NHR Compute Time Projects
Current NHR compute time projects on Alex and Fritz at NHR@FAU
Humanities and Social Sciences | Life Sciences | Natural Sciences | Engineering Sciences | Completed NHR compute time projects
Humanities and Social Sciences
Life Sciences
Basic Research in Biology and Medicine
CaSRMD: Molecular Dynamics Investigation of Calcium-sensing Receptor Dynamics (01/2024–12/2024; LARGE SCALE)
The Calcium sensing receptor (CaSR) is a class C GPCR, a homodimeric heptahelical membrane protein with am extracellular venus-fly-trap (VFT) domain and a cystein-rich region (CRR). CaSR senses calcium, physiologically present in millimolar concentration range. However, how the extracellular region senses concentration differences in these high ranges and it appropriately mediates receptor activation remains unknown. In light of new Cryo-EM structures of the CaSR, we intend to explore VFT/CRR activation dynamics by permutating occupancy of binding sites, and further simulate the full-length and G-protein bound CaSR dimer. Additional analysis by e.g. allosteric network analysis of this multi-regulated system featuring various allosteric modulators, helps to uncover the activation dynamics of the CaSR.
Scientific field: 201-02 Biophysics; 201-04 Structural Biology
University: Freie Universität Berlin
Target system: parallel computer Fritz & GPGPU cluster Alex
ORmd: Systematic molecular dynamics simulations of the odorant receptor family (10/2023–09/2024; LARGE SCALE)
ORmd aims to explore the dynamics of all the odorant receptors (ORs). To achieve this, we will employ cutting-edge High-Throughput Molecular Dynamics simulations starting from receptor models in the active and inactive states. This work will allow us to generate for the first time a database of OR structures and molecular dynamics, shedding light on this large but unexplored GPCR subfamily. The analyses of the trajectories will allow us to outline structural features relevant to the activation, which will be common to all ORs or specific for some subgroups.
Scientific field: 201-02 Biophysics
University: TU Munich
Target system: GPGPU cluster Alex
ChannelProtonation: Investigating the Influence of Protonation States on Cation Conductivity in Ion Channels through Molecular Dynamics Simulations (10/2023–12/2024)
This project focuses on investigating the dynamics of proton-coupled cation permeation in ion channels, with a specific emphasis on comprehending how pH influences the channel’s cation conduction and selectivity. Our research falls under the framework of CRC1078, which aims to enhance our understanding of protonation dynamics in protein function by fostering collaboration among various experimental and computational groups. Within this project, we have set two primary objectives:
(i) To study the Influenza A M2 proton channel, a pH-activated viroporin known for exhibiting both proton and cation transport activities.
(ii) To investigate CNG channels, which are non-selective cation channels that play a pivotal role in the signal transduction pathways of vision and olfaction.
Our approach involves the utilization of classical molecular dynamics simulations and constant-pH simulations to model the protonation patterns of titratable residues within the channel pore and the cation conductance across various protonation states. We have used similar approach before to study AMPA receptor and K2P channels.
Scientific field: 201-02 Biophysics
University: TU Berlin
Target system: GPGPU cluster Alex
DNA-glycosylase: DNA Base excision repair (06/2023–06/2026)
DNA damages are a constant threat to the integrity of the genome. Damages of the DNA bases can arise from oxidation or aberrant methylation. Deamination of cytosine bases, for example, results in U:G mismatches which ultimately lead to mutations in the encoded protein. The Base excision repair system (BER) is a machinery of enzymes, recognising and removing damaged/wrong bases and replacing them with the correct one. While most base excision repair glycosylases usually only excise the base, it was shown that some also have lyase activity and thereby circumvent AP endonuclease which is otherwise responsible for this second step of the repair mechanism. This project aims for revealing the mechanism of this strand incision, how this relates to the rather narrow specificity of glycosylases with dual activity compared to the broader set of substrates in mono-functional glycosylases.
Scientific field: 201-02 Biophysics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
AmPeL: Interaction of Antimicrobial Peptide Lugdunin with Membranes (07/2023–06/2024; LARGE SCALE)
The AmPeL project seeks to advance our understanding of the novel antimicrobial peptide lugdunin by characterizing its interaction with various bacterial and eukaryotic membrane models. Specifically, the project aims to elucidate the mechanisms of action underlying lugdunin’s antimicrobial properties against bacteria, as well as investigate the reasons for its lack of activity against erythrocytes. By shedding light on the precise molecular interactions between lugdunin and different membrane types, this research could ultimately facilitate the development of new antimicrobial therapies that are both effective and safe for human use.
Scientific field: 201-02 Biophysics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz & GPGPU cluster Alex
AMPARGating: Investigation of cation gating and conduction mechanism of an AMPAR-type glutamate receptor (07/2023–06/2024)
The understanding of the molecular details behind the regulatory mechanisms of the AMPA-type glutamate receptor (AMPAR) is of high importance. The process of cation conduction in the AMPAR ion channel offers several points of action for modulation. The divers AMPAR family exhibits in general a huge plethora of different regulations und modulations of their activity. Some examples for those effects are stabilization of the open or closed state, slow down or facilitate desensitization, influence on the conductance for calcium ions, prevent or allow the ability to be blocked by polyamines and so on. Many of these effects are mediated by auxiliary proteins which bind to the core-receptor. The atomistic details behind these interactions are not fully understood. Thus, MD Simulation represent a versatile method to investigate questions like these. In this project, we apply all-atom MD Simulations to Cryo-EM structures of the AMPAR with bound auxiliary proteins to reveal the mechanisms by which the cation gating and conduction is modulated. Our focus will lie on a subunit composition which is thought to represent a stabilized open state and a different subunit composition which has an influence on the polyamine block. We aim to elucidate molecular interactions which are crucial for AMPAR regulation.
Scientific field: 201-02 Biophysics
University: Humboldt-Universität zu Berlin
Target system: parallel computer Fritz
Dynasome3: Exploring Protein Dynamics Space to Improve Protein Function Prediction (10/2022–09/2024; LARGE SCALE)
The function of proteins is determined by their amino acid sequence and tertiary structure, but nevertheless the particular function of most proteins is unknown. In the Dynasome project we explore to what extent protein function can be predicted by protein dynamics, and explore the space of protein dynamics in general. To this aim, we perform molecular dynamics simulations for a large set of 200 proteins. We analyze these simulations, using e.g. Markov state models, to capture a ‘dynamics fingerprint’ of the studied proteins. We suggest these dynamics fingerprints as a new tool for protein function prediction and for quantitative comparison of protein dynamics.
Scientific field: 201-02 Biophysics
University: Georg-August-Universität Göttingen
Target system: parallel computer Fritz
CoupledFoldBind:Conformational presentation switching processes studied by Molecular Simulations (07/2022–12/2024)
Using Molecular Dynamics (MD) simulations, it is possible to follow conformational changes in proteins at atomic resolution and at high time resolution. MD simulation studies can supplement experimental approaches that typically allow only the structural characterization of initial or final or average structures. Especially in case of conformational switching processes such as binding induced folding an understanding of the process requires the analysis of intermediate states and driving forces for conformational changes. We apply MD-simulations and advanced sampling techniques to understand the molecular details of conformational switching processes within the SFB1035 (“Control of protein function by conformational switching”). One focus is on the transition of a DNA binding domain from a partially unfolded state to a folded state upon binding to DNA. The simulations will be used to define arrangements at which the partners start to influence the conformational switching process and to identify associated energy barriers. As a second system we study the process of collagen folding which is process that involves the coupled folding of three peptide strands and association to form a stable triple helix conformation. The general goal of our simulation studies is to understand at atomic detail how a protein or DNA surface or a specific modification of a protein can help to promote transitions towards a folded conformation of a binding partner or protein segment.
Scientific field: 201-02 Biophysics
University: Technical University of Munich
Target system: GPGPU cluster Alex
SimMediSoft: Biomolecular simulations for the efficient design of lipid nanoparticles (07/2022–12/2024)
With the incredible growth in RNA therapeutics, lipid nanoparticles (LNPs) have become an indispensable tool to deliver RNA to target cells thereby providing promising perspectives to combat life-threatening diseases such as Amyloidosis or COVID-19. We use all-atom molecular dynamics simulations to resolve the structure of these clinically relevant particles and proved molecular insights how the RNA cargo is distributed inside the LNPs. The structural information gained from our simulations will help to guide the design of LNPs with improved properties.
Scientific field: 201-02 Biophysics
University: University of Augsburg
Target system: parallel computer Fritz & GPGPU cluster Alex
GPCRSCOMPEVO: Computational models of structure, dynamics and evolution of GPCRs (07/2022–12/2024)
GPCRs constitute the largest protein family in the human genome. This genome-encoded protein repertoire of about 1000 receptors is expressed in a tissue- and organ-specific manner and transduces a large variety of extracellular signals into the cell. Whilst key residues determining coupling specificity of G proteins have been localized at the Gα-C terminus (GαCT), e.g. of Gs or Gi or at the finger loop of arrestin, a common sequence motif of GPCRs responsible for specific recognition of different GαCT or finger loops has not been detected yet. The R*-Gs/i/o arrestin complexes resolved so far do not provide a clear explanation for G protein coupling specificity. Evidence from several sources suggests the existence of transient complexes between the R* and GTP- bound G protein that may represent several novel intermediates on the way to the formation of GαsGTP and may contribute to coupling specificity.
Scientific field: 201-07 Bioinformatics and Theoretical Biology; 201-02 Biophysics
University: Leipzig University
Target systems: parallel computer Fritz & GPGPU cluster Alex
Intermediate Report (02/2023)
Heterotrimeric Gαβγ proteins are guanine nucleotide-binding proteins consisting of alpha (α), beta (β), and gamma (γ) subunits. Gαβγ proteins are activated by G protein-coupled receptors (GPCRs), the largest family of membrane receptors sharing a common seven transmembrane helix (TM) structure. Essential fundamentals of GPCR signaling at the molecular level are now considered to be well understood: Upon binding of extracellular ligands, the receptor undergoes a conformational change that opens up an intracellular pocket to which the G protein can bind. Formation of the ternary ligand-receptor-G protein complex goes along with a conformational change in the Gα subunit that ultimately triggers the exchange of GDP for GTP. For over a decade GPCRs signaling through G-prortein were over simplified looking only at the nucleotide free state. To extend our knowledge on the structural events of GPCR signaling at atomic-level details, we have used all-atom molecular dynamics (MD) simulations combined with time-resolved cryo-EM to explore dynamics of the agonist bounds GPCRs complexed with G-protein in either nucleotide bound (GTP or GDP) or nucleotide free state.
Intermediate Report (09/2023)
Heterotrimeric Gαβγ proteins, comprising alpha (α), beta (β), and gamma (γ) subunits, are key players in G protein-coupled receptor (GPCR) signaling. GPCRs represent the largest family of membrane receptors, characterized by a common seven-transmembrane helix (TM) structure. The basic principles of GPCR signaling at the molecular level have long been established: extracellular ligand binding induces a conformational change in the receptor, creating an intracellular pocket for G protein binding. This leads to the formation of a ternary ligand-receptor-G protein complex, accompanied by a conformational change in the Gα subunit, ultimately promoting the exchange of GDP for GTP. However, this over-simplified view of GPCR signaling has evolved in light of current research endeavors. Our approach combines all-atom molecular dynamics (MD) simulations with innovative experimental techniques such as time-resolved cryo-electron microscopy (cryo-EM) to explore the intricate dynamics of GPCRs complexed with G-proteins. Our investigations encompass agonist-bound GPCRs in conjunction with G-proteins, considering both nucleotide-bound (GTP or GDP) and nucleotide-free states. Our research seeks to extend the boundaries of GPCR signaling by providing atomic-level insights into the dynamic interplay between GPCRs and G-proteins across various nucleotide states. Through the integration of computational and experimental methodologies, we aim to unravel the conformational changes and molecular events governing these interactions. This approach promises to refine our comprehension of GPCR signaling and its intricacies, potentially offering novel avenues for drug discovery and therapeutic intervention.
Antivirals: Structure-based design and optimization of ligands for novel antiviral strategies (04/2022–06/2024)
Broadly neutralizing antibodies that bind to viral fusion proteins represent a promising strategy for protection from viral infections. Such antibodies can be used for passive immunization and are currently tested in clinical trials, but they are expensive and difficult to produce. As an alternative, antibody-derived peptides may be used for this purpose. In the present project, the complexes between antibodies and the viral fusion proteins from HIV-1 and CoV-2 are analyzed to identify energetic hot-spots of the interaction. This information will be used for the design of antibody-derived peptides that bind to viral fusion proteins thereby blocking viral infection. For that purpose, a computational pipeline is developed that uses molecular dynamics (MD) simulations to identify the most promising peptides for further experimental testing.
Scientific field: 201-02 Biophysics; 204-04 Virology
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Intermediate Report (04/2024)
Neutralizing antibodies that bind to viral fusion proteins represent a promising strategy for protection from viral infections. Such antibodies can serve as templates for the generation of peptides, which retain the ability to bind to viral proteins. We have screened more than 1300 antibody-antigen complexes from HIV-1 and SARS CoV-2 proteins to identify energetic hot-spots of the interaction. The energetically most favorable peptide stretches were selected for further computational characterization and experimental binding studies. Molecular dynamics simulations allowed to identify structural features that differ between binding and non-binding peptides. In particular, the formation of rather stable binding-incompetent conformations seems to be a major reason for the inability to bind. This information will now be used for an improved identification of peptides, which retain antibody-like binding properties.
