Researcher: Václav Vávra            


Camera pose estimation from bounding boxes

Karolina CPU  Alloc=190;  LUMI-G  Alloc=1775  

This project aims at finding novel methods in camera pose estimation. Specifically, a new minimal solver for camera pose estimation using only two 2D-3D correspondences was developed. It has shown some interesting properties and it seems like it may very well complement the existing state-of-the-art minimal solvers or even exceed their performance for some applications. To measure the performance of the new and baseline minimal solvers, the effort is now being made to prepare the needed data derived from the original dataset images (segmentation, local features) as well as to measure the performance of the compared solvers on large datasets (Mapillary Metropolis, ARKitScenes). Also, as part of the project, a new depth-estimating CNN will be trained to estimate depth from monocular images as the information about


Researcher: Martina Ćosićová  


Testing of diabatization using neural network approach

Karolina CPU  Alloc=1000            

In this project, I would like to focus on the use of neural networks for diabatization of electronic bases. The task consists in finding a numerically convenient (so-called diabatic) representation of the potential energy matrix that appears in the system of partial differential equations for the motion of atomic nuclei. This system is obtained by Born-Oppenheimer separation of the electron and nuclear degrees of freedom in time-dependent Schrodinger equation describing a given physical system – in this project, I would like to test the method on rare gas ion clusters. The use of neural networks for diabatization is  a new and promising approach which, if successful, could significantly reduce the computational demands of dynamical simulations of quantum systems. This problem is part of my dissertation, which I am working on as part of a cotutelle agreement in collaboration with Université Toulouse III - Paul Sabatier, the project is supported by the international Barrande Fellowship program.


Researcher: Michael Bakker       


Leveraging Clustering Algorithms for Improved Chemical Shift Prediction Through Dimensionally Reduced Ensemble Clustering          

Barbora CPU  Alloc=8300;  Barbora FAT  Alloc=400;  Barbora GPU  Alloc=500;  Karolina CPU  Alloc=2800;  Karolina FAT  Alloc=400;  Karolina GPU  Alloc=2300    

In the contemporary biochemistry research arena, intrinsically disordered proteins (IDPs) present a particular challenge. Biochemically, pharmaceutically and medicinally, they pose challenges for researchers due to their higher propensity for assembly and interaction with other proteins/targets, as well as their inherent stochasticism. Scientists have uncovered that many of these proteins are associated with devastating illnesses such as cancer, Alzheimer’s, Parkinson’s, Type 1 diabetes and many immune system diseases. [1] Post-translational modifications (e.g. phosphorylation) impart functional complexity on these proteins, enabling intricate control over gene expression, protein-protein interactions, and enzymes. For a better understanding and analysis of these phenomena and proteins, researchers need to construct a smaller set of representative structures from trajectories of the complete molecular dynamics. In this investigation we will be delving into the dynamics of two vital IDPs, hTH1 and Map2C, and comparing them with two ordered proteins, GB3 and Ubiquitin. Smaller conformational ensembles (CEs) can be generated sequentially (traditional method), but utilizing many revolutionary machine learning techniques such as dimensionality reduction (DR) and clustering algorithms (CA) can be employed to better refine these CEs. From this investigation, we will be generating new trajectories for analysis, refining the investigation of these trajectories into ensembles, and deriving a polished method of replicating this methodology.


Researcher: Elliot Michael Rothwell Perviz          


Nanofriction induced Sliding modes of 2D TMD heterostructures (NANO-SLIDE)

Barbora CPU Alloc=45800;  Karolina CPU Alloc=10900   

Transition Metal Dichalcogenides have been widely applied as thin film solid lubricants in recent years, owing to their exceptional frictional properties. These arise due to their topologically flat sandwich-like structure, where weak inter-layer bonding facilitates low shear strength allowing for crystalline sliding. However, issues such as instability in dynamic and humid conditions creates major challenges for their design, control and reliability. Previous studies have explored the possibilities of using dopants, and/or heterostructures of TMDs (pure or with dopants) to address these issues, which have lead both to improvements in the coefficient of friction and lifetime of the materials. From a theoretical perspective, we want to understand the physical mechanisms which govern energy dissipation due to nanofriction within these materials.

To this end, we will perform a computational study to determine the phonon modes and their interactions in a selection of bilayer TMD heterostructures. We will relate the phonon modes and their anharmonic effects to structural and electronic properties of the systems to understand the source of energy dissipation. Ultimately, we seek to develop rules to systematically design materials with minimal energy dissipation.


Researcher: Assia Benbihi           


3D Reconstruction and Feature Matching Benchmarks  

Karolina CPU Alloc=10000;  Karolina GPU Alloc=6800    

This project proposes two computer vision benchmarks related to 3D reconstruction and feature matching in urban environments under a “walking pedestrian” scenario. The availability of well-defined benchmarks facilitates progress on a wide array of computer vision, machine learning, and robotic problems by allowing researchers to evaluate and compare their contributions under a unified setup. These benchmarks aim to summarize real-world challenges in carefully curated datasets and provide an analysis of the state-of-the-art that simplifies experimental evaluations, encourages open-source and repeatable research, highlights new research directions relevant to the community, and spurs research progress. Benchmarks are all the more appreciated as they require considerable effort to collect data in a way meaningful for the task at hand. The data must exhibit new challenges yet to be tackled, provide annotations to enable the evaluation, and be memory and runtime-efficient. Also, the evaluation of existing methods is computationally heavy and the analysis of the results, while extremely insightful to the community, requires a serious time investment. With the release of two benchmarks, we propose a significant contribution that will benefit both the research community and the industrial applications related to 3D reconstruction and feature matching.


Researcher: Petr Valenta             


Intense and Compact Muon Sources for Science and Security     

Karolina CPU Alloc=35000;  Karolina GPU Alloc=2400    

Thanks to muon radiography, we can see inside seemingly impossible places, such as nuclear reactors, volcanoes, tsunamis, hurricanes, and Egypt’s Great Pyramid of Giza. This imaging technique uses naturally occurring subatomic particles called muons, which can penetrate far deeper than possible with x-rays through material as thick and dense as 30-meter concrete walls. But this process is also slow. Due to the low flux of naturally occurring muons, these images require exposure times on the order of months. This project aims to develop terrestrial, portable, laser-based muon emitter with much higher flux than naturally occurring muons and harness the lab’s world-class laser technology and expertise to lay the groundwork for imaging breakthroughs.


Researcher: Zdeněk Futera        


Interactions of biomolecules with metal surfaces affecting their electronic transport properties

Barbora CPU Alloc=20000;  Karolina CPU Alloc=15000   

Biomolecules such as peptides, proteins, or nucleic acids can surprisingly strongly interact with solid metal surfaces. For example, in electrochemical measurements, solvated redox proteins in contact with gold electrodes spontaneously adsorb on their surfaces. The adsorption is typically non-covalent, facilitated by stacking of organic amino-acid residues to the metal or relatively strong interaction with available amino groups. However, in the presence of cysteine residues bearing the thiol groups, even the covalent bonding, often used to anchor such proteins to the electrode surfaces, occurs. Nevertheless, the mechanism of such bonding in the case of disulfide bridges between two cysteines is not fully understood. The bridges are known to dissociate near the gold surfaces while forming Au-S bonds, but it is not clear how much these chemical changes affect the protein structures, how far are the two gold- sulfur bonds from each other, and whether these chemical processes have any effects on the conductance of the adsorbed redox proteins. Here, we investigate these phenomena by first-principles calculations within the frame of the density functional theory to get better insights into the electronic properties of bio/metallic interfaces.


Researcher: Dušan Knop             


Algorithms design and testing for theory             

Barbora CPU Alloc=10000;  Karolina CPU Alloc=1500     

The proposed project aims on experimental evaluation of algorithms and algorithmic experiments in theoretical computer science. Algorithm design unfolds on two fronts, and our aim is to closely scrutinize both aspects. The synergy between experimental evaluation and algorithmic design is crucial; a deeper understanding of a problem from one perspective should propel advancements in the other. However, this symbiosis is not always realized, particularly with algorithms that remain confined to theoretical realms. Conversely, real-world computation often relies on heuristic-based algorithms, lacking the comfort of theoretical guarantees. Our mission is to bridge these two distinct worlds, fostering a harmonious relationship between theoretical elegance and practical utility. By exploring this dynamic interplay, we seek to unravel novel insights that transcend the boundaries of traditional algorithmic paradigms, propelling us toward a more holistic and impactful understanding of algorithm design.


