Extreme weather events, such as the 2021 flooding in the Ahr Valley in Germany, are rare but have severe impacts on people and nature, causing significant financial and material damage. Reliable forecasting of such events is therefore crucial and a key focus of current research. Artificial Intelligence (AI) and Machine Learning (ML) are becoming increasingly important in weather and climate prediction. Despite recent progress, existing models still struggle to accurately predict extreme events, particularly peak wind and rainfall. This is often due to an insufficient representation of atmospheric dynamics. Foundation models, trained on large and diverse data sources, offer a promising solution. The goal of the RAINA project is to develop such a foundation model for weather forecasting. Building on AtmoRep , the first foundation model for atmospheric dynamics, RAINA aims to achieve an unprecedented resolution of 1 km and improve the representation of extreme weather events. This requires extensive training data provided by the German Weather Service (DWD) and the European Centre for Medium-Range Weather Forecasts (ECMWF) . The Jülich Research Centre contributes its expertise in high-performance computing (HPC), while the University of Bonn develops advanced methods such as test-time adaptation and diffusion models to optimise the system. A high-resolution short-term forecast for temperature, wind, and precipitation will be implemented as a prototype, aiming to outperform existing operational models. RAINA applies the successful concepts of foundation models from language and image processing to weather modeling and integrates multiple data sources into a consistent atmospheric description. For the first time, probabilistic forecasts will be combined with statistical downscaling. If successful, this project will provide essential insights for the operational application of such a model and significantly improve weather prediction.

This project will build the WeatherGenerator – the world’s best generative Foundation Model of the Earth system – that will serve as a new Digital Twin for Destination Earth. The WeatherGenerator will be based on representation learning and create a general and versatile tool that models the dynamics of the Earth system based on a large variety of Earth system data. The WeatherGenerator will be task-independent and will improve results for a wide range of machine learning applications when compared to task specific machine learning tools. It will also be more resilient for climate applications when the underlying data distributions are changing, and it will lead to a significant reduction in computational costs and faster turnaround times. To achieve this, we will: (1) Collect and use the most important datasets of Earth system science including data from Digital Twins of Destination Earth, selected observations, analysis and reanalysis datasets, and output of conventional Earth system models. (2) Build the WeatherGenerator as a novel representation learning-based machine learning tool that exploits the full potential of Europe’s largest supercomputers. (3) Engage with the wider community via services and apply the WeatherGenerator for 22 selected applications that can be integrated into the Destination Earth framework. The applications include global and local predictions, local downscaling, data assimilation, model post-processing, and impact applications in the domains of renewable energy, water, health and food. The project consortium that will build the WeatherGenerator consists of experts in machine learning, supercomputing and Earth system sciences, and includes industry, SMEs, and leading operational weather centers. The WeatherGenerator will lead to key innovations in weather and climate science and machine learning to enable Europe to establish and defend leadership with respect to machine-learning based Earth system modelling.

The Helmholtz Representation Model for Climate Science (HClimRep) proposal will develop a comprehensive foundation model for climate science, focusing on atmospheric dynamics, tracer transport and ocean coupling, and encompassing seasonal to decadal timescales. Building on the success of AtmoRep , a foundational model for atmospheric dynamics, HClimRep seeks to capture complex long-term interactions of atmosphere, ocean, and sea ice, extending the accuracy, robustness and efficiency of AI-based weather prediction to the climate system. HClimRep will leverage large-scale transformer-based Artificial Intelligence (AI) models, optimised for graphics processing unit (GPU) efficiency, to emulate first-principle climate simulations at a fraction of the traditional computational cost. It incorporates cutting-edge AI methods, such as diffusion models, where they are beneficial. This proposal highlights the potential of foundation models to revolutionise Earth system science by overcoming computational limitations, improving cost-accuracy trade-offs, and facilitating extensive numerical experimentation. By integrating advancements in AI and high-performance computing (HPC), HClimRep aspires to provide a groundbreaking tool for climate research, offering new insights into the Earth's climate system, producing large climate projection ensembles for localised impact modeling, and supporting informed decision-making in climate adaptation and policy, for example in the context of digital twins.

