36^th open access competition RESULTS

THE ALLOCATION COMMISSION DECIDED ON THE ALLOCATIONS WITHIN THE 36^th OPEN ACCESS GRANT COMPETITION AS FOLLOWS:

Researcher: Georgios Kordopatis-Zilos     

OPEN-36-1     

Learning a Universal Similarity Function for Multimodal tasks – LUSt-multimodal     

LUMI-G  Alloc=14200             

Multimedia content is indispensable in our society, necessitating effective content management. A critical aspect of this is assessing the similarity between two multimedia items, i.e., images, videos, and documents. Our mission is to learn a universal similarity function (LUSt) capable of precisely measuring similarity across a broad spectrum of multimedia domains and tasks. Diverging from traditional problem-specific approaches prevalent in current literature, we adopt a novel strategy. With LUSt, we plan to break down multimedia items into constituent parts, including image regions, video frames, and text sentences. Subsequently, a foundational model will be trained on input data comprising part similarities across various multimedia items. This strategic choice yields a universal input space with multiple advantages. In the second stage of this venture, the objective is to develop LUSt-multimodal, a modality-agnostic similarity network meticulously designed for multimodal tasks. Multimodal Large Language Models (MLLMs) have demonstrated strong cross-modal reasoning capabilities, yet their potential for similarity estimation remains underexplored. Hence, we will repurpose MLLMs for similarity estimation and build self-supervised pipelines for their effective fine-tuning on our downstream task. Also, we will expand the modalities used during training to include all available modalities, not just the image-text pairs typically used, and apply modality-tailored augmentations.

Researcher: Andrea NEDĚLNÍKOVÁ             

OPEN-36-10  

Interactions of graphene derivatives with brain cell membranes             

Karolina CPU  Alloc=14500;  LUMI-C  Alloc=2200

Neurodegenerative diseases like Parkinson’s affect millions of people, and today there is no cure, only treatment to manage symptoms. One promising approach is deep brain stimulation, where electrodes implanted in the brain deliver electrical impulses to regulate abnormal brain activity. But the metal electrodes we use today are rigid, their implantation is highly invasive, and they often cause inflammation. Researchers have decided to develop a new generation of electrodes based not on metal but on graphene - an ultra-thin, highly conductive carbon nanomaterial.  Our research group in Olomouc, as part of the MINIGRAPH project (Minimally Invasive Neuromodulation Implant and implantation procedure based on ground-breaking GRAPHene technology for treating brain disorders – HORIZON-EIC-2021-PATHFINDERCHALLENGES-01:101070865), aims to investigate the biocompatibility of graphene with brain cells both experimentally and theoretically.  In this project, we aim to validate the interactions between brain cell membranes and a variety of graphene structures - ranging from pristine graphene to graphene oxide using coarse-grained simulations, with a complex representation of the graphene electrodes proposed for the implants, including several layers of porous graphene structures.

Researcher: Matěj Kripner   

OPEN-36-11  

Advancing Out-of-Domain Reasoning in Machine Learning        

Karolina CPU  Alloc=1200;  Karolina GPU  Alloc=7200;  LUMI-C  Alloc=1800;  LUMI-G  Alloc=12000             

Reliable out-of-domain reasoning is a holy grail of current AI research, being both essential for real-world deployment of autonomous AI systems and unsolved. While large language models (LLMs) combined with chain-of-thought and trained using reinforcement learning from verifiable feedback achieved unheard-of performance in math and coding competitions, they still struggle with seemingly trivial tasks, suggesting that their reasoning capabilities do not generalize the way humans do and leading to the phenomenon commonly known as hallucinations.     This makes it difficult for humans to trust the outputs of AI systems, hindering their use as fully-fledged collaborators. It also limits their applicability to unattended long-horizon tasks where multiple steps must be performed without error to achieve a desired outcome.     The overarching goal of our project is to develop methods for reliable out-of-domain reasoning on small, self-contained, and clearly defined domains such as Automated Theorem Proving and the Abstraction and Reasoning Corpus, and subsequently to integrate these advances into general purpose systems, such as LLMs. Improvements in the reliability and depth of reasoning lead directly to the applicability of AI in virtually every field, including mathematics, physics, medicine, cryptography, software verification, manufacturing, and more.

Researcher: Adéla CHADALÍKOVÁ 

OPEN-36-12  

Insights into LNP Inner Structure: Effects of N/P Ratio    

Karolina CPU  Alloc=1700;  Karolina GPU  Alloc=3100    

Lipid nanoparticles (LNPs) are nanoscale particles composed of several lipid classes that can encapsulate and safely deliver nucleic acids or small molecules into target cells. They typically consist of artificially designed ionizable lipids (ILs) together with helper lipids. Helper lipids include phospholipids, cholesterol, and PEG-lipids. Each of these lipid classes has specific properties that influence the characteristics of the final nanoparticles. LNPs have been successfully used in approved therapeutics, such as patisiran (ONPATTRO, Alnylam Assist), which employs LNPs for siRNA-mediated gene silencing [1], and in mRNA vaccines against COVID-19, including Comirnaty (Pfizer/BioNTech) and Spikevax (Moderna) [2]. Despite their widespread use, the precise mechanism of endosomal escape and intracellular processing of LNPs remains incompletely understood.  Current theories propose that the LNP carrying the RNA enters target cells via endocytosis. Inside the gradually maturing endosome, the acidic environment protonates the ILs, increasing their positive charge and promoting interaction with the negatively charged endosomal membrane’s lipids.  It is believed that during this process, the lipids of the LNPs reorganize from a lamellar phase into an inverted hexagonal (HII) phase. This lipid phase transition results in membrane destabilization and fusion, enabling the release of mRNA into the cytosol, where it is translated into an antigenic or therapeutic protein (in the case of mRNA vaccines). This project aims to use molecular dynamics (MD) simulations to investigate the internal structure of LNPs, particularly the behaviour of the HII phase. 

Researcher: Jakub Šístek     

OPEN-36-13

Accelerated vortical structures: GPU acceleration of numerical simulations for advanced vortex identification

Karolina CPU  Alloc=2000;  LUMI-G  Alloc=5800 

The main aim of the project is performing high-resolution computational fluid dynamics simulations of prototype problems of incompressible viscous flows. The primary goal of  these simulations is to generate high-resolution 3D data with vortical structures, which will be subsequently used for development of new methods for flow-field analysis and vortex identification and visualization. Unsteady flows considering very fine computational meshes are required for this purpose. The computations will be performed using a parallel finite element solver based on multilevel domain decomposition and accelerated by GPUs.  A subsequent goal of the project is further development of the computational method and optimization of the open-source BDDCML library for large numbers of computer cores combined with GPU accelerators by AMD.

Researcher: Anna ŠPAČKOVÁ

OPEN-36-14   Modelling Phospholipid Membranes Using Artificial Intelligence

Barbora CPU  Alloc=1200;  Barbora GPU  Alloc=3000;  Karolina CPU  Alloc=1300;  LUMI-G  Alloc=1200    

Cell membranes are the gatekeepers of biology, enabling communication, energy conversion, transport, and signaling by selectively admitting some molecules while excluding others. Their behavior depends on molecular architecture. Real membranes are crowded, heterogeneous assemblies of lipids, proteins, sterols, and carbohydrates that constantly reorganize in response to their environment. This structural complexity is functional: even small changes in lipid type, packing, or lateral organization can reshape curvature, alter permeability, modulate protein activity, and influence how a cell senses and adapts. Understanding how membrane composition shapes membrane behavior is therefore essential for explaining cellular function and addressing diseases linked to disrupted membrane organization.  In this project, we develop artificial intelligence methods that can predict how membrane molecules arrange themselves. By learning patterns from large-scale molecular dynamics simulations, our models will construct realistic membranes and place each molecule according to its surroundings. This will make it easier for scientists to build complex membrane models and study how they change under physiological conditions or interact with objects such as lipid-nanoparticle vesicles. It also represents a key step toward building complex (bio)molecular assemblies. 

Researcher: Zaira Carboni  

OPEN-36-15  

Exploration of Graphene–Graphyne Hybrid Structures and Their Photoinduced Charge-Transfer Properties       

Karolina CPU  Alloc=300;  LUMI-C  Alloc=250;  LUMI-G  Alloc=2000     

In recent years, γ-graphynes have attracted significant interested due to their remarkable potential in practical applications. Their unique combination of sp^2-sp hybridized carbon atoms, extended π-congiugation, and uniformly distributed dehydrobenzo[n]annuleme pores grants γ-graphynes outstandig performance in gas separation, water purification, sensing and catalysis. Building on these properties the proposed project aims to investigate the formation of innovative carbon dots incorporating graphene (G) together with γ-graphynes (γG1), γ-graphdiyne (γG2) or γ-graphtriynes (γG3) sheets. The overall aim of this project is to explore the feasibility and functional potential of hybrid nanostructures formed by graphene (G), γ-graphynes  (γG1),  γ-graphdiyne (γG2) and γ-graphtriynes (γG3) sheets. The main objectives of the project is to perform Molecular dynamics investigation of the formation and stability of γG complexes. This includes assessing the structural compatibility of graphene with γ-graphynes-family flakes, analyzing non-covalent interactions and determining the condition under which stable assemblies or stacked hybrid  structures can form

Researcher: Frantisek Karlicky         

OPEN-36-16  

Effects of surface terminations and defects on MXenes reactivity for water splitting

Barbora CPU  Alloc=6000;  Barbora FAT  Alloc=200;  Barbora NG  Alloc=9000;  Karolina CPU  Alloc=54700;  Karolina FAT  Alloc=100;  Karolina GPU  Alloc=4300;  LUMI-C  Alloc=6400       

This project focuses on the study of the surface properties of a new class of 2D materials, MXenes. These materials are now regarded as highly suitable candidates for numerous technological applications for their robustness and wide range of physical properties achievable within a single materials class (e.g., magnetic properties, tunable band gap from metals to semiconductors), e.g., for the use in catalysis or environmental remediation, due to their exceptional chemical reactivity and tunable surface properties by virtue of the variable composition and surface functionalization. This study employs density functional theory to investigate water adsorption and water splitting on variously terminated and defective MXenes.

