38^th open access competition RESULTS

We would like to thank all applicants for their strong interest in computational resources in the recently concluded 38th Open Access Grant Competition.

A total of 80 project proposals were submitted, with a combined request for 21,405,102 normalised core hours, corresponding to approximately 3.17 million node hours. Demand for computing capacity once again significantly exceeded the available resources. For the Karolina CPU, Karolina GPU and LUMI-G systems, the combined demand exceeded twice the capacity available for allocation. In accordance with the competition rules, requests exceeding 10% of the total capacity of a given resource were reduced to the specified limit.

All projects underwent a multi-criteria evaluation, taking into account the technical quality of the proposal, scientific excellence, project readiness, publication record, expected socioeconomic impact, and the efficiency of using previously allocated resources. Publication history was assessed based on the ratio of publications to completed projects over the past three years, with at least one publication per project considered the expected standard.

Given the exceptionally high level of competition, computational resources were allocated primarily to projects with the highest overall evaluation scores and demonstrated readiness to make effective use of the allocated resources. Three projects submitted by new users received allocations corresponding to the limits set for FTA projects. One application from a new applicant was not supported due to its low evaluation score. We recognise that this approach may have placed some limitations on new users, and we apologise for any complications this may have caused.

Of the 43 projects evaluated, 24 received 25 or more points out of a maximum of 30. No project received the full 30 points.

In this round, a total of 1,842,900 node hours were allocated to 71 OPEN projects, 11 of which are multi-year projects. Nine applications were not supported because they did not meet the evaluation criteria, particularly in terms of publication results, the use of previous allocations, or the overall evaluation score.

For the second time, the grant competition also included the allocation of QPU seconds on the VLQ quantum computer. Six PIs and their teams applied for this resource, with a total of 1,318,320 QPU seconds allocated.

We congratulate all successful applicants and thank everyone for their confidence in IT4Innovations' computing infrastructure. We look forward to the scientific results and new insights that the supported projects will deliver.

THE ALLOCATION COMMISSION DECIDED ON THE ALLOCATIONS WITHIN THE 38^th OPEN ACCESS GRANT COMPETITION AS FOLLOWS:

Researcher: Michal Sykora

OPEN-38-1    

Temporal Fact Invalidation in Agentic Knowledge Bases for the Czech Legal Domain

Karolina CPU Alloc=600; Karolina GPU Alloc=5600          

Large Language Models (LLMs) are evolving into autonomous agents capable of solving complex tasks. However, they struggle to process and remember massive amounts of information, especially when facts change over time. In the legal system, when a new amendment is passed, specific sections of older laws become invalid. Current AI systems, which mostly rely on keyword and semantic searches, often get confused by this \fact invalidation\" and provide outdated, and therefore incorrect answers. Our project addresses this problem by testing and further improving advanced AI memory systems that understand relationships between facts and their timeline (temporal validity), focusing on Czech legal texts. We will evaluate how well open-weight AI models coupled with these advanced memory systems, all deployed on local hardware to ensure privacy, can accurately answer legal questions when laws change. Crucially, we will optimize these systems by tuning knowledge base configurations, refining agent skills and prompts, and enhancing linguistic analysis for Czech to maximize their performance. Our research aims to create a framework for reliable and secure AI assistants capable of navigating complex information without hallucinating outdated facts, not only for lawyers, civil servants, and the general public, but also extendable to other domains and use cases."

Researcher: Valentína Berecová   

OPEN-38-10  

Theoretical investigation of defect-induced magnetism in WO₃ thin films        

Barbora NG Alloc=6900; Karolina CPU Alloc=33800; Karolina GPU Alloc=700              

Tungsten oxide (WO₃) is a wide-band-gap semiconductor already explored in sensors, light detectors and memory devices. Its properties are highly sensitive to structure, oxygen content and physical form, making it attractive for targeted tuning. WO₃ is nonmagnetic in its ideal bulk form. Surprisingly, experiments show that thin WO₃ films can become ferromagnetic at room temperature even without doping with magnetic elements. This unusual combination of semiconducting and magnetic behaviour makes WO₃ interesting for spintronics, which uses electron spin alongside electric charge and could enable faster or more energy-efficient devices. What causes this unexpected magnetism and how can it be controlled?  This project will use quantum-mechanical simulations to reveal how atomic-scale defects alter the magnetic behaviour of WO₃ films. We will investigate tungsten and oxygen vacancies, their charge states,  concentrations and positions within the film, together with the effects of strain and substrate interfaces. Our preliminary results indicate that defects near surfaces and interfaces are crucial, with tungsten vacancies producing a much stronger magnetic response than oxygen vacancies. By identifying stable magnetic defects and determining how they respond to experimentally adjustable factors, we aim to reveal how magnetic states could be favoured during film growth. These results could guide future experiments toward tunable material for spintronic technologies. 

Researcher: Anagha Sasikumar    

OPEN-38-11  

Periodic Relativistic Magnetic Resonance Observations of Solid-State Paramagnetic Perovskites   

Karolina CPU Alloc=27100 

Magnetic resonance (MR) techniques, including Electron Paramagnetic Resonance (EPR) and paramagnetic Nuclear Magnetic Resonance (pNMR), provide vital structural insights for developing tunable optoelectronic materials, such as solar cells. Computational predictions are indispensible in these systems due to their poor spectral resolution. However, accurately modelling MR parameters in extended periodic systems remains theoretically challenging. Because a unified methodology is currently unavailable within a single periodic code, this project develops a robust cross-code solid-state DFT framework utilising periodic boundary conditions (PBC). Quantum ESPRESSO will be employed to perform the fundamental structural optimisations as well as the calculation of the EPR parameters (electronic g-tensors, hyperfine A-tensors), and pNMR shifts. Subsequently, to isolate and evaluate the impact of complex relativistic phenomena, the ReSpect program will be integrated to rigorously assess scalar-relativistic and spin-orbit coupling contributions to the NMR shifts. Using this synergistic approach, we will systematically investigate EPR parameters (electronic g-tensors, hyperfine A-tensors) and pNMR shifts in metal halide perovskites (MHPs) doped with paramagnetic transition-metal ions (e.g., Fe, Cr, Ru). By comparing theoretical simulations with experimental observations, this project provides a predictive computational strategy for deciphering complex NMR spectra, thereby driving the rational design of highly efficient functional materials.

Researcher: Michal Cifra

OPEN-38-12  

Electric-Torque-Driven Destabilization of the Microtubule Lattice: A Multi-Scale Molecular Dynamics Study 

LUMI-C Alloc=600; LUMI-G Alloc=20000  

Cells build much of their machinery from self-assembled protein polymers — in particular the microtubules of the cytoskeleton, hollow tubes that give cells shape and let them divide and move cargo. Medicine already exposes tissue to brief, intense electric-field pulses (cardiac pulsed-field ablation, cancer electrochemotherapy). We recently found, in large simulations, that such fields can twist microtubules apart: each tubulin building block carries a large electric dipole, so a strong field exerts a torque on it, and once that torque outweighs the weak bonds holding the blocks together, the tube wall opens. This project maps, for the first time, when and how the electric torque overcomes the binding between subunits — as a function of field strength and orientation relative to the tube axis — using molecular dynamics on LUMI-G across three scales: a single tubulin block, a whole microtubule ring, and a segment large enough to show how the opening nucleates and spreads. The results will identify field regimes safe versus disruptive for cell architecture, guiding safer medical and biotechnology pulse protocols, and will support two funded Czech national grants.

Researcher: Lukas Neuman            

OPEN-38-13  

Quantum algorithms for Neural Architecture Search      

LUMI-G Alloc=4300; VLQ Alloc=10000      

Deep neural networks have come to dominate virtually every area of artificial intelligence and computer vision. Their central appeal is that given enough training data, an optimisation algorithm adjusts the network parameters to maximise the target objective - classification accuracy, for instance - with no need for hand-designed intermediate representations. Yet this success is remarkably brittle. Models are still derailed by trivial inputs that simply happen to fall outside the training distribution, and even the largest of them, with hundreds of billions of parameters trained on correspondingly vast and diverse data, exhibit striking failures of generalisation and elementary reasoning. One candidate for the missing piece is the architecture itself: the inductive biases built into a network decide what it can generalise from, and unlike its parameters, they are not learned. They are chosen in advance, by hand, and largely by intuition.     Automating that choice is the goal of Neural Architecture search, which aims to automatically find the best network architectures. Contrary to popular beliefs, Neural architecture search spends most of its budget not on training candidate networks but on deciding which candidates deserve training. In predictor-guided NAS this decision reduces to a hard combinatorial subproblem: selecting the k-subset of a large ranked pool that maximises aggregate predicted performance while penalising pairwise redundancy under an architecture-similarity kernel, which is the type of problem that can be efficiently solved by quantum algorithms. 

Researcher: Pavel Krc          

OPEN-38-14  

MicroBUS simulations – final phase            

Karolina CPU Alloc=41100 

The PALM model system allows to perform detailed simulations of conditions in urban areas, mainly with respect to phenomena of urban heat island, thermal comfort and air quality. ICS team significantly contributed to the model development and validation, moreover we use PALM for testing the efficiency of urban climate adaptation measures. The goals of the TACR project MicroBUS (Microscale Based Urban Scenarios, TACR-SQ01010181) include development of new tools to assess the dispersion of traffic emissions in traffic-laden metropolitan areas at high temporal and spatial resolution. Moreover, additional model validation against measurements in a wind tunnel is planned to test if important sizes of eddies, inside and above the urban area, are correctly captured. The results will be used by the Ministry of Environment and the municipal authorities to propose measures to achieve the air quality limits.  As the model is computationally demanding, all these simulations require a large amount of parallel computation power, which exceeds the usual in-house resources. The supercomputer facilities in IT4I allow us to manage this challenge efficiently. Computational resources were tested within the project OPEN 34-31-00. Although we assumed to submit a multi-year project, we decided to request a single-year project. We also need to note that we could not use node-hours on the Barbora cluster due to technical complications of the cluster.

