Results of the seventh Finnish LUMI Extreme Scale call

The accepted projects in the seventh Finnish LUMI Extreme Scale call:

COMICS: Coupling the Outer and inner Magnetosphere with the Ionosphere via self-Consistent Simulations

PI: Minna Palmroth, University of Helsinki

The project will deliver the first self-consistent 3D ion-kinetic simulation of the coupled outer magnetosphere, inner magnetosphere, and ionosphere. It will use the hybrid-Vlasov code Vlasiator on LUMI to resolve the 6D proton velocity distribution function, capturing physics that fluid (MHD) models miss and that particle methods represent with noise. It will determine substorm-onset mechanisms by quantifying the interplay of magnetic reconnection and 3D kinetic instabilities and by resolving how magnetotail currents close through the ionosphere. Storm-time transport will be quantified by tracking bursty bulk flows and evaluating their role in energizing the inner magnetosphere and building the ring current, together with ionospheric signatures such as electro-jets and current systems relevant to geomagnetically induced currents. The outcome will be a unique dataset and publications supporting space-weather understanding and EuroHPC Plasma-PEPSC exascale metrics.

MLMD‑GNP: Machine-learning enabled molecular dynamics study of graphite nucleation from paramagnetic iron melt

PI: Jaakko Akola, Tampere University

Controlling the microstructure of structural materials is vital to achieve desired properties. A crucial step in microstructure formation is nucleation which is the process initiating solidification. In most metals of practical interest nucleation occurs heterogeneously on solid inclusions (oxides, sulfides, nitrides, etc.) pre-existing in the melt. The role of these particles in the nucleation of the new phase can be diverse – to provide a solid support for the new phase, act as a structural template for the formation of a phase with similar crystal structure, and to support nucleation through chemical affinity by attracting atoms from the liquid to adhere. Yet, despite much empirical effort, this process remains poorly understood. It is now believed that the primary role of the inoculant is acting in the liquid state. In this project we will study the role of inclusions on graphite nucleation from a ferrous melt on an atomistic scale using molecular dynamics simulations.

AILA-EXTREME: AILA-ENS extremes and impacts

PI: Marko Laine, Finnish Meteorological Institute, co-PI: Olle Räty, Finnish Meteorological Institute

Data based AI-models have initiated a paradigm change in operational weather prediction. Instead of using physics-based numerical weather prediction (NWP) models for the forecast production, it is possible, by using existing historical weather analyses, to train neural network models that have similar predictive power as their NWP counterparts. We will utilize the Anemoi framework developed at ECMWF together with national weather services and the Aila model developed at FMI using Anemoi. We will re-train the probabilistic ensemble version of Aila to better capture extreme weather and the tails of the forecast uncertainty distribution. We will utilize the multi-decoder extension developed in the previous Extreme Scale Access project. This work is part of ECMWF Machine Learning Pilot Project (MLPP).

DISLOC-CrCoNi: Predictive Multiscale Modeling of Dislocation-Mediated Plasticity in CrCoNi Medium-Entropy Alloy

PI: Mikko Alava, Aalto university, co-PI: Amin Esfandiarpour, NOMATEN Centre of Excellence

We propose a multiscale computational study of the CrCoNi medium-entropy alloy, combining density functional theory (DFT), machine-learning interatomic potentials (MLIPs), and large-scale molecular dynamics (MD) simulations. High-fidelity DFT data will train transferable MLIPs, enabling atomistic simulations of dislocation motion and depinning, short-range order, stacking fault energy, lattice distortion, and strain-rate–dependent plasticity in polycrystals. These results will be coupled with mean-field dislocation models to bridge electronic, atomic, and microstructural scales. The project will deliver predictive understanding of strength–ductility synergy in CrCoNi alloys and clarify the fundamental mechanisms governing mechanical performance in complex concentrated alloys. The outcomes will advance multiscale materials modeling methodologies and support the design of high-performance structural materials relevant to Finnish manufacturing, energy, and engineering sectors.

Q-DEFECT: Quantum Defect Engineering in Diamond via Machine Learning Modeling

PI: Tapio Ala-Nissilä, Aalto university, co-PI: Mikko Karttunen, University of Eastern Finland

Quantum defects in diamond are essential for emerging quantum technologies, including single-photon emitters and solid-state qubits. Their accurate modeling requires simulations that combine quantum mechanical-level accuracy with large length and time scales, making the problem computationally demanding. The challenge is further amplified by the vast configuration space of defects.