Plant Sciences
CEC: Convergent evolution of carnivorous plants (08/2022–09/2024)
Plant carnivory (i.e., plants feeding on trapped animals) has evolved independently at least six times. This repeated evolution implies that the genetic basis for plant carnivory is present in most – if not all – plants. By comparison of the genetic and transcriptomic landscape between carnivorous and non-carnivorous plants, we may be able to identify the common genetic elements required for plant carnivory. This will not only grant us insights into plant carnivory, but into plant evolution as a whole.
Scientific field: 202-01 Evolution and Systematics of Plants and Fungi
University: University of Wuerzburg
Target system: parallel computer Fritz
Medicine
MIA-NORMAL: Medical Image Analysis with Normative Machine Learning (10/2023–09/2024; LARGE SCALE)
As one of the most important aspects of diagnosis, treatment planning, treatment delivery, and follow-up, medical imaging provides an unmatched ability to identify disease with high accuracy. As a result of its success, referrals for imaging examinations have increased significantly. However, medical imaging depends on interpretation by highly specialized clinical experts and is thus rarely available at the front- line-of-care, for patient triage, or for frequent follow-ups. Very often, excluding certain conditions or confirming physiological normality would be essential at many stages of the patient journey, to streamline referrals and relieve pressure on human experts who have limited capacity. Hence, there is a strong need for increased imaging with automated diagnostic support for clinicians, healthcare professionals, and caregivers.
Machine learning is expected to be an algorithmic panacea for diagnostic automation. However, despite significant advances such as Deep Learning with notable impact on real-world applications, robust confirmation of normality is still an unsolved problem, which cannot be addressed with established approaches.
Like clinical experts, machines should also be able to verify the absence of pathology by contrasting new images with their knowledge about healthy anatomy and expected physiological variability. Thus, the aim of this proposal is to develop normative representation learning as a new machine learning paradigm for medical imaging, providing patient-specific computational tools for robust confirmation of normality, image quality control, health screening, and prevention of disease before onset. We will do this by developing novel Deep Learning approaches that can learn without manual labels from healthy patient data only, applicable to cross-sectional, sequential, and multi-modal data. Resulting models will be able to extract clinically useful and actionable information as early and frequent as possible during patient journeys.
Scientific field: 205-30 Radiology, Nuclear Medicine, Radiotherapy, Radiobiology; 205-12 Cardiology, Angiology; 409 Computer Science
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Medchem-Dynamics: Molecular Dynamics and Docking Studies with Multifunctional Receptor-Ligand Complexes (01/2023–03/2024; LARGE SCALE)
The establishment of suitable molecular models, based on experimental target structures, is a prerequisite for the successful design of multifunctional drugs. In all approaches of the proposed CRC (TRR 351), such models help to understand relationships of structure and function, including the interactions of the individual modules. Modeling will also guide ligand optimization. Retrospectively, molecular models can support the rationalization of the observed biological responses and integrate novel structural information obtained by biophysical methods.
Our molecular models will be important
- to understand the origins of a drug’s pharmacological effect at the atomic level. Based on existing X-ray crystallography or cryo-EM structures, our multifunctional ligands bound to their target will be investigated by long-timescale molecular dynamics (MD) simulations.
- to guide the design and development of multifunctional drugs in all project areas of the CRC. The interaction quality and stability of envisaged ligand-target complexes will be evaluated by means of docking, energy minimization and MD simulations.
- to identify new chemotypes by virtual screening. Filtering ultra-large libraries (in the order of 109 compounds) with pharmacophore models and molecular docking paves the way to novel multifunctional ligands with non-canonical receptor-ligand interactions.
- to decipher the relationships between structure and function. The elucidation of multidimensional SAR analyses allows us to predict useful modifications of lead compounds to improve affinity, selectivity and functional properties.
- to predict potential cooperativity between the modules. MD simulations of target-bound multifunctional ligands compared to their target-bound modules will guide compound development.
- to establish design principles of multifunctional biopharmaceuticals. Simulations will evaluate their stability and structure and, thus, guide their design.
Scientific field: 205-8 Pharmacy
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
Intermediate Report (03/2024)
β-Adrenergic receptors (βARs) induce a beneficial increase in inotropy of the heart in acute heart failure, strengthening myocardial contraction and, thus, heart function. β1AR-Gs and β2AR-Gi biased ligands with functional selectivity over β-arrestin recruitment to prevent receptor desensitization will be useful for the treatment of pathological hypertrophy and heart failure. Bitopic ligands are drugs or lead compounds comprising a pharmacophore for the orthosteric site and a second moiety specifically addressing an allosteric site. The design of our bitopic ligands will be assisted by computational methods. Whereas the choice of the orthosterically-binding modules will rely on recently described lead structures, the allosteric modules will be identified by docking studies in combination with simulation analyses. In particular, methods for the design of covalent ligands, including covalent docking, will allow us to find new chemotypes with a point of attachment that can be used to connect such fragments to orthosteric moieties via a suitable linker unit. For this purpose, both the non-covalent association complex and the covalent bond formation must be considered in the rational development and fine-tuning of covalent ligands. This strategy is being adopted for bitopic βAR ligands targeting nucleophilic residues at allosteric sites of the receptor. The particular challenge here is based on less nucleophilic and more surface-exposed residues. The warheads must be reactive enough to compensate for this problem without increasing the risk for off-target effects. This requires ligands that bind with good affinity and also enable optimal and stable orientation of the warhead. The simulations carried out so far helped to prioritize the addressable nucleophilic resdiues as well as the electrophilic warheads that can be suitably positioned for a covalent reaction. The calculations and experiments are ongoing.
5-HT1A receptor – The overuse of hydrocodone and inadequate pain treatment have contributed to a severe opioid epidemic in the United States, affecting millions of people across all socio-economic classes. The 5 HT1A receptor is a key target for pain suppression, but previous attempts faced challenges due to unwanted side effects.
In our recent study, we utilized cryo-EM to examine the structure of the 5 HT1A receptor bound to befiradol, a drug candidate with limiting side effects. We also introduced our novel ligand, ST171, which exhibited promising properties for clinical use as an analgesic. Molecular dynamics simulations revealed persistent macroscopic differences in the helical structure of the receptors, confirming that these changes were induced by the binding of different ligands.
Comparing the signaling profiles of ST171, befiradol, and the endogenous ligand serotonin yielded intriguing results. While befiradol and serotonin recruited Gi protein at low and moderate concentrations, both also recruited Gs protein at high concentrations, along with the effector protein β-arrestin. In contrast, ST171 exclusively recruited the Gi protein, even at high concentrations, suggesting a potentially superior safety profile. Molecular dynamics simulations indicated that amino acids Q972.65 and W3877.40 contributed to this difference in signaling, with stable interactions in the ST171-bound complex compared to significant fluctuations in the other complexes.
Finally, our investigation into the befiradol-bound complex revealed that the side chain of F1123.28 needed to adopt an unfavorable conformation to open a lipophilic pocket for befiradol binding. This pocket remained closed in both our ST171-bound complex and all other previously reported 5-HT1A structures. This led us to question whether befiradol binding is based on a conformational-selection or an induced-fit mechanism. Using the enhanced sampling method metadynamics we clarified that a spontaneous opening of this pocket is energetically unfavorable, suggesting that only an induced-fit mechanism is feasible.
GPR3 – With the rise in life expectancy, the prevalence of Alzheimer’s disease has consistently increased in recent decades. While existing treatments slow down the disease’s progression, they fall short of providing a cure. In many Alzheimer’s patients, the GPR3 receptor is overexpressed in neurons, making it a valuable target for treatment. Furthermore, the absence of GPR3 receptors has been linked to obesity, underscoring the pharmacological importance of this receptor.
In a collaborative initiative with colleagues from the Chinese University of Hong Kong, we successfully determined the cryo-EM structure of the GPR3 receptor. However, the electron density didn’t allow for a clear identification of the bound ligand. MD simulations played a crucial role in resolving this uncertainty, demonstrating that the ligand OEA adopted a more stable binding pose compared to oleamide, the other potential candidate. This finding conclusively answers the question regarding ligand binding.
Moreover, the GPR3 receptor displays an increased basal activity. To delve into these findings, we employed our well-established metadynamics protocol to investigate the activation dynamics of the GPR3 receptor. Our simulations revealed that even in the absence of a ligand, the receptor stabilizes in a conformation that lies between a typical active and inactive state. This phenomenon elucidates the favorable basal coupling of the Gs protein.
PatRo-MRI-2: Pathology-robust image reconstruction in Magnetic Resonance Imaging (10/2022–12/2026)
Medical imaging, particularly Magnetic Resonance Imaging (MRI), plays a pivotal role in contemporary healthcare, offering noninvasive diagnostic tools, guidance, and exclusive treatment monitoring and disease understanding options. The standard diagnostic medical imaging pipeline involves collecting raw data via scanner hardware, processing this data through image reconstruction algorithms, and subsequently analyzing it for pathology by radiology specialists. While this procedure has traditionally been optimized to produce visually interpretable images for human experts, it may inadvertently lead to the loss of essential patient-specific diagnostic information contained in raw sensor data.
This issue has been further compounded by recent advancements in machine learning for MRI reconstruction. Machine learning has the capacity to generate aesthetically pleasing images from minimal sensor data, which is beneficial for speedy imaging protocols that prioritize patient comfort. However, this reduction in data acquisition could cause generic image reconstruction techniques to obscure disease markers, replacing pathological features with typical healthy image features derived from the training data.
Machine learning has also shown remarkable proficiency in autonomously analyzing medical images and robustly identifying disease patterns across various modalities. However, these tools are generally detached from the image acquisition process, and are only applied to the reconstructed image data.
In this project, we aspire to amalgamate machine learning for image reconstruction and disease localization based on the image, thereby crafting an end-to-end learnable solution for image reconstruction and joint pathology detection that functions directly on raw measurement data. We predict this combination will amplify diagnostic accuracy, producing images that are ideally suited for both human experts and diagnostic machine learning models.
Our approach will lean on the capabilities of foundation models, with a particular focus on state-of-the-art diffusion models for image synthesis. This choice is motivated by their leading performance, but they require substantial GPU time for training, hence our request for more GPU compute time.
We’ve outlined three primary objectives for our project:
1. Developing a machine learning reconstruction method for accelerated MRI that withstands variations in scanners, sequences, and anatomy.
2. Designing disease detection techniques that can learn from human-generated annotations in image space while concurrently extracting information straight from accelerated raw measurement data.
3. Combining these developments to introduce end-to-end training of image reconstruction and image analysis.
We hypothesize that such an approach will prevent machine learning-based MRI interpretation from missing crucial disease indicators and offer tools to express interpretation uncertainties in image regions with limited sensor data. To this end, we will utilize our fastMRI+ dataset, which comprises fully sampled raw measurement data, images, and 20,000 specialist expert bounding box annotations for more than 50 pathology categories from over 8,000 patients.
Scientific field: 409 Computer Science;205-30 Radiology, Nuclear Medicine, Radiotherapy, Radiobiology; 205-12 Cardiology, Angiology
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
FPRMetaD: Investigating binding pathways for a diverse set of ligands with biased and unbiased simulation of the Formyl Peptide Receptor (08/2022–12/2025)
Our project includes unbiased and biased molecular dynamics simulations of the FPR receptor class and its vast array of ligands that include modified peptides as well as non-modified peptides and small molecules like Lipoxin A4 or the circular peptide Ciclosporin A. A standardized metadynamics protocol is applied to study the binding/unbinding events of the ligands to these receptors and enhance sampling. High parallel performance is achieved using multiple-walker simulations. The ligands will include agonists, partial agonist and antagonist to cover as many different binding routes as possible.
Scientific field: 205-08 Pharmacy
University: Westfälische Wilhelms Universität Münster
Target system: parallel computer Fritz & GPGPU cluster Alex
Intermediate Report (10/2023)
The Formyl Peptide Receptors (FPRs) belong to class A of G-protein coupled receptors (GPCRs); the family has three members: FPR1, FPR2 and FPR3. FPRs play a key role in the host defense against microbes because they are located on immune cells like phagocytes that belong to the innate immune system. Thus, they are also involved in inflammatory diseases like Alzheimer’s disease or cancer and are novel targets for the treatment of those diseases. Several small molecules and peptides are known which either act as agonists, partial agonists or antagonists. We selected some of the small molecules that represent these classes for further computational analyses including binding-mode identification, binding-pathway analysis and free energy of binding calculations using metadynamics simulations. First, the molecules were docked into the orthosteric binding site of FPR2 of a cryo-EM structure in complex with a known peptide agonist (PDB 7WVW). The resulting FPR2 complex structures were used for μs-time scale metadynamics simulations conducted mainly on the Alex cluster of NHR@FAU in Erlangen using Gromacs and Plumed. The simulations were run following a protocol that was already described earlier by Saleh et al.. With this approach, it was possible to simulate the binding/unbinding of the ligands and to determine the free energies of binding with sufficient accuracy. The global free energy minimum obtained corresponded well to the known binding poses and gave new insight into the binding mode of small-molecule ligands. This knowledge will be used to design novel ligands for FPR2. The results highlight the importance of a deep, hydrophobic pocket at the bottom of the orthosteric site and of three polar residues right above it. Furthermore, these data suggest comparable binding modes of small molecules and peptides, contrary to earlier results.