Researcher: Jiri Brabec        

OPEN-38-15  

From downfolded Hamiltonians to the QPU: DMRG, selected configuration interaction and quantum sampling for strongly correlated molecules 

Karolina CPU Alloc=73900; Karolina GPU Alloc=10300; VLQ Alloc=81620       

Iron-sulfur clusters, which occur in the active sites of enzymes such as nitrogenase, belong among the most challenging problems of current computational chemistry. Their electronic structure is dominated by both static and dynamical electron correlations, and an accurate treatment of such systems is expensive.     In this project we focus on how accurately strongly correlated systems can be computed, and what a quantum computer contributes, when the molecule is first compressed by a downfolding technique into an effective Hamiltonian of a few orbitals which nevertheless carries the dynamical correlation from external orbital space. Such a Hamiltonian fits into the 24 qubits of VLQ, the quantum computer operated by IT4Innovations, and it is equally suitable for the selected configuration interaction methods we run on Karolina. DMRG supplies the reference state, which is encoded into a short quantum circuit whose depth is governed by a single parameter, the bond dimension. We will prepare such states on VLQ, measure the configurations which the device returns, employ them in the selected configuration interaction calculation, and establish how deep a circuit the hardware can execute before noise destroys the signal.   

Researcher: Pavel BANÁŠ  

OPEN-38-16  

From Molecular Ensembles to Melting Curves: Quantitative Thermodynamics of Short RNA Duplexes

Karolina CPU Alloc=8500; Karolina GPU Alloc=9300; LUMI-C Alloc=6800; LUMI-G Alloc=13100 

RNA molecules regulate essential cellular processes and are increasingly important targets for diagnostics and therapeutics. As exhaustive sampling begins to cover complete RNA conformational landscapes and force-field development enters a data-assisted era, the predictive power of molecular simulations must be validated against quantitative experimental data using state-of-the-art enhanced sampling and modern data-driven approaches.  This project will investigate the formation of short RNA double helices using an advanced simulation approach called parallel-tempering on-the-fly probability enhanced sampling (PT-OPES). Starting from a pilot simulation of the four-base-pair r(GGCC) duplex, we will study a controlled series of duplexes with four to six base pairs and different sequences. This will reveal how sequence length, base stacking, pairing and end fraying determine duplex stability.  The simulated molecular ensembles will be linked directly to temperature-dependent UV–visible melting experiments. The resulting benchmark will expose where current RNA force fields correctly or incorrectly balance canonical duplexes and non-canonical motifs. The data will support systematic, data-assisted improvement of RNA models and will be prepared for reuse through the Integrated Database (IDA; https://next.ida.4sims.eu) and the ADAMS4SIMS workflow.

Researcher: Athanasios Koliogiorgos       

OPEN-38-17  

Many-Body Excited-State and Magneto-Optical Properties of CrSBr/MoS2 van der Waals Heterostructures              

Barbora NG Alloc=6500; Karolina CPU Alloc=28100; Karolina GPU Alloc=2600; LUMI-C Alloc=9500; LUMI-G Alloc=5900  

Two-dimensional van der Waals heterostructures let us stack atomically thin crystals to create materials with properties that neither layer has alone. A particularly promising building block is CrSBr, a magnetic semiconductor whose light emission is coupled to its magnetic order, so that a magnetic field can be used to tune its optical response. Stacking CrSBr on the nonmagnetic semiconductor MoS2 combines this magneto-optical behaviour with the electronic and valley band structure tunability of MoS2, opening a route to valleytronic devices. In a previous project we mapped the electronic structure of CrSBr/MoS2 at the density functional theory (DFT) level. DFT, however, apart from underestimating energy gaps, cannot describe the bound electron-hole pairs (excitons) that dominate the optical spectra of these materials, nor the way circularly polarised light couples to their magnetism. Here we apply state-of-the-art many-body methods (the GW approximation and the Bethe-Salpeter equation), including relativistic spin-orbit coupling, to predict the true quasiparticle gaps, exciton binding energies, and optical and circular-dichroism spectra of CrSBr/MoS2 and of its isolated CrSBr and MoS2 components in their competing magnetic configurations. The results will be directly compared with magneto-optical experiments carried out by our collaborators.

Researcher: Dominik Legut              

OPEN-38-18  

Strongly-correlated Magnetic Topological Materials by Band Engineering       

Barbora NG Alloc=125400; Karolina CPU Alloc=193500; Karolina FAT Alloc=300; Karolina GPU Alloc=31200; LUMI-C Alloc=31200

This project will exploit the unique combination of 5f‑electron physics and the non‑symmorphic space‑group 220 (I¯43d) to create tunable Weyl‑Kondo semimetals, topological Kondo insulators and related quantum phases in the An₃(T₃)X₄ family (An = actinide/rare‑earth, T = 3d/4d transition metal, X = chalcogen). By systematic compositional, strain‑ and pressure‑engineering we will map the full phase diagram, predict topological invariants, magnetic anisotropies and many‑body spectra, and guide the synthesis of bulk crystals and epitaxial thin films. High‑throughput DFT + U, DFT + ED and fully charge‑self‑consistent DFT + DMFT calculations will be performed on the national super‑computing facilities (Barbora NG, Karolina, LUMI). The scientific outcome will be a publicly available database of electronic, magnetic and topological properties, identification of at least three candidate materials with robust anomalous Hall/ Nernst effects verified by the experimental partners at Institue of Physics, ASCR in Prague, and pave a clear pathway toward low‑power spintronic devices.

Researcher: Michelle Tsirou            

OPEN-38-19   The electron injection problem in the context of protostellar jets              

Karolina CPU Alloc=5200    

Astrophysical plasma can be found in many systems in our Universe, from interplanetary shocks, to  supernova remnants, to jets, to galaxy clusters, amongst many others. Understanding the underlying  physics of particle acceleration of non-thermal particles in a variety of multi-scale systems constitutes  a major challenge in cosmic-ray Physics.  We use particle-in-cell (PIC) simulations to understand the microphysics of protostellar jets, launched  in high-mass stellar environments. We want to investigate the excited plasma instabilities and  understand their role in efficiently accelerating thermal electrons to higher energies. For such, we  want to simulate the shock region, and understand if the geometry and properties of shocks in  protostellar jets are constraining the particle acceleration mechanisms. We plan to use a  one-dimensional (1D) and two-dimensional (2D) spatial approach, and simulate shocks with velocities  of ~(0.1 - 0.01)c, and track the particles that gain energy, to understand how they manage to do it  within the shock region, when encountering the electromagnetic fields in the plasma where they  propagate.  This study tackles a critical, still open, electron injection problem in our field : how are particles  accelerated in astrophysical shocks?

Researcher:  Lukáš Eigler  

OPEN-38-2    

Depth-Grounded Latent Geometry for Vision-Language Models in Driving Scenes              

Karolina GPU Alloc=6000    

Driver-assistance and self-driving systems increasingly use AI models that look at a camera image and answer questions in plain language: is the cyclist on the left about to cross, how far away is the truck ahead. These vision-language models are good at naming what they see but poor at judging where things are. Trained on flat pictures and text, they have no reliable sense of distance.  We want to give a small model a sense of three-dimensional space. While it learns to answer driving questions, it must also reproduce the geometry of the scene, using dense depth maps derived from the laser scanner mounted on the same car. Instead of predicting millions of pixels, it predicts a short compressed code of the depth image, which we can decode back into a picture and look at. The extra training signal is cheap, and it lets us check whether the model has understood the road or simply guessed from context. We will train models of about two billion parameters on public driving datasets and on new recordings combining commented driving with laser, CAN-bus and GPS data, and measure whether this improves answers about distance, trajectory and risk. 

Researcher: Marketa Paloncyova 

OPEN-38-20 

pH dependent structure of lipid nanoparticles    

Karolina CPU Alloc=18600; Karolina GPU Alloc=2800; LUMI-C Alloc=4700; LUMI-G Alloc=5600

Encapsulation of RNA in lipid nanoparticles (LNPs) opened a new field of medicine, allowing for efficient development of vaccines, cancer treatment or rare diseases therapy. Though the efficiency and safety of RNA carrying LNPs have been proven, there are still many questions regarding LNP structure and function that hinder the LNP targeted design. In this project we will employ multiscale molecular dynamics simulations to model the composition-structure relationship in LNPs. We will develop a protocol to model their structure and stability with varying pH and focus on the lipid-RNA interactions and their role on the overall morphology of LNPs. Understanding of LNP structure as a function of its lipid component and surrounding pH will facilitate the design of more efficient and stable LNPs.

Researcher: Sergiu Arapan

OPEN-38-21 

A DFT study of the biquadratic interlayer exchange coupling in Co|Ru(1-x)Fe(x) multilayers     

Barbora NG Alloc=3100; Karolina CPU Alloc=3100; Karolina GPU Alloc=900; LUMI-C Alloc=3100  

The physics of interlayer exchange coupling (IEC) in layered systems has recently attracted renewed interest, driven by their promising application in high-frequency spintronic and magnonic devices, as well as by the fundamental insights they provide into magnetic interactions at the nanoscale. The strength of IEC is determined by the interlayer exchange coupling constant J, which consists of the antiferromagnetic bilinear J1, and biquadratic J2 terms. The biquadratic coupling strength is responsible for noncollinear alignment, which can enhance the MRAM and spin-torque oscillators performance. The coupling angle between the magnetic moments of two ferromagnetic layers can be controlled by inserting a specially designed magnetic metallic spacer layer between them. Here I propose a computational study of the strength of biquadratic IEC by noncollinear electronic structure calculations in Co|Ru(1-x)Fe(x) multilayers to predict its dependence on the structure of the interface and concentration x of Fe in Ru spacer.

Researcher: Martin Matys  

OPEN-38-22 

Laser-Accelerated Proton Drivers for Relativistic Plasma Mirrors           

Karolina CPU Alloc=23500 

High-power laser interactions with solid targets can generate ultrashort proton beams containing a large number of particles. Recent theoretical and simulation studies have shown that such beams are promising drivers of nonlinear plasma waves. However, existing analytical models and proof-of-principle simulations generally assume idealized particle beams with uniform density, well-defined energy, and negligible divergence. The influence of realistic beam properties, therefore remains an important open question. In this project, we will perform combined particle-in-cell simulations connecting laser-driven proton acceleration with proton-beam-driven plasma-wave excitation. First, we will simulate the generation of proton beams in laser–target interactions. The resulting realistic proton-beam distributions will then be imported into separate simulations, in which the proton bunches propagate through a plasma and drive nonlinear wake and interior waves. By considering laser-accelerated proton beams that are more realistic than those assumed in previous theoretical studies, the project will support the design of future laser–target experiments and applications in proton-driven plasma physics.

