Project Goal

---

EMP2 aims to develop a proof-of-concept for a machine learning based digital twin of the atmosphere for environmental applications. To accomplish this, the project is subdivided into two main parts. The first segment will focus on the development of a machine learning based modelling core prototype, called AtmoRep, built on the concept of large scale representation learning applied to Earth System Science. In the second phase, the modelling core will be integrated into the digital twin architecture currently under development within the CERN IT department by the InterTwin project.

Overview

---

The atmosphere and its dynamics have a significant impact on human well-being, from agricultural decision making, to policy making and the renewable energy sector. An accurate and equitable modeling of atmospheric dynamics is consequently of critical importance to allow for evidencebased decision making that improves human well being and minimizes adverse impacts for current and future generations. Very recently, AI-based models have shown tremendous potential in reducing the computational costs for numerical weather prediction. However, they lack the versatility of conventional models. The EMP2/AtmoRep project aims at developing an AI-based model of atmospheric dynamics for multi-purpose applications. The model will be implemented leveraging the concept of large-scale representation learning, so to encapsulate the information from the large amounts of available data. The implementation on the digital twin platform will make such information more accessible to the general public, allowing the users to easily develop their own applications in weather and climate.

Highlights in 2024

---

In 2024 we published the model on GitHub, making it fully open source. During summer we focused in getting the model smaller and more efficient. We moved from a 3.5B parameter model, to a 600M parameter core model which is 6 times faster and 30% more accurate than the previous published architecture. In addition, we have started implementing the roll-out mechanism to enable medium range global weather forecasting up to 10-15 days. Through the HClimRep collaboration, kicked-off in September 2024, the model is getting upgraded towards decadal time scales predictions. The latest changes include the integration of ocean and stratosphere level information into the current architecture. In the last part of the year, in collaboration with interTwin, we have settled a roadmap for the integration of the model within the interTwin digital twin platform for AI models, which will be completed in January 2025. 

Next Steps

---

The project will end in February 2025. The model is already serving as the basis for the HClimRep project in the context of the Helmoltz AI call on foundation models. Starting from February 2025, the developed core model will serve as basis for the development of the WeatherGenerator (link), an EU funded project accounting for 16 partners and lead by ECMWF. The goal of the WeatherGenerator will be to build a foundation model for a large number of tasks, from renewable energy and flood prediction to food security, health and the biosphere. In 2025 the model will be used in the context of the WFP-LIST-CERN strategic partnership on AI, as atmospheric model for seasonal predictions of crop yields in critical areas of the globe. 

More Information

---

Project Coordinators: Alberto Di Meglio, Ilaria Luise

Technical Team: Alberto Di Meglio, Ilaria Luise

Collaboration Liaisons: Michael Langguth, Christian Lessig, Martin Schultz

In partnership with: Julich Forschungszentrum, ECMWF, Universitat Magdeburg