Researcher: Jan Lehecka             


Multimodal Transformers for Speech and Text   

LUMI-G Alloc=6000       

The project's goal is to continue our computationally challenging research in the field of Artificial intelligence with a focus on speech technologies. Our team has strong experience with training state-of-the-art (SOTA) models for many speech- and text-related tasks, such as Automatic Speech Recognition (ASR), Text-to-Speech (TTS), Document Classification, etc. So far, we have been using modal-specific models to solve each of these tasks, e.g. to solve document classification tasks, we used text-only models, to solve ASR tasks, we used acoustic models, etc. However, the nature of many tasks (and also the trend in the literature) leads us unavoidably to multimodal models capable of processing and generating both text and speech in a unified way. The key idea behind multimodal models is to encode all input modalities (e.g. an input audio record together with a text prompt) into a common feature space and then, based on these encoded embeddings, to generate desired output (e.g. an answer of the model as a text or speech). Training such models requires a large amount of data from speech and text modalities and a lot of computing resources.


Researcher: Barbora Venosova


Optical and magnetic properties of MXene-based quantum dots              

Barbora CPU Alloc=30000          

The variety of compositions and structures of MXenes has led to the formation of a large and rapidly expanding family of 2D materials. MXenes have a unique combination of properties that have led to many applications. In the last six years, there has also been increased interest in MXene-based quantum dot synthesis (MXQD). Due to the quantum confinement effect, their properties such as higher chemical stability, better electronic, magnetic, and optical properties, good biocompatibility, various functionalizations, etc. can be significantly improved compared to their 2D counterparts. However, despite the growing interest in their most attractive properties (i.e., optical properties), they are still less explored than expected. One of the major challenges is to study in more detail the structure of quantum dots, i.e., how the change of metal atoms, different surface, and edge terminations and/or size of MXQDs affect the properties. Here, we propose a systematic theoretical study to model diverse types of MXQDs as a tool to tune the electronic, optical, and magnetic properties of MXQDs and thereby provide details on the potential use of MXQDs in photoelectronic, spintronic, or photocatalytic applications.


Researcher: Vladimir Petrik        


Object 6D pose estimation from images for unknown 3D models             

Karolina GPU  Alloc=7200           

Object pose estimation is one of the fundamental problems in computer vision with important practical applications in autonomous robotics, enabling robots to interact with the surrounding environment. An example of an important use case is a new generation of AI-driven robotic assistants for homes or industrial environments. The research project aims to explore 6 DoF pose estimation using image methods trained on very large-scale 3D object databases. With the recent release of Objaverse/Objaverse-XL datasets, we can now train 6 DoF pose estimators on a large set of 3D objects and aim to explore the data scale to learn truly generalizable methods that will work in zero-shot settings on novel objects when an exact mesh is unavailable. We will develop a novel large-scale 6 DoF estimation method that will focus on two things: a) learning robust estimators when the 3D models are available and b) extending this method for cases where only an approximate 3D model exists.

This is an important problem for real-world applications, especially for robotics, since modern 6 DoF estimation methods lack robustness when encountering objects that lie on the long tail of object distribution in the training data or when the 3D model of the object of interest is not readily available. By exploring large-scale mesh collections, we aim to increase the generalizability of 6 DoF models so that they can deal with rare objects out of the box and with cases when an accurate 3D model is not available.


Researcher: Jan Zemen


Modeling of martensitic transformation in Heusler alloys             

Barbora CPU Alloc=38400;  Karolina CPU  Alloc=5120    

Heusler alloys (HA) are a broad class of ternary compounds of general formula X2YZ where X is a late and Y an early transition metal and Z metalloid or semimetal. The so called martensitic transformation (MT) has been investigated in a subgroup of HA systematically. MT is a structural diffusionless and displacive transition to a low symmetry phase. It is often associated with large magnetocaloric and magnetoelastic effects. The occurrence of MT in a given alloy is extremely sensitive to chemical composition of the alloy which constitutes a great challenge to materials science. The sensitivity may originate in the interplay of chemical (dis)order with specific magnetic order which can be investigated in detail by local methods – Mössbauer spectroscopy (MS) and nuclear magnetic resonance (NMR) making use of the hyperfine interactions. Disorder can also affect the hysteresis of MT which is a crucial parameter for the efficiency of magnetocaloric materials and other energy-conversion devices. In our project we intend to study HA where X = Fe, Ni and Co, Y = Mn, and YZ = Ga, Sn and their off-stoichiometric derivatives. We will use density functional theory (DFT) to relax large supercells incorporating the chemical disorder to find the local magnetic structure and determine the likelihood of the sought-after MT by comparing the density of states (DOS) of the austenite and martensite structure. We will build on our recent DFT results showing a formation of a sharp peak in DOS close to the Fermi level induced by partial Fe-substitution. The new results will be used in future as an input of a full-potential DFT study of the hyperfine parameters to be compared to values observed in MS and NMR experiments.


Researcher: Jakub Luštinec        


Magnetoelastic coupling in Ni-Mn-Ga martensite            

Karolina CPU Alloc=5000             

The Ni2MnGa Heusler alloy has a prominent position among other alloys because it is the only system where a combination of high magnetic anisotropy and supermobility of twin boundaries in the martensitic phase has been experimentally observed. This gives this system unusual behaviour and new functionalities, such as magnetic shape memory or giant magnetostriction. Despite more than twenty years of extensive study of this alloy, the reason for its extraordinary properties is still unclear, and the search for other alloys with similar behaviour has also been unsuccessful. Here we propose the use of recent theoretical tools (relativistic density functional theory computation) to unravel this mystery more closely. We will study the stability of the martensite lattice and its dependence on the magnetic arrangement by evaluating the elastic constants of martensite for differently oriented magnetic moments. Such complex calculations have recently been made possible thanks to the development of fully relativistic pseudopotentials to solve problems of general magnetic order and thanks to large-scale computing infrastructures such as IT4 Innovations.


Researcher: Mithun Manikandan


Computational Design and screening of Self-assembling molecules driven by Hydrogen Bonding (CompHBond) 

Barbora CPU Alloc=40000          

Contemporary computer technology based on silicon transistors is pushed to its limits by the ever increasing demand for performance. Computations now consume a substantial part of energy produced by our civilization, and this share is expected to increase. Molecular switches, e.g. in the form of memristors or cellular automata, promise the ultimate level of miniaturization, dramatic reduction of power consumption and eventually even implementation of quantum algorithms. But mass production of complex computational systems from molecules remains to be an unsolved challenge. We try to address this challenge by designing photosensitive molecular templates able to combine molecular self-assembling with photolithography into a single nanofabrication method. This should be achieved by polymer templates analogous to DNA, consisting of photosensitive polydiacetylene backbone and end groups analogous to nucleobases. In order to find the optimal design of such polymer templates we conduct a computational survey of a large number of hydrogen bonded molecules searching for optimal end groups governing the selective self-assembly of these templates in an anhydrous environment on ionic substrates.


Researcher: Mireia Diez Sánchez             


Towards Robust End-to-End Diarization and Source Separation  

LUMI-G  Alloc=8000      

Speaker diarization (SD) is the task of automatically determining speaker turns in conversational audio. Despite the big advances with large pretrained models such as the popular ChatGPT on conversational text processing, conversational audio processing is yet an extremely challenging task for machine learning approaches. The recent development of end-to-end neural diarization (EEND) systems has boosted research in the field and shifted the paradigm on how to handle conversational speech. In this project we plan to move the field forward by finding methods to optimize the training strategies; by combining the benefits of well-founded generative models with the powerful end-to-end approaches and by exploiting the synergies of the related speech separation (SS) task to enhance SD performance.


Researcher: Matus Dubecky      


Electon correlation effects by fixed-node diffusion Monte Carlo method

Karolina CPU Alloc=40000          

Distinguishing between dynamic and nondynamic electron correlation energy is a fundamental and important concept in quantum chemistry. We have recently introduced a novel approach [J. Chem. Theory Comput. 19, 8147–8155, 2023] for partitioning an exact electron correlation energy into its dynamic and nondynamic components by restricting a ground-state solution to share its node with a spin-restricted Hartree–Fock Slater determinant. The fixed-node diffusion Monte Carlo method was used as a convenient tool to project out a lowest-energy state obeying such a boundary condition. The proposed approach provided unambiguous and useful insights into the effects of electron correlation that align with common knowledge and experience in the field of quantum chemistry. In the current project, we plan to extend this basic conceptual/theoretical work by providing numerical proof-of-concept examples of the method's use, in order to illustrate its wide application potential.


Researcher: Samuel Lukeš          


ERO2.0 simulations of tungsten & liquid metals on COMPASS Upgrade tokamak

Karolina CPU Alloc=5000             

With the ever-increasing energy consumption of mankind, the need to use ecologically and politically acceptable, reliable 24/365 and inherently safe sources also increases. One of the very few ways to satisfy these requirements seems to be nuclear fusion. That is why its flagship research project ITER is currently the most expensive science experiment on the planet. However, several unanswered questions still remain, one of them is the long-term reliable heat shield of the reactors. Current solid shields (W, Be) must withstand extreme heat fluxes from thermonuclear plasma in the form of charged energetic particles, which significantly degrade any crystalline lattice or overheat the matter to the point of cracking or melting. Liquid metal technology is supposed to solve this. Compared to solid metals, their liquid state is assumed even from low temperatures (Li, Sn). Capillary forces in a porous mesh (W, Mo) then hold the liquid in the desired shape.