InTimeVRSimulPatMod: In-time Virtual Reality Simulation Patient Models: Machine Learning and immersive-interactive Modeling of Virtual Patient Bodies (06/2022–05/2025)
We aim to provide high-quality body models given medical imaging data by the segmentation of relevant structures. Fast 3D modelling is relevant in clinical-radiological everyday routine. In the automatic modeling part of a DFG project, current machine learning methods and atlas-based methods are to be compared for their segmentation proposals and combined to their strengths. A major goal of this project is the development of novel, fully automatic, group-oriented deep-learning and multi-atlas segmentation estimators for the highly efficient multi-organ and simultaneous tumor segmentation in medical 3D-CT image data sets.
Scientific field: 205-01 Epidemiology, Medical Biometry, Medical Informatics; 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Aalen UAS
Target systems: parallel computer Fritz & GPGPU cluster Alex
Neurosciences
SFB 1540: Exploring Brain mechanics (EBM) Understanding, engineering, and exploiting mechanical properties and signals in central nervous system development, physiology and pathology (01/2024–12/2024)
The CRC1540 EBM project explores mechanical influences on the central nervous system (CNS), of which many processes are not yet fully understood. Mechanical signals have shown to impact CNS cell function, emphasizing mechanics’ crucial role. EBM aims to enhance CNS understanding for improved neurological disorder diagnosis and treatment. Subproject X02 focuses on leveraging the diverse data collected within EBM and employs advanced machine learning for domain adaptation and generalization.
Scientific field: 206-04 Systemic Neuroscience, Computational Neuroscience, Behaviour; 409 Computer Science
University: Julius-Maximilians-Universität Würzburg
Target systems: parallel computer Fritz & GPGPU cluster Alex
HPC-MarkovModelling: Single-channel Markov modelling of voltage-gated ion channels with simulations and implementation of the 2D-Fit algorithm on High Performance Computing Cluster (08/2022–02/2025)
In this project, we want to explore the computational power of HPC-Cluster for modelling single-channel patch-clamp data with Markov models. The 2D-Dwell-Time fit with simulations of time series captures gating kinetics with a high background of noise and can extract rate constants beyond the recording bandwidth. That makes the 2D-Fit exceptionally valuable for relating ion-channel kinetics with data from simulations of single molecules. In addition, 2D-distributions preserve the coherency of connected states. Thereby the algorithm can extract the full complexity of underlying models and distinguish different models.
Scientific field: 206-04 Systemic Neuroscience, Computational Neuroscience, Behaviour
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target systems: parallel computer Fritz & GPGPU cluster Alex
Natural Sciences
Molecular Chemistry
SpectroscopicProperties: Spectroscopic properties of molecules with unusual electronic structures (07/2022–12/2025)
The isolation of multiple bonded late transition metal complexes is a challenging task. Nevertheless, they are alleged key intermediates for important “every day” catalysis processes such as the oxidation of ammonia to nitric acid by O2 (Ostwald process) or the catalytic depletion of toxic exhaust gases. Our calculations predict how to tame these elusive molecules to study them in the laboratory. Our work focused so far on terminal imido complexes of the late transition metals (Mn, Fe, Co, Ni, Ir, Pd, Pt) and has not only allowed to understand their intriguing electronic structure, but furthermore to demonstrate first applications in small molecular activation, energy conversion, catalysis, and photochemistry.
Similarly, we use computational methods to study organic redox systems and diradicals. We have predicted how to harness the peculiar properties of carbene decorated diradicals in solar cells and demonstrated their use as singlet fission molecules. Thus, our calculations helped to discover a new class of molecules of use for solar energy conversion, quantum computing, or organic light emitting diodes (OLEDs).
Scientific field: 321-01 Inorganic Molecular Chemistry; 321-02 Organic Molecular Chemistry; 323-01 Physical Chemistry of Molecules, Liquids and Interphases, Biophysical Chemistry
University: Saarland University
Target system: parallel computer Fritz
Intermediate Report (12/2022)
Our project deals with molecules with unusual electronic structures. We aim to harness these electronic structures in green catalysis as well as for material design in organic electronics such as photovoltaics. For bringing these molecules to application, a profound understanding of their complicated electronic structures is required.
As such, we perform synthetic, computational- and spectroscopic work hand-in-hand for optimal synergy. So far, we have focused on late transition metal complexes for catalysis and carbene-derived organic materials. A recent breakthrough was the isolation of the first example of a terminal nitrene of palladium.
Another landmark finding was, how carbenes stabilize radicals for organic electronics. In subsequent steps, we wish to extend our studies to other earth-abundant elements for enhanced sustainability.
MoTrNanoMat: Molecular transport in nanoporous materials (07/2022–10/2025)
Nanoporous materials are of interest in many applications due to their high surface area and physicochemical properties. The interconnected channels can be used for “flow-through” applications such as purification of drinking water or nanoseparation of proteins or organic solvents. Here, we will evaluate the impact of (i) nanomaterial kind, (ii) pore size, (iii) pore shape, and (iv) solvent polarity on the material’s permeability using sequential multiscaling molecular dynamics (MD). We will establish molecular models of 3D structures of amorph carbonaceous materials and of diverse metal-organic frameworks. Next, variation of the pore diameter, surface functionalization and solvent polarity will enable us to systematically evaluate the impact of the individual material properties on e.g. molecular transport, solvent competition, and nanoseparation.
Scientific field: 301 Molecular Chemistry; 302-03 Theory and Modelling
University: University of Stuttgart
Target system: parallel computer Fritz & GPGPU cluster Alex
Chemical Solid State and Surface Research
Crystal23: Prediction of new ferroelectric metal fluorides (04/2024–03/2025; NHR STARTER)
Ferroelectrics are insulators with a spontaneous electrical polarization that can be reversed by an external electric field. Their properties make them sought-after functional materials with a wide range of potential applications. A basic requirement for the occurrence of ferroelectricity is that the material crystallizes in a polar space group.
In this work we want to predict new ferroelectric metal fluorides. To this end, we will compile an overview of polar metal fluorides, which we will investigate using crystallographic and quantum chemical methods. The starting point is a database screening that yields fluorides in 105 structure types that potentially crystallize in polar space groups. We will perform DFT calculations on individual representatives of the structure types to confirm their polar structure.
We will perform structural optimizations and calculate vibrational frequencies at the Γ-point in the harmonic approximation to verify that our candidates are dynamically stable at 0 K. If the DFT structural optimization converges to a nonpolar structure, we will classify the structural model of this compound as potentially erroneous and suggest a redetermination of the crystal structure. We will perform the DFT calculations with the program CRYSTAL23 using the hybrid functional PBE0 and atom-centered basis functions. We will adopt technical parameters and convergence criteria from our previous theoretical studies on fluorides. All structure types of our screening have less than 200 atoms per unit cell. From our preliminary investigations and the system sizes, we extrapolate the required computing time to about 500,000 CPU hours. For all structural optimizations, we estimate the requirement at 100,000 CPU hours, for the frequency calculations at 400,000 CPU hours. Crystal23 is effectively parallelized via OpenMP and MPI. For our system sizes, we use 16 to 64 cores and 4 GB to 32 GB RAM as standard.
Scientific field: 302 Chemical Solid State and Surface Research
University: Philipps-Universität Marburg
Target system: parallel computer Fritz
Lg_SurfCatal_AIMD_MLFF: Computational modeling of new surface catalysis systems by means of ab initio methods as well as novel machine-learning force-field approaches (07/2022–12/2024; LARGE SCALE)
Catalysis at liquid interfaces (CLINT) provides a fascinating new research area with great potential to explore more efficient and sustainable catalytic processes. Since such kind of catalysis, especially those on supported catalytically active liquid metal solutions (SCALMS) and surface catalysis with ionic liquid layers (SCILL), is still quite new, much more understanding need to be gained on how exactly the mechanistic processes are taking place, leading to know-how accelerating the development of targeted systems for several reactions. Periodic DFT simulations are able to shed a light on the exact processes taking place at the catalyst. Recently, a new machine-learning force-field (ML-FF) was developed which is able to efficiently learn on the fly from DFT data, leading to a high-level FF for metal surfaces in contact with other phases, which are very complicated to describe with FFs so far. By parametrizing these ML-FFs for SCALMS and SCILL systems, we are able to explore new time- and length scales.
Scientific field: 302-03 Theory and Modeling
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
Intermediate Report (10/2023)
Our calculations focus on the atom-resolved understanding of novel catalyst systems, namely catalysis on liquid interfaces (CLINT). In contrast to usual heterogeneous catalysis taking place on metal surfaces, the novel systems have the advantage of better control of the catalytic processes and being less prone to unwanted side-effects like poisoning and coking. Two different catalytic concepts are mainly studied: Supported Catalytically Active Liquid Metal Solutions (SCALMS) and Surface Catalysis with Ionic Liquid Layers (SCILL). Within SCALMS, active atoms like Pt are distributed into a liquid solvent metal like Ga. The Pt atoms move around in the liquid and occasionally reach the surface, where they are available to the catalysis. Our simulations focus on how this motion actually takes place on an atomistic level. Since liquid metals are studied rarely by theory on larger scales, we set up novel machine learning force fields (ML-FF) for a cheap but nevertheless accurate simulation of them, enabling averaging of the surface properties. We further consider in our ML-FFs the interaction of the metal solvents with the support material, on which droplets of them are placed in reality and the formation of oxide layers on the surface and intermetallic phases in the liquid, thus being able to study a large part of the real-world complexity of SCALMS. Electronic properties of the systems can be calculated by periodic DFT, using snapshots of the ML-FF simulations. Within SCILL, the catalytically active metal surface is coated with a thin ionic liquid (IL) film. Depending on the IL species present, the interactions with the reactive species can be fine-tuned, further, unwanted side-reactions are suppressed by the coating. Our simulations aim at a detailed description of the interactions of IL molecules with the metal surface. We optimize IL adsorption patterns by DFT and simulate scanning tunneling microscope (STM) images of them, we further study the electronic properties of the IL/metal interface by applying electric fields to model systems with CO molecules adsorbed on the surface, acting as probing species. In close collaboration with the Zahn group, we are able to link the Coulomb interactions of explicit IL molecules obtained from traditional force field simulations to effective applied electric fields in model DFT surfaces. Besides these two main branches, we use our calculation time for several smaller projects, all covering novel surface systems, like the interaction of hexagonal bornitride (h-BN) sheats on a Rh surface with Br molecules. Due to fascinating nanomesh (moire) patterns arising from a unit cell mismatch of the Rh surface and the h-BN sheet, the interaction with gaseous Br is versatile and complicated. Besides thorough DFT calculations, we are able to simulate the real-time dynamics of such a system using a ML-FF trained for that purpose.
MoTrNanoMat: Molecular transport in nanoporous materials (07/2022–10/2025)
Physical and Theoretical Chemistry
AKES: Chemical Modelling of Processes in Pharmaceutical Chemistry (11/2022–02/2024)
Covalent inhibitors currently experience a renaissance in medicinal chemistry due to their various advantages, including prolonged residence times, lower sensitivity against pharmacokinetic aspects, and high efficacy. Our work addresses the reaction mechanisms of cysteine protease rhodesain with covalent inhibitors. Our goal is to further improve the modelling of the underlying inhibition mechanisms so that we can suggest improved compounds. Furthermore, we aim to develop a protocol to treat enzyme-inhibitor reactions with various local minima.
Scientific field: 303-02 General Theoretical Chemistry
University: Julius-Maximilians-Universität Würzburg
Target system: GPGPU cluster Alex
Ion-catch: Molecular Modelling based design of ligand shells to functionalize magnetic nanoparticles for the removal of heavy metal pollutants from water (07/2022–05/2025)
This project aims at designing tailor-made functionalization of magnetite nanoparticles to bind heavy metal ions and related organometallic compounds by means of molecular modelling and simulation. While at this stage reflecting a pure theory project, we aim at developing a model-based search strategy for identifying suitable constituents and structures as guides to syntheses. To achieve this, we build on a recently established concept for the removal of oil pollution from water that takes use of super-paramagnetic nanoparticles functionalized by self-assembled monolayers (SAMs) of n-alkyl phosphonic acids. While the association of hydrocarbons to the alkyl chains of such SAMs is driven by a simple hydrophobic segregation mechanism, well-chosen SAMs of tailor-made chelating residues are needed for efficiently binding and removing heavy metal pollutants from water. This calls for in-depth understanding of molecular recognition and the tuning of structural motifs – which shall both be achieved from molecular simulations.
Scientific field: 303-02 Theoretical Chemistry; 406 Materials Science
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target systems: parallel computer Fritz & GPGPU cluster Alex
Condensed Matter Physics
SuperSEM: Simulation of unconventional superconductors with long-range interactions beyond the mean-field approximation (10/2023–09/2025)
Long-range interactions in unconventional superconductors are of high importance as they can give rise to topologically nontrivial phases of matter within the mean-field approximation. In this work, we use the T-matrix approach to derive a generalized gap equation that goes beyond the mean-field approximation. We subsequently investigate how the phase diagram of a 2D superconductor changes as higher-order quantum corrections are included. Going further, we explore whether the Higgs mode, which can be stabilized due to long-range interactions in the mean field limit, remains stable in the T-matrix approach.