Researcher: Sumit Ghosh  

OPEN-38-23 

Study of microscopic origin of the dynamical properties of non-trivial magnetic textures

Karolina CPU Alloc=12700 

The enormous requirement for computational resources to process today’s huge data have pushed us to go beyond the conventional magnetic memory. Among different new proposals, magnetic memories based on magnetic texture are the most promising one which already has shown their experimental feasibility. Among different magnetic textures, the localised and isolated magnetic textures, such as magnetic skyrmions, are of particular interest from both theoretical and experimental perspective. These configurations behave as point like topological objects which can be moved with a charge current. Such behaviour make them an ideal candidate for binary memory. Our recent studies have revealed the possibility of even more complex magnetic configurations with high topological indices which also enhances the possibility of their applications as multi-level memory devices. To exploit their full potential, we need a proper understanding of their response properties and dynamics which is the focus of this project. The main objective of this project is two fold - (i) to obtain the microscopic description of the response properties of different magnetic texture and (ii) use their response and topological features to control their dynamics. These results would instrumental to build an effective theory of their complex behaviour which will be beneficial for their practical applications.

Researcher: Martin Friak    

OPEN-38-24 

Next-Generation Machine Learning Interatomic Potential for Large-Scale Simulations of Tin     

Barbora NG Alloc=18000; Karolina CPU Alloc=13500; Karolina GPU Alloc=500; LUMI-C Alloc=4500; LUMI-G Alloc=6300  

Tin is a key engineering material used in electronics, energy technologies, and advanced manufacturing. However, atomistic simulations of its behaviour are still limited by the accuracy of existing computational models. This project aims to develop a next-generation machine-learning interatomic potential (MLIP) for tin with almost quantum-mechanical accuracy. We will generate an extensive quantum-mechanical database covering large atomic structures, lattice defects, high-temperature configurations, and interfaces between different crystal phases. These training data will be used to train a highly transferable Atomic Cluster Expansion (ACE) potential capable of accurately describing complex atomic processes beyond the capability of conventional interatomic potentials. The resulting model will enable realistic simulations of atomistic processes in tin, e.g. phase transformations, defect evolution, and mechanical behaviour in systems containing millions of atoms, providing a valuable computational tool for future materials research and engineering applications.

Researcher: Thibault Derrien         

OPEN-38-25 

Polarization- and Orientation-Resolved Analysis of Electronic Excitations in Band-Gap Materials (POLARIS)       

Karolina CPU Alloc=104200; Karolina FAT Alloc=140; Karolina GPU Alloc=11000; VLQ Alloc=217000   

We employ predictive quantum formalisms for enabling the development of novel applications based on usage of intense laser processing of solids and nanomaterials. In recent years, PI activities established the time-dependent density functional theory (TDDFT) as an excellent prediction tool for supporting the design of future photonic & optoelectronic components along with their functionalization. The national MSMT-OPJAK project SENDISO (2024-2028) and the EU project Q-NEKO (2026-2030) aim at developing both fundamental and applied research for enabling production of prototypes of new kinds of sensors and qubits using intense light. In this context, the present HPC project focuses on describing the excitation of electrons in poly-atomic crystals that are relevant to sensors and qubits, when they are irradiated by intense linearly polarized waves. This project will also require access to the VLQ machine. Finally, this project will also support the training of young researchers in using advanced theoretical techniques adapted to the problems met in the engineering field of wave-matter interaction.

Researcher: Carlos Manuel Pereira Bornes          

OPEN-38-26 

A foundational equivariant graph neural-network model for NMR tensors of inorganic porous materials 

Karolina CPU Alloc=223000; LUMI-G Alloc=8000

Solid-state nuclear magnetic resonance (NMR) is one of the most powerful experimental methods for providing atomistic insights into materials and molecules. However, turning NMR spectra into a picture of the structure, dynamics and interactions in a sample usually requires ab initio simulations that are extremely expensive, limiting how much chemical diversity can be studied. Machine learning has transformed atomistic simulation by reaching ab initio accuracy at a fraction of the cost, yet most models predict only energies and forces, which cannot be measured directly. We recently showed that equivariant graph neural networks can instead predict full NMR tensors, and hence spectra, for entire families of porous aluminosilicates. Here we take the next step and train foundational NMR models spanning inorganic porous materials, including metallosilicates, aluminophosphates and silicoaluminophosphates and the guest molecules they host. These models will predict NMR tensors thousands of times faster than quantum calculations, enabling direct comparison with experiment and accelerating the design of materials for clean energy, carbon capture and green chemistry.

Researcher: Veronika Šritterová    

OPEN-38-27 

Quasi-harmonic models of negative thermal expansion of metal-organic frameworks

Karolina CPU Alloc=10100; Karolina GPU Alloc=600       

Computational modelling offers microscopic insight into solid-state materials, linking macroscopic properties to atomic-scale structure and dynamics. It is particularly powerful for studying thermal expansion, whose magnitude and sign often require an atomistic description. While most solids expand upon heating, many metal-organic frameworks (MOFs) exhibit pronounced negative thermal expansion (NTE). This counterintuitive behavior is technologically interesting, enabling the design of composites with tailored or near-zero expansion for optics, electronics, and biomedicine. This project aims to establish a robust computational description of NTE in MOFs to support the rational design of materials with controllable thermal properties. The focus is placed on a systematic evaluation of density functional theory (DFT) functionals and the quasi-harmonic approximation (QHA) to accurately capture both MOF structures and their vibrational response, specifically identifying the microscopic modes responsible for NTE.

Researcher: Ondrej Chrenko          

OPEN-38-28 

Formation of 3D radiative disks from the ejecta of mass-losing binary stars   

Karolina CPU Alloc=3600; Karolina GPU Alloc=4200; LUMI-C Alloc=900; LUMI-G Alloc=9300    

In this project, we will study the evolution of gas ejected from binary stars undergoing Roche lobe overflow and mass loss through the outer Lagrange  point L2. To this end, we will perform three-dimensional radiation hydrodynamic simulations using the Fargo3D code. Our primary focus will be on configurations  in which the ejected material remains gravitationally bound, falls back toward the binary, and collides with the ongoing outward L2 stream. We will investigate  whether this hydrodynamic interaction can lead to the formation of a circumbinary decretion disk. By determining the density, temperature, and flow structure of such disks, we will place valuable theoretical constraints on the evolution of mass-losing binary systems and on the observable properties of luminous red nova transients. This project is carried out within the framework of the GAČR Standard Grant No. 24-11023S.

Researcher: Martin Golasowski    

OPEN-38-29 

OpenWebData Index              

Karolina CPU Alloc=96000; Karolina GPU Alloc=9200; LUMI-C Alloc=9200; LUMI-G Alloc=12800            

Open Web Data is vital for European AI development and enables not only LUMI and Czech AI Factory users to access large-scale, transparent, reusable data for training, evaluating, and operating advanced AI systems. The completed OpenWebSearch.eu Horizon Europe project established a European Open Web Index, demonstrating web data crawling, processing, and indexing on federated HPC infrastructure. Outputs include cleaned content, structured metadata, sparse lexical indices, and dense vector representations. This project builds on that foundation, focusing on sustained operation and systematic improvement. Rather than operating a general search engine, it maintains the Open Web Index, enhancing quality, quantity, coverage, and freshness. Tasks include continuous crawling, regular reprocessing, improved filtering/metadata extraction, and regeneration of sparse/dense indices. CPU HPC resources handle large-scale preprocessing and sparse indexing; GPU resources manage embedding generation and intensive enrichment. For the LUMI and Czech AI Factories, this links European supercomputing capacity with open web-data infrastructure. Researchers, startups, public sectors, and industry users gain access to documented, reusable datasets without relying on closed commercial indices. The program strengthens Europe’s AI development and evaluation capacity while maintaining realistic goals for coverage, legal compliance, sustainability, and operational costs.

Researcher: Ondřej Maršálek         

OPEN-38-3    

Machine-learned Interatomic Potential for a Cobalt Porphyrin Cage in Water

Karolina CPU Alloc=27100; Karolina GPU Alloc=1800    

The increasing atmospheric CO₂ levels and their impact on the global climate have stimulated efforts to develop technologies for its capture, storage, and processing. Electrochemical reduction with the help of suitable catalysts is one of the most promising approaches. In this project, we will develop a highly accurate atomistic description of a cobalt porphyrin cage solvated in water, a catalyst for the decomposition of carbon dioxide. Understanding its solvation and interaction with surrounding ions at finite temperatures is crucial for optimizing its catalytic performance. Because ab initio molecular dynamics (AIMD) with electronic structure calculated on the fly is computationally prohibitive for a system comprising approximately 3,000 atoms over nanosecond timescales, we employ machine learning to construct an equivariant neural network interatomic potential. An optimal reference training dataset will be efficiently extracted using an iterative multigenerational active learning methodology. The resulting computational model will enable extensive molecular dynamics simulations to study the carbon dioxide decomposition process with the accuracy of density functional theory. Ultimately, the atomistic insights gained from these simulations will facilitate the rational design of efficient carbon capture and utilization technologies.

Researcher: Mukeshkumar Prakash Khanore      

OPEN-38-30  Machine Learning Accelerated Simulation of Ferroelectrics Perovskites.

Karolina GPU Alloc=500; LUMI-G Alloc=600          

Ferroelectric perovskites are the working material of piezoelectric sensors, actuators, ultrasonic transducers, high-k dielectrics, and microelectromechanical systems. The best of them are lead-based, and finding lead-free replacements is, at heart, an atomic-scale problem. Best candidates with superior electromechanical properties are usually mixed perovskites, particularly in the region of the so-called morphotropic phase boundary (MPB). It is a meeting point of two symmetries, usually a vertical line in the composition-temperature phase diagram. Material is frustrated, leading to softness and an enhanced response. Their enhanced properties arise from local polarization coupled to cation order in chemically complex solid solutions. Investigation of their physics requires finite-temperature dynamics on supercells far beyond DFT's reach, where empirical potentials fail. We propose to address this with machine-learning potentials built on GRACE, a graph atomic cluster expansion framework, fine-tuned on compact datasets from VASP's on-the-fly ML force fields. Our results already validate the pipeline: fine-tuning on 40-atom datasets for PbTiO₃, BaTiO₃, and CaTiO₃ cuts energy and force errors by an order of magnitude on independent DFT-MD test sets; the model generalizes to mixed (Ca, Ba)TiO₃, including two chemical orders unseen in training, and reproduces the ferroelectric phase transitions of PbTiO₃ and BaTiO₃ in 5000-atom MD at a cost 3–4 orders of magnitude below DFT. The potentials inherit the underlying functional's accuracy: they extend DFT's reach. We will now target the (Ba, Ca)TiO₃ family, with lead-based references, using 10⁴–10⁵ atoms supercells: phase transitions, hysteresis, and local polarization reconstructed from trajectories and validated against DFT, thereby establishing fast descriptors of piezoelectric response for future composition screening.