This project simulates the erosion & transport of liquid metal heat shields in the new COMPASS Upgrade tokamak built in Prague. 3D Monte Carlo code ERO2.0, which solves the transport and plasma-wall interactions of impurities in the region between the confined plasma and the solid wall of current fusion devices, is used for this purpose.


Researcher: Jan Pokorny             


Lattice Boltzmann Simulations of Transport and Deposition of Fiberous Particles in Human Respiratory Tract       

Karolina CPU Alloc=16900;  Karolina GPU Alloc=100       

Computational Fluid Dynamics (CFD) is an indispensable source for understanding the transport and deposition of various particles in the respiratory tract. Locating deposited particles experimentally proves challenging, especially concerning non-spherical particles. This project focuses on developing the Lattice Boltzmann Method for simulating the transport and deposition of fibrous particles, utilizing the Euler-Lagrange Euler-Rotation approach on a realistic geometric model of the women's respiratory tract. The simulations will be validated against conventional CFD methods and experimental data, offering insights for potential applications of the Lattice Boltzmann Method in medicine. Analyzing results across different age and gender geometries will provide valuable comparative data.


Researcher: Vojtěch Pánek         


Visual Localization using State-of-the-Art Machine Learning Methods    

Karolina CPU Alloc=200;  Karolina GPU Alloc=2200         

Our project deals with visual localization, a method that estimates the precise location and orientation of a camera from an image. Ground robots or drones use visual localization to determine their exact location, allowing them to plan their motion through the environment safely. Extended reality devices, such as headsets or smart phones, can estimate their pose with visual localization to properly align the real and virtual worlds. Standard localization pipelines usually consist of several blocks, which all initially started as hand-crafted algorithms but are gradually replaced by machine-learning solutions in recent years. One of the blocks is also the environment representation, a cornerstone of every localization algorithm that is also being replaced by neural-based implicit methods. In our research, we are looking into improving these pipelines by adjusting the learned blocks and by making them work well together.


Researcher: Jiri Klimes  


Accuracy and precision for extended systems XII              

Barbora CPU Alloc=36000;  Karolina CPU Alloc=18000;  Karolina GPU Alloc=20;  LUMI-C Alloc=7000        

Computer simulations have become indispensable for understanding the majority of experimental observations and thus understanding our world. This is especially true at the atomic and molecular level where experimental observations can become very difficult to interpret. However, if simulations are to be useful, they need to model the system reliably. For example, if there are different pathways for a chemical reaction, the simulation needs to be able to correctly predict their relative importance so as to increase our understanding of the system. The world at the atomic level is governed by the complex laws of quantum mechanics. In practice, this means that increasing the reliability of simulations, i.e., using less approximations, increases the computational cost substantially. While we understand well the impact of some of the approximations, there are some for which our knowledge is limited. In some cases, these can lead to unexpectedly large loss of the reliability of calculations. In our project we analyse some of the problematic cases occurring for calculations of intermolecular interactions, either in small clusters, or in molecular crystals.


Researcher: Petr Strakos             


Scalable Parallel Astrophysical Codes for Exascale (SPACE)           

Barbora CPU  Alloc=8400;  Barbora VIZ  Alloc=100;  DGX-2  Alloc=700;  Karolina CPU  Alloc=28100;  Karolina FAT  Alloc=600;  Karolina GPU  Alloc=4300;  Karolina VIZ Alloc=100;  LUMI-C Alloc=23500;  LUMI-G Alloc=7000            

In Astrophysics and Cosmology (A&C) today, High-Performance Computing (HPC)-based numerical simulations are outstanding instruments for scientific discovery. In the upcoming exascale computing era, the efficient and effective exploitation of such computing capabilities is a challenging task. The primary aim of SPACE CoE is to make a set of A&C community applications exascale-ready, allowing them to take full advantage of emerging infrastructures. In addition, SPACE will address the high-performance data analysis of the data produced by exascale A&C simulation applications leveraging machine-learning and visualization tools. Integrating with European astrophysical communities around exascale computing is ensured by adopting software and data standards and interoperability protocols.

IT4Innovations, within the scope of SPACE CoE, is responsible for performance and energy assessments and optimization, which allows efficient usage of resources in terms of performance and energy consumption. IT4Innovations is also involved in various tasks related to A&C data visualization.


Researcher: Jakub Velímský       


Swarm for Ocean Dynamics: A parametric study

Karolina CPU Alloc=50000          

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. The goal of this project is to provide a parametric study of OIMF calculated for different ocean circulation and electrical conductivity models, and thus establish the sensitivities needed for further interpretation of satellite magnetic data. Each parameter set leads to a high-resolution solution of the motionally-driven electromagnetic induction equation, spanning more than a decade of satellite observations.


Researcher: Jennifer Za Nzambi


Discovery of Economic Opinions and Causal Relationships from Textual Data       

Karolina GPU Alloc=18400          

Social media platforms are akin to an untapped gold vein harbouring a reservoir of public opinions, attitudes, and sentiments which, if realised, could revolutionise how public opinions are gathered and interpreted. This project introduces a novel method for extracting opinions about economic indicators and factor impacting society from social media texts by fine-tuning large language models, on datasets comprising of social media posts, comments and more. Through fine-tuning, language models can acquire the ability to understand and mimic the economic discourses within posts published on social media. This project’s value is threefold. First, it amalgamates carefully curated datasets through which the model can effectively learn domain-specificities and economic understanding on an advanced level. Second, it devises metrics based on perplexity comparisons of opposing statements which validate the model’s comprehension of economic texts, thereby measuring the model’s alignment with datasets it was fine-tuned on, and finally applies said models to datasets garnering results indicating that the model-based approach can rival, and in some cases outperform, survey-based predictions and professional forecasts in predicting trends of economic indicators. Beyond the scope of this study, the methods and findings presented could pave the way for further applications of language model fine-tuning as a complement, or potential alternative, to traditional survey-based methods.


Researcher: Jan Rezac  


Benchmarks for validation and extension of a hybrid QM/ML computational model        

Karolina CPU Alloc=22500          

Our aim is to develop novel approximate quantum-mechanical computational chemistry methods applicable to very large systems such as biomolecules and complex materials.  Currently, we are working on a hybrid method combining semiempirical quantum chemistry with a machine learning correction, combining the strengths and avoiding the weaknesses of both these approaches. Validation of the resulting model requires large amount of reference data computed at higher level, i.e. state of the art DFT calculations, and additional high-quality benchmarks. In this project, we devise building a database of such calculations rationally mapping intramolecular conformational degrees of freedom in non-covalent complexes. The project also includes work on reparametrizing the underlying semiempirical method which should improve the accuracy further.


Researcher: Petr Macha              


GBS simulation of the COMPASS tokamak (II)    

Karolina CPU Alloc=35100          

The proposed project is a continuation of the previous work focused on the first GBS simulations of boundary plasma turbulence in the COMPASS tokamak. Turbulent transport is one of the most important research topics in tokamak physics due to its direct impact on plasma confinement and the heat load on plasma-facing components (PFCs). To explore this topic, we carry out nonlinear simulations using a three-dimensional, self-consistent, global, two-fluid GBS code, which facilitates both interpretative and predictive analyses of plasma turbulence. In our previous project, we demonstrated an excellent agreement between numerical simulations and experimental data, particularly in plasma profiles at the edge and divertor regions, as well as in power decay length, a key parameter in tokamak studies. These results were presented at the Joint EU-US Task Force Transport meeting in 2023. The current project aims to finalize the validation of the GBS code against experimental results from the COMPASS tokamak, within the framework of the GACR project (22-03950S). This involves advancing our numerical simulations to a steady state, where comprehensive statistical turbulence analysis is available, and conducting detailed blob analysis to further enhance our understanding of turbulent transport in the boundary region of tokamaks.


Researcher: Martin Beseda        


Variational Quantum Eigensolver for Topological systems            

Barbora CPU Alloc=8400;  Barbora GPU Alloc=2000        

This project aims to combine two hot-topics in the broad field of computational sciences – Quantum Computing and Topological Insulators. Considering the contemporary, pre-NISQ phase of quantum computers, it is necessary to focus on algorithms, which could a) deal with the important, real-world problems and b) are efficient without the need for large infrastructures.

The algorithm efficiency on smaller infrastructures needs to be evaluated not only by its runtime, but also w.r.t. our options to deal with unavoidable quantum noise, i.e. naturally-occurring decoherence. Difficult problem by itself, it may be necessary to employ several noise-mitigation and quantum-suppression methods to make up for the lack of error-resistance included in assembled circuits, which would be very well possible having more qubits both to introduce input-data redundancy for higher safety and combine it with stabilizer circuits offering both the options of detection and localization of the errors.

