ZOZNAM ÚSPEŠNÝCH PROJEKTOV VO VÝZVE 2/25:
Study of kinetic pathways of structural phase transitions in crystals by metadynamics simulations with machine learning potentials
Roman Martoňák
The project focuses on computer simulations of pressure-induced structural phase transitions employing metadynamics with generic collective variables (such as coordinations numbers and volume) and machine learning (ML) based potentials. We plan to study the compression of α-cristobalite silica (SiO2) under hydrostatic and non-hydrostatic conditions.
Theoretical prediction of vibrational spectra and thermodynamics of interacting systems
Monika Gešvandtnerová
Within this project, we shall investigate vibrational spectra beyond simple harmonic approximation. To this end, the use machine learning based strategies will be tested. Among others, we shall investigate experimentally measured change of experimental IR spectra during isobutanol conversion into butenes, or Raman spectra of perovskite-based semiconductors. Furthermore, we shall study finite-temperature thermodynamic properties of various systems of technological relevance beyond simple static approach. In particular, we will develop and apply a computational strategy based on machine learning perturbation theory to study effect of presence of one or more extra water molecules on isobutanol transformations catalyzed by acid zeolite.
Self-adhesivity of asmorphous Al2O3 oxide. Phase transition between dense metallic liquid and low-density amorphous semiconducting phases of Germanium.
Marek Mihalkovič
Upon pressurization and annealing, Al nanoparticles covered by 5 nm thick layer of amorphous Al2O3 form a compact material with remarkable mechanical properties. The critical ingredient of this process is self-adhesivity of amorphous Al2O3 surface layer. The structural link of this property is unclear: what is the role of amorphous structure? Clarification of this aspect is important for design of this class of materials.
We will model the self-adhesivity of amorphous Al2O3 in atomistic simulations at pressures of 1-2 GPai, and in collaboration with Peter Krizik from UMMS SAV, correlate simulation results with experimental evidence.
In the second part of this project, I will build upon unpublished preliminary result of my previous project p373-23-1 – observation that ultrathin layer of Ge undergoes phase transition from liquid metallic into amorphous semiconducting phase. In collaboration with M. Widom from CMU Pittsburgh, we will carefully prove this transition for bulk Germanium, by performing series of molecular dynamics simulations with varying volume/pressure.
The two parts of the project will share the same simulation approach: molecular dynamics and machine learning force fields, all within the package VASP. In addition, we will explore experimental option of LAMMPS, implementing VASP MLFF interaction models.
Next-Generation Energy Materials: High-Entropy Perovskites for a Sustainable Future
Imrongnaro Longkmer
Perovskite oxide materials have attracted significant interest for thermoelectric applications due to their low toxicity, environmental compatibility, and abundance of constituent elements offering distinct advantages over lead- and tin-based chalcogenides. However, their thermoelectric efficiency is limited by inherently low electrical conductivity and relatively high thermal conductivity. These shortcomings are primarily due to short phonon mean free paths, which make substantial suppression of lattice thermal conductivity challenging. This project aims to systematically study a range of perovskite oxide structures through doping at both A- and B-sites with multiple elements. By evaluating the thermoelectric properties of these doped systems across various molar ratios, we seek to identify optimal candidates for thermoelectric applications. The computational component of the project will involve detailed electronic structure analysis within the framework of Density Functional Theory (DFT), utilizing the VASP code. Electronic transport properties will be assessed using BoltzTraP, which employs the Boltzmann transport formalism under the constant relaxation time approximation. Lattice thermal conductivity will be evaluated using ShengBTE and/or almaBTE, supported by harmonic and anharmonic force constants derived from the finite displacement method. In parallel with simulations, high-entropy perovskites (HEPs) with varying dopant compositions will be synthesized using a combination of mechanical alloying and spark plasma sintering techniques.