Researcher: Martin PYKAL 

OPEN-38-31 

Molecular Mechanisms of Carbonic Anhydrase Immobilization in Functionalized Mesoporous SBA-15 for Efficient CO₂ Capture     

Karolina CPU Alloc=16700 

Carbon dioxide capture using enzymes immobilized inside porous materials offers a promising strategy for reducing atmospheric CO2 while avoiding the high energy demands of conventional capture technologies. In this project, we will use molecular dynamics (MD) simulations to investigate how the pore structure and surface chemistry of aminopropyl-functionalized mesoporous SBA-15 influence the adsorption, stability, and catalytic activity of carbonic anhydrase, as well as the transport and organization of CO2 inside the pores. The simulations will provide an atomistic picture of processes that cannot be directly observed experimentally and will support the design of more efficient materials for enzymatic CO2 capture. The proposed work naturally follows our previous IT4Innovations project, which focused on simplified silica surfaces to understand the fundamental interactions between the enzyme, the functionalized support, and CO2. Building on these insights, we now move to realistic porous SBA-15 systems that closely resemble the materials used in experiments, allowing direct comparison between simulations and measurements.

Researcher: Erik Andris      

OPEN-38-32 

Development of implicit solvation models to complement universal ML force fields  

Karolina CPU Alloc=28900; Karolina GPU Alloc=600       

We aim to develop new ways to handle solvation phenomena in machine learning (ML) force fields. So far, the force fields have been trained to describe molecules in vacuum and solvation effects had to be modeled explicitly – by simulating the studied molecule and its surrounding solvent molecules together. This explicit solvation treatment adds substantial resource consumption and speed penalty for such calculations. If we need to perform many simulations to confirm properties of drug candidates, the use of explicit solvent may not always be the best choice. Traditional quantum mechanics methods use so-called implicit solvent models, which do not model solvent molecules directly, but often model the solvent as a continuum, which greatly simplifies the solvation treatment. We plan to combine some of the most successful implicit solvation approaches – COSMO and COSMO-RS with the current gas-phase foundational ML force fields, to make the large scale property calculations possible.

Researcher: Ivana Miháliková        

OPEN-38-33 

Comparing quantum algorithms for Gibbs state preparation in geometrically frustrated spin systems: The kagome Heisenberg model             

Karolina CPU Alloc=12900; VLQ Alloc=18000      

Every material has a natural tendency to settle into a state that matches its temperature, what physicists call a thermal, or Gibbs, state. Predicting these states is essential for understanding magnets, superconductors and other materials, but for complex quantum systems the calculation quickly outgrows even the most powerful classical supercomputers, since the information needed grows exponentially with the number of particles involved.  Quantum computers offer a possible way around this bottleneck, as they can directly encode and manipulate the same kind of quantum information that makes the problem hard. In this project we compare two very different strategies for preparing thermal states on a quantum computer. We test both on a known difficult case, that is a triangular arrangement of interacting magnetic particles known as the kagome lattice. It is a textbook example of geometric frustration, where competing interactions prevent the system from settling into a single preferred configuration. By identifying how quickly each method converges and where it breaks down, we aim to clarify which strategy suits materials with this kind of built-in conflict, that can help us to guide future quantum algorithms for simulating magnetism and other finite-temperature phenomena. 

Researcher: Adéla CHADALÍKOVÁ              

OPEN-38-34 

Factors influencing the stability of inverted hexagonal phase (Hii) of Lipid Nanoparticles

Karolina CPU Alloc=7700    

Lipid nanoparticles (LNPs) are widely used as delivery systems for gene therapies. Although they became well known through COVID-19 mRNA vaccines, their development has continued. Companies such as BioNTech are now developing LNP-based therapeutics for the treatment of many other serious diseases. [2] However, despite the widespread use of LNPs, several key molecular processes remain poorly understood, including cellular uptake, endosomal escape or intracellular RNA release. These are highly dynamic processes that cannot be fully captured by experimental techniques such as cryo-TEM or NMR, which primarily provide static structural information. Molecular dynamics simulations offer a unique opportunity to investigate these processes at atomistic resolution, providing insights into the mechanisms governing LNP behaviour inside target cells. Such an understanding could contribute to the rational design of more efficient LNP formulations with improved intracellular delivery and RNA release.  This project focuses on lipid phase transitions, which are believed to play a crucial role both in the formation of lipid nanoparticles and in the intracellular release of their therapeutic cargo. This phase transition is driven by a decrease in pH, which leads to the protonation of ionizable lipids. Ionizable lipids are the key component of LNPs because of their pH-sensitive behaviour. They are nearly neutral at physiological pH, which minimises toxicity during circulation, but become positively charged in the acidic environment of the endosome. This change in charge promotes interactions with negatively charged membrane lipids and is believed to trigger the lipid phase transition from the lamellar phase at neutral pH to the inverted hexagonal (Hii) phase at acidic pH. 

Researcher: Ondrej Meca  

OPEN-38-35 

FALCON – Optimization of high-aspect ratio wing             

Karolina GPU Alloc=1900    

Reducing the structural weight of aircraft is one of the most effective approaches to lowering fuel consumption and CO₂ emissions, making it a key objective for achieving the European Union's climate goals. The design of lightweight aircraft structures, however, requires accurate prediction of complex fluid–structure interaction (FSI) phenomena, which directly influence aerodynamic performance, structural loads, and aeroelastic stability.  The proposed project focuses on structural weight optimization using high-fidelity fluid–structure interaction simulations. The objective is to investigate optimization strategies that reduce airframe weight while maintaining the required aerodynamic and structural performance. The requested computational resources will enable large-scale optimization studies involving repeated high-fidelity FSI simulations, supporting the development of predictive multidisciplinary design methodologies for future lightweight and environmentally sustainable aircraft within the FALCON project.

Researcher: Ondrej Meca  

OPEN-38-36 

FALCON – Optimization of aeroacoustics

LUMI-G Alloc=4600  

Aircraft noise is one of the major environmental challenges facing modern aviation. Reducing aeroacoustics emissions is essential for the development of more sustainable aircraft and for meeting the European Union's long-term environmental objectives. Achieving this goal requires accurate numerical simulations capable of capturing complex fluid–structure interaction phenomena.  The proposed project focuses on aeroacoustics optimization using high-fidelity simulations based on the Lattice Boltzmann Method (LBM). The objective is to investigate strategies for reducing noise emissions through numerical optimization supported by large-scale simulations of fluid–structure interaction. The requested computational resources will enable extensive optimization studies and high-resolution aeroacoustics analyses, contributing to the development of quieter and more environmentally sustainable aircraft.

Researcher: Oldřich Plchot              

OPEN-38-37 

Large-Scale Multilingual Attribute Supervision for Speaker Verification and Audio Language Models     

Karolina GPU Alloc=4500; LUMI-G Alloc=10900  

Automatic voice verification is in daily use: banks and call centers confirm a caller's identity by comparing the voice on the line with a stored recording, and law enforcement searches for suspects in large datasets from surveillance. These systems return a single similarity score, and the score does not say which properties of the voice matched. Nor can they check a simple written description such as \a deep-voiced man with an Irish accent\". In our earlier work, we built SpeakerCard-1M, a public resource that provides short, checkable descriptions for more than 10,000 real-world speakers. Every stated property, such as the pitch range or the accent, is derived from a recorded acoustic measurement; a language model only turns the measurements into sentences and never hears the audio. When we tested eight recent audio-capable AI assistants with these descriptions, they rejected a description with a wrong pitch range in only 49–77% of trials, compared with 89% for our purpose-built system. On LUMI, we will expand the resource with speakers from several European languages, train models to compare a written description with a recording, and adapt open-source audio AI models to report measurable voice properties more reliably. The annotation tools, evaluation sets, and permitted models will be published. The result is a practical way to test what voice systems actually hear, so that developers and auditors can find unsupported claims before such systems are trusted with everyday decisions. Our methods will also contribute to creating an explainable speaker verification system."

Researcher: Zaira Carboni

OPEN-38-38 

Computational Study of Carbonization via Machine Learning-Based Interatomic Potentials       

Karolina CPU Alloc=300; Karolina GPU Alloc=2600; LUMI-C Alloc=200; LUMI-G Alloc=1700      

Carbon dots are promising carbon-based nanomaterials with tunable optical properties, but the relationship between synthesis conditions, atomic structure, and final properties is still poorly understood. In this project, we aim to investigate the carbonization process of organic precursors through large-scale atomistic simulations using machine learning interatomic potentials.  Machine learning potentials will enable the exploration of reactive structural transformations with improved accuracy compared to classical force fields, while maintaining the computational efficiency required for long-timescale simulations. The study will focus on the evolution of molecular precursors into carbonaceous structures, analyzing bond rearrangements, carbon network formation, and the development of sp^2-sp^3 domains.

Researcher: Suresh Ravisankar    

OPEN-38-39 

Electric Field Effect on Nanofriction in Transition metal dichalcogenide hetrostructures using deep potential molecular dynamics (EFENT)     

Barbora NG Alloc=600; Karolina CPU Alloc=5800             

This project investigates how an applied electric field influences nanoscale friction in transition-metal dichalcogenide (TMD) heterostructures, using machine-learning force fields (MLFF) built on a deep neural network. This method will be applied to TMD monolayers sandwiched between Au substrate, with a Si tip probing from the top. Molecular dynamics simulations will be used to examine the structural behaviour and frictional characteristics of these systems under an electric field. A comprehensive analysis of the phonon spectra will be carried out to evaluate the dynamical stability of the heterostructures and to investigate how electric field induced changes in lattice vibrations correlate with the observed frictional response.

Researcher: Jai Bardhan     

OPEN-38-4    

Spatio-Temporal World Models for Robotic Applications             

LUMI-G Alloc=15500

Robots remain far behind humans at everyday manipulation, largely because they cannot imagine what their actions will do before doing them. World models learn that ability from data: trained on large collections of real robot interaction, they simulate the physical world, predicting how objects move, deform and make contact in response to a robot’s actions, so that a robot can plan ahead, catch its own mistakes and handle objects it has never seen. Today’s world models are trained on camera images alone, and it shows: their predictions look plausible frame by frame but do not add up to a consistent physical scene — objects change shape, pass through one another, or drift out of place as the prediction runs on. This project teaches world models the two things that pictures alone leave out. First geometry: how far away every surface is, in real units, so that what the robot imagines is a scene rather than a picture of one. Then motion: following the same physical points on an object through time, so that a cup which has moved is still recognisably the same cup. Alongside this we will test a new generation of much faster geometry networks in place of the model’s current core. The aim is a robot that can rehearse a task in its head accurately enough for the rehearsal to be worth trusting — a prerequisite for putting robots to work outside the laboratory. 