Researcher: Lukas Grajciar

OPEN-36-17  

Towards chemically accurate hybrid-DFT machine learning potentials for zeolite synthesis

Karolina CPU  Alloc=32900;  LUMI-G  Alloc=2100              

Zeolites are indispensable materials for numerous industrial applications, including catalysis, water purification, and ion-exchange. Despite their technological significance, a comprehensive atomic scale understanding of the complex reactive processes governing their synthetic routes remains elusive. Historically, atomistic simulations using ab initio methods have been severely constrained by high computational costs. This limitation has precluded the study of the large unit cells and long timescales necessary to accurately model the dynamics of zeolite synthesis. Recent progress in Machine Learning Interatomic Potentials (MLIPs) has enabled the exploration of broader chemical spaces and longer timescales, successfully retaining the accuracy of their training data. However, achieving the chemical accuracy required to describe the highly reactive bond-breaking and bond-forming events central to zeolite synthesis often necessitates training with high-level theoretical methods, a process that is itself computationally demanding. This project addresses this critical methodological gap by combining state-of-the-art hybrid periodic DFT with MLIPs. This combined approach will be applied to investigate the complex, multi-step reaction pathways of zeolite synthesis precursors. The resulting unprecedented atomic insight will significantly accelerate the rational design of new and improved synthetic routes for zeolites.

Researcher: Fatemeh Kayanikhoo

OPEN-36-18   Fatemeh Kayanikhoo              

Karolina CPU  Alloc=18100

We aim to investigate the neutron star  X-ray binaries through numerical simulations in general relativity and using the magneto-hydrodynamics method. The goal of this project is to understand the fundamental physics behind the characteristic behaviours of ultraluminosity X-ray sources powered by supercritical accretion onto neutron stars.

Researcher: Marek Hrúz       

OPEN-36-19  

Sign Language Recognition, Translation, and Interpretability     

LUMI-C  Alloc=500;  LUMI-G  Alloc=13800

Sign languages are vital forms of communication for millions of people in Deaf and hard-of-hearing communities worldwide. However, unlike spoken and written languages, sign languages are visual and highly nuanced, with gestures, facial expressions, and body movements all contributing to meaning. This complexity makes translating sign language into spoken or written language a challenging task - but one with immense potential to foster inclusivity and break down communication barriers.  Our project focuses on designing and interpreting an automatic sign language translation system. Instead of relying on handcrafted linguistic alignments, we aim to build a model that can learn to understand sign language directly from large scale data. This means teaching the system to interpret visual input - hand shapes, motion trajectories, facial expressions, and body posture—and map it to meaningful spoken or written language output. By studying how the system internally represents and processes these visual cues, we gain insights into how artificial intelligence can reason about complex, expressive human communication.  The ultimate goal is to make communication seamless between sign language users and speakers of other languages, creating technologies like real-time sign language interpretation and improved accessibility in education, workplaces, and public services. By bridging this gap, we hope to empower the Deaf community and bring us closer to a world without communication barriers.

Researcher: Michal Hradiš 

OPEN-36-2    

Fast models for historic Czech documents            

Karolina GPU  Alloc=900;  LUMI-G  Alloc=2000    

Libraries and archives digitize their collections which are then accessed by scholars and the general public alike. However, current access systems lack AI-enabled functionality we are all starting to expect. One of the reasons is definitely a slow pace of adoption, but others include the lack of models suitable for the often specific and older language of the content, prohibitive operating cost of large AI models and sometimes the requirement for local on-site processing.  In this project, we will prepare a set of models enabling three specific functionalities relevant to libraries and archives which will be released under a permissive license which will enable their utilization not only in Česká digitální knihovna and Kramerius, but also in other applications including commercial media archives and text analysis tools.  The models will include text embedding models suitable for document retrieval of Czech and specifically historic documents, extreme summarization models (query-based short title generation), and multi-purpose text tagging models, for example capable of localizing topics, entities, styles and arguments in the text. These models will range from 100M to 3B parameters and will be trained mostly by distilling state-of-the-art large embedding models or large language models. 

Researcher: Debora LANČOVÁ        

OPEN-36-20 

Simulation of accretion disks in black hole ultraluminous X-ray sources          

Barbora NG  Alloc=5000;  Karolina CPU  Alloc=43500     

Ultraluminous X-ray Sources (ULXs) are extremely bright, off-nuclear X-ray sources observed in nearby galaxies. Their high apparent luminosities originally suggested accretion onto intermediate-mass black holes. However, the detection of coherent pulsation in several ULXs pointed to the presence of a neutron star in some of them. Because accretion onto neutron stars and black holes differs fundamentally -through surface emission, magnetic fields, and boundary conditions - clarifying the nature of ULXs requires detailed theoretical and numerical modelling. In this project, we employ state-of-the-art general relativistic radiation-magnetohydrodynamic simulations to systematically compare super-Eddington accretion onto neutron stars and black holes. Building on recent simulation results for neutron-star ULXs, we will explore the corresponding parameter space for intermediate-mass black holes. By examining the resulting energy output, radiative efficiency, and observable spectral and timing properties, we determine whether black hole accretion can reproduce observed characteristics of ULXs, such as candidate black-hole ULX M82 X-1.

Researcher: Michael Bakker              

OPEN-36-21 

Computational Elucidation of Mechanosensing in Vascular Remodeling: Force-Dependent Conformational Dynamics of PECAM1 and PLXND1        

Barbora CPU  Alloc=1000;  Karolina CPU  Alloc=3000;  LUMI-G  Alloc=15000

Blood vessel formation and remodeling, essential processes in health and disease (e.g., wound healing, cancer, atherosclerosis), are precisely controlled by the physical forces exerted by blood flow, known as hemodynamic forces. Our cells sense these forces via specialized \mechanosensing\" proteins like PECAM1 and PLXND1 on the endothelial cell surface. This project utilizes molecular dynamics (MD) simulations to decode how these two proteins mechanically sense force, leading to a conformational change that acts as a molecular \"on/off\" switch. Ours is a multi-scale strategy, first using all-atomistic (AA) MD simulations to characterize dynamics of individual protein domains and disordered inter-domain linkers of PECAM1 and PLXND1. The AAMD data will then be used to top-down parameterize and a unique coarse-grained (CG) force field specifically for large extracellular domains. This CG force field enables us to simulate the application of hemodynamic force over the necessary μs timescales to map the proteins' complete transition pathways and identify critical mechanical \"catch bonds\" or \"slip bonds\". This will allow us to predict targeted mutations that could enhance or block force sensitivity, paving the way for novel therapeutic strategies against vascular diseases or tumor angiogenesis."

Researcher: Jakub Šebesta

OPEN-36-22  Low dimensional molecular magnets         

Barbora NG  Alloc=1000;  Karolina CPU  Alloc=600;  Karolina FAT  Alloc=50;  Karolina GPU  Alloc=100       

Low-dimensional Heisenberg magnets represent systems with a rich phase diagram and unusual magnetic properties that stem from geometrical frustration and short-range magnetic interactions pointing only in certain directions. In the present project, one focuses on magnetism of Cu-based molecular magnets, studying the nature of the magnetic ordering and related magnetic exchange interactions.

Researcher: Marketa Paloncyova  

OPEN-36-23  Endosome Maturation as a Crucial Step in RNA Delivery              

Karolina CPU  Alloc=10900;  Karolina GPU  Alloc=300;  LUMI-C  Alloc=7100;  LUMI-G  Alloc=1800

The delivery of RNA into cells is based on using the endosome maturation as a trigger for the RNA release from lipid nanoparticles (LNPs). During the endosome maturation, the pH drops and LNP lipids change their protonation state and induce the RNA release. The in-silico studies of this process are limited to discrete states of the bioenvironmental models, either a plasma membrane or a matured endosome. The size of the relevant models that is difficult to downgrade also limits the design of necessary protocols for simulating these processes. In this project we would like to design in-silico protocol for modelling gradual exchange of lipid composition in lipid vesicle, allowing for the simulations of LNP in gradually maturing endosome. The designed protocol will then be used to study the RNA release process, allowing the future study of its efficiency and assisting the targeted design of LNP composition.

Researcher: Jiri Klimes          

OPEN-36-24 

Accuracy and precision for extended systems XV              

Barbora CPU  Alloc=27000;  Barbora NG  Alloc=3000;  Karolina CPU  Alloc=14900;  Karolina GPU  Alloc=100;  LUMI-C  Alloc=7000     

The development of new or improved pharmaceuticals makes a substantial use of methods of materials science. One of the examples is the formulation of the solid form of pharmaceuticals. A pharmaceutically active compound, typically a molecule, has often several functional groups that ensure that it ends up on the right place in body and does the right thing. This can make the molecules complicated and when they form crystals, these can have different structures. This is called polymorphism. Knowledge of these polymorphs is crucial for the pharmaceutical industry as some of them might have undesirable properties. Simple methods, often based on quantum mechanics, are used to identify the possible polymorphs so that the landscape of structures is known. However, our data show that the methods that are currently used rely on cancellations of errors. This means that the results can look fine but there can be situations in which unexpectedly large errors occur. We are trying to find out how large problem the error cancellation is by developing a large dataset of reliable energy differences between polymorphs. Moreover, we are getting so-called many-body energy contributions to the energies of the structures which allows one to understand the accuracy of the different theoretical methods in detail. This knowledge can be then exploited to develop methods with improved accuracy, reducing thus the uncertainty of theoretical predictions.

