researchers at ECMWF’s Code for Earth
In partnership with
Geospatial CDT Students Develop NLP Chatbot for Weather Data at ECMWF’s Code for Earth
Last week, I had the opportunity to present an exciting project at the European Centre for Medium-Range Weather Forecasts (ECMWF) in Reading. Along with my fellow PhD candidates, Kaitlyn Reiss and Nadia Skifa, from the Centre for Doctoral Training (CDT) in Geospatial Systems, we worked on developing an innovative large language model (LLM) based project as part of the Code for Earth initiative.
What is Code for Earth?
The Code for Earth initiative, a collaborative program led by ECMWF, brings together individuals from diverse backgrounds to create open-source software solutions addressing environmental challenges. It’s supported by Copernicus, Destination Earth, IFAB, and Helmholtz-Zentrum Hereon.
Each year since 2018, participants from earth sciences, computer science, and software development gather to collaborate on projects that advance environmental research.
Challenge 21 – Polly: An NLP Interface for Weather Datacubes
Under Stream 2: Machine Learning for Earth Sciences applications, our team tackled Challenge 21 – Polly: An NLP Interface for Weather Datacubes. The goal was to develop a machine learning “chat” interface that allows users to retrieve custom queries from the Destination Earth Weather Extremes digital twin using Polytope.
Polytope is a new data access mechanism implemented by ECMWF. It enables users to extract precise and complex shapes of data—such as 2D country outlines or 4D flight paths—from numerical weather prediction datacubes, rather than downloading entire datasets.
The project was guided by ECMWF mentors Mathilde Leuridan and Adam Warde, whose expertise was instrumental in integrating machine learning with existing data extraction tools. Their support helped the team navigate technical challenges and refine the solution.
This project also deepened our understanding of the challenges involved in moving from proof-of-concept LLMs projects to production-ready systems. We learned a great deal about the complexities of fine-tuning LLMs, especially for specialised tasks like weather data queries. Developing agent-based workflows and refining prompt engineering were crucial parts of the process as was the evaluation and use of tools like Ollama and Open Web UI.
I would like to extend the team’s gratitude to ECMWF for hosting the event and to partners like Copernicus, Destination Earth, IFAB, and Helmholtz-Zentrum Hereon for their support. The Code for Earth initiative provided an excellent platform for collaboration.
The full recording of the presentations is available on ECMWF’s YouTube channel.
For more insights into our work or to connect directly with us, you can view our profiles, along with those of our fellow PhD researchers, here on the Geospatial CDT website.