The family of Variational Quantum Eigensolver algorithms offers all of the above-mentioned necessary properties. These methods are applicable to many real-world problems, mainly from the field of Physical Chemistry. Also, they are hybrid in nature, i.e. they do not rely on quantum infrastructures for the whole runtime, avoiding much of the potential decoherence, but they’re off-loading computations of operators’ expectation values to them, thus trying to leverage the quantum advantage in its strongest aspect.

Topological insulators are, on the other hand, is one of the fields promising spintronic devices allowing for building and improvement of superconducting quantum bits. Such a quantum bits would be great for general applications, as they would tend to possess an inherent resistance to outer sources of decoherence up to a much greater degree, than allowed by the current construction approaches. Moreover, they are also predicted to offer much faster switching times.

While their construction currently lies in the area of hybrid processors, unprepared for production-level computations, it is definitely worth researching such materials in-depth, thus making it possible to exploit their potential in the future. In this project we plan to perform computations utilizing 1-D Su-Schrieffer-Heeger (SSH) model as a first building-block allowing us to identify shortcomings of our proposed approach and to deal with decoherence- and measurement-related problems in practice, not needing to tackle the complexity of the model itself so much.

That said, in this project the VQE solver will be implemented, allowing us to model material properties both in topologically trivial and non-trivial phases utilizing different possible backends from the straight-forward statevector simulation, over shot-based simulators with different noise models up to computations on real (IBM-provided) quantum processors.


Researcher: Lukas Neuman        


Inductive Bias of Deep Neural Networks for Computer Vision     

DGX-2 Alloc=1550          

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: Jakub Planer            


Tunable Charge Injection Layers for Organic Semiconductors      

Barbora CPU Alloc=6000;  Karolina CPU Alloc=7000        

In recent time, light weight, flexibility, low cost and tunability of organic semiconductors (OS) have drawn the attention of a broad scientific community and semiconductor industry, making them usable in many applications such as active-matrix organic light-emitting diodes (OLEDs) in smartphones, organic field-effect transistors (OFETs), solar cells, memories, photoswitches and sensors. Therefore, tremendous effort has been made to improve OS functionality. The development has reached the point where the contact resistance with the metal electrode, especially in OFETs, is becoming the main parameter limiting the performance of OS-based devices. Recent research in the Host Group has revealed that molecular monolayers (MoMs) composed of aromatic carboxylic molecules can be deprotonated on a silver substrate in a controlled manner. This finding opens a way towards developing entirely new highly efficient Charge Injection Layers for organic semiconductors. In this project I will employ state-of-the-art computational methods to characterize interfaces between Ag substrate and MoMs composed of aromatic carboxylic acid molecules, and design tunable CILs based on the controlled deprotonation. If successful, our collaboration can help the OFETs finally reach their full market potential.


Researcher: Martin Fajčík           


Language Transfer through Continued Pretraining of Large Language Models      

Karolina GPU Alloc=9600            

While current large language models (LLMs) are trained on multilingual corpora, the majority of these corpora are English (e.g., 95% for LLAMA). The quality of LLMs such as LLAMA2, Mistral7b, or ChatGPT3.5 on tasks in a different language (though medium-resourced, like Hindi, Chinese, and Spanish) is known to be mediocre.

In this project, we will continue with the research on continued pretraining (CP) of medium-resource language adoption of LLMs pretrained on trillions of tokens available in English corpora (e.g. LLAMA2 is pretrained on 2T tokens). Such a quantity of tokens is not available for most non-English languages. We seek to experiment with methods based on weight alignment and latent manifold mixup to prevent these models from catastrophic forgetting during their CP on Czech. Ultimately, we seek to release 7B and 13B transfer-learned LLMs for Czech.

We have already investigated vocabulary swap techniques known from machine translation and pretrained language models and released a (currently largest available) 1.5B Czech language model on our own collected 16b token large Czech web corpus.

Exploiting CP procedures is a necessary step to achieve transfer learning from English to languages without abundant resources. Furthermore, continued pretraining is working only with the target corpus (in the order of billions of training tokens), thus requiring much less training compared to training from scratch, saving resources and reducing CO2 emissions.


Researcher: Riccardo FUSCO     


Machine Learning Accelerated In-Silico Screening of Multi-Component Accessible Molecules

Karolina CPU Alloc=511;  LUMI-G Alloc=2285     

Our project ambitiously integrates classical computational methods such as Docking and Molecular Dynamics (MD) with advanced Deep Learning to achieve groundbreaking advancements in Drug Design. Utilizing Multi Component Reactions (MCR) for targeted compound library creation, we focus on generating a vast array of compounds. Our methodology is widely applicable, and in this case, we will be using it as a first-case scenario on immunological targets PDL-1 and CBL-B, particularly in the broader context of the AMADEUS ERC advanced project and the Era Chair ACCELERATOR grant won by PI Alexander Domling (project nos. 101087318, 101098001). On a more general path our work will provide a pipeline to fully (semi) automate and miniaturize virtual screening, synthetic and biological screening. This will be designed to provide a general acceleration of the process of drug discovery. To achieve this, we present the first piece of our project: a Machine Learning Accelerated Virtual Screening Pipeline capable of efficiently screening billions of easy-to-make compounds. In addition to our scientific objectives, our project places a strong emphasis on valorization, aiming to translate our innovative drug design methodologies into tangible applications. We anticipate that the successful implementation of our approach will not only accelerate drug discovery processes but also provide vaporizable results. The potential socio-economic impact is profound, as the development of novel drugs targeting immunological factors such as PDL-1 and CBL-B holds promise for more effective treatments in cancer, ultimately improving patient outcomes and reducing healthcare burdens.


Researcher: Michael Komm       


Influence of charge exchange collisions on sheath in partially detached plasmas

Karolina CPU Alloc=17200          

Future thermonuclear reactors such as DEMO will operate with edge plasma at significantly higher plasma and neutral densities then contemporary tokamaks. This will lead to a very high collisionality of the edge plasma, effectively causing its partial de-magnetisation. The aim of this project is to study such plasma regimes by means of 2D3V and 3D3V particle-in-cell collisions and to determine the influence of intensive charge exchange collisions on the properties of the sheath, especially on the distribution of heat and particle fluxes onto the components of the first wall.


Researcher: Ján Michael Kormaník         


Design of catalytic metal-binding peptides          

Karolina CPU Alloc=45000;  Karolina GPU Alloc=1400    

Metal ions commonly appear in natural protein scaffolds. They are essential for life and often appear within enzymes, allowing for various chemical reactivity, such as complicated bond breaking or oxidation reactions. The diverse reactivity of metal ions makes them attractive for potential applications within chemical and biochemical industry, as they can catalyze complicated reactions at room temperature and standard pressure. If we can design protein scaffolds to selectively bind metal ions and catalyze desired reactions and properly assess their viability for their intended purpose, this could prove to be a big opportunity for the industry.


Researcher: Karel Sindelka         


Computer simulations of colloidal solutions       

Barbora CPU Alloc=28000;  Karolina CPU Alloc=16000

Aqueous solutions are omnipresent in nature, industrial processes, and daily life. Understanding their behaviour in inhomogeneous environments (self-assembled or confined systems) in equilibrium and non-equilibrium (shear flow) conditions is important in many applications from medicine to environmental protection. This project investigates interactions of structured fluids (micellar solutions, surfactant bilayers) with surfaces; these play important role in industrial or household products (e.g., fabric softeners, cosmetic products). We use mesoscopic simulations to provide molecular-level insights into chemical and physical behaviour of these systems in and out of equilibrium.


Researcher: Jan Hůla     


Efficient Language Models II      

Karolina GPU Alloc=1900            

This project is focused on cost-effective and efficient methodologies for developing and adapting language models, addressing the high financial and computational costs associated with current models such as Llama or ChatGPT. Focusing on reducing dependency on extensive training data and large model sizes, the research leverages previous work by Mikolov et al. (2011) to develop models with complementary strengths. Key strategies include the use of efficient fine-tuning, distillation pipelines, and active learning within a teacher-student framework. This approach aims to democratize LLM technology by significantly lowering training/inference costs, thereby bridging the gap between academic and industrial applications.