Researcher: Jiri Brabec          

OPEN-36-25 

New DMRG-DUCC-TCCSD Approach for Strongly Correlated Systems              

Barbora CPU  Alloc=5000;  Barbora NG  Alloc=10400;  Karolina CPU  Alloc=54700;  Karolina GPU  Alloc=6400    

Strongly correlated molecular systems, such as transition metal complexes, polycyclic aromatic hydrocarbons, or many intermediates in chemical reactions, remain the most challenging areas in quantum chemistry. These systems cannot be accurately described by traditional single-reference methods due to significant multiconfigurational character. Recently, we developed a Density Matrix Renormalization Group approach based on the Double Unitary Coupled Cluster downfolding (DMRG-DUCC), which integrates dynamic and static correlation effects through an effective Hamiltonian formalism. However, our latest results indicate that for strongly correlated systems, DMRG-DUCC suffers from a bias originating from single-reference CCSD amplitudes used in the downfolding step.   To address this, we propose a new hybrid framework - the DMRG-DUCC-TCCSD method - which utilizes a Tailored Coupled Cluster (TCCSD) to optimize the external virtual space. Here, internal amplitudes are obtained from the DMRG matrix product state on a downfolded effective Hamiltonian, while external amplitudes are partially relaxed to include dynamic correlation without destroying the internal static correlation structure. The resulting scheme is expected to yield high accuracy in describing strongly correlated systems, advancing computational chemistry toward predictive modeling of complex materials. The method will be applied to a wide range of strongly correlated systems, and the resulting high-level data will form a benchmark database for machine learning. 

Researcher: Jiří Jaroš              

OPEN-36-26 

Closed-loop individualized image-guided transcranial ultrasonic stimulation IV         

Barbora CPU  Alloc=2000;  Barbora FAT  Alloc=50;  Barbora GPU  Alloc=50;  Barbora NG  Alloc=3000;  Karolina CPU  Alloc=700;  Karolina FAT  Alloc=50;  Karolina GPU  Alloc=900;  LUMI-C  Alloc=500;  LUMI-G  Alloc=100       

Disorders of the brain, including neurological and psychiatric diseases, affect one in four people. New treatment options are needed with enhanced efficacy and reduced side-effects, costs, and invasiveness. Neurostimulation techniques that modulate the electrical activity of the brain have evolved as an important class of second-line treatments for pharmacoresistant cases. What is needed is a non-invasive brain stimulation technique that can stimulate brain targets with high anatomical precision, unlimited penetration depth, full reversibility, and low risk-profile. This can be achieved using the newly emerging technique of low-intensity focused transcranial ultrasonic stimulation (TUS) for neuromodulation.  This project focuses on the development of closed-loop individualized image-guided transcranial ultrasonic stimulation, under the Horizon Europe CITRUS project. The ultimate goal of the CITRUS project is to develop a fully functional prototype of a medical device that integrates an ultrasound transducer system possessing advanced 3D steering capabilities with a custom-built magnetic resonance receiver, enabling high-resolution transcranial neuromodulation with unprecedented flexibility and sensitivity.   The computational resources will be used for preoperative MR-based brain imaging, personalized ultrasound treatment planning including temperature mapping, and validation of fast real time re-planning software based on advanced mathematical models and artificial neural networks.

Researcher: Cristina-Ioana Balaban           

OPEN-36-27 

Reconstructing Quaternary mountain glaciation and palaeoclimate in Central Europe with the Parallel Ice Sheet Model          

Karolina CPU  Alloc=13300

Glaciers are sensitive to climate change, making them powerful indicators of past climate conditions. This project aims to decipher the climate conditions that grew glaciers in the mountains of Central Europe, where plants, animals, and humans could have taken refuge between the large Fennoscandian and Alpine ice sheets during the Pleistocene ice ages. Using the computationally efficient Parallel Ice Sheet Model, our objectives are to understand how cold and how dry or wet the climate was to form any glaciers, grow them to their maximum size, and assess their response to past climate changes in the region. By identifying the appropriate climate conditions to grow glaciers under each objective, potential climate gradients may be reconstructed across Europe, with crucial implications not only for atmospheric circulation changes over millennia but also for the dispersal of biota and humans across the continent.

Researcher: Petr Hyner          

OPEN-36-28 

LLM-based Evolution for Automating Machine Learning Research         

LUMI-G  Alloc=4300

We plan to build “coding agents” – computer programs that can design their own machine-learning experiments. Instead of a human researcher manually trying out different model architectures, training tricks, and data settings, our agents will read a short description of the task and a prepared project template, then automatically propose many variants: they will design network architectures, loss functions, data augmentations, training schedules, and sampling strategies. These proposed experiments will be sent to the Karolina supercomputer and run at scale, allowing us to test a large number of ideas quickly and systematically. By analysing which combinations work best, we hope to discover more efficient ways to train models with less trial-and-error from humans. In the long term, this research could make the development of AI systems faster, cheaper, and more accessible to smaller research teams and companies, and it will also provide insight into how to safely and reliably use AI tools to help automate scientific and engineering work.

Researcher: Jana Precechtelova     

OPEN-36-29 

A Computational-First Strategy for NR2F6: Design of Covalently-Stabilized Peptidomimetics to Disrupt the NSD1 Immune Checkpoint Interaction             

Karolina CPU  Alloc=1000;  LUMI-G  Alloc=7600 

This project targets the orphan nuclear receptor NR2F6, a transcriptional regulator. We have structurally characterized the receptor into its constituent domains: the N-terminal domain (1-53; NTD), DNA-binding domain (53-130; DBD), Hinge region, and ligand-binding domain (164-404; LBD). Targeting the DBD is non-viable due to poor specificity, as this domain shares 97% sequence homology with family members NR2F1 and NR2F2. The presents a challenge as is an orphan receptor with no known endogenous ligand, and its canonical ligand site is blocked by Helix 12 (H12) in the constitutively autorepressed state. Despite the LBD being structurally inaccessible, NR2F6 is functionally active through its protein-protein interactions (PPI). These PPIs include homodimer and heterodimer formation with other family members, as well as binding to essential coregulators like NSD1 and FOG3 (Table 1). The specific binding interface of the coregulator has been defined atomistically via X-ray crystallography (PDB ID: 8C5L) as a small residue peptide. This structural information establishes two primary therapeutic strategies: disrupting dimerization or leveraging the coregulator binding site via peptidomimetics. As NR2F6 is a key intracellular immune checkpoint, successful therapeutic agents could be used to augment -cell responses for applications in immuno-oncology, particularly in treating cancer. More broadly, the multiscale computational protocol developed herein is directly applicable for designing PPI modulators against other challenging orphan nuclear receptors.

Researcher: Michal Novotny             

OPEN-36-3    

Comparative study of the effect of mixed composition and vacancies on MXenes and MXene-based quantum dots II          

Barbora CPU  Alloc=12000;  Karolina CPU  Alloc=27000;  Karolina GPU  Alloc=5700

MXenes represent a wide and rapidly growing class of two-dimensional (2D) materials with a unique combination of properties such as high conductivity, chemical stability, magnetism, and versatile surface functionalization. These properties make MXenes very promising for a wide range of applications in nanoelectronics, spintronics, and photocatalysis. In recent years, their quantum dot derivatives (MXQDs), which combine the advantages of 2D materials with the additional effects of quantum confinement, have also received increasing attention. These nanostructures exhibit enhanced electronic, optical and magnetic properties that can be further tuned by targeted structural modifications.  Although interest in MXQDs has grown considerably in recent years, detailed understanding of their structural characteristics - such as surface functionalization and vacancy formation - remains limited. However, these structural features are expected to play a key role in defining the physicochemical properties of MXQDs. A deeper understanding and achieving precise control over the surface terminations and configurations of the vacancies could allow tuning their electronic, optical, and magnetic behavior, leading to the attachment of MXQDS with the desired properties.  In this context, systematic theoretical research will be carried out to model different types of MXQDs and to investigate how structural factors - such as vacancies and mixed surface functional groups - affect their electronic, optical and magnetic properties. The resulting findings will support the design of suitable types of MXQDs for potential applications in photoelectronics, spintronics and photocatalysis.

Researcher: Petr Macha       

OPEN-36-30 

GRILLIX SIMULATION OF THE COMPASS UPGRADE TOKAMAK 

Karolina CPU  Alloc=28300

The COMPASS-Upgrade tokamak [1,2], now under construction at IPP Prague, is a high-magnetic-field, ITER-like device that will play a key role in European fusion research. Its high density and diverted geometry place it in a regime where edge and scrape-off-layer (SOL) turbulence strongly affects confinement and heat-exhaust performance, making predictive modelling essential before first plasma. This project will deliver the first full-size, self-consistent 3D fluid simulation of COMPASS-Upgrade using the GRILLIX code [3]. GRILLIX evolves turbulence self-consistently and, thanks to its flux-coordinate-independent (FCI) formulation, enables realistic full-device simulations in diverted geometries at manageable computational cost. The results will provide predictions of global transport, heat-flux widths, divertor loads, and fluctuation properties, key inputs for defining safe operational windows and guiding early experiments. GRILLIX has already demonstrated reliable performance on multiple machines, and we have validated our workflow on the present COMPASS tokamak, where simulations reproduced key experimental trends. This experience significantly reduces risk and ensures that full-device modelling for COMPASS-Upgrade is both feasible and timely.

Researcher: Zdeněk Futera 

OPEN-36-31

Optimization of Computational Approaches to Biomolecular Electronics        

Barbora CPU  Alloc=50000;  Barbora FAT  Alloc=200;  Barbora NG  Alloc=10400;  Karolina CPU  Alloc=33200;  Karolina Alloc=130;  LUMI-C  Alloc=5000

Molecular electronics is a rapidly developing research area with promising scientific and industrial applications. Single-molecular junctions are manufactured and their conductance properties studied by scanning tunnelling microscopy (STM) techniques. However, theoretical and computational modelling is often required to interpret the measured data and bring deeper insight into the electronic structure of these nanoscale electronic devices. These are challenging as an accurate quantum description, typically based on first-principles methods, must be applied to obtain quantitative data. However, the computational cost and unfavourable scaling limit these calculations to small systems only. Therefore, we focus on various approximate approaches that enable applicability to larger biomolecular junctions relevant to real-world applications.