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: Universität des Saarlandes
Target system: parallel computer Fritz
PCL-topology: Optical and vibrational signatures of structural and electronic topology in planar carbon lattices (07/2023–06/2026)
Planar carbon lattices (PCLs), such as graphene nanoribbons, patterned graphene, 2D metal- organic/covalent-organic frameworks (MOFs/COFs), 2D polymers and their nanoribbons, combine the versatility of carbon-based molecular building blocks with the long-range order and symmetry-imposed properties of a 2D lattice. In recent years, precision synthesis methods have made tremendous progress in creating predefined PCLs on surfaces or at liquid interfaces. In this project, we will investigate the physical properties of such PCLs by numerical methods and predict new PCL structures with intriguing topological and correlated electronic properties.
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
FRASCAL P12: Quantum-to-Continuum Model of Thermoset Fracture (06/2023–05/2026)
FRASCAL improves understanding of fracture in brittle heterogeneous materials by developing simulation methods able to capture the multiscale nature of failure. With i) its rooting in different scientific disciplines, ii) its focus on the influence of heterogeneities on fracture at different length and time scales as well as iii) its integration of highly specialised approaches into a “holistic” concept, FRASCAL addresses a truly challenging cross-sectional topic in mechanics of materials.
In particular, sub-project P12 (Quantum to Continuum Model of Thermoset Fracture), will develop a concurrent multiscale modelling approach to study the interplay and coupling of process on different length scales (e.g. breaking of covalent bonds, chain relaxation processes, fibril formation and crazing at heterogeneities,…) during the fracture of an exemplary thermoset and its dependence on the (local) degree of cross-linking.
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz & GPGPU cluster Alex
SELRIQS: Series expansions for long-range interacting quantum systems (06/2023–12/2024)
We investigate the quantum cooperativity of strongly correlated light-matter systems in the presence of spatially competing matter-matter interactions. We aim at a deeper understanding of quantum phases and quantum phase transitions, where competing long-range interactions are expected to result in unconventional correlations and interesting entanglement properties. To this end we employ linked- cluster expansions on white graphs in combination with classical Monte Carlo integration to determine relevant physical quantities in the thermodynamic limit. This allows us to determine the notoriously complicated quantum-critical properties in quantum many-body systems with long-range interactions.
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
UltrafastDyn: Ultrafast electron dynamics and Kerr rotations (01/2023–06/2025)
This project focuses on simulating and understanding topological materials and their excitations under ultrafast laser driving; it is motivated by experimental activities towards “lightwave spintronics” in Dirac systems. The main diagnostic observable will be the optical and magnetic responses, in particular the high harmonic regime and Kerr rotations.
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: University of Regensburg
Target system: parallel computer Fritz
DMFT2TBLG: DMFT study of a heavy-fermion model for twisted bilayer graphene (12/2022–03/2024)
Twisted bilayer graphene (TBLG) has recently captivated the interest of the condensed matter community, for its capability of hosting a wide variety of peculiar phenomena such as superconductive and correlated insulating phases, as well as topological features. The aim of this project is to apply Dynamical Mean-Field Theory (DMFT) to a recently developed heavy-fermion model for TBLG, investigating the onset of correlated insulating phases for different occupations and temperatures, as well as dynamical screening effects.
Scientific field: 307 Condensed Matter Physics
University: Julius-Maximilians-Universität Würzburg
Target system: parallel computer Fritz
Intermediate Report (03/2024)
The first stage of the DMFT2TBLG project has focused both on Twisted Bilayer Graphene proper and on an equivalently interesting bilayer setup, the 1T-1H TaS2 heterostructure.
Regarding bilayer graphene, we have elucidated the behavior of the local moment fluctuations across a wide range of temperatures and fillings, determining how the heavy electrons in the flat bands are not efficiently screened up to very low values of temperature. We have further analyzed the charge compressibility curve of the Song-Bernevig model for MATBLG, highlighting a number of negative compressibility branches in accordance with recent experiments.
On the topic of TaS2, we have described the physics governing its peculiar spectral properties, showing a narrow peak near zero bias in various experimental realizations: we have determined it to be due not to heavy-fermion physics, but to a doped Mott scenario, by which the Mott insulating 1T layer is depleted by a work function gradient with respect to the metallic 1T layer, acting as an electron reservoir. We have supported our claim with a DFT+DMFT study, analyzing the properties of the Fermi surface and the local spin susceptibility.
nqsQuMat: Neural quantum states for strongly correlated quantum matter (12/2022–09/2024; LARGE SCALE)
The efficient numerical simulation of quantum systems constitutes a key challenge for current computational methods. In this project, we employ a modern method named neural quantum states which has shown great potential in studying various quantum problems. We plan to target both ground states as well as the time evolution of cutting-edge quantum models to gain new insights on novel quantum phenomena.
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: University of Augsburg
Target system: GPGPU cluster Alex
Intermediate Report (07/2023)
In this project, we utilize so-called neural quantum state method, which provide a compressed representation of the quantum many-body state into artificial neural networks. Up to now this method has been limited shallow networks with only around thousands of parameters, which has been caused by the computational complexity of optimization.
We have developed a new optimization method named MinSR that allows us to train deep neural quantum states of unprecedented size with up to 64 layers and 1 million parameters. This has allowed us to significantly improve the variational accuracy of the ground state for several paradigmatic quantum systems.
We expect that the deep neural quantum state trainable with MinSR will open up a new route for the numerical studies of strongly correlated many-body quantum systems. Apart from the current applications on quantum spin liquids, it can also be applied to fermionic systems or quantum dynamics in the future.
FRG: Functional Renormalization Group calculations for material analysis (11/2022–12/2024)
Current ab-initio theory for solid state materials excels in the prediction of electronic band structures. The secondary part of any full description – the interaction between electrons – is beyond the scope of those methods. To remedy this omission the functional renormalization group (FRG) has the expressed goal of deriving effective low-energy interaction models. This method has successfully been used to calculate some simple systems, but the goal of this project is to widen the scope of possible applications and include ever expanding classes of materials. We hope to uncover new effects such as pair-density waves, nematic instabilities and thus extend the predictive power of FRG.
Scientific field: 307-2 Theoretical Condensed Matter Physics
University: RWTH Aachen
Target system: parallel computer Fritz
Intermediate Report (01/2024)
Within this project, we have investigated models for common material classes such as square, triangular, and hexagonal lattice structures. Already for this seemingly simple model systems, the effective low-energy theory revealed a broad range of highly unconventional phases in certain parameter regimes. The close competition and cooperation between different symmetry breaking propensities often omits a direct intuitive understanding of the phase diagram.
Our calculations utilized the functional renormalization group, an iterative approach to the prediction of quantum mechanical effects in many-electron systems. With this numerical method, it is possible to successively incorporate higher order screening processes for the electron-electron repulsion provided by the crystallographic environment in a controlled way and monitor the emergence of different ordering tendencies when approaching lower energy scales. Within each step of the iterative procedure, electrons of a distinct energy scale are included resulting in feasible calculation times.
This profile promotes the functional renormalization group as a ideal tool to provide accurate numerical estimates for the parameter dependence of the emergent low temperature phases and simultaneously a detailed understanding of the physical origin governing the encountered phase transitions. This allowed us to uncover guiding principles for the emergence of long sought after pair density waves in the Kagome materials, a new route toward correlated topological physics in graphene and establish a direct connection between spin-orbit coupling and geometrical frustration governing the phase diagram of the triangular lattice Hubbard model.
ALFQMCsim: Emergent and critical phenomena in correlated electron systems: Quantum Monte Carlo simulations (10/2022–12/2024; LARGE SCALE)
Correlation effects in the quantum mechanical many body problem are fascinating since there seems to be an unlimited richness in emergent phenomena. The common feature of all the subjects presented in this grant proposal is a model Hamiltonian that describes the physics at high energy. The aim is to understand the emergent low-lying and critical phenomena. Generically since the many body quantum mechanical problem does not allow for an exact analytical solution, there is no way to unambiguously elucidate the nature of the low lying excitations. Hence numerical simulations. The numerical simulations that we will carry out here are exact: for a given dimension of the Hilbert space (i.e. volume V ) and inverse temperature, β, we can carry out an unbiased calculation. However, since critical and emergent phenomena is very often a property of the thermodynamic limit, the bigger the system size the more conclusive and impactful our results. This essentially explains why we are dependent on supercomputing resources. In fact without access to supercomputing resources, we would not be able to address the questions posed in this proposal.
Our tool is a general implementation of the so called auxiliary field quantum Monte Carlo algorithm. This approach is triggered at solving systems of correlated electrons that couple to bosonic modes such as lattice vibrations. For a class of models that do not suffer from the so called negative sign problem, this approach allows us to compute properties of systems in thermodynamic equilibrium at polynomial cost. Generically, for short ranged interactions, the computation time scales as βV3.
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: Julius-Maximilians-Universität Würzburg
Target system: parallel computer Fritz
Intermediate Report (10/2023)
The ALF library implementation of the auxiliary field quantum Monte Carlo method allows for simulations of lattice fermions. Using this approach we have investigated emergent and topological phenomena in quantum matter. Highlights include a novel formulation for the simulation of the canonical Su-Schrieffer-Heeger model of lattice vibrations coupled to electrons on a square lattice. This model shows emergent Dirac fermions as well as Landau forbidden continuous transitions between different broken symmetry states. We have equally pursued our work in the domain of magnetic impurities in metals. By adopting a setup where there is a dimensional mismatch between the magnetic impurities and conduction electrons we can realize metallic magnetic phase transitions and analyze the critical behavior.
Optics, Quantum Optics and Physics of Atoms, Molecules and Plasmas
DaREXA-F: Datenreduktion für Exascale-Anwendungen in der Fusionsforschung (06/2023–11/2025)
For the efficient usage of HPC applications in the exascale era, we need to improve the scalability on large and heterogeneous systems. This requires a variety of components, such as efficient processing, data storage, software, and algorithms. The goal of this BMBF-funded project is to develop new methods for reducing data traffic between compute nodes with distributed memory and storage in file systems on supercomputers. For this purpose, a co-design approach will be used to develop solutions for variable-precision computation, data compression and novel data formats. These solutions will be used to improve GENE, a program used worldwide for the simulation of plasma turbulence, and will be validated using GENE.
Scientific field: 308-01 Optics, Quantum Optics, Atoms, Molecules, Plasmas
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz & GPGPU cluster Alex
Particles, Nuclei and Fields
ETH: SU(2) real time evolution on a classical computer (04/2023–11/2025)
We plan to study the thermalization of SU(2) gauge theory using techniques developed for quantum computing but on a classical computer, namely “Fritz”. More precisely we want to find out whether SU(2) (as prototype for all SU(N) gauge theories) fulfills the Eigenstate Thermalization Hypothesis (ETH). This is a highly relevant problem for many subfields of physics. It can be addressed on small lattice volumes and thus is one of the best candidates for an early demonstration of quantum supremacy. For this reason ingenious techniques have been developed (Natalie Klco, Duke University, is one of our collaborators) which allow us to check the validity of ETH for SU(2) on Fritz.
Scientific field: 309 Particles, Nuclei and Fields
University: Universität Regensburg
Target system: parallel computer Fritz
GAMMAML: Deep learning with the H.E.S.S., CTA, and SWGO gamma-ray telescopes (01/2024–12/2024)
The High Energy Stereoscopic System (H.E.S.S.) telescopes in Namibia observe the very-high-energy gamma-ray sky from 20 GeV to 10s of TeV. H.E.S.S. operates by observing the Cherenkov light emitted when such a gamma-ray creates a particle cascade in the atmosphere; the nature, origin, and energy of the incident particle can be determined using image processing techniques. State-of-the-art deep learning methods allow for high-sensitivity image analysis at high-speed, by exploiting patterns in the data which are currently not being exploited, making them an attractive analysis prospect for such gamma-ray telescopes (DFG FU 1093/13-1).
The future Cherenkov Telescope Array (CTA) and Southern Wide-Field Gamma-Ray Observatories (SWGO) will also greatly benefit from these new analysis techniques. With this project, we aim to develop a variety of new analysis methods for the gamma-ray observatories H.E.S.S., CTA, and SWGO to improve event classification, reconstruction, and simulation.
Scientific field: 309 Particles, Nuclei and Fields; 311 Astrophysics and Astronomy
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
CLS3pt: Hadron structure observables on lattices at low pion masses (10/2023–09/2024; LARGE SCALE)
In current and upcoming experiments (like the Electron Ion Collider) nucleon structure is extensively studied. Within the framework of lattice QuantumChromoDynamics (QCD) nucleon structure can be investigated from first principles utilizing large computational resources. Using this method we propose to compute (and improve on) the matrix elements related to Beyond-the-Standard-Model (BSM) interactions (which are not accessible experimentally) and (lower) moments of the parton distribution functions.
Scientific field: 309 Particles, Nuclei and Fields
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
DPDa: Quark Double Parton Distributions of the Nucleon (04/2023–06/2025; LARGE SCALE)
A better understanding of Double-Parton Distributions (DPDs) in the proton is vital to make full usage of the discovery potential of the LHC (CERN). We have pioneered corresponding lattice simulations with quite encouraging results, and we want to perform additional simulations with different lattice constants, which will allow for a continuum extrapolation. Without a reliable continuum extrapolation lattice results have, strictly speaking, no physical relevance.