Researcher: Pierre Koleják         


Chiral biomolecules studied by spin-driven terahertz technologies           

Karolina CPU Alloc=2800;  Karolina GPU Alloc=100          

Many molecules essential for pharmaceuticals, food processing, and industrial applications exhibit a unique optical response to the rotation of light, known as chiroptical activity. These biomolecules strongly interact with electromagnetic radiation, especially in the terahertz range, through vibrations and rotations of specific molecular components. Our investigation into these substances involves employing quantum chemical computations (so-called ab initio) to gain insights into the mechanisms of light interactions, including absorption, scattering, and chiroptical response. Understanding the quantum properties of these molecules through ab initio methods is crucial for accurate predictions and interpretations in the realm of chiroptical activity. In parallel, we're developing terahertz spintronic sources that harness electron spin for precise control over polarization. We simulate their ultra-fast dynamics, performance and optimization to apply them directly in our laboratories for the in-depth study of biomolecules and chiroptical activity.


Researcher: Radim Vavrik           


Performance optimization and energy efficiency analysis of academic and industry codes within the EuroHPC POP3 project

Barbora CPU Alloc=18700;  Karolina CPU Alloc=24900;  Karolina GPU Alloc=17900;  LUMI-C Alloc=8000;  LUMI-G  Alloc=8000        

The key element to ensure efficient use of HPC infrastructures is to optimize the performance and efficiency of the applications. The Centre of Excellence on Performance Optimization and Productivity (POP CoE) [1] was initiated in October 2015 with the fundamental purpose of assisting a broad community of HPC application developers and users in both science and industry domains helping them to understand the performance-related issues of their applications and thus improve their efficiency and productivity. This fundamental purpose is reached by externally and objectively auditing the performance of the codes to all interested users, by providing not only qualitative but also quantitative analysis through the use of POP tools and methodology.

The current project, Performance Optimization and Productivity 3 (POP3), is articulated in 3 main pillars: services, users, and co-design. POP services mainly focus on performance assessments with the goal to evaluate code performance and scaling, identifying the main sources of inefficiency and providing some insight and recommendation about how to improve it. POP3 also provides second level services that have been extended and include proof of concepts, energy efficiency and advisory studies. Although POP3 will continue targeting all scales and types of users, with POP3 we will focus our efforts on larger scales, assessing the execution on the European HPC facilities of some of the European flagship HPC applications for other CoEs. However, POP3 will still provide services to prescribers and SMEs to promote an efficient usage of the computing resources. The co-design will be covered in two dimensions. Internally in POP3 we will co-design the tools and methodology to be able to analyse the applications on the selected platform at the selected scale. Externally, POP3 will identify best practices and kernels that will be offered to other European projects as well as the wide audience of parallel applications.


Researcher: Marta Jaroš             


Automated Tuning of Workflows Executions on Remote Computational Resources           

Barbora CPU Alloc=1000;  Barbora GPU Alloc=100;  Karolina CPU Alloc=2000;  Karolina GPU Alloc=300  

In recent years, therapeutic ultrasound has diverse applications like tumor ablation and targeted drug delivery. Optimal outcomes require precise, customized preoperative planning. A challenge is accurate, safe, and noninvasive ultrasound energy delivery to the target region. Computation-intensive models for treatment estimation use high-performance computing (HPC). Despite the significance of HPC, clinical end-users lack efficient utilization expertise. The k-Plan software simplifies HPC use without specifying parameters, dependencies, or monitoring. It addresses parameter selection challenges and scaling issues, critical for calculation cost and execution time.

Having deployed k-Plan for initial workflows, this project aims to (1) develop and test a new GPU code for thermal simulation, (2) apply real clinical and biomedical workflows, (3) customize task submission planning logic for IT4Innovations clusters with machine learning, and (4) explore methods for k-Plan to auto-tune execution parameters for tasks. Creating a publication covering experiments is a key goal.


Researcher: Michal Langer         


Establishing the relationship between the structure and photoluminescence of carbon dots       

Karolina CPU Alloc=31200;  LUMI-C Alloc=30000             

Creating connections between the structure and photoluminescence (PL) of carbon dots (CDs) is a significant challenge in the field. While it's accepted that different structural aspects may develop during CD preparation, we're still far from understanding the specific features that dictate their overall PL. Although the core, surface, and molecular states are commonly recognized as the three primary sources of PL, the extent of their interaction and mutual influence is not well-established. The communication among various PL centers is anticipated to rely on their arrangement and the type of connections formed. Recently, time-dependent density functional theory (TD-DFT) calculations were performed for several (N-doped/O-functionalized) polyaromatic hydrocarbons (PAHs) as representative models for the core/surfaces PL states and the prototypical molecular fluorophore (MF) 5-oxo-1,2,3,5-tetrahydroimidazo-[1,2-α]-pyridine-7-carboxylic acid (IPCA), with the aim to demonstrate if any communication between core/surface/molecular states are present. In this project, we will employ slightly different approach than using small ad-hoc small models. Classical molecular dynamics (MD) simulations will be employed to generate possible models of large CDs, whose structure will not contain crystalline graphitic core as in the most common theoretical studies, but polymeric CDs or amorphous CDs will be studied. Therefore, we will focus on those CDs, which can be experimentally formed under the conditions which do not allow full carbonization of the structure. Subsequently, we will use the quantum mechanics/molecular mechanics (QM/MM) approach to quantify the PL of many different configurations, which will be obtained from the MD simulations. Overall, our results will provide an insight into the structure/PL relationship of polymer/amorphous CDs, thus, adding a piece to the overall puzzling conundrum of structure-PL of CDs.


Researcher: Pavlo Polishchuk    


CACHE challange             

Karolina GPU Alloc=3600            

CACHE challenges focus on specific protein targets of biological or pharmaceutical relevance. Participants should predict hits and CACHE will validate these hits experimentally. Each competition includes a hit-finding and a hit expansion round of prediction and experimental testing after which all data, including chemical structures, will be made publicly available without restrictions on use.

The target of the first CACHE Challenge was Leucine-rich repeat serine/threonine-protein kinase 2 (LRRK2), the most commonly mutated gene in familial Parkinson's Disease (PD). PD-associated LRRK2 mutations tend to promote LRRK2 filament formation and enhance LRRK2 interaction with microtubules. Recent structural data reveals that only compounds stabilizing the open form of LRRK2 antagonize the pathogenic formation of LRRK2 filaments in cells, but most kinase inhibitors stabilize the closed form of LRRK2. An alternative and so far overlooked strategy is to pharmacologically target the WDR domain of LRRK2, which is juxtaposed to the kinase domain. The WDR domain in LRRK2 may be important for recruiting LRRK2 signalling partners or for binding to tubulin.

On the previous stages of the challenge we identified active hits and to enable their rational optimization we have to establish their binding modes that can be achieved by application of advanced modeling techniques.


Researcher: Pavlo Polishchuk    


Fragment-based de novo design and structure optimization       

Karolina CPU Alloc=27500          

Exploration of chemical space is chemical space is extremely difficult due to a vast number of potential drug-like molecules – 1036. Approaches which can generate compounds on-the-fly and adaptively explore and navigate in this space will substantially increase search speed of potentially active molecules, improve their novelty and greatly expand the knowledge about favorable and unfavorable regions of chemical space. All this will increase efficacy of searching of new hits, leads and drug candidates. This project is focused on development of such computational tools based on the previously developed fragment-based structure generation framework CReM which the main advantage is generation of synthetically accessible molecules that will improve outcomes of medicinal chemistry projects.


Researcher: Jan Hecko


Drift effects in the COMPASS tokamak SOL with SOLPS-ITER simulations

Barbora CPU Alloc=2100             

The COMPASS tokamak is an experimental fusion energy device utilizing magnetic confinement of plasma. Under ideal conditions, the magnetic field constricts the movement of charged particles to a direction parallel to the field lines, confining the superheated plasma and levitating it inside of the vacuum vessel. In a real device, there is also a cross-field transport driven by diffusion, turbulence, and drift effects. Drifts are unintuitive velocity components perpendicular to the magnetic field lines predicted by plasma physics, whose exact influence is often hard to attribute. The final experimental campaign of the COMPASS tokamak yielded a series of discharges with reversed direction of the toroidal magnetic field, which implies reversed direction of drifts. Drift-driven cross-field transport has been long suspected to play a significant role in the COMPASS scrape-off layer (SOL) plasma region and this dataset provides an excellent opportunity to investigate this hypothesis by modelling the SOL using the SOLPS-ITER code.


Researcher: Ondrej Olsak           


Transcranial ultrasound stimulation        

Barbora CPU  Alloc=5500;  Barbora FAT Alloc=50;  Barbora GPU Alloc=100;  DGX-2 Alloc=50;  Karolina CPU Alloc=1000;  Karolina FAT Alloc=50;  Karolina GPU Alloc=1000;  LUMI-C Alloc=300;  LUMI-G Alloc=100              

Disorders of the brain, including neurological and psychiatric diseases, affect one in four people. There are approaches for treating or alleviating the symptoms of these disorders however, there are still many goals to achieve in areas like side effects and cost reduction, invasiveness, efficiency and so on.  Neurostimulation techniques that modulate the electrical activity of the brain have evolved as an important class of second-line treatments for pharmacoresistant cases. A non-invasive technique for stimulating brain targets with high anatomical precision, unlimited penetration depth, complete reversibility, and a low-risk profile is essential in this field. This can be achieved using the newly emerging technique of low-intensity focused transcranial ultrasonic stimulation (TUS) for neuromodulation. This project focuses on the development of closed-loop individualized image-guided transcranial ultrasonic stimulation under the Horizon Europe CITRUS project. The computational resources will be used for preoperative MR-based brain imaging, personalized ultrasound treatment planning, including temperature mapping, and validation of fast real-time re-planning software based on advanced mathematical models and artificial neural networks.