Researcher: Petra KÜHROVÁ            

OPEN-36-32 

Building an RNA Benchmark: Short Duplexes and Motifs as Standard Test Systems 

Karolina CPU  Alloc=2900;  Karolina GPU  Alloc=2800;  LUMI-C  Alloc=7100;  LUMI-G  Alloc=15000             

RNA is a key molecule of life. It helps read and regulate our genetic information and is at the heart of new medicines such as mRNA vaccines and therapeutic oligonucleotides. To study RNA in detail, scientists use computer simulations that follow the motion of every atom. These simulations are powerful, but different research groups currently test their models on different RNA fragments and under different conditions, which makes results difficult to compare.  In this project we will build a small, well-defined set of RNA test systems that can serve as a common reference for the whole community. A central role will be played by short RNA double helices, for which high-quality experimental data on thermal stability exist but which are rarely used as standard tests. We will complement them with a few simple RNA loops and other motifs.  Using large-scale simulations on the Karolina and LUMI supercomputers, we will characterize how these RNAs fold, unfold and respond to changes in conditions. All input data, protocols and analysis tools will be made available so that other groups can directly reuse the benchmark. In the long term, this should make computer models of RNA more reliable and help accelerate the development of RNA-based therapies and biotechnological applications.  

Researcher: Šimon Vrba       

OPEN-36-33 

Modelling of ELM transport in the ITER SOL            

Barbora NG  Alloc=10000;  Karolina CPU  Alloc=69100  

One of the hot topics of the Magnetic Confinement Fusion research is the estimation and reduction of power loads to the plasma facing components and the related erosion of these elements. Among the most critical threats are power loads during the so-called Edge-Localized Modes (ELMs) in tokamaks, when a significant amount of hot plasma is quickly released from the confined region into the scrape-off layer (SOL), where it directly propagates towards the divertor. These increased divertor fluxes can cause cracking or even melting of the surface material, and significantly reduce its lifetime. ELMs can be also strong sources of high-Z material erosion (W), leading to contamination of the plasma core and even terminating the discharge. ELMs are particularly threatening for future large machines such as ITER and DEMO, because the heat loads scale with the machine size. Therefore, different ELM mitigation techniques have been developed e.g. impurity seeding, where by injecting impurity particles (Ne), a portion of the ELM power can be radiated away before hitting the divertor. The aim of this project is to perform a fully kinetic predictive modelling of the ELM transport in the SOL of the ITER tokamak during a high-performance Ne seeded discharge. We will study the evolution of the ELM-related divertor heat loads, buffering via impurity seeding, tungsten sputtering rates, and plasma profile evolution. These simulation predictions are vital for ensuring the safe operation of future machines.

Researcher: Ondrej Meca    

OPEN-36-34 

Research and Development of Libraries and Tools in the INFRA Lab IV               

Barbora CPU  Alloc=1000;  Barbora FAT  Alloc=200;  Barbora GPU  Alloc=200;  Barbora NG  Alloc=3000;  Karolina CPU  Alloc=21200;  Karolina FAT  Alloc=100;  Karolina GPU  Alloc=1600;  LUMI-C  Alloc=21700;  LUMI-G  Alloc=10000             

As members of the Infrastructure Research Laboratory, our goal is to bring improvements and extensions to available tools that support the users of the IT4I clusters and their research. The key topics of our research are Energy efficiency, the development of scalable libraries for engineering applications, MESIO and ESPRESO, and the development of Visualization tools. Developments in these areas will also serve to meet the objectives of the individual projects in which members of the laboratory are involved.

Researcher: Jakub Hromádka          

OPEN-36-35 

Kinetic modelling of the propagation of the edge-localized modes in the scrape-off layer of the COMPASS-U tokamak           

Barbora NG  Alloc=10300;  Karolina CPU  Alloc=70600  

Magnetically confined fusion (MCF) plasma devices represent a promising way to a practically unlimited and environmentally friendly energy sources.  At present, number of MCF experimental devices are under operation and some are under construction – one of such devices is the COMPASS-U tokamak being built at the Institute of Plasma Physics of the Czech Academy of Sciences. COMPASS-U is a middle size, high magnetic field tokamak, where among others one of the main problems of MCF will be studied – plasma and power exhaust. The subject of the proposed project is predictive and fully kinetic modelling of the edge-localized modes (ELM) and their propagation in the scrape-off layer of the COMPASS-U tokamak, when highest power and plasma fluxes to the plasma-facing components are expected. Modelling of plasma transport in the COMPASS-U SOL will directly contribute to the machine design and optimization of the discharge scenarios and will provide a rare study of the ELM propagation in high density plasma. For the simulations we intend to use the 1D3V electrostatic particle-in-cell Monte Carlo code BIT1.

Researcher: Tereza Neuwirthová    

OPEN-36-36  Prediction of Amino Acid Dispensability in Model Organisms   

Barbora GPU  Alloc=1700;  Karolina CPU  Alloc=59100  

Does current life need all 20 amino acids? Let's assume that the answer is no. We think that organisms with fewer amino acids likely existed in the past and life's protein alphabet evolved to its current 20-amino-acid form only shortly before the Last Universal Common Ancestor (LUCA). Building on this assumption, we ask two more questions. Which current amino acid would be easiest to replace? What types of replacements would work best? These questions are important for two reasons. First, we don't understand why all life uses the same 20 amino acids despite the wild variety of life forms. Second, the answers could guide future protein design and artificial life projects.

Researcher: Pavla Hrušková              

OPEN-36-37 

Optimized Matrix Decomposition in Parallel BEM Computations Using Combinatorial Designs (part II)     

Karolina CPU  Alloc=3600   

This application builds on our previous application, Optimized Matrix Decomposition in Parallel BEM Computations Using Combinatorial Designs. We continue to develop the parallel library BEM4I [5], specifically focusing on more efficient distribution of computational load between processors. The originally implemented approach used cyclic decomposition [6] in combination with a heuristic approach [3], which, however, may not be ideal in terms of memory requirements (many small blocks) and communication between processors. We now propose to use decompositions of complete graphs into cliques K₃ and K₄ [1], [2], and higher orders in the future. This approach should lead to more efficient distribution (with fewer large blocks), reduced communication requirements, and thus shorter computation times for complex engineering simulations, especially for larger matrices.

Researcher: Lukáš Eigler      

OPEN-36-38 

LLM-Based Synthetic Data Generation for NLP Metric Validation           

Karolina CPU  Alloc=100;  Karolina GPU  Alloc=6500       

Evaluating generative AI, such as tools for translation or summarization, is currently a costly process because it depends on human experts to judge quality [1]. This project, which forms the basis of a Master's thesis, aims to make this evaluation faster and cheaper by replacing human judges with high-quality synthetic data.  Following up on a Fast Track Access (FTA-25-74), this work uses IT4Innovations' supercomputers to fine-tune [2, 13] Large Language Models (LLMs) on existing datasets containing human feedback [10, 11, 12]. Instead of training models to be perfect, we are training them to generate specific, \damaged\" text, such as translations with controlled errors that mimic human imperfections.  The resulting data will be used to test automatic evaluation metrics [3, 4]. By checking if these metrics react to our synthetic errors the same way they react to real human errors (a method called meta-correlation [5, 6]), we hope to validate a new, efficient way to test NLP systems. This would allow researchers to develop and improve models much faster, without the bottleneck of organizing human evaluation campaigns. 

Researcher: Varun Burde     

OPEN-36-39 

Learning 3D Manipulation Affordances from Egocentric Videos via Semantic Correspondence       

Karolina GPU  Alloc=2900;  LUMI-C  Alloc=7400;  LUMI-G  Alloc=12700            

Manipulation is the core capability of robots, yet performing operations on new objects remains challenging. Humans effortlessly pick up unfamiliar objects by understanding their shape, function, and properties can we give robots this intelligence? In this project, we investigate whether robots can learn manipulation skills by observing human video demonstrations. We create 3D affordance maps capturing how humans interact with objects from egocentric footage. Our approach combines semantic understanding (what the object does) with geometric reasoning (its shape), enabling robots to transfer manipulation skills to new objects intelligently. This promises flexible manufacturing that adapts without reprogramming, with applications in healthcare robotics, warehouse automation, and recycling, making advanced robotics more accessible.

Researcher: Ondrej Chrenko             

OPEN-36-4    

3D radiation hydrodynamic simulations of circumplanetary disks        

Barbora GPU  Alloc=1500;  Karolina CPU  Alloc=6000;  Karolina GPU  Alloc=4900;  LUMI-C  Alloc=1800;  LUMI-G  Alloc=4300

Protoplanetary disks swirling around young stars are the birthplaces of planets. Around a newborn gas-giant planet, gravity can gather gas and dust into a miniature sub-disk, a so-called circumplanetary disk (CPD). The CPD plays a pivotal role in mediating the planet’s accretion, facilitating moon formation, and producing potential observable signatures of planet formation. In this project, we will study (i) the thermal structure of CPDs and their ability to retain dust grains, and (ii) the properties of CPDs (or envelopes) of short-period sub-Neptunes capable of carving gaps in their parent disk. The aim of (i) is to determine whether the dust-to-gas ratio in the CPD is sufficient to drive detectable radio continuum emission, and whether the local abundance of molecular tracers such as CO can increase due to volatile sublimation. The results of (ii) will help us understand the formation pathways of the most common type of exoplanets. Our investigation will be based on 3D radiation hydrodynamic simulations run on the GPU clusters of IT4I. These grid-based simulations will cover a large portion of the parent protoplanetary disk while resolving the circumplanetary environment of an embedded planet using a non-uniform grid spacing. The project is related to the solution of the standard GACR 25-16507S.

Researcher: Karel Sindelka

OPEN-36-40 

Pectin gelation: Computer simulations of polysaccharide-based annealed polyelectrolytes              

Barbora CPU  Alloc=18000;  Karolina CPU  Alloc=14500

Pectin, a natural polysaccharide, plays an essential role in food processing as a gelling agent and is increasingly used in biomedical applications, such as drug delivery. This project investigates pectin gelation mechanisms via computer simulations, addressing a significant gap in understanding the behaviour of pectin-based materials. Pectin gelation reles on electrostatic crosslinking with calcium ions, resulting in stable “egg-box” structures. We therefore use mesoscopic simulations with explicit electrostatics to study interactions between many pectin chains of realistic lengths.