Lean Neural Networks
Vladimír Boža
Neural networks have demonstrated remarkable success in various domains, such as computer vision, natural language processing, protein folding, drug interaction prediction, and many others. However, most cutting-edge applications rely on large neural networks with billions of parameters, demanding substantial computational resources (e.g., expensive GPUs) for training and inference. Large neural networks cannot be easily used in embedded or low-power devices, where the amount of available computational power is limited. Running a neural network in offline systems is beneficial in many cases, especially in privacy-sensitive environments.
Even though widespread use of large neural networks is likely inevitable, specific tasks can be handled by much leaner networks that can be deployed on regular computers, mobile devices, etc. This should expand the possibilities for applications and lower deployment costs. This project aims to design and develop innovative techniques for training lean neural networks that maintain high accuracy while drastically reducing required computational resources. Thus, these networks will be accessible for deployment on regular computers and mobile devices. We will mainly focus on improving the efficiency of sparse neural networks. Our primary targets are large language models, but our methods are applicable to any neural network.
Evaluating and improving the training efficiency of AI models
Róbert Belanec
This project aims to evaluate and improve the training efficiency of artificial intelligence models, particularly Large Language Models (LLMs), with a focus on faster convergence and parameter efficiency. Motivated by the growing computational demands of state-of-the-art models, the research addresses three key challenges: evaluating optimization methods, benchmarking parameter-efficient fine-tuning (PEFT) techniques, and analyzing multi-task training factors. The project will involve extensive experiments across diverse tasks in natural language processing, computer vision, and reinforcement learning, using models such as LLaMa, Mistral, BERT, and GPT-2. A novel benchmarking framework for PEFT methods will be developed and released as open-source, alongside a public leaderboard. The expected outcomes include new insights into the efficiency of PEFT methods, improved optimization strategies, and knowledge about multi-task learning. This work aims to reduce the computational cost of training LLMs, to allow more sustainable and accessible AI research.
Designing Defect-Engineered Perovskites for Enhanced Charge Transport and Thermal Stability
Inga Zhukova
We propose a high-throughput computational study to explore defect engineering in perovskite materials aimed at improving their electronic and thermal properties. Perovskites, with the general formula ABX₃, exhibit wide tunability in their composition, making them prime candidates for applications in photovoltaics, thermoelectrics, and catalysis.
While the stability and performance of classical perovskites such as SrTiO₃ and MAPbI₃ have been extensively studied, emerging interest lies in tailoring defect chemistry and mixed-metal compositions to overcome limitations in charge mobility, degradation, and thermal resistance. Our study focuses on simulating point defects, doping strategies, and lattice strain effects on novel perovskite systems using high-performance computing resources.
Insights into the preaggregation conformation and early aggregation stages of intrinsically disordered protein tau
Ondrej Cehlár
As a member of the class of intrinsically disordered proteins (IDPs), tau protein in physiological state doesn’t acquire a clearly defined 3D structure, which makes the process of finding potential inhibitor of tau aggregation radically difficult. Therefore we aim to simulate molecular dynamics of truncated tau proteins capable to attain pre-aggregation tau conformation in monomeric state as shown by interaction with specific antibodies. Further we plan to simulate the early steps of aggregation of truncated tau proteins. MD simulations will be run using Gromacs.
In silico search for efficient small molecule drug targeting tau protein in the pathogenesis of Alzheimer’s disease
Vladimír Garaj
The fact that the only disease modifying therapy for Alzheimer’s disease being expensive biological therapy with anti amyloid B antibodies represents a significant burden for the healthcare systems worldwide. Therefore the obtainment of small molecule inhibitor of tau pathology may be a groundbreaking scientific and pharmaceutical result. In this project we aim to find a small molecule inhibitor of tau pathology out of libraries of available molecules using molecular docking, and characterize its interaction with disordered tau protein. The target structures of tau protein for docking will be obtained from extensive molecular dynamics simulations. The best lead compounds will be tested experimentally and will be further improved for higher affinity and better selectivity and ADMET properties.