Researcher: Martin Žonda          


Neural network quantum states for many-body systems

Barbora CPU Alloc=3000;  Barbora GPU Alloc=600;  Karolina CPU Alloc=5100;  Karolina GPU Alloc=400  

The neural network quantum states (NQSs) have recently emerged as a promising alternative to common trial states in variational Monte Carlo (VMC) studies of quantum many-body systems. These include frustrated lattice spin models and correlated electron systems which are notoriously difficult to solve with other numerical methods.  The research of NQSs is driven by the fact that neural networks (NNs) are universal function approximators. As such, they are expected to be a suitable alternative to traditional variational ansatze. Several recent publications showed that already relatively simple NQSs  can outperform standard trial states in the variational search of the ground-state energies for various frustrated spin systems and some models of correlated electrons.

The project's goal is to apply these techniques in the search of ground-stated and dynamical properties of two distinct experimentally motivated systems. The first one is the compound CeMgAl11O19. Our preliminary results suggest that this material contains a quantum spin liquid phase resulting from the triangular lattice and the anisotropic dipolar interactions. However, this must be confirmed by model calculations. The second system is a nanodevice, a multilevel quantum dot with significant electron-electron interaction, coupled to superconducting leads. Here we plan to focus on investigating the role of the so-called Yu –Shiba-Rusinov (YSR) bound states in the charge and spin-transport.


Researcher: Marek Hrúz              


Large Sign Language Models      

LUMI-G  Alloc=10000    

In recent years, there has been notable progress in Sign Language Translation (SLT), i.e., the translation of a sign language video directly to its spoken-form counterpart. SLT encompasses several areas of research: Natural Language Processing, Machine Translation, Visual Processing, Multi-modal data representation and fusion, and SL linguistics. As such, it is a solid candidate for bringing together experts from these fields to work on a common goal. In this project, we will pre-train strong baselines for an image and pose encoder based on the Transformer model. Then, we want to leverage the potential of Large Language Models (LLMs). A pre-trained LLM already encodes the knowledge about a language (including sign language) and thus can be used as an efficient decoder of the translation model. It should be able to guide an image/pose encoder to a strong linguistic representation of the input sequences.


Researcher: Diana Sungatullina


Backpropagating through Minimal Solvers          

LUMI-G  Alloc=8000      

We aim at developing an approach to backpropagating through minimal problem solvers in end-to-end neural network training. Traditional methods relying on manually constructed formulas, finite differences, and autograd are laborious, approximate, and unstable for complex minimal problem solvers. Through this project, we want to develop fast, simple, and reliable alternatives to the aforementioned methods. We will leverage the Implicit function theorem and explore how to use it to compute derivatives for backpropagation through the solutions of a minimal problem solver. The resulting solution will be applied to a wide variety of geometric problems both on synthetic and real data to show its broad applicability and evaluate its speed, accuracy, and stability. If successful, our method would open up a promising direction for efficient backpropagation through hard minimal problems and bringing more efficient optimization in deep learning.


Researcher: Camille Flore Marie Landri 


Impact of binary evolution on the outcome of common envelope evolution in massive stars

Barbora CPU Alloc=120000        

An important fraction of stars are found in binary systems, especially massive stars. Binaries made of massive stars are known to evolve in short-period binary systems that often become sources of astrophysical transients and/or gravitational waves and therefore, understanding their evolution is crucial.

It is well-established that stars in binaries evolve differently than single stars. For instance, binary stars are known to exchange matter during the so-called mass transfer phase, and may later enter common envelope evolution, during which the lighter star plunges into the envelope of its giant companion star. Common envelope evolution is one of the mechanisms by which a binary system can become a short-period binary, and as such, it has been extensively studied. However, the impact of the previous phases of evolution of the binary on later common envelope evolution has yet to be properly assessed.

In this project, we aim to explore how preceding phases of binary evolution, such as mass transfer phases, may impact the outcome of common envelope evolution. To do so, we will perform a combination of simulations where we will first simulate the evolution of a binary system until the onset of common envelope evolution using 1D stellar evolution methods. We will then perform state-of-the-art 3D hydrodynamics simulations of the common envelope phase to characterise how it is affected by previous binary evolution processes.


Researcher: Tomáš Hrivnák        


Modelling of Electronic Transitions in Multichromophoric Molecular Assemblies in Carbon Dots

Karolina CPU Alloc=9625             

Carbon dots (CDs) have emerged as one of the most studied carbon-based nanomaterials due to their attractive photophysical properties, chemical stability, low toxicity and high biocompatibility. Numerous studies have been dedicated to describe the underlying structure-property relationships, yet due to the high complexity and variability of both their structure and molecular composition significant scientific effort still needs to be invested to elucidate the principles governing their properties. Here, computational chemistry can provide indispensable insights into the structure as well as photophysics of CDs. Whereas for isolated molecular systems, the “golden-standard” and reliable approaches can be applied, the large heterogeneous systems like CDs are much more intricate to model. Particularly for the multichromophoric systems, the theoretical modelling remains a significant challenge.

In this project, we aim to perform a series of calculations to validate the performance of recently developed simplified quantum-chemical (QCh) methods for selected polycyclic aromatic hydrocarbons (PAHs) and their complexes as model systems of CD fragments. The validated methodology will then be used for larger aggregate systems to shed light into the nature of their optical processes. Understanding the role of different structural and chemical modifications will help to design novel photoluminescent and photocatalytic devices based on CDs.


Researcher: Jiří Pittner 


Machine learning assissted molecular dynamics

Karolina CPU  Alloc=19700         

Molecular dynamics in excited states including the non-adiabatic and spin-orbit effects is an important theoretical tool for the simulation of photochemical processes which play an important role in nature and technology, like e.g. photosynthesis, phototherapy, photovoltaics, etc. Its computational cost when ab-initio or DFT methods are employed is very limiting in both size of the molecules treated and the length and number of computed trajectories. Machine learning (ML) has recently become very popular thanks to its widespread applications in many areas of science, industry, and commerce. Recently the machine learning methods have been successfully employed to speed up the molecular dynamics in the ground state, and some progress was done also on the ML of excited states and non-adiabatic effects. The aim of this project is to implement ML methods for MD with nonadibatic and spin-orbit couplings and perform ML-accelerated molecular dynamics in excited states undergoing both internal conversion and intersystem crossing processes, with application to systems like (modified) DNA bases and inverted ST-gap OLED emitters. The application is related to solving the grant no. 23-06364S awarded by the Czech Science Foundation.



Researcher: Mario Vazdar          


Molecular dynamics simulations of uncoupling proteins UCP3 – UCP5 in the inner mitochondrial membrane      

Karolina CPU Alloc=15000          

Uncoupling proteins (UCPs) 3, 4, and 5, integral to the inner mitochondrial membrane, remain structurally elusive and their experimental structures are not determined. UCP3 is localized in skeletal muscle and brown adipose tissue and influences fatty acid oxidation and heat production. UCP4, expressed in the central nervous system has pronounced neuroprotective properties while UCP5 is predominantly found in the brain. This proposal aims to uncover the structure and dynamics of uncoupling proteins (UCPs) 3, 4, and 5 using molecular dynamics (MD) simulations. Utilizing Gromacs codes with CHARMM36m force fields, this state-of-the-art MD approach will address the absence of structure and dynamics for UCP3 – UCP5, providing crucial insights into their properties at the molecular level. This research aims to reveal fundamental knowledge about UCP3, UCP4, and UCP5 structures and dynamics in membranes, which will be compared with existing experimental structures of UCP1 and UCP2 together with previously reported MD simulations. Taken together, we aim to provide a more detailed insight into bioenergetics and signaling in mitochondrial membranes by UCPs, with implications for future biomedical research in metabolic and neurological disorders.