This project develops machine-learning interatomic potentials (MLIPs) trained on density functional theory (DFT) generated data to enable large-scale molecular dynamics (MD) simulations of defect formation, stability, and dynamics. This approach allows efficient exploration of defect configurations, electron-phonon coupling, and excitation-related processes beyond the reach of direct quantum-mechanical methods. Q-DEFECT combines physics-informed ML with high-performance computing to accelerate nanoscale materials modeling (MD and hybrid DFT), enabling efficient modeling of quantum defects in diamond.

mammoth-Tx: Multitask Encoder-Decoder Language Models at Scale

PI: Tiedemann Jörg, University of Helsinki

Large language models are the backend for most AI applications, yet training and deploying massive models across diverse tasks and languages present computational and efficiency challenges. In this project, we continue our work on modularization in which task-specific and partially-shared components enable flexible and efficient inference. Our framework supports distributed training of such modular architectures at scale, and this project will provide the computational resources to train the essential multi-task setup across a diverse selection of languages and tasks. The project will serve the ERC proof-of-concept project MARMoT and follows up on the EU-funded project HPLT and GreenNLP funded by the Research Council of Finland. Our resources, tools and models will be released with permissive open licenses and we emphasize integration in community-driven frameworks for maximal reuse and sustainability.

AA-Vesicles: All-atom MD simulations of vesicles — from fundamental membrane biophysics to understaniding extracellular vesicles

PI: Matti Javanainen, Tampere University

LUMI-G enables all-atom molecular dynamics simulations of lipid vesicles of dozens of nanometers in diameter, reaching physiologically relevant sizes. Such vesicles are also used in experiments that we can now replicate in silico, providing a bi-directional validation for both the fitted models used in experiments and the simulation protocols and force fields. We will develop a simulation and analysis pipeline and validate it using simple vesicles, after which we increase complexity in terms of lipid composition, asymmetry, and protein content towards a faithful model for extracellular vesicles (EVs). EVs are small and dynamic and they are released from all cells for signalling, transport, and waste management purposes. Our simulations will provide quantitative insight into how the compositional complexity couples to membrane stress, curvature, height fluctuations, lateral diffusion dynamics, and defect formation, which together determine the stability, targeting and uptake of EVs.

MoRTAOx: Modeling of Room-Temperature Plastic Amorphous Oxides

PI: Erkka J. Frankberg, Tampere University, co-PI: Jiahui Zhang, Tampere University

Amorphous aluminum and gallium oxides are novel amorphous oxides that have shown great promise for applications ranging from smartphone screens to modern electronics, including potential uses in optoelectronics, flexible electronics, photovoltaics, single-electron transistors, and battery technologies. These materials allow for a wide range of tailored, functional properties, from full dielectrics to tuned semiconductors coupled with visible light transparency, and good chemical and thermal stability. They exhibit exceptional room-temperature plasticity under mechanical loading, as we observed previously. Improving the mechanical property will allow a much broader usage of these materials. With the help of computational methods, we can have a deeper understanding of the formation and deformation mechanism in amorphous materials, providing theoretical support and guidance to future experimental investigating, and improve the designing paradigms.

EXDEM: Explainable and data-efficient multimodal content reasoning and comprehension

PI: Jorma Laaksonen, Aalto University, co-PI: Abduljalil Radman, Aalto University

One of the most significant challenges in AI is developing systems capable of explainable visual and multimodal understanding, providing human-like depth and nuance while clarifying their reasoning within and between the input modalities. While advancements in Vision Large Language Models (VLLMs) have significantly improved multimodal analysis, enabling innovative applications in comprehending audio-visual-textual data, they still struggle with complex, long-duration tasks such as associating relevant events over extended time spans across modalities. To address the shortcomings of the existing techniques, the EXDEM project proposes a transformative approach where aural and visual content descriptions are integrated in a way that enables them to exchange learned uni- and multimodal representations by using High Performance Computing (HPC), the largest available VLLMs, the most recent Group Relative Policy Optimization (GRPO) training techniques, and huge multimodal data collections.

ComPatAI: AI-Enabled Computational Pathology for Massive Routine Diagnostic Dataset

PI: Pekka Ruusuvuori, University of Turku, co-PIs: Leena Latonen, University of Eastern Finland, and Artturi Mäkinen, FIMLAB laboratories & Tampere University Hospital

With the digitalization of pathology, exceptionally large and high-quality datasets are being accumulated. Population aging and the increasing incidence of cancer are expanding the need for analysis, creating pressure to develop new, scalable, and more precise analytical methods. Artificial intelligence, particularly deep learning, offers the possibility of identifying biological features and associations in tissue samples that are not detectable by the human eye. The ComPatAI project utilizes infrastructures for digital pathology and high-performance computing, leveraging digital data generated by healthcare for the development of diagnostic decision support tools and novel computational pathology methods for virtual staining. We build foundation AI models using a population level cohort of up to 1M whole slide images from routine diagnostics from Finnish university hospital.

GALPMSSD+: Small scale galactic dynamo at high Reynolds numbers

PI: Maarit Korpi-Lagg, Aalto University, co-PI: Frederick Gent, Stockholm University

Magnetic fields (MF), integral to galaxy structure and dynamics, in the interstellar medium (ISM) have energy equipartition with its kinetic energy. The turbulent part of MF is observed as at least twice the strength of any mean MF, which aligns over kiloparsec scales. In contrast, advanced simulations of high Mach number turbulent dynamo yield only about 5% energy equipartition and large-scale dynamo simulations yield turbulent MF much weaker than their mean MF. To understand the implications for observables and galactic structure we must understand how the turbulent and mean dynamos combine to create the MF as observed. Our latest LUMI simulations affirm that turbulent MF strength is apparently asymptotic at Pm within our numerical limits. Our highest resolution simulations need to continue to a persistent statistical steady state to confirm that the low field strength due to turbulent dynamo is physical and that the observed high ratio of turbulent MF arises from tangling of mean MF.