High-Resolution Deep Learning Volumetric Precipitation Nowcasting for the Slovak Region
Peter Pavlík
Deep learning has rapidly advanced atmospheric science by enabling data-driven modeling of complex weather systems. Our research focuses on precipitation nowcasting—a critical capability as climate change increases the frequency of extreme rainfall events. Traditional 2D radar-based methods struggle to capture the vertical structure of convective systems, limiting forecast accuracy. We propose a volumetric nowcasting approach using 3D radar reflectivity and deep learning. Building on our Lagrangian model LUPIN, we aim to generate true 3D motion fields for more physically consistent predictions. We also plan to train models on sub-kilometer radar data, which has shown promise in enhancing resolution and accuracy. This work targets improved short-term forecasts for Slovakia, where training data remains sparse, and aims to support public safety and climate resilience.
Tailoring High-Entropy Perovskites for Next-Generation Thermoelectrics
Subrahmanyam Bandaru
Perovskite oxide materials have been extensively studied for thermoelectric applications due to their low toxicity, eco-friendliness, and high elemental abundance, offering advantages over tin and lead chalcogenides. However, their relatively poor thermoelectric performance is hindered by low electrical conductivity and high thermal conductivity. This limitation arises from the intrinsically low phonon mean free paths in these perovskites, making it difficult to achieve significant thermal conductivity suppression. This project aims to compute various perovskite oxide structures and investigate the impact of doping with multiple elements at the A- and B-sites. The thermoelectric properties of doped perovskite oxides with varying molar ratios will be analyzed and compared to identify the most suitable candidate for thermoelectric applications. The general objective of the present work is to perform the extensive, systematical studies of the electronic structure calculations in the framework of the Density Functional Theory (DFT) of High entropy perovskites (HEPs) materials to select the best candidates of desirable thermoelectric properties. The theoretical study to analyze the electronic structure will be made using the computer code VASP based on the DFT. The electronic transport properties will be analyzed using the code BoltzTraP which is based on the Boltzmann transport theory within the constant relaxation time approximation. The lattice thermal conductivity will be analyzed by ShengBTE and/or almaBTE. HEPs with different compositions of dopants will be synthesized by the combination of mechanical alloying spark plasma sintering in parallel to the simulations.
Interactions of albumin with poly(2-methyl-2-oxazoline) coated graphene
Zuzana Benková
The prevention or control of protein adsorption onto surfaces has significant applications diverse fields, ranging from medicine to industrial coatings. Despite the important role of surface coating by end-grafted polymer chains, the underlaying mechanism of protein repulsion remains elusive. Currently, there is no theoretical framework that adequately clarifies the inconsistencies observed in experimental findings. To bridge the gap between theoretical and experimental approaches, atomistic molecular dynamics simulations will be performed to scrutinize the interactions of poly(2-methyl-2-oxazoline) (PMOX) chains grafted onto graphene and human serum albumin (HSA), which is characterized by a secondary structure abundant in α-helices. To explore the role of water molecules in these interactions, simulations will be executed both in aqueous environment and in a vacuum. A large range of grafting densities of PMOX chains will be considered to cover different conformational regimes of the grafted polymer layers. Additionally, attention will be focused on the effect of these conformations on the interactions with HSA and on the secondary structure of HSA. Specifically, the inquiry will be addressed whether the soft layer of grafted chains serves to stabilize or destabilizes the α-helices of HSA. The systems will be studied in two setups: one that simulates biological conditions in living organisms (setup I) and another that corresponds to the surface force apparatus technique (setup II). A key interest lies in comparing whether the adsorption/repulsion of HSA in setup I coincides with the presence/absence of a negative minimum in the free energy profile of HSA in setup II. In other words, this comparison aims to reveal whether the two setups yield generally consistent results or if there are specific conditions under which they diverge. The propensity of the grafted PEO layers to either repel or attract HSA will be evaluated across all simulated systems.