Researcher: Robert Vacha  

OPEN-36-41 

Peptides Controlling Membrane Curvature             

LUMI-G  Alloc=14600

The project investigates self-organization of peptides and proteins on cell membranes. Cell surface is highly dynamic and composed of regions of changing local curvature. Some molecules can sense the curvature, which can lead to their self-organization, while other molecules can generate curvature to reshape the membrane surface. This capability is essential for processes like endo- and exocytosis, cell division, and signaling. However, viruses can also exploit protein organization and membrane architecture to hijack cellular machinery, facilitating their budding, infection, and replication. Despite the crucial importance of membrane curvature, we still lack understanding and thus control of protein localization at curved membranes due to the complex interplay between protein (shape and sequence) and membrane (curvature and composition) characteristics. We will use computer simulations to elucidate the interactions at molecular level to both predict and be able to design peptides sequences which are able to sense or generate membrane curvature. Our main objective is to better understand curvature-driven protein organization in both physiological processes like cell signaling, and pathological processes like infections caused by enveloped viruses.

Researcher: Rabindranath Lo           

OPEN-36-42 

Free Energy Investigation of Hydrogen-Bonded Systems in Solvent Environments     

Barbora CPU  Alloc=5000;  Barbora GPU  Alloc=100;  Karolina CPU  Alloc=26300;  Karolina GPU  Alloc=1900;  LUMI-C  Alloc=1000;  LUMI-G  Alloc=5200 

The project aims to improve our understanding of how solvent polarity influences the stability of hydrogen-bonded complexes using computational methods. Hydrogen-bonded systems play a crucial role in many chemical, biological, and material processes, and understanding their behavior in solvent environments is essential for accurate modeling. Free energy analysis provides insights into the stability, interactions, and dynamic properties of such systems, particularly in complex environments like solvents. The computational analyses will be closely integrated with experimental work conducted by experimentalists utilizing state-of-the-art techniques. This collaboration has the potential to significantly contribute to our comprehension of how solvents impact the stability of complexes, with broad practical applications across diverse fields. Utilizing specific DFT functionals, the project will assess the electronic properties of different complexes, considering both implicit and explicit solvent environments.

Researcher: Jiri Tomcala      

OPEN-36-43 

Gaussian Boson Sampling for Earth Observation (GBS4EO)      

Barbora CPU  Alloc=18000;  Barbora FAT  Alloc=90;  Barbora GPU  Alloc=900;  Barbora NG  Alloc=2700;  Karolina CPU  Alloc=23000;  Karolina FAT  Alloc=90;  Karolina GPU  Alloc=2800

The purpose of this project is to test the developed GBS4EO libraries in an HPC environment. In the future, these libraries, running on HPC, together with the quantum part, will be integrated into the overall hybrid algorithm of GBS4EO project.   The primary objective of GBS4EO project is to build end-to-end solutions that benefit from the Gaussian Boson Sampling (GBS) [1] for satellite data analysis. GBS outputs cannot be simulated by classical computers in polynomial time, but recent experiments with programmable optical circuits have demonstrated GBS's advantages [2]. Consequently, GBS presents a validated method for surpassing classic capabilities, with tangible implementations. It is distinct amidst ongoing efforts to realize practical quantum computers. While GBS does not provide a foundation for universal QC, it excels in specific tasks (e.g., graph problems [3]), which we build upon in GBS4EO.    [1] Aaronson, Scott, and Alex Arkhipov. Proc. of ACM symposium on Theory of computing. 2011.  [2] Madsen, Lars S., et al., Nature 606.7912 (2022): 75-81.  [3] Yu, Shang, et al., Nature Computational Science 3.10 (2023): 839-848. 

Researcher: Orsolya Morvai              

OPEN-36-44 

Advanced Optimization of Betatron Radiation via Nanoparticle-Assisted Electron Injection and Tailored Laser Pulses 

Barbora CPU  Alloc=6500;  Karolina CPU  Alloc=4700     

Given the growing demand for advanced imaging and analysis tools, there is an increasing need for compact, high-brightness X-ray sources for applications such as medical imaging and materials characterization. Developing compact alternatives to large synchrotron and FEL facilities has therefore become a major scientific goal. One promising route is betatron radiation from laser wakefield acceleration (LWFA). Nanoparticles embedded in an underdense plasma, together with tailored laser pulses, offer a way to control electron injection into relativistic plasma waves and influence the acceleration mechanism and electron motion, with the potential to improve beam emittance, injected charge, and bunch stability. These advances directly affect the resulting betatron radiation. We suggest that nanoparticle-induced microbunching and resonant betatron dynamics may open opportunities for coherent soft X-ray emission. This project will use Particle-In-Cell simulations to uncover how nanoparticles and tailored pulses shape LWFA and its X-ray emission. The required computational resources will enable accurate modeling of the plasma-wave dynamics and resulting radiation, providing a framework for stable, tunable LWFA-based X-ray sources and supporting future experimental development at ELI Beamlines. To improve challenging and time-consuming parameter scans, we are considering machine-learning techniques, specifically Bayesian optimization, to increase efficiency.

Researcher: Dominik Čáp   

OPEN-36-45  Multi-parametric optimization of laser-plasma accelerators for enhanced betatron X-ray radiation          

Barbora CPU  Alloc=2650;  Karolina CPU  Alloc=3900;  Karolina GPU  Alloc=1300     

Demand for compact particle accelerators and X-ray sources is rapidly increasing in fields such as medical imaging, materials science, and chemistry. Large synchrotrons provide high-brightness X-rays but are expensive to build and operate. Laser wakefield acceleration (LWFA) offers a promising alternative: an intense laser pulse drives a plasma wave that accelerates electrons to relativistic energies over millimeter–centimeter scales. These electrons emit betatron radiation – bright, broadband, femtosecond X-ray pulses generated by their oscillations in the plasma wake.  LWFA betatron sources already show potential for phase-contrast imaging and micro-CT, but their photon flux and tunability remain below those of synchrotrons. Many enhancement methods have been proposed. This project focuses on accelerator-radiator schemes, which separate acceleration and radiation in two plasma stages, and on parametric resonances, where periodic density modulation can amplify electron oscillations.  Optimizing these complex configurations has led to growing use of machine learning in LWFA research. We will apply Bayesian optimization, a probabilistic technique, to improve X-ray flux, spectral control, and source stability. The work will support ongoing experiments and strengthen collaboration with external users of the betatron beamline at ELI Beamlines in Czechia. 

Researcher: Matúš Kaintz   

OPEN-36-46 

Ab Initio Modelling of Diamond-Based Photovoltaic and Spintronic Devices (AIM-PhotoDiaSpin)  

Barbora CPU  Alloc=15700;  Barbora NG  Alloc=3500;  Karolina CPU  Alloc=4100      

As demand for clean energy continues to grow, developing sustainable and highly efficient energy-harvesting technologies is increasingly important. Intermediate band (IB) photovoltaics offers a promising avenue to surpass the efficiency limits of traditional solar cells by harnessing sub-band gap photons[1-3]. In this project, we explore diamond as a platform for such devices, making use of phosphorus-divacancy defect complex that introduces multiple IBs within the wide band gap.  Moreover, in connection with our previous research, this enables us to analyse the performance of a photovoltaic PIN device, in which all regions are doped with the same dopant. This may significantly simplify device manufacturing[4].       Beyond their optical properties, the phosphorus-based defects exhibit stable spin-polarisation, making them attractive for applications in quantum computing and nanoscale magnetic-field sensing. Using advanced ab initio calculations including GW corrections, Bethe-Salpeter equations, and electron-phonon coupling, we will evaluate their optical, electrical, and spin-related properties. This comprehensive approach enables us to assess the potential of such diamond defects for both high-efficiency photovoltaics as well as quantum-technology applications. 

Researcher: Oldřich Plchot

OPEN-36-47 

Novel objective functions for biometric verification leveraging alpha divergence        

LUMI-C  Alloc=1900;  LUMI-G  Alloc=12700           

We propose to address the growing need for robust embedding extractors designed for face and speaker verification systems by exploring alternatives to dominant margin-based softmax losses, such as CosFace and ArcFace.   Recent studies have proposed a promising alternative to cross-entropy, the most popular loss function for classification tasks, namely the α-divergence loss. The new alternative recovers cross-entropy and other loss functions as special cases. Moreover, it allows for a non-uniform reference measure (encoding prior class probabilities), making it more flexible. In face recognition and speaker verification, margin-based loss functions are a key component for training performant contemporary embedding extractors. Integrating angular margin into α-divergence loss, therefore, holds promise. However, the integration remains a challenge. Our primary goal is to successfully combine the two paradigms and demonstrate improvements over current best practices, with a particular focus on reducing false acceptance rates (FAR)—a key metric in high-security applications, such as banking authentication. By systematically exploring and validating these methods, this research aims to advance the state of the art in embedding extraction designed for secure biometric verification systems.

Researcher: Martina Ćosićová        

OPEN-36-48

Testing of ANN diabatization approach on (N₂He)⁺             

Karolina CPU  Alloc=3200   

This project is focused on the use of neural networks for approximation of the diabatic representation of the potential energy matrix (DPEM), which forms an essential part of system of partial differential equations for the motion of atomic nuclei. This system is obtained by Born-Oppenheimer separation of the electron and nuclear degrees of freedom in time-dependent Schrodinger equation. Construction of the DPEM is a complicated problem and most of the standard methods are based on knowledge of various physical properties and often ‘taylored’ to a specific system under study. Neural networks diabatization (DNN) is an alternative approach which could significantly reduce the computational demands of dynamical simulations of quantum systems.  In this project, I would like to test the DNN method on examples of one-dimensional and two-dimensional cuts of  (N2He)+ energy. This problem is part of my dissertation, which I am working on as part of a cotutelle agreement in collaboration with Université de Toulouse, the project is supported by the international Barrande Fellowship program.

