For most of its history, understanding what the planet is about to do has meant running enormous physics-based models, equations describing atmosphere, ocean, land and ice, computed step by step on some of Europe's largest supercomputers. That approach still sits at the core of TerraDT, the Horizon Europe funded project that is building the digital twin of the Earth system for the cryosphere, land surface and related interactions. Machine learning is however starting to be adopted widely. Machine learning is a technique that lets a computer model learn, from past examples, to recognise patterns and make predictions on its own. In TerraDT, this means models that have learned from years of climate data can then approximate the outcome of a full physics-based simulation.
On 17 September, the project held its first Tech Talk, a webinar format being launched now that it is roughly 40 per cent of the way through its four-year timeline and mature enough, as the moderator Maria Giuffrida (Trust-IT) put it in her introduction, to start discussing the technical models and methods that underpin the project with a broader audience.
More than 80 people from 21 countries joined the event, five speakers took the floor, each described one way machine learning is being put to work.
1. Speed, without losing realism
Devaraju Narayanappa (CSC), TerraDT's technical coordinator, opened with the idea that ran through the whole session: "physics-based models provide realism, and our machine learning provides the speed and flexibility." Sea ice modelling alone can take up a fifth of total computing time, so TerraDT is training AI models to reproduce that physics much faster, while staying close to what the full model would predict.
2. Sharper maps of city heat and carbon
Inês Girão (+ATLANTIC CoLAB) showed how machine learning turns coarse weather data into detailed, 200-metre maps of heat and carbon exchange within a city. The models work well overall, though they still lose some accuracy on the very hottest days, which is exactly when this information matters most.
3. Rebuilding decades of land use data
Amirpasha Mozaffari (Barcelona Supercomputing Center) is using machine learning to reconstruct high-resolution land use and vegetation data stretching from 1850 to 2100, filling gaps where satellite records do not reach. This will feed a new open benchmark dataset, Terranostra, due for release later this year.
4. Getting AI and physics to talk to each other
Benjamin Rodenberg of the German Climate Computing Centre (DKRZ) tackled a more practical question: how do you actually connect an AI model to a traditional physics-based one? He described a few options, and introduced YAC, an open-source coupling tool his team maintains, already used in ICON, the climate model behind last year's Gordon Bell Prize-winning simulation.
5. Generating weather from scratch
Finally, Peter Dueben (ECMWF) introduced WeatherGenerator, TerraDT's sister project, where machine learning is not just a helper but the whole idea. A single model learns a shared representation of the atmosphere, and can eventually be asked to generate plausible weather states on demand, almost like a weather-focused version of a generative AI model.
The recording and all five sets of slides will shortly be available on the TerraDT website, and anyone who attended and needs a certificate of participation can request one at info@terradt.eu.
For those who want to go beyond watching, TerraDT continues this conversation twice in October: on the methods side, at the Workshop on Machine Learning for Earth System Science in Ostrava (20 to 21 October), a hands-on event where participants can train and benchmark models directly on EuroHPC infrastructure; and on the application side, at Digital Twins for Land Ice: connecting users across projects in Tromsø (27 October), run jointly with the SvalbardDT project and focused on how land ice digital twins are actually being used.
Follow TerraDT on LinkedIn and Bluesky for both registrations and further Tech Talks in the coming months.