The world is warming. Near surface air temperatures in the first two decades of the twenty first century were 0.75°C warmer than they were over the thirty year period between 1951 and 1980 and about 1°C warmer than during the last two decades of the nineteenth Century. Much less clear, than our planet’s warming and its cause, is the detailed trajectory of this warming – even when conditioned on the carbon emission scenarios – and the quantitative implications of this warming for human and natural systems. High-performance computing is reaching a crescendo with the promise of Exascale machines just beyond the horizon. Within five years, computing systems capable of computing ensembles of trajectories of the climate systems over several decades and on spatial scales of 1 km, will become operational. The German national project WarmWorld by harnessing advances in information technology will compute and evaluate climate warming trajectories, make the information contained within them transparent to application communities, and help harmonize national efforts to provide, disseminate, and use the highest possible quality of climate information. The first phase of the WarmWorld project comprises four modules, with JSC contributing to Modules 2, 3 and 4. Module 2 (WarmWorld "Faster", September 2022 – August 2026) is responsible for transforming the numerical weather prediction and climate model ICON into an open, scalable, modularized and flexible code named ICON-C (“ICON-consolidated”). Specific objectives of Module 2 are to refactor ICON with the goal of scalable development to enable portable performance improvements – ultimately making ICON faster. Module 2 activities will progressively redefine the ICON code structure to expose areas of performance improvement for targeted exploration of new programming concepts. To achieve those goals, the work exploits the analogy of space exploration, with three structured themes dubbed “Ignition”, “Lift-Off”, and “Stage Separation”. Work in the “Ignition” theme will transform the existing ICON code base into a modular design. In “Lift-Off”, separate granules will be connected by an event-driven control flow. These two themes will largely be based on the current Fortran version of ICON and will take advantage of groundwork in the preparatory project preWarmWorld. At the same time, fundamental rewrite of individual granules, using other programming languages than Fortran, domain-specific languages, or even approaches based on machine learning is organized in the “Stage Separation” theme. Module 3 (WarmWorld "Easier", March 2023 – February 2027) will develop novel methods to make information visible, accessible, and interoperable. Specific of objectives are to prepare a data-centric ICON workflow to manage the exascale data streams envisioned in WarmWorld, our `Mission to MARS’ which will establish harmonized access to simulation systems and their output at DKRZ, JSC, and ECMWF through the use of common interfaces and analysis back-ends, as well as prototype exascale analysis workflows that abstract the analysis from the information source, including observational ESA/EUMETSAT and measurement networks such as ACTRIS and IAGOS. Module 3 activities will develop the federated data system, and the tools to exploit it around the highly successful and efficient Field Database (FDB) system developed at ECMWF. This will provide access to data and output hosted by ECMWF without interfering with their operations, and benefit from external projects, such as ACROSS, funded to further develop this system. Exploitation will focus on developing software systems and standards that allow users to explore and analyze information hosted by FDB instances or produced by in-situ analysis in dedicated simulations. From the user point of view the eventual aim of the WarmWorld’s workflow concept is to have a common interface for ICON simulations and data stored in an FDB instance. These are advanced and developing concepts, hence this activity group also contains considerable prototyping designed to guide developments in the second phase of the WarmWorld project. Module 4 (WarmWorld “Smarter”, until May 2027), aims to enhance model performance and workflow efficiency by involving applied mathematics and informatics communities. It focuses on data compression for faster analysis, simulation acceleration through advanced algorithms and machine learning, and in situ analysis to track features during simulations. The goal is to improve computational throughput without compromising simulation quality. FZJ is collaborating with ECMWF on this module.

The Pilot Lab Exascale Earth System Modelling (PL-ExaESM) is a "Helmholtz-Inkubator Information & Data Science" project which aims to lay the foundation for new breakthroughs in earth system modelling (ESM) on future exascale computer architectures. There is increased societal demand for very high-resolution simulations of our environment as well as a better integration of the information value chain, from observations to coupled simulations and impact assessments. Recent technology trends, which lead to more heterogeneous HPC architectures and rapidly growing data volumes, require fundamentally new programming paradigms and call for a coherent strategy for creating next-generation earth system models based on close collaboration between domain scientists and computer experts.

MAELSTROM will develop Europe’s supercomputing architecture of the future and enable the Weather and Climate (W&C) community to make efficient use of new Machine Learning (ML) capabilities on exascale supercomputers. MAELSTROM will implement a co-design cycle that will bring together system designers, ML experts and W&C domain scientists to develop compute system designs, a software framework, and large-scale ML applications that are customised for W&C science. MAELSTROM will connect the current developments of next-generation high-performance computer(HPC) architectures and high-performance data analysics (HPDA), and the needs of the W&C community for customised ML solutions that can make use of exascale supercomputers. This is realised via the joint development of artificial intelligence(AI) applications for W&C science and the design of HPC solutions that perform optimally for those applications. MAELSTROM will also develop the software environment that is required to efficiently train and apply ML solutions on exascale supercomputers using data sets of the size of tens of terabytes and develop dedicated concepts for data processing.

Artificial intelligence (AI) methods are currently experiencing rapid development and are also being used more and more frequently in the context of environmental data. However, this use is often in the context of isolated solutions. The systematic use of modern AI methods is not yet established in environmental and Earth system sciences. In particular, there is often a discrepancy between the requirements of a solid and technically sound environmental data analysis and the applicability of modern AI methods such as deep learning for researchers. The KI:STE project (AI strategy for Earth system data) closes this gap with a sophisticated strategy that combines the development of diverse AI applications on different socially relevant aspects of environmental and Earth system research with a strong training and network concept. It creates the technical prerequisites to make high-performance AI applications on environmental data portable for future users and to establish environmental AI as a key technology.

The IntelliAQ project develops novel approaches for the analysis and synthesis of global air quality data based on deep neural networks. The foundation of this project is the world’s largest collection of surface air quality measurements, which was recently assembled by the principal investigator and plays a pivotal role in the ongoing first comprehensive Tropospheric Ozone Assessment Report (TOAR). This database is complemented with data from the world’s leading effort to collect global air pollutant measurements in near realtime and combined with high-resolution geodata, weather information, and satellite retrievals of atmospheric composition in order to characterize individual measurement locations and regional air pollution patterns. State-of-the-art deep learning methods are applied to this unprecedented dataset in order to fill observation gaps in space and time, provide short-term forecasts of air quality, and assess the quality of air pollutant information from diverse measurements. The combination of diverse data sources is unique, and the project has been the first to apply the full potential of deep neural networks on global air quality data. The achievement of the three IntelliAQ objectives will shift the analysis of global air pollutant observations to a new level and provide a basis for the future development of innovative air quality services with robust scientific underpinning. Due to the heterogeneity of the multivariate data, lack of structure, and generally unknown uncertainty of the input data, the project also poses challenges for existing deep learning methods, and thus leads to new developments in this field. Direct outcomes of the project are a substantial improvement of global air quality information including methods to assess the quality of air pollution measurements, and a new data-driven method for forecasting air quality at local scales.