Scientific field: 309-01 Nuclear and Elementary Particle Physics, Quantum Mechanics, Relativity, Fields
University: Universität Regensburg
Target system: parallel computer Fritz
Intermediate Report (03/2024)
The Large Hadron Collider at CERN, Geneva has so far failed to find “Beyond the Standard Model” physics (BSM). Therefore, its luminosity (basically its collision rate) is in the process of being much increased, to substantially enhance its sensitivity to BSM physics. As so far, one has not identified BSM physics, its effects, if they exist, must be very small. Consequently, it is of crucial importance to understand all normal, Standard Model, backgrounds to ever higher precision. In this context, Multi-Parton Interactions and in particular Double Parton Interactions are a major concern. These are processes in which two hard QCD reactions take place in the same proton-proton collision, such that (yet unknown) wave function effects become relevant. These are parameterized by Double-Parton Distributions (DPDs) and what we do is to calculate so-called invariant functions which in turn parameterize these DPDs. Because each quark can have 3 flavors and 3 spin orientations and one has two such quarks taking part in a double hard interaction there are many different DPDs and invariant functions, each of them being a function of several variables. Thus, DPDs contain very much information on the quark-gluon structure of a nucleon, of which one cannot predict which will become important for BSM searches but which makes them very interesting in their own right. We have pioneered the theoretical formalism which allows to calculate properties of DPDs using lattice QCD, which is the standard numerical method to calculate properties of specific hadrons from first principle Quantum Chromo Dynamics (QCD). As very little is known so far about DPDs, we focus for the time being on their general properties.
Fig.1 shows typical results for spin-independent invariant functions for the flavors down-down (dd), up-up (uu) and up-down (ud) combinations of quark flavors, as well as one interference DPD (yellow). Fig.2 shows the same when one of the quarks is spin polarized transversely to the proton momentum. The message of both figures is that we can determine all of these leading invariant functions of DPDs rather precisely. A proton has two valence up quarks but only one valence down quark, which explains why the up-up DPD is largest. Note that the interference DPD is not smaller in size than the down-down DPD. Fig.3 compares results for one specific invariant function for two lattices with different lattice constants. Its message is that the dependence on the lattice constant is very weak. Fig.4 compares the results for different quark masses.
The message here is that it is important to extrapolate the quark masses to their physical values what we will focus on next. In the long term the task would then shift to calculating the invariant functions of higher Mellin moments of DPDs, but this is not part of the present project.
addlight: The spectrum of charmonium and glueballs: adding the light hadrons (01/2023–04/2026; LARGE SCALE)
Experiments at particle accelerators have discovered new particles called XYZ states, which do not fit into the spectrum of conventional mesons made of a charm and an anti-charm quark (charmonium). Many of the new states are close to thresholds for strong decays. States made of gluons only, so called glueballs, are predicted by the theory of strong interactions (Quantum ChromoDynamics) but their existence still awaits an experimental confirmation. In this project we plan to study charmonium and glueballs by simulations of QCD on a lattice. The novelty of our study is the inclusion of light hadrons into which these states can decay. Our software has excellent scaling behavior on HPC systems.
Scientific field: 309-01 Particles, Nuclei and Fields
University: University of Wuppertal
Target system: parallel computer Fritz
Intermediate Report (04/2023)
The charmonium spectrum is of great importance to both particle physics and astrophysics, as it provides crucial information about the properties of the strong force and the nature of quarks. In particular, the hyperfine splitting in charmonium has long been a challenge for theoretical calculations, and accurate predictions are essential for interpreting experimental data and testing the standard model. By using cutting-edge techniques and state-of-the-art supercomputers, we aim to obtain results that are much more accurate than previous calculations and include effects that were so far neglected like mixing with light hadrons. Our results will provide new insights into the fundamental properties of the strong force and the nature of quarks, and will enable more accurate interpretation of experimental data from particle accelerators.
Statistical Physics, Soft Matter, Biological Physics, Nonlinear Dynamics
ArrhythmiaControl: Investigating termination mechanisms of chaotic spiral and scroll wave dynamics underlying cardiac arrhythmias by using hypothesis-driven and AI-driven termination approaches (02/2024–03/2027)
Scientific field: 310 Statistical Physics, Soft Matter, Biological Physics, Nonlinear Dynamics
University: Technische Hochschule Nürnberg Georg Simon Ohm
Target system: GPGPU cluster Alex
CLINT-M02: Multiscale modelling of SILP and SCILL catalysis (07/2023–06/2025)
Project M02 focuses on multi-scale simulations of Interface-enhanced SILP and Advanced SCILL catalysis of hydrogenation reactions, to allow for intelligent design and optimisation of these technological systems. To this end, we will combine hybrid quantum mechanics calculations (chemical transformations), and molecular dynamics simulations (structuring & diffusivity), each thoroughly validated against the wealth of experiments performed in CLINT, to tune the functionality of the catalytic complex, the IL, and the support material. The aim is to identify, characterise and enhance effects, occurring on a multitude of molecular time and length scales, which favourably affect the overall reaction’s turn over.
Structuring of the IL and gas/liquid and liquid/solid interfaces creates a unique dynamic environment for the catalyst. We will therefore focus on developing an understanding of the structural and transport properties of functionalised ILs and reactant molecules within the film using molecular dynamics simulations. This will permit us to boost the spatial coordination of the reactants with the solid and the gas interfaces, and address issues of solubility, viscosity as well as wetting. Building on this understanding of the interface environment, we will devise a hybrid quantum mechanics/molecular mechanics (QM/MM) approach to embed local quantum characterisation of the reactive centre into inhomogeneous IL films.
Scientific field: 310 Statistical Physics, Soft Matter, Biological Physics, Nonlinear Dynamics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
CRC1411D04: Design of Particulate Products: Modelling particle aggregation and assembly into optimal structures (03/2023–06/2024)
Robust self-assembly of nanoparticles can create nanostructures that have applications as structural color pigments and scaffolds for heterogeneous catalysis. This project develops and applies new methods to study not only the self-assembly process itself but also the properties of the self-assembled nanostructures. The computations in this proposal closely interact with experimental work that is conducted in the framework of the Collaborative Research Centre Design of Particulate Products.
Scientific field: 310 Statistical Physics, Soft Matter, Biological Physics, Nonlinear Dynamics; 403/01 Chemical and Thermal Process Engineering
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz & GPGPU cluster Alex
EnSimTurb: Towards ensemble simulations of fully developed turbulence (12/2022–04/2026)
The computing time allocation supports two ongoing third-party funded projects (ERC & DFG) in our group, in which we investigate the statistical properties of fully developed turbulence. In the ERC project, we aim at developing new theoretical and computational approaches to better understand and model fully developed turbulence. The goal of the DFG project is to capture the large-scale dynamics of turbulent flows. Both projects use our code TurTLE, a pseudo-spectral solver of the Navier-Stokes equations which also features particle tracking capabilities.
Scientific field: 310 Statistical Physics; 404-03 Fluid Mechanics
University: University of Bayreuth
Target system: parallel computer Fritz & GPGPU cluster Alex
Astrophysics and Astronomy
GAMMAML: Deep learning with the H.E.S.S., CTA, and SWGO gamma-ray telescopes (01/2024–12/2024)
CosmosTNG: galaxy formation and evolution with constrained cosmological magnetohydrodynamical simulations at cosmic noon (07/2023–06/2024; LARGE SCALE)
CosmosTNG is a first-of-its-kind constrained cosmological hydrodynamical galaxy formation simulation. It is designed to explore the formation and evolution of galaxies within the large-scale structure of the high-redshift Universe. An observed patch of the sky is first used to reconstruct the density field of the early Universe, which is then evolved with a state-of-the- art structure formation code using the IllustrisTNG galaxy formation model. The outcome enables (i) a direct comparison between simulations and observations in the COSMOS field, (ii) a new tool to explore galaxy formation in constrained environments, and (iii) a probe of the physics and distribution of gas in and around galaxies – from the circumgalactic to intergalactic medium – at cosmic noon.
Scientific field: 311 Astrophysics and Astronomy
University: Heidelberg University
Target system: parallel computer Fritz
ECOGAL_MW: Full disk modeling of the Milky Way (05/2023–04/2024)
Does a young planetary system carry the imprint of its birthplace in the diversity of our Galactic ecosystem?
How does a specific environment or location in the Milky Way determines the properties of our Sun and solar system?
ECOGAL will create the essential synergic framework where the extensive observational datasets available for our Galaxy can be analyzed back-to-back with state-of-the-art theoretical simulations of all Galactic environments, also using innovative machine-learning and decision-making tools.
ECOGAL has the ambitious goal to build a unifying predictive model of the Milky Way ecosystem able to trace the properties of planet-forming disks back to the environmental conditions active in their birthplace.
This project interweaves the unique expertise of four European research groups at the forefront of science in the four pillars of astronomy and astrophysics: observations, theoretical simulations, instrumentation and data analysis.
Scientific field: 311 Astrophysics and Astronomy
University: Universität Heidelberg
Target system: parallel computer Fritz
Transient3body: Transient formation in three-body encounters between stars and black holes (03/2023–05/2024)
Many astrophysical environments, from stellar clusters to disks of active galactic nuclei, are characterized by frequent interactions between stars and stellar-mass compact objects. In particular, encounters involving three objects (three-body encounters) are frequent in dense stellar environments. Previous work on such encounters has mostly focused on understanding the orbit evolution and outcome assuming the encountering objects are point particles. However, the point-particle approximation cannot be used to study observable phenomena produced in those interactions where the size of the meeting objects and gas interactions between them become an essential role in generating electromagnetic radiation. Such phenomena include tidal disruption events of stars by black holes, which can have a significant impact on the black hole mass growth and the subsequent long-term interactions with the surrounding debris. To investigate the impact and outcomes of dynamical interactions we perform hydrodynamics simulations of various types of three-body encounters using the moving-mesh magnetohydrodynamics code AREPO.
Scientific field: 311 Astrophysics and Astronomy
University: Ludwig Maximilian University of Munich
Target system: parallel computer Fritz
Mathematics
PDExa-CFD: Efficient simulation of fluid flow with high-order finite element methods (02/2024–09/2025)
The goal of the project is to perform scale-resolving simulations of turbulent flow in wall-bounded geometries to gain new understanding of the flow behavior near boundaries and support future method development. The project PDExa-CFD uses high-order discontinuous Galerkin finite element discretizations in space and splitting implicit/explicit methods in time to discretize the incompressible Navier-Stokes equations. As test case, direct numerical simulation of the flow over a periodic hill will be computed at a Reynolds number Re=19,000 as well as limit studies towards even higher Reynolds numbers.
Scientific field: 312 Mathematik; 404-3 Fluid Mechanics
University: Ruhr University Bochum
Target system: parallel computer Fritz & GPGPU cluster Alex
Bio-FROSch: Modeling and simulation of pharmaco-mechanical FSI for an enhanced treatment of cardiovascular diseases and non-Newtonian micro-macro blood flow simulations (07/2023–06/2024)
This project delivers the necessary computing time for two different research projects, both concerned with modelling and simulation of different aspects of arteries and blood flow using the same software framework.
In the first project, we aim at the robust computational modeling of coupling fluid-structure interaction with pharmacological effects towards an enhanced treatment of cardiovascular diseases. In the future, this could lead to a virtual laboratory to assist medical doctors and patients in their decisions. The main goal is to develop a robust numerical framework including suitable models for the computational simulation of the effects of drugs on the complex bio-chemo-mechanical processes in arterial walls.
In the second project, we aim to develop an efficient multiscale model for blood flow simulation. The main focus is on implementing a computational fluid model combining two different scales: a macroscopic scale which models blood as a continuous fluid, and a mesoscopic/microscopic scale capturing the behavior of red blood cells (RBCs) within the flow. We aim to improve computational efficiency of the multiscale model by identifying and implementing areas where advanced machine learning techniques can be applied.
The development of robust numerical coupling schemes and solver strategies and the parallel implementation and integration thereof into the software libraries FEDDLib, FROSch, and Trilinos is essential for both projects.
Scientific field: 402 Mechanik; 312 Mathematik
University: Universität zu Köln
Target system: parallel computer Fritz & GPGPU cluster Alex
StroemungsRaum: Neuartige Exascale-Architekturen mit heterogenen Hardwarekomponenten für Strömungssimulationen (03/2023–09/2025)
Future Exascale computing architectures will feature a high number of heterogeneous hardware components that are comprised of special-purpose processors or accelerators. The corresponding realization of Computational Fluid Dynamics (CFD) application software as a central core component of nowaday’s CFD simulations in industrial context requires high-scaling methods. These methods solve high-dimensional and unsteady (non)linear systems of equations and need to be integrated into application software such that they can be used by non-HPC experts from the industry. Especially the open-source software FEATFLOW , is a central component of the StroemungsRaum platform that is successfully used by the industrial partner of the project IANUS for years. In the context of the whole project, FEATFLOW will be extended methodologically and by parallel near-hardware implementations.