Researcher: Anahí Villalba Pradas              

OPEN-38-40 

Impact of urban areas on regional scale rain and cloud patterns in convection permitting

Karolina CPU Alloc=28900 

It is well known that urban areas affect the warming rate over cities through the so-called urban heat island (UHI) but also induce changes in other relevant meteorological variables. In a previous study already published, we showed how these areas affect the diurnal distribution of precipitation and cloud, especially in summer, when combined with different parameterizations of the microphysics and convection (these simulations ran in Karolina thanks to the OPEN-28 project). This study is a continuation of the aforementioned one and aims to evaluate the impact that different combinations of urban and microphysics have on a number of meteorological variables, including temperature, wind and those related to cloud/rain microphysics. Simulations are performed using the WRF model with a domain in 3km horizontal resolution centered over Prague covering central Europe for the 2008-2017 period. The length of the period ensures statistically robust results. The urban schemes include bulk, the single-layer urban canopy model (SLUCM), and the multilayer urban models (BEP-BEM) with a building energy model including anthropogenic heat due to air conditioning. We further consider another scenario in which urban land use is fully replaced by a rural one (NOURBAN). Besides, we also used three different options for microphysics but in this case, no convective schemes are used due to the high resolution of the simulations (convection is explicitly resolved by the model). We expect to show that the inclusion of urban canopy schemes leads to an increase in temperature and a decrease in wind speed as well as to changes in other relevant meteorological values such as cloudiness and precipitation, which also depend on the microphysics. We will further evaluate expected differences in the downwind of cities in terms of precipitation and cloud patterns.

Researcher: Mohit Lohani  

OPEN-38-41 

3D-geodynamic models of Hellenic slab  

Karolina CPU Alloc=21700; Karolina FAT Alloc=80            

The subduction of the African Plate beneath the Aegean Plate is the primary driver of seismic activity in the Greece region. Among the many distinctive characteristics of the Hellenic subduction zone (HSZ), one of the most intriguing is the transition in the stress regime within the subducting slab. Seismological observations indicate that trench-parallel (along-strike) compression changes to down-dip compression at depths of approximately 120 km. Although the mechanisms responsible for this transition remain poorly understood, several plausible hypotheses could be proposed.  In this project, we will investigate these hypotheses using three-dimensional numerical geodynamic models of the Hellenic subduction zone. The models will allow us to quantify the effects of toroidal mantle flow, spherical geometry, the Clapeyron slope of mantle phase transitions, and along-trench lithospheric heterogeneities on subduction dynamics, slab morphology, stress distribution, and trench curvature. The geodynamic results will be integrated with petrophysical constraints and discussed in collaboration with experts in the USA to evaluate the physical plausibility of the best-fitting models.

Researcher: Tomáš Hrivnák            

OPEN-38-42  Computational Discovery and Design of PAH Donor-Acceptor Assemblies for Photocatalytic Applications           

Karolina CPU Alloc=8100    

Predictive modelling of photoactive carbon nanomaterials remains challenging because of their structural complexity and heterogeneous composition. Carbon dots (CDs) exemplify this problem and  despite their excellent optical properties, high photostability, and low toxicity, their ill-defined molecular structure complicates both experimental characterization and theoretical modelling. To address this limitation, polycyclic aromatic hydrocarbons (PAHs) with controlled size, heteroatom doping, and functionalization have emerged as well-defined molecular models that capture the key structural and electronic features of CDs. The proposed project will investigate stacked PAH donor-acceptor (D-A) complexes identified through ML-guided screening of a library of potential molecular building blocks. Candidate systems will be prioritized using multi-objective optimization based on predicted photocatalytic performance, aggregation propensity, excited-state charge separation, and model uncertainty. Selected assemblies will then be studied using static quantum-mechanical (QM) calculations and classical and excited-state molecular dynamics (MD) simulations to characterize charge-transfer processes, relaxation pathways, redox properties, and aggregation behavior. The resulting data will provide fundamental insight into the structure-property relationships governing PAH-based photocatalysts and support the rational design of efficient metal-free photoactive materials.

Researcher: Robert Vacha 

OPEN-38-43 

Design Principles of Fusogenic Proteins for Lipid Vesicles          

LUMI-G Alloc=34400

Modern medicine increasingly relies on delivering macromolecules into our cells, as shown by mRNA vaccines and by new cancer therapies. These molecules cannot readily enter cells on their own, so various delivery strategies have been developed. For example, mRNA can be packaged into tiny fatty droplets called lipid nanoparticles. Cells take up these particles, and the therapeutic cargo then reaches its destination via cellular processes. However, this strategy is inefficient: a significant amount of the cargo is lost along the way, and only a small fraction ever reaches the place where it can act. A major bottleneck is membrane fusion, a process during which the membrane of the carrier merges with a cell membrane and releases the cargo into the interior of the cell. Nature solved this problem long ago: enveloped viruses and our own cells use specialized proteins that force two membranes to fuse. However, we still do not fully understand which properties make a protein fusogenic or how to design functional miniature versions of these proteins. Therefore, new fusogenic proteins are often developed by trial and error. We will use computer simulations to follow membrane fusion at the molecular level, identify the protein properties that drive it, and design new protein sequences that make drug carriers fuse with cell membranes. Such proteins could improve the delivery of therapeutics and, consequently, the treatment of serious diseases such as cancer.

Researcher: Ivo Oprsal        

OPEN-38-44 

Final Phase of 3D Seismic Modeling and Risk Assessment for Nuclear Facility Infrastructure

Barbora NG Alloc=500; Karolina CPU Alloc=900 

Earthquakes have major social and economic impacts, causing casualties and damage to buildings and critical infrastructure worldwide. This includes urban areas, transportation networks, dams, power lines, pipelines, and strategic energy facilities. The planned expansion of nuclear energy in the Czech Republic, including the site selection and seismic hazard assessment for small modular reactors (SMRs), further highlights the importance of reliable seismic hazard assessment, where local geological amplification can significantly influence earthquake ground motions. This project focuses on advanced numerical modelling of seismic wave propagation in representative geological environments with complex three-dimensional structures. Building on previously validated methodologies, the simulations will be integrated with geological and geophysical observations to improve regional-scale characterization of local site effects and assessment of external seismotectonic hazards. The high-performance computing facilities of the IT4Innovations National Supercomputing Center will enable realistic large-scale simulations that cannot be performed on conventional computing platforms, providing more accurate predictions of seismic wave amplification under a wide range of earthquake scenarios.

Researcher: Michal Langer

OPEN-38-45 

First-Principles Design of Local Chemical Environments for Nitrate-to-Ammonia Conversion and Ammonium-Ion Storage  

Barbora NG Alloc=29700; Karolina CPU Alloc=58500; LUMI-C Alloc=9900     

Sustainable ammonia production and safe energy storage depend on controlling how ions and molecules interact with chemically complex materials. This project will use supercomputer simulations to reveal the atomic-scale factors that govern nitrate conversion to ammonia and reversible ammonium-ion storage. Chemically diverse solid-state environments will be compared to identify features that promote selective reactions, efficient charge transfer and rapid ion transport. Advanced density functional theory, solvent-aware models and transition-state calculations will determine how local bonding, charge redistribution, structural flexibility and kinetic barriers influence experimentally measurable catalytic and battery performance. The resulting material-specific design rules will guide the development of more efficient catalysts and durable aqueous electrodes. They will also reduce unnecessary experimental screening and accelerate progress in sustainable chemical production and energy-storage technologies.

Researcher: Tahir Wahab   

OPEN-38-46 

First principles study of chalcogen vacancy effect on the optoelectronic and photocatalytic properties of transition metal dichalcogenides monolayers    

Barbora NG Alloc=2300; Karolina CPU Alloc=6200          

In working conditions, chalcogen vacancies spontaneously occur in two-dimensional transition metal dichalcogenide (TMDCs) monolayers, affecting their optoelectronic and photocatalytic properties [1]. To study how chalcogen vacancies affect such properties, we will use quantum mechanical calculations considering prototypical MX2 (M = Mo, W, X = S and Se) TMDC monolayers. We will first use ATAT to determine the most favourable chalcogen vacancy concentration, followed by PBE and HSE06 calculations to examine the effect of the resulting vacancies on the electronic properties, including the band structure and band gap. We will assess the capability of the defective structures to perform the hydrogen and oxygen evolution reactions for water splitting, as well as their potential for photocatalytic CO2 reduction. The results will indicate whether vacancy-containing TMDC monolayers offer prospective optoelectronic and photocatalytic applications, guiding the experimental synthesis of vacancy-engineered MX2 monolayers.

Researcher: Prashant Dwivedi       

OPEN-38-47 

Multiscale Prediction of Cumulative Dust-Impact Damage in Helium-Exposed Polycrystalline Tungsten (W-IMPACT)         

Karolina CPU Alloc=19000; Karolina GPU Alloc=700       

Tungsten is the leading material for the most heavily loaded surfaces of future fusion reactors, but its  lifetime can be shortened by fast tungsten dust generated during plasma transients. Our recent  simulations showed that subsurface helium bubbles and platelets can strongly amplify cratering and  material ejection. W-IMPACT addresses the next reactor-relevant step: repeated dust impacts on  polycrystalline, helium-damaged tungsten. Large-scale molecular-dynamics simulations will follow  hundreds of millions of atoms during impact, local melting, crater growth and defect formation.  Dislocation networks extracted from these simulations will then be transferred to GPU-accelerated  discrete dislocation dynamics to predict how damage relaxes and accumulates beyond the short times  accessible to atomistic simulation. The project will identify when repeated impacts change from local  plastic deformation to persistent cracking or delamination and will provide quantitative damage  descriptors for future lifetime models.

