New Finnish LUMI projects announced: weather forecasting, next-gen LLMs and more

Sixteen new Finnish research projects have been chosen in the sixth Finnish LUMI Extreme Scale call, sharing resources from Finland’s country share of the LUMI resources. The projects span various scientific fields, from physical and space sciences to information sciences, and linguistics. In the selected projects, the researchers will be, for example, developing AI models for weather forecasting, studying materials for future fusion power plants, training next-generation large language models (LLMs), and advancing research related to cardiovascular diseases.

LUMI’s GPU resources were in high demand: there was a threefold oversubscription for LUMI-G resources in this call. The selected projects went through a rigorous technical and scientific review.

As this was the only Finnish LUMI Extreme Scale call of the year, we encourage Finnish researchers to make use of the EuroHPC JU access calls for any larger future needs until the next Finnish extreme scale call – likely to take place in early 2026.

Examples of the accepted projects below and full list available here.

AFMI: Anemoi decoder training

Principal Investigator (PI) Marko Laine, Finnish Meteorological Institute, co-PI Olle Räty, Finnish Meteorological Institute

Data-driven AI models are transforming weather forecasting by replacing traditional physics-based numerical weather prediction models with neural networks trained on historical weather data. The research group will use ECMWF’s Anemoi framework, developed with national weather services, along with MET Norway’s Bris model, built on Anemoi. These tools will train decoders to forecast lightning, cloudiness, and visibility using in-situ observations and satellite-based cloud analyses. This work is part of ECMWF’s Machine Learning Pilot in collaboration with MET Norway.

IrRePoT: Irradiation Response of Polycrystalline Tungsten

PI Fredric Granberg, University of Helsinki, co-PI Aslak Fellman, University of Helsinki

Tungsten is used in key parts of fusion reactors, where it faces extreme radiation and heat. Understanding its behavior under these conditions is crucial for developing fusion as a CO₂-free energy source. This project uses large-scale atomistic simulations with advanced machine learning models to study tungsten material evolution at high radiation exposure at experimentally relevant scales, offering new insights into its performance in future fusion reactors.

OpenEuroLLM Design: OpenEuroLLM Design

PI Sampo Pyysalo, University of Turku, co-PI Jonathan Burdge, AMD Silo AI

OpenEuroLLM is a major collaboration between Europe’s top AI groups to develop fully open large language models for all European languages. This project powered by LUMI will explore key architecture and data decisions for efficient multilingual training, including data filtering, cleaning, and optimal language mix. Thousands of models of varying sizes will be trained and evaluated to identify the best designs and establish scaling laws for future large-scale models.

Read more about all accepted projects in this call here.