Researcher: Marek Matas          


Ab initio simulation of a quantum dot dark matter detector (QDmatter)

Karolina CPU Alloc=35475          

In this project, we will model the electronic structure of quantum dots using ab initio techniques to guide the experimental effort of constructing a dark matter detector here on Earth. While the existence of dark matter is supported by a set of evidence of its gravitational interaction with standard model matter, its non-gravitational interaction has eluded detection to this day. Some of the most sensitive techniques aiming at tackling this mystery rely on building an earth-based detector and observing excitations caused by the incident flux of dark matter particles dispersed in the Milky Way halo, the so-called “direct detection”. The absence of a convincing signal to this day has led the community to shift the paradigm of a search for nuclear-induced excitations to lighter targets, such as electrons. To understand the electronic excitations caused by incident dark matter particles, one needs to describe correctly the underlying electronic structure and the form of its wave functions. In this work, we turn to density functional theory techniques encoded within the Quantum ESPRESSO software package to model quantum dots suspended in a liquid, a promising sub-GeV dark matter detection technique (currently in development) with good scalability, capable of testing the origin of signal excess reported by other experimental collaborations, such as by the DAMIC project (https://damic.uchicago.edu/).


Researcher: Damien Lucien Michael Gagnier      


Common Envelope Evolution: Gas Dynamics and Binary Evolution with 3D Local MHD Simulations

Karolina CPU Alloc=82000          

Common envelope evolution (CEE) is a crucial phase of binary systems' evolution, where two stars interact within a shared envelope. This process is believed to give rise to various close binary systems, including cataclysmic variables, X-ray binaries, progenitors of Type Ia supernovae, and planetary nebulae nuclei. Moreover, it precedes most stellar mergers detected by ground-based gravitational wave (GW) detectors. Despite substantial efforts over the past decade, global hydrodynamic simulations have failed to replicate the orbital properties of observed post-common envelope binaries. These simulations represent the two cores as point masses with artificial gravitational softening, introducing non-physical flow structures in the binary's vicinity and resulting in uncertain post-CEE orbital parameters. In particular, the final orbital separations are always too large for a GW merger to occur within the age of the Universe. In this project, I will conduct the first 3D magnetohydrodynamical simulations resolving the individual cores, and focused on the interaction between the binary system and its local environment. I anticipate the outcome of this project to dramatically affect our comprehension of the ensuing planetary nebulae morphology and to play a crucial role in reconciling the observed final orbital separation and those obtained with numerical simulations.


Researcher: Jan Kuneš 


Magnetic Excitations in Antiferromagnetic Metals           

Karolina CPU Alloc=31250          

Antiferromagnetic (AFM) spintronics is a rapidly developing field of physics with a span from fundamentally new phenomena to technological applications. It is driven by discoveries of new materials and new physical phenomena. Recently, a new class of AFM materials – altermagnets – was identified. Unlike conventional AFMs, altermagnets host the anomalous Hall effect, linear magneto-optical effects or spin polarized bands, previously associated with ferromagnetism. The magnetic order gives rise to collective fluctuations of magnetic moments. Dynamics of these fluctuations and theeir impact on electrons is the key to understanding these materials. Theoretical description of these effects is notoriously difficult.

We have recently developed a method to do so and applied it successfully to models as well as a real material SrRu2O6. Since SrRu2O6is a large gap insulator the magnetic excitations take place at different energy scale than the electronic excitations. In this project we want to apply our method to metals where the magnetic and electronic excitations overlap and therefore strongly interact with each other. We plan to study both the simplified Hubbard model in its metallic phase as well as RuO2, which has become the prototypical altermagnetic metal. This project will not only answer some outstanding physical questions, but also expand the capabilities of the method -- the dynamical mean-field theory -- which has hundreds of users around the world.


Researcher: Ctirad Červinka       


Towards sub-chemical accuracy of ab initio atomistic simulations of fusion enthalpies for molecular crystals        

Barbora FAT Alloc=1500;  Karolina CPU Alloc=72000;  LUMI-C Alloc=24000          

Organic molecular materials are ubiquitous, spanning diverse areas from pharmaceuticals, over fertilizers, explosives to semiconductors. Fabrication of final products in those fields often requires a sufficient solubility of precursors, and subsequently, availability of a crystallization scheme of the target compound from solution. Development of accurate and widely transferable first-principles models of the melting parameters for molecular materials will significantly contribute to the material research.

Modeling the heats of fusion represents a first step towards predictions of the melting temperature and goes hand in hand with high-throughput predictions of solubility of molecular materials. Since existing protocols based on classical all-atom molecular-dynamics simulations fail to reach the chemical  accuracy of the predicted heats of fusion, incorporation of the first principles into the computational methodology seems to inevitable to advance such predictions both in terms of their accuracy and range of applicability. Current proposal will address the incorporation of ab initio molecular dynamics simulations into the computational protocol for prediction of the heats of fusion, namely concerning the questions of its feasibility and optimum computational setup in terms of the size of simulated ensemble, applied ab initio level of theory, and relationships to the classical simulation and its outcomes.


Researcher: Martin Čadík           


Deep-Learning Approach to Geo-localization      

Karolina CPU  Alloc=300;  LUMI-G Alloc=3000   

How often do we ask ourselves while looking at a beautiful photograph, where was it taken? An automatic system which could use just visual cues to provide this answer would enable us with exciting new opportunities. For example, we could enrich large internet photo collections with geo-tags and organize them, or we could travel to a photograph's location through an augmented reality application. A real-time localization system could enable navigation in areas without GPS signal, such as a surface of a distant planet. Despite being actively researched for many years, visual geo-localization in natural environments remains a challenging task. Vegetation, illumination, and seasonal changes all contribute to the complexity of localizing an image in a constantly changing environment. This work builds upon our previous experience with deep learning driven visual geo-localization in mountainous regions. Our aim is to improve localization accuracy and research new methods which would provide end-users with localization uncertainty estimates.


Researcher: Rostislav Langer     


Modelling of single atoms in specific organic-based systems for catalysis, gas separation and electrochemistry    

Karolina CPU Alloc=120000       

In the ever-evolving landscape of nanoscience and in the demand for clean and renewable eco-friendly materials, our focus lies in the intricate modeling of single atoms (SAs) which reduce costly and scarce bulky noble metals extensively used for industrial applications, such as energy conversion or chemical production. The ability to model individual atoms allows us to design new materials with tailored properties. We aim to decipher the intricate interplay of chemical bonding, structural arrangements, electronic states, and other phenomena within SAs anchored at diverse two-dimensional organic-based materials which have been investigated as a promising material in various industrial fields including catalysis, supercapacitors, or gas separation. To be more specific, our objectives encompass the investigation of stability-activity relationships and structural preferences of SAs embedded in doped graphene, to elucidate the principles behind the noble gas separation by using graphene-based system, to provide a crucial theoretical support for experimental observation in hydrosilylation reactions catalyzed by SA embedded in organic matrices and in SA–organic electrode interactions. To address these multifaceted goals, our approach involves the high-throughput quantum mechanical calculations by using VASP and Gaussian software. The outcomes of this research could play a pivotal role in the design of new economically viable materials with enhanced properties.


Researcher: Jan Heyda 


Drug-surfactant interactions in aqueous solutions           

Karolina CPU Alloc=30000;  Karolina GPU Alloc=2900    

Most active pharmaceutical compounds and potent drugs are poorly soluble molecules in water, where organic chemistry and engineering tricks are used to bypass this limitation. In this proposal, motivated by our in-house experimental collaborators, our objective is to investigate the impact of five approved biocompatible surfactants on the solubility of four poorly water-soluble drugs. These surfactants form micelles of different sizes and their non-polar interior is capable of drug accumulation. Micelles present a drug reservoir, improving the bioavailability of the drug in the solution.

In this project, we aim to provide a microscopic picture of the drug binding in the interior of the micelle, evaluating binding free energies. This allows  to connect the role of chemical motifs (functional groups) of drugs and surfactants in improving drug solubility in aqueous solutions.


Researcher: Radim Špetlík          


Applications of Iterative Alpha Deblending Models          

Karolina GPU Alloc=3800;  LUMI-G Alloc=8000  

Machine learning is crucial for computer vision. It enables automatic feature extraction, accurate object detection and recognition, precise semantic segmentation, image captioning, video analysis, and transfer learning. It empowers computers to understand visual data like humans, leading to applications in autonomous driving, surveillance, medical imaging, augmented reality, and more.

The iterative alpha deblending method (IAD) [4] is a simpler approach compared to diffusion probabilistic models for image generation. It operates by sequentially estimating and subtracting the contributions of individual sources within an image. Unlike diffusion probabilistic models (DDPMs) [1] that involve intricate probabilistic frameworks and complex computations, iterative alpha deblending simplifies the process by iteratively refining the estimated sources through an alpha parameter, representing the contribution of each source. By repeatedly subtracting these estimated contributions from the original image and updating the alpha values, this method gradually separates overlapping sources. Iterative alpha deblending offers a more straightforward and iterative way to disentangle multiple sources within an image.

The training of IAD presents several technical challenges. These include the need for large-scale datasets with paired clean and noisy images, the computational complexity of training deep neural networks, and the difficulty of balancing denoising performance and computational efficiency. Simply put, training IADs is unthinkable without cluster computing capabilities."