Researcher: Anastasia Ostapenko             

OPEN-38-48 

Domain Adaptation of NVIDIA Cosmos for Controllable Driving-Scene Video               Karolina GPU Alloc=6400    

Generative world foundation models offer a promising approach to driving-video generation and transformation, but pretrained models may struggle with domain-specific data and preserving scene structure and temporal consistency. This project uses the NVIDIA Cosmos family of generative world models to develop a controllable driving-video generation and editing pipeline. Preliminary work on the Karolina supercomputer established the software pipeline and showed that domain-specific fine-tuning improves video reconstruction, while multimodal structural controls can preserve scene geometry. In the next phase, an extended driving-video dataset provided by Valeo will be prepared and used for large-scale LoRA post-training and systematic experimentation. The study will compare training configurations, conditioning modalities, and editing strategies for tasks such as weather and time-of-day transformation, while evaluating visual quality, temporal consistency, controllability, and preservation of road-scene structure. The results may support more efficient creation of diverse and safety-relevant driving scenarios for autonomous-driving research and testing.

Researcher: Jakub Šístek   

OPEN-38-49 

Vortical flow benchmarks: GPU-accelerated numerical simulations for advanced vortex identification 

Karolina CPU Alloc=4300; LUMI-G Alloc=7100     

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: Jan Brezina     

OPEN-38-5    

Graph U-Net Surrogates for Hydro-Mechanical Upscaling in Fractured Media

Karolina CPU Alloc=500; Karolina GPU Alloc=300            

Simulation of hydro-mechanical processes in fractured rocks is important for applications such as groundwater management, geothermal energy production, and the long-term safety assessment of radioactive waste repositories. However, modelling the interactions among thousands of fractures within complex fracture networks, as well as their interaction with the surrounding heterogeneous rock, is computationally expensive. This is particularly limiting when repeated model evaluations are required for multiscale simulations, parameter calibration, and predictions under uncertainty.   This project will develop neural-network surrogate models to accelerate the determination of the equivalent hydraulic conductivity tensor under mechanical loading. One surrogate will represent the mechanical response of the fracture network, including changes in fracture apertures, while a second surrogate will predict the corresponding upscaled hydraulic conductivity tensor. While the main goal is the effective conductivity tensor, the mechanical surrogate will eventually be used to accelerate the mechanical component of a coupled hydro-mechanical numerical model. 

Researcher: Vladislav Pokorný       

OPEN-38-50 

DFT+U and DFT+DMFT study of correlated 4f electrons and phonon spectra in cerium and CeH10   

Karolina CPU Alloc=31000 

Cerium and its hydrogen-rich compounds are difficult materials to understand. Their electrons interact strongly with one another, often pushing conventional computational methods beyond their limits. At the same time, some cerium hydrides become superconducting under extreme pressure, making them important systems for studying the relationship between electron correlations and superconductivity.  We combine density functional theory (DFT) with two methods that add a local Hubbard interaction: DFT+U (a static approach) and dynamical mean-field theory (DMFT). We use these to calculate how the materials' atomic vibrations change with temperature and pressure, which conventional methods handle poorly. The calculations will run on the IT4Innovations supercomputers and start from first principles, and we will test the results against experiments.   The aim is to develop reliable computational protocols that clarify why existing models succeed or fail, and to support the search for new superconductors. The results also bear on questions beyond fundamental physics, including future energy technologies and high-performance electronics that depend on understanding strongly correlated materials.

Researcher: Lukas Neuman            

OPEN-38-51 

Inductive Bias of Deep Neural Networks for Computer Vision   

Karolina GPU Alloc=4700    

Deep Neural Networks models have in the recent years dominated virtually all areas of Artificial Intelligence and Computer Vision. Their main advantage is that, given enough training samples, a training algorithm can automatically update network parameters to directly maximise given objective, such as image classification accuracy. Despite the recent success, the models are easily confused by trivial samples not present in the training set and even the largest models lack basic generalisation and reasoning abilities despite having hundreds of millions of parameters and despite being trained on millions of very diverse data samples -- suggesting that a fundamental piece of understanding is still missing.    We propose that one of the missing pieces in current models compared to humans is an appropriate inductive bias -- the set of prior assumptions used to generalise and make a prediction based on a finite set of training samples. In this project, we want to exploit this observation and search for new inductive biases to incorporate them into modern Deep Neural Networks used in common Computer Vision tasks. This will result in Deep Neural Network models which require less parameters, which are more efficient, which are less confused by out-of-distribution data samples and which require less training data, as using an appropriate inductive bias is likely equivalent to even exponentially less training data. 

Researcher: Michael Bakker           

OPEN-38-52 

Structural Mechanisms of CHCHD2 Dysfunction: Multi-scale Modeling of Pathogenic Mutations, Hinge Stability, and Complex Formation             

Karolina CPU Alloc=800; LUMI-G Alloc=5500       

CHCHD2 is a highly interactive, bi-organellar hub protein that functions as a mitochondrial metabolic gatekeeper and a nuclear transcription factor. The disordered tail contains hotspot for missense mutations linked to neurodegenerative diseases, including Parkinson’s disease (T61I) and Amyotrophic Lateral Sclerosis (P14L). Additionally, overexpression of this protein is a key driver in the metastasis and survival of aggressive malignancies like lung and breast cancer. Understanding how these single-point mutations propagate structural failure is challenging because the dynamic NTD undergoes intermediate chemical exchange and promotes aggregation, rendering it invisible to solution-state NMR. This project utilizes a multi-scale computational framework to map the sequence-to-structure relationships of CHCHD2. By deploying extended all-atomistic molecular dynamics (AAMD) on specific regions of the protein, evaluating fibril core thermodynamics, and simulating the complete 152-residue monomer alongside its physiological quaternary structures (specifically the CHCHD2 homodimer and CHCHD2-CHCHD10 heterodimer) with coarse-grained molecular dynamics (CGMD), we will determine how localized sequence variations drive divergent molecular phenotypes, and the normal wild-type function of the protein which could be exploited for oncological treatment, providing a unified structural rationale for CHCHD2-mediated pathogenesis and targetable interfaces. Specific allocations are to execute full-scale simulations following successful preliminary exploratory simulations.

Researcher: Martin Čadík  

OPEN-38-53 

Visual Camera Orientation Estimation in Natural Environments             

Karolina CPU Alloc=200; Karolina GPU Alloc=800; LUMI-G Alloc=200 

When looking at a photograph of a natural landscape, a fundamental question arises: \Where was this photo taken?\" In practice, photographs rarely contain location metadata, leaving the visual content of the image as the sole available source of information. Answering this requires visual geo-localization to recover 3D geographic coordinates. However, full 6-DOF spatial localization remains incomplete without camera orientation as an indispensable component. Visual orientation estimation is critical in natural environments—such as deep mountain valleys or search-and-rescue zones—where satellite signals are unavailable, degraded, or spoofed. It also supports autonomous navigation, digital forensics, and computational re-photography.    We recover camera orientation by matching RGB query images against synthetic views rendered from Digital Elevation Models (DEMs), bypassing the lack of dense 3D reconstructions in natural environments. To address visual ambiguities from changing weather, foliage, and snow on static terrain, our approach integrates Vision Transformers (ViTs) with relational structural feature learning to extract landscape invariants. Operating independently of camera Field of View (FOV) metadata, our method generalizes to arbitrary photographs. This work directly extends the OPEN-36 project, which established the feasibility of transformer-based orientation recovery.  "

Researcher: Petr Slavicek  

OPEN-38-54 

Triggering and probing ultrafast reactions with high energy photons: advancing computational methods      

Karolina CPU Alloc=9500    

Modern X-ray free-electron lasers and high-harmonic sources can trigger chemical reactions and follow them on femtosecond timescales. These experiments reveal charge migration, proton transfer, ionization, electronic decay and bond breaking, but their interpretation requires simulations that describe several electronic states, nuclear motion and the surrounding solvent.  We will develop computational protocols for reactions initiated by XUV and X-ray photons and for their observation by time-resolved photoelectron spectroscopy. Non-adiabatic molecular dynamics, time-dependent quantum chemistry, core-level electronic-structure calculations and multiscale solvation models will be applied to molecular clusters and liquids. The resulting methods will support experiments at international X-ray facilities and clarify the earliest stages of photochemical and radiation-induced processes. 

Researcher: Gaurav Joshi  

OPEN-38-55 

Predicting Metal-Complex Stability Constants: Protocol Development and Method Benchmarking         

Karolina CPU Alloc=28200; Karolina GPU Alloc=45          

Metal-ligand interactions drive processes across medicine, catalysis, and biology, yet accurately predicting how tightly these components bind—their stability constant—remains a persistent challenge. Traditional experimental measurements are slow and resource-heavy, while standard theoretical calculations often fall short of real-world accuracy. This project addresses the problem by building a direct, high-throughput computational pipeline. Using ARCHITECTOR and CREST, we explore the full conformational space of metal complexes in solution, then pair these structures with benchmarked density functional theory (DFT) calculations and machine learning methods like UMA. By targeting and removing systemic computational errors, this protocol will deliver accurate absolute stability constants that match experimental benchmarks, paving the way for targeted molecular design without the trial-and-error.

Researcher: Filip Jozefov    

OPEN-38-56 

DreaMS De-Novo: cycle-consistent generation of molecular structures from mass spectra

LUMI-C Alloc=4000; LUMI-G Alloc=15500

Fewer than one in ten of the small molecules in the human body and the plant kingdom has been structurally characterised. Among the unknowns are future drug leads, diagnostic biomarkers, and environmental contaminants that routine screening does not detect. Mass spectrometry is the primary technique for finding them: it fragments each molecule and records the masses of the fragments, producing a spectrum. Reconstructing the original structure from that fragmentation pattern, known as structure elucidation, remains largely unsolved. Machine learning models can be trained for the task, but only on annotated spectra, whose molecule has been confirmed experimentally. Newly released public libraries have raised the annotated total tenfold, to roughly 300,000 molecules, and no model has yet been trained on all of them; we intend to be the first. We will train on those together with the hundreds of millions of unannotated spectra held in public repositories, which no supervised method can use. Our model proposes a structure for a measured spectrum, a forward model then simulates the spectrum that structure would produce, and the mismatch between simulation and measurement gives a training signal that needs no annotation. Using both resources together will make structure elucidation possible for molecules absent from every reference library, with direct consequences for natural product drug discovery, clinical diagnostics, and environmental monitoring.