Researcher: Dana Nachtigallova   

OPEN-36-49  Computational Study on Photocatalytic Water Splitting over TiO₂ Surfaces

Barbora CPU  Alloc=51300;  Barbora FAT  Alloc=350;  Barbora NG  Alloc=10400;  Karolina CPU  Alloc=52200;  Karolina FAT  Alloc=100;  Karolina GPU  Alloc=7;  LUMI-C  Alloc=31700;  LUMI-G  Alloc=13900 

Hydrogen production via photocatalytic water splitting has garnered significant attention in pursuing sustainable energy solutions. Despite the immense efforts, most of the existing photocatalysts suffer from low activities, a narrow range of absorption, and low solar energy conversion efficiency. The main obstacle to obtaining an effective TiO¬2 surface variant is the lack of knowledge of the complex nature of different surface types, including their detailed geometry and electronic character. Our studies aim to fill this gap by collaborating closely with experimental partners using state-of-the-art techniques. This collaboration has the potential to contribute to the realization of the hydrogen economy with great applicability in various fields and recognize the challenges that must be dedicated the drive the area further.

Researcher: Jhacson Andres Meza Arteaga            

OPEN-36-5    

Dynamic 3D reconstruction of natural scenes      

Karolina CPU  Alloc=1000;  Karolina GPU  Alloc=2400    

Reconstructing the 3D world as it changes over time is essential for technologies like virtual and augmented reality, robotics, and autonomous driving. Today’s methods can capture moving people, cars, or animals, but they largely ignore one of the most common sources of motion around us: nature itself. Moving trees, flowing water, and shifting vegetation remain almost completely unexplored in current research. This project aims to fill that gap by developing new methods to model and even predict how natural scenes evolve. By building fast, modern 4D reconstruction techniques such as Gaussian Splatting, we will learn how these natural motions behave, and we will predict the appearance of them under any season or lighting condition. Because many of these movements are roughly periodic, we will also explore the idea of learning probability distributions that capture these features, enabling realistic forecasts of how scenes will change in the future. Modeling changes in nature is important because it can advance realism in AR/VR environments and support digital preservation of natural sites. We can also introduce new algorithms and datasets for research that can be useful for the computer vision community.

Researcher: Martin Čadík    

OPEN-36-50 

Camera Orientation Estimation using Machine Learning Methods        

Karolina CPU  Alloc=400;  Karolina GPU  Alloc=1500       

Many times, we look at a photo and ask: Where was this photo taken? A related task is camera pose estimation, which finds a camera’s position and orientation in a known scene, while geolocalization finds global coordinates. Both require estimating camera orientation, described by three angles. In this project, we aim to solve this orientation estimation problem.    We use two inputs: a synthetic DEM and a real query image showing part of that DEM. The task is to determine the query image’s location inside the DEM using pitch, yaw, and roll. Mountain scenes make this difficult because conditions such as snow and seasonal changes significantly alter the appearance of the terrain.    The current SOTA method from 2018 has not been surpassed. Our goal is to develop a hybrid CNN–transformer or pure transformer model that achieves higher accuracy and speed. Unlike the SOTA method, our approach does not require FOV, which is often missing in freely available photos. 

Researcher: Debashree Manna      

OPEN-36-51 

Computational Modeling of Solvent Effects on Non-Covalent Interactions     

Barbora CPU  Alloc=5000;  Barbora GPU  Alloc=100;  Karolina CPU  Alloc=8200;  Karolina GPU  Alloc=4900;  LUMI-C  Alloc=8400;  LUMI-G  Alloc=3900 

Non-covalent interactions play a crucial role in catalysis, biochemistry, and supramolecular chemistry. Their strong directional nature makes solvent effects particularly significant in non-covalent systems.1,2 However, accurately describing these systems theoretically is quite challenging. Conventional implicit solvation models often struggle to capture these directional effects, especially in hydrogen-bonding interactions. Therefore, explicit solvent models, particularly for the first solvation shell, are essential in these cases. While low-level DFT, QM-MM and semi-empirical approaches can be effective with a sufficiently large number of explicit solvent molecules, achieving high accuracy with more advanced ab initio methods is often difficult.3,4 To enhance accuracy in large systems, we propose utilizing recent advancements in Machine Learning Perturbation Theory. By integrating quantum chemistry and machine learning, this work aims to establish the foundation for next-generation computational solvation models for non-covalent systems. This approach is expected to deepen our understanding of non-covalent interaction dominated enzyme catalysis, molecular recognition, and functional materials.

Researcher: Tomáš Hrivnák

OPEN-36-52 

Data-Driven Insights into Structured PAH Models for Carbon Dot Photophysics         

Karolina CPU  Alloc=5200;  LUMI-G  Alloc=3300 

The design of novel metal-free photocatalytic materials relies on a clear understanding of the photophysical processes that govern their performance. In recent years, carbon-based nanomaterials, particularly carbon dots (CDs) and carbon quantum dots, have attracted significant attention owing to their low toxicity, high photostability, good water dispersibility, and tunable optical properties. However, their practical application remains limited by structural complexity and varied molecular composition, which hinder detailed characterization and predictive modeling.  To address this, polycyclic aromatic hydrocarbons (PAHs) of varying size, heteroatom doping, and functionalization are increasingly used as model systems to mimic the structural and electronic properties of CDs. Building on this concept, our project proposes stacked PAH donor–acceptor (D–A) complexes as well-defined molecular analogues of CDs. By elucidating the photophysical behavior of these D–A systems, we aim to advance understanding of the fundamental processes and structure–property relationships that drive photocatalytic activity in carbon-based materials. Integrating machine-learning (ML) tools for large-scale data analysis will enable the rational design of PAH-based materials with tailored photoactive properties and inform strategies for precursor selection and post-synthetic modification toward efficient, light-driven energy conversion.

Researcher: Martin Zelený   

OPEN-36-53  Machine-Learning Potentials for Ni–Mn–Ga Shape Memory Alloys        

Barbora CPU  Alloc=20000;  Barbora GPU  Alloc=1000;  Karolina CPU  Alloc=34000;  Karolina GPU  Alloc=3700;  LUMI-C  Alloc=5500  

Using machine-learning methods, it is now possible to perform realistic simulations of complex materials at finite (non-zero) temperatures with an accuracy comparable to that of computationally intensive ab initio approaches. This project will apply machine-learning techniques to investigate thermally induced structural changes in Ni–Mn–Ga magnetic shape-memory (MSM) alloys. These alloys exhibit spontaneous macroscopic deformation under an external magnetic field, known as magnetic-field-induced strain (MFIS), which enables applications in actuators, sensors, energy harvesters, and magnetic refrigeration systems. Recent experimental studies suggest that MFIS is closely linked to the temperature-dependent evolution of elastic instabilities and lattice modulation in the martensitic phase of the alloy, manifested through the shuffling of specific lattice planes. The aim of this project is to develop an interatomic potential capable of accurately describing the lattice modulation and its mechanical response under loading at finite temperatures. Such a potential will enable large-scale, long-timescale simulations that provide deeper understanding of the microscopic mechanisms governing MFIS and guide the design of improved MSM materials.

Researcher: Miroslav Voznak           

OPEN-36-54 

AI-assisted Post Quantum Cryptography for NextG Communication Networks            

Karolina CPU  Alloc=11500;  Karolina FAT  Alloc=100;  Karolina GPU  Alloc=1700       

This project builds a practical testbed to evaluate post-quantum cryptography (PQC) for AI-enabled IoT devices in future 6G networks. Over three years, we will (1) survey and select NIST-standard PQC schemes, (2) implement and measure their latency, energy, and memory on representative hardware (Raspberry Pi, ESP32, STM32), run large-scale simulations, and (3) develop and produce models and deployment guidelines. Methods combine literature review, controlled laboratory experiments, large-scale simulations, mathematical performance modeling, and integrated end-to-end emulation. Deliverables, such as open benchmarks, device-specific configurations, and policy recommendations, will help telecoms, manufacturers, and regulators secure data against future quantum threats while  keeping services fast and energy-efficient.

Researcher: Michael Matějka           

OPEN-36-55 

Impact of the ongoing climate change on temperature and moisture of Antarctic permafrost           

Karolina CPU  Alloc=25400

The project is focused on projections of future development of permafrost (soil or rocks frozen for ≥2 years) in the Antarctic Peninsula region. This region is known for high sensitivity to climate fluctuations and has already experienced pronounced warming, glacier melt and ice shelf break-ups since late 20th century. Besides these changes, the shifts in permafrost temperature and moisture impact not only the soil itself, but also hydrological processes and sensitive polar ecosystems. In this project, evolution of permafrost and its active layer (the top layer which thaws in the polar summer) until the year 2100 will be assessed. The investigation will consider multiple climate change scenarios using a very high-resolution configuration of the Weather Research and Forecasting model. The model output will be validated with in-situ observations from James Ross Island, available through the Czech Antarctic Research Programme. The project is expected to significantly improve our understanding of rapidly changing environment of the Antarctic Peninsula region.

Researcher: Lukasz Rafal Bujak     

OPEN-36-56 

HPC -Accelerated SOFI-Like Processing of Ultrafast Label-Free iSCAT Microscopy Data

Karolina GPU  Alloc=2600   

Imagine being able to watch tiny moving parts inside living cells in real-time, without needing any labels. Our project will do exactly that using interferometric scattering (iSCAT) microscopy, a cutting-edge imaging method that detects the faint light scattered by nanoscale objects inside cells. This label-free technique offers nanometer-scale sensitivity and can capture events at ultra-high speeds (up to millions of frames per second) – far beyond the reach of traditional microscopes. We focus on live MCF7 cancer cells, where we record movies of intracellular motion (such as motor proteins moving along filaments) at unprecedented frame rates. The challenge is making sense of these enormous movies, each containing millions of frames (tens of gigabytes of data). We will tackle this by applying super-resolution optical fluctuation imaging (SOFI)-like analysis – a computational method that finds hidden details by analyzing random intensity fluctuations in the image sequence. By computing statistical measures (correlations and cumulants) of each pixel’s intensity over time, we can highlight dynamic structures and potentially achieve resolution beyond normal limits, all without phototoxic dyes or photobleaching. This means we can observe delicate cellular processes for longer durations and at higher speeds than fluorescent microscopes allow. To process such massive datasets quickly, we will harness powerful GPU-based supercomputers. In summary, this project merges an innovative label-free microscope with advanced data analysis and high-performance computing to reveal the nanoscale dance of life inside cells, opening new avenues for biological discovery and medical research.