Scientific field: 404-03 Fluid Mechanics; 312 Mathematics
University: TU Dortmund
Target system: parallel computer Fritz & GPGPU cluster Alex
Atmospheric Science, Oceanography and Climate Research
ATMOS: Numerical atmospheric modeling for the attribution of climate change and for model improvement (08/2022–03/2025)
The project “Detection and Attribution of climate change for the glaciers on Kilimanjaro: Targeting the processes at regional and local scales” is attempting to break down the concept of Detection & Attribution to the local scale, by using the case study of glacier change on Kilimanjaro. This is because (a) the long research record for this place and (b) the unique climate indicator potential of these glaciers on a freestanding peak in the tropical mid troposphere. Results will reveal how human influence on the climate is transferred to the site-specific level.
The project “Exploring the potential of coralline algae as climate proxy and for climate model evaluation: a Southern Hemisphere case study of New Zealand” aims to explore a novel climatic indicator, namely crustose coralline algae that grow in shallow ocean waters (CCA), for the purpose of improved global climate model (GCM) evaluation. In particular, we will target the improvement potential with regard to sea surface temperatures in the Southern Ocean and the effect of these ocean conditions on the regional high-mountain climate of New Zealand.
Scientific field: 313-01 Atmospheric Science
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
Intermediate Report (03/2024)
In the first 1.5 years, several successful simulations were conducted in the project ATMOS. These were mostly based on the numerical atmospheric model WRF, a well-known community model. The main progress has comprised (i) the development of new modules for WRF to implement important (yet thus far, neglected) processes, especially with regard to cold climates; (ii) the creation of a new regional climate data set for New Zealand over the period 2005-2020 with a focus on the Southern Alps; and (iii) insights into cloud processes that are tied to radiation effects and the isotopic composition of water.
Physical Chemistry
SpectroscopicProperties: Spectroscopic properties of molecules with unusual electronic structures (07/2022–12/2025)
Theoretical Chemistry
ELTRANS: Electron transfer in organic and inorganic light-converting systems (10/2022–09/2026)
The project aids the design of new electrodes for photoelectrochemical water oxidation by studying the underlying electronic processes. It collaborates with experimental partners in Sonderforschungsbereich 1585. The electronic coupling between different molecules and nano-particles, as well as the charge transport through the particles and the interfaces will be studied using the real-time propagation approach to time-dependent density functional theory. By simulating electron dynamics in real-time and on a real-space grid, the project aims at unravelling charge-transport pathways in complex, interface-governed materials. Effort will also be devoted to use and develop exchange-correlation approximations that combine computational efficiency with a reasonable accuracy for long-range charge-transfer.
Scientific field: 327-01 Electronic Structure, Dynamics, Simulation
University: Universität Bayreuth
Target system: parallel computer Fritz
Engineering Sciences
Mechanics and Constructive Mechanical Engineering
Bio-FROSch: Modeling and simulation of pharmaco-mechanical FSI for an enhanced treatment of cardiovascular diseases and non-Newtonian micro-macro blood flow simulations (07/2023–06/2024)
StroemungsRaum: Neuartige Exascale-Architekturen mit heterogenen Hardwarekomponenten für Strömungssimulationen (03/2023–09/2025)
FLINSENOI: Flow induce self-noise (01/2023–06/2024; LARGE SCALE)
In this project, numerical studies are carried out to investigate the influence of surface modifications on the near-wall turbulence of external and internal wall-bounded flows. The surface modifications consist of streamwise aligned grooves in combination with regions of strong transverse curvature between the grooves. The aim of the surface modifications is to reduce the near-wall turbulent fluctuations that can lead to flow-induced vibrations of the adjacent structure and/or to flow-induced sound which is undesired in many applications. Recent investigations show that such surface modifications could lead to a 20 % decrease in the average wall shear stress and result in local reductions in the turbulent intensities, Reynolds stress, the temporal velocity spectrum, and the turbulent dissipation rate. The analysis within the anisotropy-invariant space revealed a local tendency towards flow relaminarization. Improvements of the surface modifications should be investigated in a numerical study using the finite volume code OpenFOAM. The size of the improved surface modifications are in the order of the viscous sublayer. Transient and scale-resolving simulations are needed to extraxt the quantities of interest, such as turbulent dissipation, anisotropy of turbulence, Reynolds stress, flow-acoustic source terms, velocity and pressure fluctuations as well as temporal spectra. Therefore a high-resolution mesh in combination with multi-node parallelized simulations are required.
Scientific field: 404-03 Fluid Mechanics; 402-04 Acoustics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
FRASCAL-FE: Computational continuum mechanics simulations at LTM for FRASCAL (11/2022–12/2026)
FRASCAL improves understanding of fracture in complex heterogeneous materials by developing simulation methods able to capture the multiscale nature of failure. With i) its rooting in different scientific disciplines, ii) its focus on the influence of heterogeneities on fracture at different length and time scales as well as iii) its integration of highly specialized approaches into a “holistic” concept, FRASCAL addresses a truly challenging cross-sectional topic in mechanics of materials.
Specifically, project P8 (“Fracture in Polymer Composites: Meso to Macro”) aims to study the influence of different mesoscopic parameters, including micro-structure morphology, on the macroscopic fracture properties of nano-particle reinforced polymers.
Scientific field: 402-2 Mechanics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
Intermediate Report (11/2023)
Modelling and Computation of Fracture in Silica-Polymer Composites: From Meso to Macro Scale
Motivation: The mechanical properties and the fracture toughness of polymers can be enhanced by adding silica particles or other fillers. These nano- or micro-particles increase the elastic and especially the fracture properties of the material. However, the elastic properties exhibit a size-effect that cannot easily be modelled with classical continuum mechanics and homogenization approaches. The experimentally observed increase in the yield strength and the fracture properties is due to different mechanisms. Complex crack paths develop on the mesoscopic scale, accompanied by localized shear bands, particle debonding, and void growth, which are strongly influenced by the distribution of the particles. Modelling and simulation of the mesoscopic fracture behaviour are thus essential for the determination of macroscopic fracture properties.
Vision and Objectives: The aim of the project is to study the influence of silica particles and other fillers on the elastic and fracture properties of polymer composites. Therefore, a multiscale simulation approach will be developed to derive the macroscopic properties of the composite, depending on the underlying mesostructures. A numerical homogenization approach to determine the elastic properties, including size effects, was developed and will be extended to determine the macroscopic fracture toughness of polymer composites based on mesoscopic crack simulations.
State of the Art: The mechanical properties and the fracture toughness of polymer composites depend strongly on their underlying mesostructure. In particle-reinforced thermosets, different mechanisms are responsible for the observed increase in fracture toughness and yield stress, e.g. localized shear bands, debonding of particles, and subsequent void growth in the matrix material. The increases in strength and fracture energy have been shown to depend on various mesoscopic characteristics, including the volume fraction and the geometric arrangement of the particles. Simulation of the fracture behaviour on the mesoscale requires a flexible method that can represent crack propagation in heterogeneous materials, the emergence of new cracks, and crack bifurcations. The phase-field method for fracture offers these possibilities and can be coupled with regularized cohesive models to represent debonding. Multiscale approaches for fracture mainly use the variational multiscale method or related approaches, which will also be applied in the present project, or they are based on homogenization methods.
Work plan: An extended homogenization method to account for size effects in polymer composites has been successfully developed. This method includes interfaces or interphases of a finite thickness with their own mechanical properties around the particles on the mesoscopic scale. We showed that for materials with stiffer particles only the interphase-approach yields the desired size effect. The phase-field approach for fracture has been extended by means of different spatial adaptivity methods. Error indicators based on phase-field threshold values, the gradient of the phase-field, and configurational forces are introduced and compared in terms of accuracy and efficiency. Next, in the project, a multiscale simulation approach for fracture of polymer composites (epoxy with silica particles and cellulose-polymer composites) will be developed and applied considering the mentioned research results.
AkuRad: Fluid-Structure-Acoustic Interaction of Enclosed Radial Fans (09/2022–08/2025)
The research project comprises the acoustics of a radial fan operated in a spiral housing. Radial fans have a wide range of applications in a wide variety of fluid engineering applications, in which the requirements for sound radiation are becoming increasingly dominant.
The goal of the present application is to investigate the multiphysical interrelationships of flow-related sound radiation of radial fans in volute casings using a combined, experimental, simulation-based approach. It is important to treat the excitation paths for the casing structure separately: There are the fluid dynamic pressure fluctuations as well as the flow induced acoustic pressure field within the casing.
The numerical simulation methods are verified and validated against experimentally obtained data. Subsequently, developed methods are used to investigate the sound generation in the fan impeller and the resulting sound radiation through the spiral housing. This is done by means of variant studies. The variants are formed by different inflow conditions to the radial impeller, which influence the flow-induced sound generation in the impeller and thus also the casing excitation and sound radiation into the far field.
The innovation of the research work results from the fact that for the first time it is possible to develop a simulation method for enclosed rotating systems, which maps the entire chain of fluid-structure-acoustic interaction. In a complementary approach with experimental investigations, a conscious separation of different mechanisms in sound generation and their complex interactions results in knowledge with a high general validity that is not yet available in literature for enclosed radial fans. The findings allow conclusions to be drawn as to which part of the machine (impeller flow, casing structure, etc.) changes the radiated sound with which influencing parameters. In the development process, this results in a specific treatment of the acoustic problem areas.
The basic-oriented investigations have a high general validity and form the basis of future design rules for noise-reduced radial fans. The coupled simulation method provides a calculation tool that allows the flow- induced sound generation and its propagation in radial impellers to be analyzed in the frequency and time domain.
Future work will focus on the use of optimization algorithms for noise reduction, which consider not only the inflow conditions to the impeller and their interaction with the leading edge of the blade, but also the material parameters and the shape of the casing structure.
Scientific field: 404-03 Fluid Mechanics; 404-04 Hydraulic and Turbo Engines and Piston Engines; 402-04 Acoustics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
FRASCAL-MD: Particle-based computing at LTM for FRASCAL (09/2022–09/2024)
FRASCAL improves understanding of fracture in brittle heterogeneous materials by developing simulation methods able to capture the multiscale nature of failure. With i) its rooting in different scientific disciplines, ii) its focus on the influence of heterogeneities on fracture at different length and time scales as well as iii) its integration of highly specialized approaches into a “holistic” concept, FRASCAL addresses a truly challenging cross-sectional topic in mechanics of materials.
In particular, sub-project P6 (“Fracture in Thermoplastic Polymers: Discrete-to-Continuum Coupling”) provides a link between the level of atoms and the continuum with specific interest in the multiscale modelling and simulation of polymer fracture.
Scientific field: 402 Mechanics and Constructive Mechanical Engineering
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz & GPGPU cluster Alex
Intermediate Report (09/2023)
The first focus is on further development and enhancement of the domain-decomposition multiscale “Capriccio” method, which couples a particle-based domain treated by molecular dynamics (MD) with a continuum solved by the finite element (FE) method. In contrast to classical MD procedures, the particle- based domain is subjected to non-periodic boundary conditions which enables application of non-affine deformations typically not feasible in MD. In this regard, extensive parameter studies have been performed to reveal an optimal setting of the coupling parameters both for the equilibration and deformation phases of the numerical samples. Specific attention has been paid to the performance of the particle-based domain under the non-periodic boundary conditions. These studies during equilibration have been recently published. Further activities to gain deeper understanding of the coupling and the effect of parameters on the results compare the fully coupled three-dimensional system to a simplified one-dimensional set-up. By these comparative studies, the relevant factors of influence can be separately investigated and strategies to reduce deviations introduced by the coupling can be developed. In this context, one project thesis has been completed and a master thesis is currently in progress. Depending on the findings of the latter one, one or two manuscripts will result from these studies. Additional, preliminary work to extend the Capriccio method to thermal coupling has been done by thoroughly investigating the time-temperature superposition in amorphous thermoplastics at large strains by means of pure MD simulations. An associated manuscript is currently under review. In parallel, the original version of the Capriccio code has been carefully revised, which resulted in a significant increase of efficiency. In summary, these activities shall eventually result in a user-oriented user-oriented guideline for setting up coupled simulations together with an appropriate choice of parameters, which is an essential precondition to make the Capriccio method publicly available in an open repository.
The second focus is on fracture simulations of thermoplastic polymers and comprises several sub-projects. In the first one, numerical studies have been performed to specify appropriate criteria for bond rupture in thermoplastic polymer systems. Due to their complexity, these investigations are still ongoing. The second sub-project concentrates on double-notched numerical specimens inspired by classical fracture mechanical set-ups. Here, we were able to obtain trends of mechanical properties of polymer nanocomposites similar to those known from experiments. The related manuscript is currently in preparation. Furthermore, fracture simulations using the Capriccio method are currently prepared. Here, only the vicinity of the crack tip is resolved at molecular level, whereas the remaining part is treated as a continuum. By this set-up, flexible application of boundary conditions will become possible in order to run simulations mimicking typical set-ups used in fracture mechanical applications. In addition, we performed studies to integrate domain adaptivity into the Capriccio method, such that finely resolved MD domains can move throughout the entire domain and have to be used only on parts of high loads.
Our third focus is on accelerating the coupled MD-FE simulations by choosing generic polymer models at molecular level. Such generic models may use simplified formulations for the interactions between molecules and atoms, but still cover the main characteristics of polymer materials. By our studies on the HPC cluster, we were able to reveal trends observed in experiments when investigating the influence of the average molar mass, the dispersity, the filler particle size as well as the filler particle content on the mechanical properties of thermoplastics and their nanocomposites. In this context, two manuscripts are currently in preparation.