CavChemOTF: On-the-fly mesoscale simulation of cavity modified chemical dynamics

PI: Gerrit Groenhof, University of Jyväskylä, co-PI: Arkajit Mandal, Texas A&M University

Over the past decade, experiments have reported dramatic changes in ground-state chemical reactivity when placing molecules inside an optical cavity. Specifically, it is observed that when confined radiation inside optical cavities couples strongly to the molecular vibrations, various ground-state chemical reactivities are modified even in the absence of external fields. When fully realized, this phenomenon has the potential to unlock a new paradigm of chemical transformation. However, despite extensive theoretical efforts, the microscopic understanding of this remarkable phenomenon remains elusive. The key shortcoming of prior theoretical efforts is the use of simplified model systems that do not capture the molecular complexity, the mesoscale nature of the cavity-coupled molecular systems, and the multimode nature of the confined radiation modes. Using on-the-fly mesoscale simulations, we will address these drawbacks and provide a microscopic understanding of this phenomenon.

OpenEuroLLM Design 2026: OpenEuroLLM Design 2026

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

OpenEuroLLM is an unprecedented collaboration where Europe’s leading AI companies, research institutions, and HPC centres combine forces to develop fully open next-generation large language models (LLMs) for all European languages. The OpenEuroLLM Design project will address critical questions in the design of the OpenEuroLLM models, exploring key architecture decisions for efficient training of large scale multilingual models, and identifying the optimal composition and preprocessing of data for training multilingual models, addressing considerations such as data filtering and cleaning as well as optimal data mixtures taking into account factors such as language, register, and topics. The project will train thousands of models of various sizes that will be comprehensively evaluated to identify the most effective design for the largest OpenEuroLLM models and establish scaling laws to predict their performance.

extremeXPS: Computational XPS spectroscopy of atmospheric adsorbates on aerosol particles

PI: Milica Todorović, University of Turku, co-PI: Nønne Prisle, University of Oulu

Atmospheric aerosols influence weather, climate, air quality, and human health, and their molecular coatings govern the relevant chemical interactions. However, aerosol molecular adsorbates and surface structures remain poorly understood, given demanding X-ray photoelectron spectroscopy (XPS) experiments. This supercomputing project combines expertise from materials science, atmospheric chemistry and computer science to close this scientific gap.

Advanced atomistic simulations will be combined with AI and high-throughput workflows to compute XPS spectra for molecules on aerosol surfaces. Without extreme scale high performance computing, such highly resource-intensive studies would remain unfeasible. This project will result in a rare dataset of computational XPS spectra for periodic surfaces for future AI applications and in fundamental knowledge of how molecular adsorbates modify aerosol XPS signals measured in experiments.

DMAXX: Dark matter axion mass from extreme scale simulations

PI: Mark Hindmarsh, University of Helsinki, co-PI: Kari Rummukainen, University of Helsinki

Most of the matter in the Universe consists of an unknown particle which so far has eluded detection. A leading candidate for this “dark” matter is the axion. Experimental searches for the axion require accurate predictions of its mass to achieve sufficient sensitivity, which in turn requires accurate calculations of the density of axions in the universe today. Such calculations require large-scale numerical simulations of the extended string-like structures which produce the axions up to around a millisecond after the Big Bang.

To perform these simulations, we have developed a massively parallel code AXHILA, which uses the parallel framework HILA to run on multiple parallel systems, included GPUs. It has already run successfully on LUMI-G to produce the world’s largest fixed-grid axion string simulations. With new Extreme Scale access and upgrades to the code we can run faster in more than twice the volume, reducing the statistical and systematic uncertainties on our mass prediction.

PF-FORGE: PF-FORGE

PI: Anssi Laukkanen, VTT Technical Research Centre of Finland, co-PI: Tatu Pinomaa, VTT Technical Research Centre of Finland

In PF-FORGE, we predict how additively manufactured (3D-printed) metal materials form their microstructure via competitive evolution of dendritic crystals, whose multiscale patterning determine strength and durability of the printed parts. These patterns emerge during rapid solidification under dynamic thermal conditions, and accurate prediction is industrially important but difficult. PF-FORGE addresses this by combining extreme-scale phase-field simulations, generative AI, and atomistic simulations. With machine learning surrogate models, we will speed up phase field simulations while preserving the underlying physics.

To accurately parametrize the phase field model, and to study the local rapid solidification defects, large scale atomistic simulations are needed. Using LUMI’s computation, we will run high-resolution 2D- and 3D-simulations to generate high resolution PF-data with different thermal conditions and train denoising diffusion surrogate models to accelerate prediction.