Researcher: Paulo Miguel Guimarães da Silva     

OPEN-36-57 

Traffic Modelling for Improved Urban Mobility: Intelligent Routing and Scalable Computation          

Barbora CPU  Alloc=3000;  Barbora NG  Alloc=7000;  Karolina CPU  Alloc=3600;  LUMI-C  Alloc=7000       

Urban traffic systems are complex, large-scale and difficult to analyse without efficient simulation tools. High-performance computing plays a central role in enabling realistic modelling of routing behaviour, congestion dynamics and the effects of real-time information on citywide mobility. This proposal builds directly on the outcomes of our previous project “Intelligent Traffic Modelling for Improved Urban Mobility: Towards Traffic Routing Equilibrium”, during which we optimised key components of the Ruth traffic simulator, developed the ACE distributed execution library, and produced results now presented in our SC25 poster “Scalable Alternative Route Computation with ACE: A C++17 Library for HPC Traffic Simulations”. We are now requesting computational time to continue the optimisation work and expand the scientific studies enabled by the traffic simulator. We will extend ACE to additional functions, refine memory and data-handling structures and ensure suitable performance on systems such as LUMI and the new Barbora-NG. We will also generate extensive simulation data to train machine-learning models that can later be integrated into the simulator as additional decision-making options, complementing traditional routing logic. We will study how routing strategies, re-routing dynamics and partial access to live traffic information influence congestion across diverse cities. We will continue improving our extended Betweenness Centrality approach to better represent network dynamics on different HPC architectures. The project will deliver a faster, more scalable simulator and new insights to support improved mobility, reduced congestion and lower emissions.

Researcher: Samuel Lukeš 

OPEN-36-58 

ERO2.0 simulations of tungsten and liquid metal heat shields on COMPASS Upgrade

Karolina CPU  Alloc=400      

With the ever-increasing energy consumption of mankind, the need to use ecologically and politically acceptable, reliable 24/365 and inherently safe sources also increases. One of the very few ways (if not the only way) to satisfy these requirements seems to be nuclear fusion. That is why its flagship research project ITER is currently the most expensive science experiment on the planet. However, several unanswered questions still remain, one of them is the long-term reliable heat shield of the reactors. Current tungsten (W) shields must withstand extreme heat fluxes from thermonuclear plasma, which erode, degrade and melt any crystalline lattice. Melting could be solved by replacing W with liquid metal (LM) technology in the most loaded areas. However, both options can still pollute the confined plasma, where even at very low concentrations of impurities from heat shields most of the energy supplied to the plasma will be radiated away and the ongoing nuclear fusion will be interrupted.    This project simulates the erosion and transport of W and LM heat shields in the new COMPASS Upgrade tokamak built in Prague. Parameters of fully W COMPASS Upgrade corresponds to future reactors, and any research on it in this sense will be globally unique. 3D Monte Carlo code ERO2.0, which solves the transport and plasma-wall interactions of impurities in the region between the confined plasma and the solid wall of current fusion devices, is used for this purpose.

Researcher: Jan Rezac           

OPEN-36-59 

Extending the applicability of hybrid quantum-mechanical / machine learning computational method for biomolecules        

Barbora CPU  Alloc=32000;  Barbora NG  Alloc=200;  Karolina GPU  Alloc=4400         

We are developing a novel computational chemistry method based on semiempirical quantum-mechanical (SQM) calculations and state-of-the-art machine learning. The present version of the method, PM6-ML, has already been published. In terms of accuracy, PM6-ML outperforms both all the existing SQM methods, as well as the standalone ML potentials, and opens applications to large molecular systems. We are now extending it towards the description of biomolecules in solution with direct applicability in computer-aided drug design. Currently, no other ML method has these capabilities, which makes our approach unique. In the previous IT4I project, we computed the core training set necessary for its development and trained a prototype of the method. In this project, our goal is to expand the training set by using an active learning loop over large molecular databases. This will extend the scope of the method to a more diverse chemical space and increase its robustness.

Researcher: Martin Friak      

OPEN-36-6    

Exploring limits of the hydrogen storage capacity in La-Ni-Sn class of materials         

Barbora CPU  Alloc=39000;  Karolina CPU  Alloc=50000

Hydrogen is considered one of the clean fuels of the future, but storing it safely and efficiently remains a major challenge. Our proposed research is focused on a promising material LaNi₅, which can absorb quite significant amounts of hydrogen but yet more is needed. We suggest achieving the higher hydrogen absorption capacity by altering the chemical composition. By replacing a small portion of the nickel atoms with tin atoms, we hope to understand (at the level of individual atoms) how these substitutions affect the material’s ability to store hydrogen. Using advanced quantum-mechanical simulations, we will explore the fundamental limits of how much hydrogen this material can hold and how thermodynamically stable it remains. Insights from the planned research could help guide the development of safer, more efficient hydrogen-storage technologies that support a cleaner and more sustainable future energy industry. Our project is a part of Operational Program Jan Amos Komenský (OP JAK), in particular the project called “Materials and technologies for sustainable development” with the Technical University in Ostrava (VŠB-TUO) being the coordinator and the Institute of Physics of Materials of the Czech Academy of Sciences (IPM-CAS) in Brno being one of the partners responsible for both experimental and theoretical research of hydrogen-storage materials.

Researcher: Taoufik Sakhraoui        

OPEN-36-60 

Graphene allotropes and goldene 2D materials  

Barbora CPU  Alloc=25000;  Karolina CPU  Alloc=14500

In the modern age of nanotechnology, the discovery of graphene has opened up the way to study and develop of several novel 2D materials due to their unique physical and chemical properties, which makes them superior to the commercial bulk materials used in various applications. It was already shown that the size of a material affects its properties and behavior. Hence, the electrical, chemical, mechanical, and optical properties of a material improve drastically when the dimensions reduce to nanoscale. Currently, intensive research into 2D materials is expected to lead to the discovery of new materials with enhanced properties that will benefit the industry and society at large. On the other hand, accurate theoretical description of 2D materials remains challenging. Due to the presence of multiple configurations and magnetic alignment for each of them, and the need for accurate predictions of their properties, supercells with hundreds of atoms are required. Recently, a new scheme has been pioneered and combined in the DFTB method with the use of the extended tight binding Hamiltonian (xTB). The general applicability together with the excellent cost-accuracy ratio and the high robustness make the xTB family of methods very attractive for various fields of computer-aided chemical research.

Researcher: Dominik Farka

OPEN-36-61 

Effects of Functionalization and Adsorption onto Functional Surfaces of Conductive (Bio-Inspired) Polymers II     

Barbora CPU  Alloc=35100;  Karolina CPU  Alloc=17800

This is a follow-up project to the project submitted in the 34th call as the requested hours did not suffice.  Interrim report: With 1800 hours we were able to run calculations for one manuscript and significantly improve the assembly of adamantyl-substituted polythiophenes to the point where we understand their self-assembly types. We now ask for significantly more hours to be able to fulfil the needs of our bio-inspired conductive polymers for our future project and finish the commenced work from the first project.  Two parts are to be solved: a follow-up study on our recent joint experimental and theoretical work on polythiophene and polypyrazine based systems and secondary will also want to extend our investigation in terms of conductive polymers to the indole-ring system.  The first part, our follow-up study will allow for a more systematic, in-depth study of PEDOT (polyethylendioxythiophene), PEDTT (polyethylendithiathiophene) and PTP (polythioenopyrazine) polymers. We intend to use quantum chemical methods to investigate their frontier orbitals, conformational stability, and their charge-transport. We want to focus on the effects of different functionalizations and/or doping. Further, we want to elucidate the effects of interaction of these polymers with different surfaces, which we have shown experimentally to have tremendous impact. In our recent experimental work, we have investigated oxidative chemical vapour deposition (oCVD) of PTP and inquire to gain insight into the effects of various surfaces (see below). These are expected to act as deposition templates to our polymer.  The second part of the project will investigate the indole-system for its reactivity and potential as a conductive polymer. Substitution effects will be studied by applying the study to various serotonine-like molecules (serotonine, triptans, etc.). The same properties will be studied as in the abovementioned polymers. This is in part to support a submitted GAČR-project (Standard).  This project is meant to be a complement for/basis of the experimental research on these materials. 

Researcher: David Barina    

OPEN-36-62  Recursive sufficiency for the Collatz conjecture and computational verification         

Barbora CPU  Alloc=34700;  Barbora NG  Alloc=5900;  Karolina CPU  Alloc=11100;  LUMI-C  Alloc=17200

In 2025, Mohammad Ansari published a new method for verification of Collatz conjecture. This method has significantly lower computational requirements than the algorithms used so far. We decided to use this method together with supercomputers to further increase the limit below which the convergence of the Collatz conjecture is proven (which is currently 2^71).

Researcher: Štěpán Sklenák             

OPEN-36-63 

Splitting dinitrogen   

Karolina CPU  Alloc=27100;  Karolina GPU  Alloc=2500;  LUMI-C  Alloc=5000

Zeolite based catalysts are the most important industrial catalysts. Zeolites are crystalline microporous aluminosilicates with a unique microporous nature, where the shape and size of a particular pore system exerts a steric influence on the reaction, controlling the access of reactants and products. Periodic DFT methods permit investigations of properties of zeolite-based catalysts which are needed for their fine-tuning. DFT calculations are complementary to experimental examinations and together they can provide more complex knowledge of the properties of the studied catalysts and the reactions they catalyze.