Machine Learning and Geometric Methods in 3D Vision
Viktor Kocur
The goal of this proposal is to enable the researchers at the Department of Applied Informatics, Faculty of Mathematics, Physics and Informatics, Comenius University Bratislava to acquire access to the Devana supercomputer. The scientific goal of this project is to research novel methods in the field of 3D vision. 3D vision is a subfield of computer vision and mostly studies methods and algorithms enabling machines or computer systems to perceive and understand the three-dimensional structure of objects and environments. Following the general trends in the broader field of computer vision, compute-intensive methods such as deep learning are often employed in 3D vision making the utilization of specialized HW (such as GPUs) necessary. Geometric methods also play an important role within the current 3D vision research. Evaluating the accuracy and computational efficiency of such methods may require evaluation on large-scale datasets requiring significant computational resources. More specifically, in this research project we aim to investigate various ways in which modern machine learning or deep learning based approaches may be utilized in combination with geometric methods. We plan to publish our findings at top-level computer vision conferences and journals.
Development and applications of numerical modelling of seismic motion, seismic rupture and seismic ambient noise
Martin Galis
One of the important factors contributing to the minimization of earthquake damage is a sufficiently accurate estimation of the seismic motion during future earthquakes. Seismic motion can be anomalously amplified or prolonged by surface sedimentary structures. In the absence of seismic observations, numerical modeling is an irreplaceable tool to better understand the conditions determining seismic motion and to refine the estimation of seismic motion. Given the evolution of computational resources, the increasing demands for model accuracy and frequency range, it is essential to adequately develop sufficiently accurate and efficient numerical methods. Therefore, the project is primarily focused on the development of numerical methods for modelling of seismic motion. In the project we will also address simulation of fault rupture propagation and seismic ambient noise.
Multi-physics modeling of high-field magnets and superconducting motors
Enric Pardo
Superconductors enable to transport huge electric current with no resistance when temperature is below a certain threshold, Tc. Thanks to this, superconducting electromagnets can generate high magnetic fields like in ultra-high field medical magnetic resonance, reaching above 11 T, and electromagnets for material research, reaching above 40 T. High temperature superconductors like REBCO can operate at higher temperatures than classical Low-Temperature Superconductors (LTS), such as NbTi or Nb3Sn. This opens the gate for electric power applications for decarbonized economy, such as propulsion motors for zero-emission hydrogen-electric aircraft for passenger flights. Liquid hydrogen acts as both fuel for electric fuel cells and coolant for the motor. This project will enable to make parametric studies for the design and optimization of both high-field magnets and motors for aviation. Computations will use our own software, which use innovative mathematical methods. This work will bring greener technology closer to reality, such as zero-emission air transport, wind electric generators, or clean fusion energy.
Molecular Mechanisms of Carbohydrate-Mediated Protection of Human Hemoglobin during Freeze-Drying: An MD Simulation Approach
Ivan Klbik
Freeze-drying (lyophilization) is a key method for preserving therapeutic proteins such as human hemoglobin, but it imposes significant stress that can destabilize protein structure. Carbohydrates like glucose, trehalose, and raffinose are widely used as stabilizers during this process. Experimental evidence shows that while their protective effects vary on a molar basis, they become comparable when normalized by weight fraction—suggesting that protection may stem from non-specific, physical interactions with the protein surface.
This project aims to elucidate the molecular mechanisms by which these carbohydrates stabilize hemoglobin during freeze-drying using classical molecular dynamics (MD) simulations. By modeling different phases of lyophilization (hydrated, cooling, frozen, and desiccated), we will examine sugar–protein interactions and monitor the structural integrity of hemoglobin. Key analyses will include hydrogen bonding, surface coverage, sugar clustering, and conformational changes in the protein.