Researcher: Jan Benáček  

OPEN-38-57 

QED pair cascades and coherent radio emission in pulsar polar caps Karolina CPU Alloc=76300     

Pulsars are rapidly rotating neutron stars that sweep narrow radio beams across the sky like lighthouses. More than fifty years after their discovery, we still cannot explain the origin of their coherent radiation. The magnetic field of the star, a trillion times stronger than that of any magnet on Earth, drives short electric discharges above the stellar magnetic poles. In such a discharge, plasma particles produce gamma-rays, which eventually turn into matter and antimatter particles. The avalanche that follows the discharge powers both the radio pulses and the gamma-ray emissions.    This project models one such discharge from first principles on the Karolina supercomputer. To our knowledge, this will be the first systematic calculation of both coherent radio Poynting flux and intrinsic high-energy emission from the polar-cap discharge using the implemented quantum-electrodynamic (QED) pair cascade model. The method provides a self-consistent route to connecting cascade microphysics with radio and high-energy source properties.    The results will give physical templates for radiative transfer codes through the neutron star magnetosphere and wind, for eventually interpreting data from radio telescopes and from gamma-ray observatories. The project is part of my Marie Skłodowska-Curie Fellowship (Horizon Europe).

Researcher: Rostislav Langer         

OPEN-38-58 

Atomistic Design of Single-Atom Catalysts for Sustainable Chemistry

Barbora NG Alloc=22400; Karolina CPU Alloc=64300     

The transition to sustainable chemical technologies depends on our ability to design catalysts that combine high activity, selectivity, and long-term stability. Achieving this goal requires an atomistic understanding of catalytic processes under realistic reaction conditions. In this project, we will employ advanced first-principles simulations to investigate the stability, electronic structure, and reaction mechanisms of single-atom catalysts across three representative catalytic transformations of environmental and industrial relevance: (1) CO2 conversion on transition-metal single atoms embedded in C3N5, (2) hydrosilylation catalyzed by Ni and Pt nanoparticles supported on g-C3N4, and (3) nitrate-to-ammonia electrocatalysis on Au-decorated NiCo layered double hydroxides containing atomically dispersed active sites. By combining density functional theory, transition-state calculations, and electronic structure analysis, the project will establish general design principles for next-generation single-atom catalysts and support ongoing experimental collaborations.

Researcher: Tomas Blejchar           

OPEN-38-59 

Digital twin of an AEM electrolyser

Karolina CPU Alloc=30400; LUMI-G Alloc=6000  

Anion Exchange Membrane (AEM) electrolysis is a promising route to green hydrogen because it combines a compact membrane-electrode architecture with the possibility of using less expensive, non-platinum-group catalysts. Its industrial deployment is nevertheless constrained by coupled transport, electrochemical and degradation phenomena that are difficult to observe directly and expensive to investigate experimentally. The project will create a digital twin of a commercial 2.4 kW pressurised AEM electrolyser operated in the CEETe hydrogen laboratory. High-fidelity CFD simulations, experimental polarisation curves and operational measurements will be combined to identify uncertain internal parameters and to train fast AI surrogate models. The resulting digital twin will reproduce the electrolyser response over varying current, temperature and pressure, estimate hidden physical parameters and state-of-health, and support rapid what-if studies without repeated time-consuming CFD calculations. HPC resources are essential because the planned uncertainty analysis and design-of-experiments campaign requires tens of thousands of independent multiphysics simulations, followed by GPU-based training and validation of neural-network models.

Researcher: Vladimír Sobota          

OPEN-38-6    

Computational modeling of alcohol-induced atrial arrhythmias            

Karolina CPU Alloc=18100 

Cardiac arrhythmia is a disturbance of the natural rhythm of the heart. It can range from a harmless rhythm irregularity to a major electrical disorganization that prevents the heart from pumping blood and leads to death. Alcohol consumption elevates the risk for developing an arrhythmia. Yet, despite the known detrimental effects of alcohol on human health and decades of cardiac electrophysiology research, the mechanisms by which ethanol promotes arrhythmias remain poorly understood. To unravel these mechanisms, this project utilizes anatomically accurate computational models of the human atria. These virtual heart models incorporate state-of-the-art computational descriptions of cardiac cellular electrophysiology, allowing realistic three-dimensional simulations of the heart’s electrical activity. The simulations will be performed across a range of blood alcohol concentrations to reflect differences in drinking behavior and alcohol metabolism between males and females. The results will improve our understanding of the mechanisms underlying alcohol-induced arrhythmias and help identify cardiac conditions that increase vulnerability to arrhythmias in the presence of ethanol.

Researcher: Matouš Soldát             

OPEN-38-60 

Generative AI for product-specific terpene synthase design      

Karolina CPU Alloc=14600; Karolina GPU Alloc=6800    

Enzymes are the catalysts of life, and designing them on demand is a long-standing goal of biotechnology. We focus on one remarkable family - terpene synthases - which build the scaffolds of terpenoids, the largest and most diverse class of natural products on Earth (fragrances, flavors, biofuels, and medicines such as artemisinin and Taxol). Our goal is an AI model that designs entirely new terpene synthases producing a terpene product we specify.  We have already shown, with wet-lab confirmation in living yeast, that AI can design novel, functional enzymes of a target terpene class. This project takes the two next steps. First, we will systematically train and compare many state-of-the-art AI model architectures for enzyme design and test the most promising ones experimentally - establishing, in the laboratory, which approaches truly work. Second, we will develop models that also use the enzyme’s three-dimensional shape to steer the reaction it catalyses - and thus which product it makes - with progressively finer control. This moves us from designing the right kind of enzyme toward producing any chosen terpene on demand - with broad benefits for sustainable, green manufacture of high-value natural products.

Researcher: Lukas Vojacek             

OPEN-38-61 

DMA4AI Repository  

Cloud services together with AI data storage        

The DM4AI project (this is part of the Open Science II project (CZ.02.01.01/00/24_030/0015041)) will deploy a domain-specific repository designed to support the long-term management, discovery, sharing, and reuse of artificial intelligence data. Built on the established DSpace repository platform, DM4AI will extend standard repository functionality with services tailored to the needs of the AI community, including inference previews, Lexis integration, and enhanced search and discovery capabilities.  Following deployment, the repository will be populated with curated, high-quality AI datasets prepared according to the FAIR principles, ensuring that the deposited resources are findable, accessible, interoperable, and reusable. Each published resource will receive a persistent identifier, enabling reliable citation, long-term access, and transparent tracking of data use.  DM4AI will provide researchers, developers, institutions, and external stakeholders with a trusted environment for accessing and sharing AI-related data. By combining robust repository infrastructure with AI-specific tools and professionally curated content, the project will improve data visibility, support reproducible research, and encourage collaboration across the wider AI ecosystem. 

Researcher: Frantisek Prinz             

OPEN-38-62 

Holistic approach to Rotating Packed Bed (RPB) Carbon Capture process by means of 3D CFD with focus on gyroidal packed rotor   

Karolina CPU Alloc=16100; Karolina GPU Alloc=800       

To combat climate change, industries need fast, efficient ways to trap carbon dioxide CO2 before it enters the atmosphere. Rotating Packed Beds (RPBs) are promising devices that use strong centrifugal forces—literally spinning gas and liquid together at high speeds—to capture CO2 inside a machine a fraction of the size of traditional equipment. Even though these spinning machines work, scientists don't fully understand how each individual internal part contributes to capturing the gas. Because the equipment spins so fast, it is nearly impossible to place sensors or cameras inside to see where the liquid flows, how it sprays, or where it forms films on the casing walls. Current engineering relies on looking at the machine as a \black box,\" focusing only on the overall output rather than optimizing the internal components. This project uses advanced 3D computer simulations—essentially building a high-precision \"digital twin\"—backed by targeted lab experiments to look deep inside the spinning device. By mapping the exact behaviour of liquid as it moves through the liquid sprayers, internal gyroidal structure, and outer walls, the team will reveal how each segment directly impacts CO2 absorption. Full understanding of the occurring fluid dynamics will allow engineers to redesign and optimize every section of the machine, maximizing carbon-capture efficiency and bringing smaller, cheaper, and cleaner industrial tech closer to real-world deployment."

Researcher: Safa Mert Akmeşe     

OPEN-38-63 

Modelling the Mechanisms of Terpene Biosynthesis using Deep Learning       

Karolina CPU Alloc=39000; Karolina GPU Alloc=9700    

Enzymes are biological catalysts essential to biotechnology and synthetic biology. Terpene synthases (TPSs) generate the core hydrocarbon scaffolds of terpenoids—the largest and most diverse natural‑product family, including hormones and pharmaceuticals3,4. Predicting precise enzyme function from sequence alone remains a “holy grail” unmet by conventional methods1,2.  In “Modelling the Mechanisms of Terpene Biosynthesis using Deep Learning,” we will build neural networks that embed biochemical principles as inductive biases and model every carbocation rearrangement in the TPS active site. Leveraging our curated dataset5 of thousands of annotated TPS reaction mechanisms, the models will predict substrates, intermediates, and final terpenoid products from sequence and structural features. By this, we will screen large unannotated sequence databanks to mine TPSs with the goal three goals: (1) Discovery of novel TPSs and biochemistry, (2) Discovery of novel terpene scaffolds and study their chemical meaning, (3) Investigate evolutionary diversity of TPSs and understand their role throughout evolution of life.  This project integrates computational forecasting with experimental validation: engineered TPS variants will be tested in yeast to confirm predicted product profiles and stereochemistry. Achieving accurate, rapid computational annotation and de novo enzyme design will unlock sustainable production of natural and new‑to‑nature terpenoids and establish a general framework for enzyme function prediction.

Researcher: Christopher Heard    

OPEN-38-64 

Chemoselective Aldehyde Hydrogenation over Noble Metal Catalysts

Karolina CPU Alloc=64300; LUMI-G Alloc=4300  

The chemoselective hydrogenation of α,β-unsaturated aldehydes over later transition metals (Pt, Pd, Au) is a globally important class of catalytic reactions, with longstanding worldwide applications in diverse fields, including fine chemical and pharmaceutical industries. To achieve selectivity towards high value saturated aldehydes and unsaturated alcohols, control over preferred adsorption modes, reactant availability and the geometric and electronic structure of the catalyst is required. However, the subtle and interconnected contributions of various parameters make rational catalyst design a significant challenge for experiment. Theoretical treatments have thus far focused heavily on idealised model surfaces, in contrast with the reality of ensembles of nanoparticles of various shapes, which have been shown to display vastly different binding and reactive properties to pristine single crystal surfaces. In this work, we will bridge the gap between experiment and theory and examine the roles of i) nanoparticle shape and ii) presence, concentration and location of rare earth metal dopants, in controlling hydrogenation selectivity, through a combined density functional and machine learning study. We aim to isolate the factors which control both activity and selectivity towards C=C and C=O hydrogenation, and derive functional catalyst design principles alongside concomitant testing provided by a 1:1 collaboration with experiment.