Process Engineering, Technical Chemistry
CRC1411D04: Design of Particulate Products: Modelling particle aggregation and assembly into optimal structures (03/2023–06/2024)
Fluid Mechanics, Technical Thermodynamics and Thermal Energy Engineering
PDExa-CFD: Efficient simulation of fluid flow with high-order finite element methods (02/2024–09/2025)
FLINSENOI: Flow induce self-noise (01/2023–06/2024; LARGE SCALE)
EnSimTurb: Towards ensemble simulations of fully developed turbulence (12/2022–04/2026)
AkuRad: Fluid-Structure-Acoustic Interaction of Enclosed Radial Fans (09/2022–08/2025)
AOTTP-DFG18-1: Characterization of molecular diffusion in electrolyte systems (08/2022–07/2024)
Electrolytes conduct electric current by the movement of ions while blocking the free movement of electrons. For applications of electrolytes as working fluids, the transport of ions is important and can be subdivided into diffusion, convection, and, in the presence of an electric field, migration. The diffusion process has shown to be of special interest since it is a limiting factor for the performance of applications where large concentration gradients can build up when electro-active ions react at electrode surfaces. Therefore, exact knowledge of molecular diffusion coefficients is required for the design and optimization of efficient processes involving electrolytes.
Scientific field: 404-01 Energy Process Engineering; 404-02 Technical Thermodynamics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz & GPGPU cluster Alex
Materials Science
ASiMENiS: Atomistic simulation of microstructure elements in Ni-based superalloys (01/2023–04/2024)
Single-crystal Ni-based superalloys play a key role for the efficiency of gas turbines for aviation and power generation. The optimization of these high-performance structural materials requires an understanding of the influence of the elements of the microstructure on the mechanical properties. The goal of this project is to understand dislocation mobility and the formation of local defect phases in the microstructure. The atomistic simulations involve high-throughput ab-initio calculations and large-scale molecular-dynamic simulations with machine-learning potentials based on the atomic cluster expansion.
Scientific field: 406 Materials Science
University: Ruhr-Universität Bochum
Target system: parallel computer Fritz & GPGPU cluster Alex
Materials4.0 - AITDB: Ab initio thermodynamic database development (10/2022–01/2025)
Phase diagrams have revolutionized materials development by providing the conditions for phase stabilities and transformations, and thereby a thorough thermodynamic understanding of materials design. However, the majority of today’s phase diagrams are based on scarce experimental input and often rely on daring extrapolations. Every multicomponent phase diagram relies on a fragile set of phase stabilities as very recent studies show.
Within this project, materials design is raised to the next level by providing a highly accurate first principles thermodynamic database. First principles, alias ab initio, approaches do not require any experimental input and can operate where no experiment is able to reach. However, they have been limited to zero-Kelvin or low-temperature approximations which are not representative of phase diagrams.
Utilizing our unique expertise in high-accuracy finite-temperature ab initio simulations, we have developed an extremely efficient procedure including novel methods accelerated by machine learning potentials. This procedure facilitates a highly efficient determination of Gibbs free energies, thereby allowing convergence also of properties depending on the second-order derivatives, such as the heat capacity and thermal expansion.
With these tools, we have now reached a stage where a large ab initio thermodynamic database will be computed for elements across the periodic table. The main focus will be on phase stabilities of various phases, including dynamically unstable ones. Having acquired such a database, the phase stabilities can be put into practice by re-parametrizing binary phase diagrams and studying the implications on multicomponent phase diagrams.
Scientific field: 406 Materials Science
University: Universität Stuttgart
Target system: parallel computer Fritz
Intermediate Report (01/2024)
1) Thermodynamic properties on the homologous temperature scale from direct upsampling: Understanding electron-vibration coupling and thermal vacancies in bcc refractory metals
In this work, we present thermodynamic properties (heat capacity, thermal expansion, bulk modulus; see example figure below) of four bcc refractory elements (V, Ta, Mo and W) up to melting point to full density-functional-theory (DFT) accuracy (including anharmonic interactions) within the PBE exchange-correlation functional. The results are also analyzed on a self-consistently determined homologous temperature scale, and we get remarkable agreement to experimental data from literature. Such an accuracy to high-temperature experimental data has never been achieved before for refractory systems, and it exemplifies the strength of the recently proposed methodology (direct upsampling) and the possibilities for developing ab initio thermodynamic databases.
Additionally, we have calculated vacancy formation Gibbs energies, and included the effect of vacancies on bulk thermodynamic properties. We also formulate and individually calculate, for the first time, the coupling between thermal vibrations and electronic excitations—a crucial concept that refines our understanding of free energy contributions.
2) Machine-learning-accelerated ab initio simulations of dynamically stabilized phases
Imagine a top that becomes stabilized in an up-right position by spinning on a table. Similarly, in materials physics, certain phases have their atoms vibrating around the static lattice positions to achieve a stabilization at elevated temperatures (see the figure below). In this study, we develop an advanced method accelerated by machine-learning potentials to compute thermal properties of such dynamically stabilized phases with quantum-mechanical accuracy. We apply it to the bcc phase in the prototype systems Ti, Zr, and Hf, with a particular focus on the solid phase transition from hcp to bcc. Our results resolve a long-standing discrepancy at 0 K between experimental and theoretical properties, providing valuable insights for both experimental studies and engineering applications.
3) Performance and limits of finite temperature DFT for SiO2
The direct upsampling approach is able to accurately predict the thermodynamic properties of materials at the full accuracy level of density-functional theory (DFT), which includes contributions of higher order than the harmonic approximation. Our approach has been applied to face-centered cubic (fcc), body-centered cubic (bcc), and hexagonal closed-pack (hcp) systems so far. The investigated materials include refractories, as well as dynamically unstable systems. However, this approach has been applied only to unary systems. Here we apply it also to the binary SiO2 system, specifically to the crystalline polymorphs. The DFT level accuracy is achieved for a series of exchange-correlation functionals. In particular, the flat energy-volume curve, which correlates with the low bulk modulus, makes the description of the thermodynamic properties very challenging. The theoretical calculation of the temperature dependent volume expansion agrees well with the experiment, but only when an appropriate exchange-correlation functional is used. The phase stabilities in comparison with CALPHAD show the results of the current exchange correlation functionals, their predictions and limitations, and call for further development of the functionals for finite temperature free-energy calculations.
Ion-catch: Molecular Modelling based design of ligand shells to functionalize magnetic nanoparticles for the removal of heavy metal pollutants from water (07/2022–05/2025)
Systems Engineering
Voxray: Quantenrekonstruktion für industrielle Computertomographie (07/2023–06/2024)
Innerhalb der zerstörungsfreien Werkstoffprüfung ist die Computertomographie (CT) der Goldstandard, um Bauteile schnell und präzise zu vermessen und auf Fehler zu analysieren. Ein typischer Inspektionsablauf ist in der folgenden Abbildung dargestellt. Die bislang etablierte Rekonstruktions-Software kämpft regelmäßig mit Problemen, bei denen das 3D- Modell nicht mit dem realen Objekt übereinstimmt. Diese treten besonders bei Verbundmaterialien aus Kunststoff und Metall auf und machen die Qualitätssicherung schwierig und zeitaufwendig. Zugleich scheitert bestehende Software an Spezialverfahren wie Roboter-CT oder der sogenannten Laminographie, die beispielsweise in der immer wichtiger werdenden Batteriezellenentwicklung angewandt wird. Wir entwickeln aktuell die neuartige Quanten-Rekonstruktionstechnik (QRT), die den Rekonstruktionsschritt in der Industrie-CT revolutioniert. Mit dieser neuen Technik wird die Ausbreitung der einzelnen Röntgenphotonen (der Quanten) simuliert und anschließend invertiert. Mithilfe eines iterativen Optimierungsalgorithmus wird so eine Rekonstruktion des Objektes erstellt.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation; 407-02 Measurement Systems
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Electrical Engineering and Information Technology
RTSE: Enhancing Speech Communication Using Real-Time Target Speaker Extraction (12/2023–11/2024)
In today’s fast-paced world, clear and effective communication is crucial, especially in real-time scenarios like teleconferences, online classes, and virtual assistance. However, this communication is often hindered by the presence of multiple speakers and background noises, which can lead to miscommunication and frustration. Our project addresses this challenge by focusing on the extraction of a target speaker’s voice from complex audio environments in real time. Leveraging the power of deep neural networks, our approach uses an enrolment recording of the target speaker. This enrolment recording serves as a reference, enabling our system to effectively identify and isolate the target speaker’s voice from a mixture of sounds.
The core of our research lies in developing an advanced algorithm that not only distinguishes the target speaker from others but also actively suppresses background noise, resulting in a clearer and more intelligible audio output. This technology holds immense potential for improving the quality and efficiency of various real-time communication applications, from virtual meetings to voice-controlled interfaces. By providing a solution that allows for clear transmission of a specific speaker’s voice, our project aims to facilitate smoother, more effective communication in diverse and noisy environments, making it an essential tool for both professional and personal settings. This research not only advances the technical field of speech and acoustic signal processing but also offers tangible benefits for everyday communication.
Scientific field: 408 Electrical Engineering and Information Technology
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
MMSP: Multi-Microphone Speaker Separation (07/2023–06/2024)
In a scenario with multiple active sound sources, noise, and reverberation, the task of extracting and separating sources is known to be challenging. Using several spatially distributed microphones, the task can be formulated as a multiple-input multiple-output problem. Our research focuses on separating speech sources in predefined spatial regions, for example, related to the seats in a car or the seats in a conference room. We employ deep neural networks for the separation process.
Scientific field: 408 Electrical Engineering and Information Technology
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
Computer Science
SysGART: Systematic Generalization for Abstract Reasoning Tasks (03/2024–02/2025; NHR STARTER)
Abstract reasoning is a fundamental aspect of human intelligence. It allows us to solve new, unfamiliar problems without the need for prior knowledge (Gilead et al. 2014). Developing machines that can reason this way (an ability closely related to the concept of fluid intelligence (Cattel 1963)), has been a long-standing goal in the field of artificial intelligence (Barrett et al. 2018). However, most modern AI systems (including large language models) do still lack general abstract reasoning capabilities (Marcus 2018, Gendron et al. 2023). In 2019, François Chollet introduced the Abstraction and Reasoning Corpus (ARC), a collection of visual puzzles designed to evaluate human-like general fluid intelligence in AI systems (Chollet 2019). While humans can robustly solve around 80% of ARC tasks, current algorithms markedly underperform with a maximum success rate of 31% (Johnson et al. 2021). This disparity underscores the necessity for continued development in the field of automated abstract reasoning. Various challenges have been created to motivate developers to solve ARC. This project intends to tackle the ARC benchmark by leveraging insights from recent findings on systematic generalization. For this, we will exploit the compositional nature of ARC tasks. The aim is to develop an algorithm capable of systematic generalization, extrapolating from specific examples to novel test scenarios. This approach aims not only to enhance the performance of AI systems on the ARC benchmark, but also to contribute to the broader understanding of abstract reasoning in artificial intelligence.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Ludwig Maximilians Universität München
Target system: GPGPU cluster Alex
GMFM: Generative Multimodal Foundation Model Pretraining and Finetuning (01/2024–06/2024; NHR STARTER)
Multimodal models have demonstrated remarkable capabilities in various natural language understanding and computer vision tasks. Previous approaches typically employed separate vision encoders to extract image features and text encoders for multimodal understanding and reasoning. In this project, we propose a novel approach that utilizes a decoder-only architecture for direct alignment of pixel-level images and text in a generative framework. Unlike previous works that relied on vision encoders, our method streamlines the process, improving efficiency and effectiveness in multimodal tasks, without incorporating additional large vision encoders.
Our research focuses on developing a state-of-the-art multimodal foundation model by leveraging a decoder-only architecture. This approach allows us to bridge the gap between textual and visual information more directly, eliminating the need for separate vision encoders. By aligning text and images at the pixel level, we aim to enhance the model’s ability to understand and generate multimodal content.
The proposed model will be pretrained on a large-scale multimodal dataset, capturing rich and diverse textual and visual information. We will investigate novel techniques for aligning image and text representations during pretraining, enabling the model to learn meaningful cross-modal relationships. This will result in a powerful multimodal foundation that can be fine-tuned for various downstream tasks.
Our project’s primary goal is to demonstrate the superiority of our decoder-only architecture in handling multimodal data. We will conduct extensive experiments and evaluations on a range of tasks, such as image captioning, visual question answering, and multimodal sentiment analysis, to showcase the model’s improved performance over existing methods.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Ludwig Maximilians Universität München
Target system: GPGPU cluster Alex
SFB 1540: Exploring Brain mechanics (EBM) Understanding, engineering, and exploiting mechanical properties and signals in central nervous system development, physiology and pathology (01/2024–12/2024)
DeepVehicleVision: Simulating Sensors for Vehicle Vision in Virtual Environments (01/2024–12/2024)
Developing autonomous driving and advanced driver-assistance systems is getting more important for the automotive industry. In addition to developing these systems, it is also necessary to guarantee their safety, which may require covering billions of driven kilometers. Virtual testing environments can pose solutions if they reach the necessary realism. This project tries to achieve this by training neural networks on real sensor data and transferring the trained systems to the virtual driving simulation.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
InvRadarSim: Inverse Radar Simulation and Rendering for Scene Parameter Reconstruction using Gradient Descent (01/2024–12/2026)
We aim to transfer concepts from inverse rendering to radar simulation and reconstruction, thereby bridging the gap between recent research from the fields of computer graphics and millimeter-wave multiple-input-multiple-output (MIMO) imaging radar. We leverage gradient descent based optimization to match the scene parameters of a radar simulation to observations of a real radar and derive shape or material properties to improve simulation and reconstruction fidelity. Due to the computationally extensive requirements of this technique, it is well suited for acceleration on the NHR.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
DiffSE: Diffusion models for Speech Enhancement (12/2023–12/2024)
Speech enhancement seeks to recover clean speech from audio recordings affected by different sources of distortions, such as environmental noise. Diffusion models, a subset of generative models, learn a prior distribution over clean speech data, capturing inherent spectral and temporal structures. This prior knowledge enables enhancing degraded speech into natural clean speech estimates. However, enhancing with diffusion models during inference is computationally expensive, therefore we aim at reducing the computational costs of diffusion models by reducing the architecture footprint and modifying the diffusion process itself.