Researcher: Klára Kalousová     


Meltwater generation and transport in the deep ice layers of water-rich exoplanets        

Barbora CPU Alloc=3500             

Within the last two decades, the observations by the Kepler telescope and the TESS mission led to the discovery of more than five thousands of confirmed exoplanets orbiting stars beyond our Solar System. Some of these planets may have hydrosphere atop the rocky core. It was suggested that these planets, that can only form far from their star beyond the snow line, could migrate closer to their star to end in the habitable zone. Depending on their

mass and the pressure-temperature conditions, they can be large rocky planets covered by thin global oceans or planets with thick hydrosphere, where the liquid oceans are sandwiched between the outer ice crust and a deep, high-pressure (HP) ice layer. Such planets (e.g., some of the Trappist planets) may represent an opportunity to explore new classes of potentially habitable planets.

The presence of the HP ice layer is often seen as a barrier for material exchange between the core and the ocean. In our previous work tailored to Solar System icy moons, we have demonstrated that such an exchange may be possible if melting occurs at the core-hydrosphere interface. Furthermore, organics may be present in the rocky cores, some of them being the building blocks of life. The interaction of organics with liquid water is essential for assessing the exoplanets habitability potential. The goal of this project is to study the transport of organic-rich liquids through the HP ice layer of water-rich exoplanets.


Researcher: Tomas Martinovic  


Biodiversity Digital Twin development II

Karolina CPU Alloc=300;  Karolina GPU Alloc=100            

Project aims to support development of the biodiversity digital twins, which can be computed on HPC clusters enabling better precision, or increased scale of the models. These models will be possible to execute through the LEXIS Platform in a specially prepared custom web portal. This will enable researchers in Europe to calibrate their models and test hypothesis very quickly leveraging HPC resources without increased difficulty. In case of this project only selected invited users will be able to use this resources to test the models.


Researcher: Jakub Elcner             


Holistic approach to Rotating Packed Bed (RPB) Carbon Capture process with the use of 3D CFD

Karolina CPU Alloc=10000          

The project will study the mechanisms that take place in a rotational absorber for CO2 capture (Rotational Packed Bed RPB). The RPB combines hydrodynamic and process mechanisms. The major issues are liquid solvent breakup and distribution inside the rotating packing, where a main interphase interface is formed accompanied by the liquid spray formed on the periphery of the rotating packing in the outer cavity. Since both domains are inherently linked we can´t solve each domain separately. It is extremely difficult to study these phenomena experimentally due to limited optical access, therefore a computational simulation is a kind of “replacement” approach. However we deal with two phase flow, i.e. with continuous liquid breakup to a dispersed flow with droplets formations, which necessitates mainly using Volume of Fluid model. For correct results to obtain, a 3D approach is needed, which however requires extremely high computational power. The complexity of the situation is even aggravated by rotation of the packing introducing large centrifugal force and almost impossibility to treat only a symmetrical segment of the solution domain since there is in fact no symmetry. And moreover, the internal rotating packing geometry is quite intrinsic made - in case of knitted wire mesh - of very small diameter wires requiring a very fine mesh. All these ingredients lead to using a very high computational power which is available only at Ostrava IT4I.


Researcher: Vladislav Pokorný  


Fast and effective solvers for superconducting impurity problems            

Barbora CPU Alloc=11000          

Superconducting quantum electronics gained a lot of interest recently for the increasing number of promising applications in quantum computing and sensor technology. The fast pace of development of quantum technologies requires designing new devices with increasing complexity and number of active elements. Understanding the complex interplay among the various quantum-mechanical phenomena which take place in nanoscale is a necessary step in developing a new generation of such devices. Supercomputers are a necessary tool that allows us to build the theoretical  understanding of the underlying processes and explain the available experimental results before such devices can be reliably used to extend the abilities of the current silicon-based electronics.


Researcher: Jun Terasaki             


Vertex correction for nuclear matrix element of neutrinoless double-β decay     

Karolina CPU Alloc=46000          

The goal of my project is to establish the calculation of a reliable nuclear matrix element (NME) of the neutrinoless double-β (0νββ) decay of nuclei. If this decay is found, it proves the existence of a new type of neutrino. This finding has tremendous impacts on neutrino physics, particle physics, and astrophysics because a crucial foundation is given to the theories to solve the mystery of the current matter-prevailing universe. The NME is a theoretical physical quantity controlling the decay probability of the 0νββ decay. Thus, the difficulty of the experiments to observe the decay is influenced. The NME is also important for determining the neutrino mass scale parameter, which has been unknown except that it is very small. Since the decay is extremely rare, the reliable prediction of the NME is as challenging a subject as the experimental observation of the decay. The improvement of the accuracy of the calculation is one of the most important subjects in neutrino physics. I have been studying this improvement for more than several years, and it turned out that the key to the improvement is the higher-order terms of the transition operator beyond the approximation used so far. In this computational project, I intensify the calculation of these higher-order terms and enable the reliable prediction of the NME of the 0νββ decay.


Researcher: Fabien Jaulmes       


Computational modelling of fast ion orbits and their consequences in tokamak 

Barbora CPU Alloc=23000;  Karolina CPU Alloc=1000     

Nuclear fusion will enable us to generate energy without releasing large amounts of greenhouse gases into the atmosphere or leaving behind us long lived radioactive waste. The tokamak concept involves the use of magnetic fields to confine plasma hot enough to sustain fusion within itself. Fusion is now supported as a way to revert climate change and was discussed at the COP26 : https://unfccc-cop26.streamworld.de/webcast/iter-organization-fusion-energy-the-state-of-art. COMPASS Upgrade (COMPASS-U) will be a large magnetic field (5T) tokamak that will allow the scientific investigation of various physics issues related to the operation of the future ITER. In particular, an 80keV Neutral Beam Injection (NBI) system is planned to heat up the plasma with 4MW of external power. Such a unit was tested on the COMPASS tokamak before its shut down and our modelling contributed to the interpretation of the results. The study and modelling of NBI-born particle behavior is of great relevance: it might influence future design of the system and its integration in the overall reactor design. We request computational time for the modelling of the interaction of the fast particles with the background plasma. Our code, EBdyna, with its new collisional features, was benchmarked against the NUBEAM code on several test cases. Publications in the Nuclear Fusion and Journal Of Fusion Energy journals summarize the results of our initial modelling effort.


Researcher: Sergiu Arapan         


A computational study of magnetoelastic transitions in Fe2Hf(1-x)Ta(x) Laves phase for magnetocaloric applications       

Barbora CPU Alloc=9400;  Karolina CPU Alloc=15000;  Karolina GPU Alloc=3200;  LUMI-C Alloc=11700   

Magnetic refrigeration is an environmentally friendly cooling technology that can, providing high efficiency, replace the conventional gas-compressed refrigeration. It is designed on the basis of the magnetocaloric effect, which is a heating or cooling of the magnetic material under the influence of the applied magnetic field. The magnetocaloric effect is large at temperatures where phase transitions take place. In the Fe2Hf1-xTax a magnetoelastic transition from a ferromagnetic to antiferromagnetic state occurs with temperature, that can be tuned by Ta content x. This magnetoelastic phase transformation becomes a first-order phase transition at certain concentration of Ta which increases the adiabatic temperature change due to decreased thermal hysteresis. In this study we will investigate the properties of the Fe2Hf1-xTax alloy with Ta concentration x to predict the magnetoelastic transition temperature and quantify the magnetocaloric effect based on electronic structure and phonon calculations.


Researcher: Karel Carva               


Manipulating magnetic systems using electromagnetic fields on femtosecond timescale

Karolina CPU Alloc=72000;  LUMI-C Alloc=7200

Strong photon pumping can drive matter to unique transient states and drastically change  its magnetic properties. This effect is often mediated by phonons coupled strongly to specific photon energies. We intend to employ first principles calculation methods to reveal optical and magneto-optical properties properties of the strongly nonequilibrium pumped state. We will study how is the ground state and exchange interactions affected by the presence of coherent phonon population with emphasis on nonlinear behavior. We also attempt to address open questions connected to circularly polarized phonons. This type of phononic mode calls for non-trivial approach in connection to existing simulation codes. Our method requires to calculate large number of relatively large supercells, which is numerically demanding. We will investigate microscopic aspects of nonlinear spin-phonon interactions leading to magnetic order modifications in order to understand how to tune this process and improve efficiency. The project should reveal optimal combinations of specific materials and phonon modes for control of magnetism.


Researcher: Eugen Hruska          


Predicting drug candidate metabolism  

Karolina CPU Alloc=478;  Karolina GPU Alloc=300            

Drug metabolism estimation plays a vital role during the development of pharmaceutically active compounds. To improve the reliability of computational drug metabolism prediction, high-throughput physics-driven ab initio calculations will be performed.