Researcher: Miroslav David

OPEN-36-64 

Benchmarking and Fine-Tuning Vision-Language-Action Models for Multimodal Human-Robot Collaboration

Karolina CPU  Alloc=400;  Karolina GPU  Alloc=1700       

Robots are becoming increasingly important partners in industrial and everyday environments, yet most still struggle to understand human commands naturally. Humans express intent not only through speech, but also gestures, gaze, and facial expressions. This project aims to enable robots to interpret such multimodal commands by leveraging Vision-Language-Action (VLA) models. We will benchmark and fine-tune open-source VLA models using multimodal datasets from real and simulated collaborative tasks. One part of the project will focus on integrating LLMs with analytical trajectory planners to inject semantic intent into geometrically precise motion - e.g., modulating CAD-based solvers using human guidance. The goal is to create a framework that allows robots to combine visual, linguistic, and contextual information to act safely and intuitively around humans. The research contributes to the EU Horizon project ELLIOT, advancing multimodal foundation models for robotics. Results will improve natural communication between humans and robots, fostering safer and more efficient collaboration in industry and service domains.

Researcher: Alejandro Hernandez Tanguma          

OPEN-36-65 

Selective Membrane-Binding Peptides: Computational Design and Experimental Validation for Advanced Biotechnological Applications 

LUMI-G  Alloc=14600             

Membrane-binding peptides have significant potential applications, including use as antimicrobials, biosensors and in the purification of extracellular vesicles. However, to ensure the effective implementation of these peptides in biotechnological applications, it is essential to prioritize their selectivity. This project aims to develop peptides derived from Opi1 that exhibit selective affinity for membranes with varying compositions. As Opi1 has previously been demonstrated to bind to phosphatidic acid (PA), this proposal posits that it is feasible to obtain an Opi1 variant that interacts selectively with CL, as the latter is analogous to PA and is present in some bacterial membranes. The project entails the implementation of in silico mutations of the Opi1 sequence, in order to identify novel sequences with affinity to different membranes. The optimization and analysis of peptide-membrane interactions will be conducted using atomic-resolution molecular simulations, alchemical transformations, and experimental biophysical techniques. This interdisciplinary project integrates biology, physics, and chemistry, leveraging advanced computational and experimental techniques to achieve objectives. The interdisciplinary integration ensures robust outcomes, with potential applications in targeted drug delivery, diagnostics, and synthetic biology, ultimately advancing our understanding of peptide-membrane interactions.

Researcher: Jun Terasaki     

OPEN-36-66 

Establishing the predicted nuclear matrix element of neutrinoless double-β decay  

Karolina CPU  Alloc=28400

There are two kinds of elementary particles, that is, particles and antiparticles. They are expected to exist by the same amount because of the symmetric properties of this pair. However, the reality is different; the antiparticles do not exist stably in the universe. The Majorana neutrino is a hypothetical particle introduced in a theory to solve this mystery of the universe. The weak-interaction decay of an atomic nucleus that does not emit a neutrino (the neutrinoless double-β decay), if found, proves that the Majorana neutrino exists. This decay does not occur without this neutrino. More than 30 experiments in large facilities around the world are in progress or preparation to prove the existence of this epoch-making new particle. Theoretical physics carries the important task of providing the decay probability, which is necessary to learn the properties of the Majorana neutrino when that decay is found. This decay is extremely rare; its half-life is certainly much longer than 10^21 years for Xe-136. Because of this extreme rarity, the reliable prediction of the decay probability has been a challenging problem for theorists for many years. Recently, Dr. Civitarese in Argentina and I (the primary investigator) made a breakthrough to solve this problem in the collaboration. Our long-term project is now at the stage to firmly establish the prediction of the decay probability by finding many more satisfactory examples of our new approach.

Researcher: František Fňukal           

OPEN-36-67 

Validation of the CSD crystal structure database using the DFT method (II)    

Karolina CPU  Alloc=10900

The Cambridge Structural Database (CSD) is the largest repository of organic and metal-organic crystal structures including pharmaceuticals. It can be used in studying molecular and crystallographic phenomena, predicting crystal structures or designing and optimizing pharmaceutical compounds. In our work, we discovered that the CSD potentially contains a large number of erroneous entries. The issues found range from minor inconsistencies to serious errors in crystal structure determination. This is due to poor data validation consequential to the rapid growth of the CSD. To validate a crystal structure, we perform a quantum chemical calculation based on the plane-wave density functional theory (DFT) method. The results of such calculations allow us to see discrepancies indicated by the quantum theory. The goal of our work is to design a validation scheme for the contents of the CSD and ensure that its users can rely on trustworthy, high-quality data.

Researcher: Simona Sajbanova     

OPEN-36-68 

The use of computationally expensive DFT functionals for investigation of salt-cocrystal continuum area II.

Karolina CPU  Alloc=7200   

Pharmaceutical solid forms such as salts and cocrystals play a crucial role in pharmaceutical applications. The difference between salt and cocrystal is given only by the position of single hydrogen (1), making it essential to develop precise techniques for identifying this position.  Differentiation between salt and cocrystal compounds holds significant importance within the pharmaceutical industry, both for regulatory purposes and overall quality control. The Food and Drug Administration has explicitly outlined in their 2018 guidelines the necessity for accurate identification of pharmaceutical phase (2). We are developing a computational method for such hydrogen position determination. The method does not require crystallographic data from high-quality monocrystal, and it work even with data from powder samples.

Researcher: Michal Sykora 

OPEN-36-7    

Fine-Tuning Embedding Models to Improve Retrieval for the Czech Legal Domain using Synthetic Data    

Karolina CPU  Alloc=300;  Karolina GPU  Alloc=6400       

Large language models (LLMs) may struggle with accuracy or outdated information, which can be problematic in tasks where details are crucial, like interpreting legal texts. Retrieval-augmented generation (RAG) is a technique addressing this issue by incorporating a retrieval of documents relevant for the user's query and including them into the context of the LLM. This approach also allows the LLM to reference new data without the need for costly retraining. To find the relevant documents, embedding models are used to generate high-dimensional vectors that semantically capture the contents of the documents, and it is then searched for vectors similar to the vector of the query. [1]  This project focuses on optimizing the retrieval part of this system, specifically for the Czech legal domain. As a baseline, we use existing foundation embedding models, which are multilingual and not trained for a specific task [2,3,4]. We fine-tune these foundation models to improve their performance on Czech legal texts and evaluate them. In addition to existing legal texts, our work includes using powerful LLMs to generate synthetic data, which may serve as a high-quality \ground truth\" for training and evaluation of these models [5,6]. By adapting these technologies, we aim to develop an AI assistant that can provide precise, reliable, and well-cited answers in the Czech legal domain.  "

Researcher: Martin Kiss        

OPEN-36-8    

Fine-Tuning Vision–Language Models for Captioning Historical Images             

Karolina GPU  Alloc=900      

In the research project Orbis Pictus, we develop methods and software tools for processing non-textual elements – such as images, photographs, maps or diagrams – in digitized historical documents, for example those available at digitalniknihovna.cz. Today, these elements are hard to search and often lack explanation. One of the goals of the project is to make them semantically searchable and to automatically generate short textual descriptions that help users understand what they see when browsing digitized pages. In both cases, textual description of these elements is either directly needed, or it can improve current methods. Using large multi-GPU computational nodes allows us to fine-tune existing large models for image captioning to make the models more accurate on historical images. The resulting software and models will be used in libraries and archives in the Czech Republic, improving access to cultural heritage for researchers, students and the general public.

Researcher: Patrik Beliansky            

OPEN-36-9    

Efficient Adaptive Tetrahedra-based Neural Scene Representations    

Karolina CPU Alloc=300;  Karolina GPU  Alloc=6000        

Accurate and efficient 3D scene reconstruction is a fundamental problem in computer vision, with applications in robotics, augmented reality, entertainment, and cultural heritage preservation. Neural Radiance Fields (NeRFs) have recently emerged as a powerful learning-based method for photorealistic 3D reconstruction, but they are slow to train and to render. Hence, the field has mostly moved to Gaussian Splatting, which is faster to train and render, at the cost of rendering quality. Our goal is to investigate ways to make NeRFs more efficient without reducing rendering quality. We will build on geometry-aware scene representations that use available geometric priors. They are represented using triangles and can be rendered very efficiently. In this project, we aim to exploit this property and make rendering of neural radiance fields (and thus training them) significantly faster.

Researcher: Martin Kostelník            

FTA-26-5          

Pretraining of Deep Neural Network for Document Understanding       

Karolina GPU  Alloc=500;  LUMI-G  Alloc=200       

Historical documents hold invaluable insights into our cultural and societal past, but many of them remain locked away in scanned pages that are difficult for computers to read and understand. Our project aims to develop an advanced artificial intelligence model that can automatically analyze and interpret these documents using state-of-the-art methods in multimodal machine learning. Specifically, we will pre-train a transformer-based model capable of understanding both the visual layout and textual content of historical Czech documents. The training will be done in a self-supervised manner. This research is part of an ongoing collaboration with Czech libraries, which have provided large collections of scanned materials. The results will contribute to better tools for historians, librarians, and the general public, enabling easier access, search, and understanding.

Researcher: Tobiáš Kořán    

FTA-26-6          

A Quest for Secondary Structure in Short Peptides: Combining (Machine-Learned) Quantum Mechanics with NMR and CD Spectroscopy     

Barbora CPU  Alloc=1400;  Barbora GPU  Alloc=60           

Proteins and peptides are biological molecules with vital roles in all living organisms. Protein chains comprised of amino acids (AAs) must fold into their 3D structure, uniquely determined by the sequence of AAs, to function properly. Consequently, this folding determines many properties of proteins, including interactions with other molecules, binding of metal ions or catalytic activity. The druggability of proteins is also assessed through understanding of the final structure.  The goal of the project is to quantify the patterns of secondary structure formation in smaller units, oligopeptides, from first principles. The bottom-up approach to computational 3D protein structure prediction presented in this project consists of large-scale quantum chemical calculations coupled with modern solvation methods, augmented by machine-learned neural networks. Our newest NequIP-based model for predicting conformational peptide energies in solvent drastically spurred initial phases of secondary-structured short peptide design. However, numerous selected candidates must undergo expensive rigorous quantum chemical calculations. The verified peptide potentials will enable accurate conformational energy estimations of hundreds of millions of structures at a fraction of the usual processing power. As it stands, the project promises a computationally tractable way to significantly expand the understanding of protein folding with downstream applications in the chemical and biochemical industry.