Simulations will be performed using GROMACS, optimized for GPU acceleration, and executed on the Devana HPC cluster, which enables efficient high-throughput execution through parallel job scheduling. The findings will offer mechanistic insights into carbohydrate-mediated stabilization, guiding formulation strategies in both biomedical and aerospace contexts.
Chiral interplay in knotted supercoiled double stranded dsDNA via Ox-DNA
Renáta Rusková
This project harnesses High-Performance Computing to explore the intricate interplay between the intrinsic chirality of DNA (right-handed double helix), the chirality of DNA supercoiling (predominantly negative in living organisms), and the chirality of knotted DNA structures. By leveraging molecular dynamics simulations with OxDNA—a unique coarse-grained model that faithfully represents the double-helical structure and includes sequence-dependent interactions—we aim to provide new insights into how these chiral properties interact and influence DNA topology and dynamics.
Lunar Surface Gravitational Maps from Satellite, Topography and 3D Density Data
Blažej Bucha
The project develops high-frequency components of the lunar surface gravitational field by modelling gravitational effects implied by the shape and density of the Moon. This is achieved by decomposing the lunar topography into simple elementary bodies called tesseroids. Unique to this project is the use of realistic 3D crustal density models as opposed to simplistic constant-density-based approaches from previous studies. After high-pass filtering in the gravitational domain, the topography-implied gravitational effects are finally combined with satellite gravitational information to deliver full-scale gravitational maps as the main project’s goal. The output lunar surface gravitational maps can find their applications in, for instance, scanning for prospective landing sites of future lunar exploration missions, designing entry, descent and landing of spacecraft or adopting the Earth’s concept of mean-sea-level heights to accurately reference the lunar topography.
MaxEnt-based Habitat Suitability and Predictive Modelling to Define Biome Boundaries
Dušan Senko
Understanding distribution limits will be essential for delineating biome boundaries. We will use habitat suitability and ecological niche modelling with MaxEnt to evaluate distribution patterns. Based on curated occurrence datasets and environmental predictors, including climate and geology, we will identify key environmental factors shaping species distributions and test the temporal stability of their niches. Our models will demonstrate high predictive performance and will reveal consistent habitat preferences, with limited potential for expansion beyond current ecological envelopes. This approach will highlight the value of collection-based data and niche modelling in advancing biome-scale ecological understanding.
In silico study of mutations in the allosteric pathways of RyR and Wfs1
Alexandra Zahradníková
Ryanodine receptors (RyR1, RyR2) and Wolframin (Wfs1) are membrane ion channels indispensable for cell function. Mutations in their genes cause debilitating or even potentially fatal diseases: malignant hyperthermia (MH, RyR1), catecholaminergic polymorphic ventricular tachycardia (CPVT, RyR2), and Wolfram syndrome (WS, Wfs1). Over two hundred pathogenic mutations were observed in each of the genes ryr1, ryr2, and wfs1. Due to the complexity of the experimental approaches, the mechanism of their effect is not sufficiently clear for the development of effective therapies. A more effective pathway of investigation is provided by structural and computational biology and their in silico tools.
The project aims to reveal the action mechanisms of selected mutations of the above proteins. Protein structures, taken from databases or created by Alphafold3, will be mutated in silico and subjected to MD simulations. The effect of the mutations on the overall protein dynamics and the occurrence and dynamics of allosteric pathways will be analyzed and correlated with the dynamics of important hydrogen bonds and inter-domain contact formation in molecular dynamics (MD) trajectories of the proteins‘ closed, open, and inactivated states. Subsequently, we will determine the effect of mutations on ion channel activity using thermodynamic models based on statistical mechanics. We will formulate a hypothesis on the possibilities of pharmacological rectification of the mutation-compromised channel function, potentially leading to reduced health consequences.
The HPC part of the project requires extensive MD simulations using the Gromacs software. The remaining in silico methods are established locally at the principal investigator’s laboratory using OHM, VMD, Chimera, CHARMM-GUI, OPM, and in-house software.