Researcher: Ivan Kološ        

OPEN-38-65  Numerical modeling of load of structures in quasi-static effect of wind

Karolina CPU Alloc=900       

The project is focused on numerical modeling of flow around objects in the atmospheric boundary layer. This issue is complicated mainly due to the atmospheric turbulence, which requires the use of advanced numerical models of the flow coupled with detailed computational mesh of the domain. This research will contribute to bigger efficiency in design of building structures.

Researcher: Petra KÜHROVÁ          

OPEN-38-66 

From RNA Benchmarking to Data-Driven Force Field Development: Reweighting, Enhanced Sampling and Community Reference Ensembles     

Karolina CPU Alloc=11700; Karolina GPU Alloc=8600; LUMI-C Alloc=7000; LUMI-G Alloc=12100

RNA molecules are essential for life and underpin technologies such as mRNA vaccines and RNA-based therapeutics. Atomistic simulations can reveal how RNA folds and moves, but their predictive value depends on the accuracy of the molecular force field.  Our previous IT4Innovations Open Access project laid the foundation for a systematic RNA benchmark and produced two complementary advances: practical criteria for obtaining converged RNA ensembles by enhanced sampling [1], and a reweighting strategy that uses such ensembles to evaluate force-field changes efficiently [2].  The proposed project will test statistical reweighting between independently developed RNA force fields using experimentally characterized oligonucleotides. We will establish where reweighting is reliable, where new simulations remain necessary, and how the workflow transfers to larger systems and emerging force fields. Trajectories, reference ensembles and analysis data will be standardized for integration into IDA and ADAMS4SIMS, enabling their reuse in future benchmarking and data-driven force-field development. 

Researcher: Peter Huszár  

OPEN-38-67 

Regional climate projections of impact of non-CO₂ forcers and landuse on climate and air quality using RegCM5-Chem        

Karolina CPU Alloc=18100 

This project builds on our previous project OPEN-35-28 and aims to finalize the tasks within the Horizon Europe FOCI initiative, focusing on the climate and air quality impact of non-CO₂ radiative forcers such as aerosols, methane, and tropospheric ozone and the interaction of these impacts with land use, mainly urban one. Despite well established knowledge of CO₂, uncertainties remain around non-CO₂ forcers due to complex atmospheric/chemical interactions and variability. To address this, the project employs high-resolution regional climate simulations using the RegCM5-Chem model, which integrates atmospheric chemistry and their feedback into the climate system while modulated by the land-use change. Simulations are focusing on future climate scenarios for the middle of the century, driven by the global climate model EC-Earth3-AerChem with telescoping nesting from continental 27 km x 27 km resolution to urban scale 3 km x 3 km. This project specifically focuses on these high-resolution urban scale simulations. The outcomes aim to improve the detailed understanding of non-CO₂ impacts on climate, air quality on urban scale including public health consequences.

Researcher: Lukas Drahnik              

OPEN-38-68 

Challenges in discovering topological structures in big data with hybrid computing

Karolina CPU Alloc=4560; Karolina GPU Alloc=500; VLQ Alloc=127700            

Scientific datasets are growing rapidly in both size and complexity, making it increasingly difficult for conventional computational methods to identify meaningful topological structures hidden in the data that may highlight biologically relevant phenomena.    We will evaluate hybrid quantum-classical workflows that combine established TDA methods with state-of-the-art quantum algorithms. These workflows introduce a novel approach for integrating information obtained from multiple executions of quantum algorithms to enable the analysis of large-scale datasets.    Because previous research into these algorithms has typically focused only on evaluations in highly specialized use cases that did not require dealing with any interference in the outputs, this project may represent their first application to datasets of this scale and complexity. By benchmarking hybrid quantum-classical workflows against established classical methods on the same large-scale datasets, we aim to identify both their advantages and limitations. The expected outcome is a better understanding of the challenges involved in inferring information about global topological structures from locally derived features within hybrid quantum-classical workflows. 

Researcher: Jan Kuneš        

OPEN-38-69 

Orbital moments in Co-based anti- and altermagnets   

Karolina CPU Alloc=18100 

Transport of orbital moment in solids, spin-to-orbital- and orbital-to-spin-current conversion have recently become a hot topic in spintronics. Driven by technological applications, this research addresses fundamental questions such as the definition of the current of a non-conserved quantity (orbital moment), the magnitude, spatial orientation and dynamics of orbital moments in materials of interest.  In this project we will investigate the magnetic order of CoO and CoF2 and its signatures in core-level spectroscopies such as x-ray magnetic circular and linear dichroism or circular dichroism in resonant inelastic x-ray scattering. Unlike Mn2+ and Ni2+ in high symmetry (quasi-cubic) environment with predominantly spin moments, Co2+ high-spin state is orbitally degenerate, which allows appearance of orbital moments with magnitude comparable to their spin counterparts. The compensated nature of these materials makes it difficult to measure these moments directly. They remained relatively unexplored, partly due to lack of interest in compensated magnets.  This has radically changed with the proposals of antiferromagnetic and altermagnetic spintronics. We will combine the density functional theory with the dynamical mean-field theory study the magnetic order in  CoO and CoF2, determine the size and orientation of local spin and orbtial moments as well as the coresponding core-level spectra. The results will provide guidance for experiments run by our collaborators.

Researcher: Denys Biriukov             

OPEN-38-7    

Physics-based inverse design of new peptides selective for the bacterial peptidoglycan

Karolina GPU Alloc=11100; LUMI-G Alloc=40800

Antibiotic-resistant bacteria are a rapidly escalating threat to human health, creating an urgent need for new ways to identify and eliminate them. This project will develop synthetic peptides that can recognize and specifically bind to the protective peptidoglycan layer of methicillin-resistant Staphylococcus aureus (MRSA). These peptides could serve as delivery agents that guide antibacterial enzymes or diagnostic markers directly to bacterial cells. We will first build detailed atomistic models of the MRSA-specific peptidoglycan structure and study how naturally occurring bacteria-killing enzymes called endolysins recognize it. High-performance all-atom molecular simulations, evolutionary algorithms, and free energy calculations will then be combined to design new peptides with the desired binding strength and selectivity. The most promising candidates will be synthesized and experimentally tested using fluorescence microscopy. The project will provide new molecular insights into bacterial cell-wall recognition, establish an efficient computational strategy for peptide design, and deliver peptide scaffolds for future antibacterial treatments and diagnostic tools.

Researcher: Lingping Kong

OPEN-38-70 

Reproduction, Benchmarking, Scaling and Improvements of AlphaFold 3 for Biomolecular Complexes    

VLQ Alloc=864000   

AlphaFold 3 is a powerful new AI that can predict the exact 3D shape of proteins together with DNA, RNA, drugs or ions — much more accurately than any previous tool. This helps scientists understand how cells work and design better medicines quickly. In this short project we will repeat the main tests from the original 2024 Nature paper, run the model on new real-world examples, and try simple improvements to fix known small problems (e.g. wrong molecular handedness or fake “hallucinated” loops). Using Czech supercomputers we will generate and compare thousands of structures. The results will give Czech research groups their own tested version of this breakthrough technology and open the door to faster drug discovery and biology research.

Researcher: Jakub Velímský           

OPEN-38-71 

Adjoint sensitivity kernels of the ocean induced magnetic field

Karolina CPU Alloc=72500 

The Earth’s oceans cover more than 72% of the planet’s surface. The global ocean circulation system is tightly connected to the evolution of the weather  and climate, with profound consequences for life on Earth and human society. The seawater has high electrical conductivity and as it moves through the  geomagnetic field, the charge carriers are subject to a Lorentz force, perpendicular both to the ambient field and the flow direction. This results in an  electric current, which interacts with its surroudnings through galvanic channelling and inductive coupling, giving rise to the ocean-induced magnetic  field (OIMF). Observations of the magnetic fields by low-orbit satellites, such as the ESA mission Swarm, can thus in principle convey the information about the ocean electrical conductivity, and the vigour and pattern of the ocean flow. Following our previous parametric study, where we have predicted the OIMF for individual ocean models, the goal of this project is to calculate the 4-D sensitivity kernels which directly relate the satellite-observed OIMF to local values of ocean flow velocity and electrical conductivity (and indirectly to ocean temperature and salinity). These kernels represent a prerequisite to direct assimilation of satellite-observed OIMF in the ocean models.

Researcher: Evangelos Kazakos   

OPEN-38-8    

Language-conditioned model-predictive control over a 3D diffusion world model for robotic manipulation      

Karolina GPU Alloc=3700; LUMI-G Alloc=6500    

Robots still struggle to manipulate objects reliably in unstructured, everyday environments. A promising approach is to let a robot simulate the consequences of its actions before executing them, and then act on the plan that looks best. This is the purpose of a world model: a learned simulator of the physical world. Trained on large amounts of robot interaction data and RGB-D scenes, such models allow a robot to imagine plausible futures and plan against them. This project builds and tests a robot controller that plans over a world model which predicts the future as explicit, measurable 3D geometry seen from several cameras, and specifies the task in ordinary language (for example, “put the cup on the shelf”) instead of by hand-tuned rewards. Because the imagined future is real geometry, the controller can check physical facts, such as whether the object is actually resting on the surface, rather than trust an unverifiable score. We will measure whether the imagined futures faithfully follow the commanded actions, and compare planning over explicit geometry, predicted appearance, and the model’s internal code, to establish when physical 3D grounding makes a robot more trustworthy. Applications include assistive robotics, logistics, and flexible manufacturing.

Researcher: Subhasmita Ray         

OPEN-38-9    

Selective Phonon Modulation of Magnetic Anisotropy in CrX3 Monolayers (X = Cl, Br, I): A van der Waals Functional Benchmarking Study  

Karolina CPU Alloc=49900; LUMI-C Alloc=7500  

Lattice vibrations directly affect magnetic anisotropy and exchange in two-dimensional van der Waals magnets (CrCl3, CrBr3, and CrI3), allowing for phonon-driven (as opposed to electrical) magnetic control. Using a frozen-phonon DFT approach (VASP) with Phonopy for eigenvector determination, this project targets the zone-center (Γ-point) Raman-active optical phonon modes (Eg, Ag) of CrX3 monolayers, the vibrations most directly accessible by Raman spectroscopy, and investigates how displacement along these modes modulates magnetic anisotropy energy and FM/AFM exchange. The calculations are efficient since no supercell is needed because Γ-point modes maintain the unit cell. The study also benchmarks a number of cutting-edge vdW-corrected functionals since multilayer vdW magnets are quite sensitive to the choice of dispersion correction. This ensures that any trend in magneto-phonon coupling is physical rather than a functional artifact.