Scientific field: 409 Computer Science
University: Universität Hamburg
Target system: GPGPU cluster Alex
Odeuropa Image Processing: Image Processing strand of the EU Horizon 2020 Odeuropa Project entailing the automated recognition of smell references in historical artworks (12/2023–11/2024)
The goal of this WP is to reconstruct the lost scents and odours of the past through their visual traces in art. By means of selected objects, motifs and iconographies, we will first analyse what meaning and design smell had in 17th–20th century works of art. At the same time, objects strongly defined by smell are to be automatically recognised via computer vision in large-scale European digital collections and the appearance and patterns of their use analysed. We thus combine a classical approach based on corpus formation and qualitative interpretation with a data-driven approach based on ‘distant viewing’. By doing so, we explore and develop new uses of mixed methods and computing in the fields of digital art history and computer vision. Finally, we build an interactive notebook-based demonstrator to show how images can be retrieved by means of their olfactory references.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
OLMAP: Adapting LLMs to Align with User Preferences (11/2023–10/2024; NHR STARTER)
We are planning to develop and evaluate approaches for fine-tuning large language models with the goal of combining different learning objectives i.e. preference learning and language modeling.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Technische Universität Dresden
Target system: GPGPU cluster Alex
TriFORCE: Learning adaptive reusable skills for intelligent autonomous agents (11/2023–10/2024)
This project will extend the current limits of Deep RL to improve its practicality in realistic high- dimensional problems. Our contributions will be a step forward in unleashing the full potential of Deep RL by making it more efficient and easier to use in real-world environments. We will achieve our goal by addressing the following fundamental problems of Deep RL: Computational and data requirements, human supervision and tuning, and task overfitting. Although agents trained with Deep RL perform excellently, they have a strong tendency to overfit the available data.
In this project, we will show that these seemingly different problems have similarities and common causes, and we will address them by introducing a new unifying view that will guide our research for more efficient and practical Deep RL methods that can be used in everyday applications.
We will demonstrate the critical importance of our methodological advances in learning locomotion skills in a leg-operated wheelchair. In this application, the wheelchair needs to move robustly and safely at different speeds, avoid obstacles and adapt to different terrains. In addition, the gait pattern should be adapted to the needs and preferences of the human user. We will demonstrate the importance of our algorithmic solutions in a realistic problem where multiple skills are required, such as navigating to a destination, walking up stairs and avoiding obstacles along the way.
The TriFORCE team is made up of outstanding scientists, an advisory board with renowned researchers, female leaders from academia and partners from industry.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Julius-Maximilians-Universität Würzburg
Target system: GPGPU cluster Alex
MIA-NORMAL: Medical Image Analysis with Normative Machine Learning (10/2023–09/2024; LARGE SCALE)
MAGNET4Cardiac7T: Physics-Informed Deep Learning Algorithms for Modeling Electro-Magnetic Fields in the Human Thorax (10/2023–09/2024)
A major obstacle to the widespread use of ultrahigh-field magnetic resonance imaging (UHF-MRT) is the complex distribution of electromagnetic waves in the patient’s thorax, which has a very negative impact on image quality and also carries the risk of undesirable overheating of the tissue. Conventional computational methods for calculating the 3D field distribution require several days and thus cannot be used on patients. Therefore, a new approach is developed in the present research project: Physics Informed Neural Networks (PINN) will be used to train neural networks based on Maxwell’s equations, which will allow to calculate the energy deposition in the body within a few minutes. Together with the German Center for Heart Failure, PINNs, which have so far only been described for two-dimensional applications, will be further developed for the three-dimensional case.
Scientific field: 409 Computer Science
University: Julius-Maximilians-Universität Würzburg
Target system: GPGPU cluster Alex
Voxray: Quantenrekonstruktion für industrielle Computertomographie (07/2023–06/2024)
XXL-CT-Segmentation: Instance segmentation of XXL-CT Volumes (06/2023–06/2024)
XXL-Computed Tomography (XXL-CT) is able to produce large scale volume datasets up to and above 10,0003 voxels which can relate up to many terabytes in file size and can contain multiple 10,000 of different entities of depicted objects with varying properties (shapes, densities, materials, compositions) which are hard to delineate. This project aims to develop deep learning based instance segmentation algorithms to delineate these entities for the data exploration process.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
EmpkinSA01: Multimodal Bodyshell Camera (05/2023–10/2025)
Sub-project A01 of EmpkinS explores a novel, multimodal sensor concept for high-precision, non-contact detection of the envelope of the human body and the velocity vector of each point on this envelope. For this purpose, a micro-doppler aperture synthesis radar is combined with an optical depth camera and the strengths of the sensors are combined. The body envelope and its movements are detected with outstanding precision and at a high measuring rate. With these properties, the system is a central basis for the research programme of the CRC EmpkinS. To be able to detect a human body and capture its pose in the radar and the depth camera, a large variety computer vision algorithms is required, ranging from simple Region-of-Interest detection of the human in an image, to segmentation, 3D reconstruction, and human pose estimation algorithms. In terms of quality, Deep Learning approaches have proven outstanding performance with respect to traditional optimization-based approaches. Therefore, this project aims at leveraging Deep Learning approaches, which operate on the previously-mentioned sensor data such as RGB ((depth-)camera), RGB-D (depth-camera) and radar data.
Scientific field: 409 Computer Science
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
digiOnko: digiOnko – Mit digitaler Medizin gegen Brustkrebs, AP5 - Histopathologie (12/2022–09/2024)
The digiOnko project (https://www.digionko-bayern.de/) aims at improving the screening, early detection, diagnosis, treatment and aftercare for women with breast cancer. One aspect of this project is the use of machine learning methods in digital pathology. Deep learning methods will be used to analyze digital Whole Slide Images (WSI), in particular to automatically calculate scores and improve treatment decisions.
Scientific field: 409-05 Materials Engineering
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
DeepPano: Erzeugung von Panoramabildern aus 3D-Laser-Punktewolken und Kamerabildern (12/2022–04/2024)
Today, it is possible to capture the 3D-Geometry of real-world objects and rooms in high quality by using modern sensors and algorithms for 3D reconstruction. These virtual models open the door for many applications, e.g., floor plans can be generated, and distances measured virtually. However, for virtual walkthroughs, the resulting models are usually not sufficient yet to achieve high-quality renderings from arbitrary views. With 360°-photographs (panoramas) it possible that a user explores a scene in photo- realistic quality, but only by looking around, and not moving freely. Image-based rendering is a well- researched approach to lift this restriction: photographs of the scene are reprojected to novel views, based on a coarse 3D model of the world. Yet, image-based rendering often suffers from artifacts, and most often does not deliver photo-realistic quality.
In the project we research novel approaches based on deep neural networks that improve the quality of the images generated by image-based rendering and that can remove or reduce typical artifacts. Focus of the project is in the first phase the generation of panoramas from a set of nearby inside-out take images that exploit such techniques to generate artifact-free panoramas from few input images. Later on, we will focus on approaches that allow the free movement of a user in real-time, making it possible to walk through a scene virtually.
The work in the project is centered around data obtained by the 3D-indoor-scanning systems of the industrial partner NavVis (Munich), who can deliver high-quality 3D point clouds and registered input images. We will thus focus on novel neural rendering techniques that as point clouds as proxy geometry, on the generation of panoramic images, and the research of possibilities to achieve free-viewpoint video, based on this data, in real-time.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
PatRo-MRI-2: Pathology-robust image reconstruction in Magnetic Resonance Imaging (10/2022–12/2026)
IRRW: Scaling Inverse Rendering to the Real World (08/2022–07/2025)
How do we best represent objects and their variations for inverse rendering? Can a combination of classical and novel techniques increase photorealism whilst retaining a low dimensional and interpretable representation? And given such object models: How do we efficiently infer the scene graph from visual input based on object-agnostic and object-specific processing? And ultimately can we explain the extreme generalization capabilities of the human visual system with an inverse rendering account?
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
ICETHICKNESS: Machine learning-based retrieval of ice thickness / internal structures from radargrams (07/2022–06/2026)
The IDP “Measuring and Modelling Mountain glaciers and ice caps in a Changing ClimAte (M³OCCA)” aims to combine cutting edge technologies with climate research. We will develop future technologies and transfer knowledge from other disciplines into climate and glacier research. Within the doctoral project “Machine Learning on Radargrams”, we aim at using and modifying machine learning techniques from medical imaging as well as natural language processing and apply those to glaciological radargrams to extract information on ice thickness and internal structures of ice bodies.
Scientific field: 409-05 Interactive and Intelligent Systems, Image and Language Processing, Computer Graphics and Visualisation
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Intermediate Report (10/2023)
So far, several cutting-edge machine learning models have been adapted to radar imagery and have been shown to exceed prior works. A comparison study is being performed to identify the best working approach. Moreover, a model that accounts for uncertainty is being developed, and learning strategies that do need fewer human annotations are considered.
InTimeVRSimulPatMod: In-time Virtual Reality Simulation Patient Models: Machine Learning and immersive-interactive Modeling of Virtual Patient Bodies (05/2022–05/2025)
Completed NHR compute time projects on Alex and Fritz at NHR@FAU
Humanities and Social Sciences
Humanities
Pose22: Pose Estimation on Russian International News Media (10/2022–12/2023; LARGE SCALE)
This NHR project is tied to an existing DFG/AHRC-funded joint project by FAU Erlangen-Nürnberg and the University of Oxford studying the mechanisms used for disinformation, in particular viewpoint manipulation, by Russian state-sponsored media. The purpose of this sub-project is to create a dataset fully annotated with pose information of the English-language programs found on the now-banned YouTube presence of RT (Russia Today) both for the purpose of our own analysis and to create an annotated open dataset.
Scientific field: 104 Linguistics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target systems: parallel computer Fritz & GPGPU cluster Alex
ProtCTRL: A conditional transformer for à la carte protein sequence generation (07/2022–06/2023)
The design of proteins with tailored functions will tackle many societal challenges. Traditionally, the protein design process relied on searching the global minima of multidimensional energy functions, a process that required significant computational times for each run. Recent advances in NLP have produced protein language models capable of generating fit protein sequences within seconds. However, these current models lack control over the design process, thus preventing user-defined design. The proposal described here will train a joint model, ProtCTRL, to capture the sequence-function relationships. ProtCTRL will be capable of generating sequences upon a user prompt, ultimately enabling tailored protein design. The model will be publicly released.
Scientific field: 201-07 Bioinformatics and Theoretical Biology; 104-04 Applied Linguistics, Experimental Linguistics, Computational Linguistics
University: University of Bayreuth
Target system: GPGPU cluster Alex
Life Sciences
Biology
GPCRSIM: Metadynamics simulations of ligand binding/unbinding and receptor activation/deactivation for G-protein coupled receptors (08/2022–04/2024; LARGE SCALE)
GPCRSIM uses classical (force-field) molecular dynamics simulations to determine binding sites, binding free energies and activation/deactivation free-energy profiles for predominantly class A G protein coupled receptors. In order to be able to observe rare events such as binding or activation, metadynamics enhanced sampling simulations employ standard optimized simulation profiles that have been developed in predecessor projects. Multiple-walker simulations ensure high parallel performance in metadynamics simulations that use standardized collective variables and funnel constraints to restrict the conformational space in the extracellular medium.
Scientific field: 201-02 Biophysics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target systems: parallel computer Fritz & GPGPU cluster Alex
ProtCTRL: A conditional transformer for à la carte protein sequence generation (07/2022–06/2023)
DNARepairTDG - DNA Repair by Thymine DNA Glycosylase (06/2022–09/2023)
Thymine DNA glycosylase (TDG) is an important enzyme involved in DNA repair, which removes mispaired or modified DNA bases and thus ensures genetic integrity. We have investigated possible reasons for its substrate specificity in our previous work and we now intend to extend the range of possible forms of the substrates to achieve a deeper understanding of the situation in the protein substrate complex prior to the chemical reaction. For this reason, we will investigate the possible role of imino-tautomeric forms of the damaged DNA bases flipped out into the enzyme active site as well as the effect of different protonation states of the substrate bases and an important histidine residue in the binding pocket.
Scientific field: 201-02 Biophysics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
ImmunoDomains: Interplay of immune receptors and lipid environment in signaling (04/2022–09/2023; LARGE SCALE)
The function of immune receptors expressed on the surface of a variety of cells is typically characterized by the sensing of an external signal, followed by signal modulation and transmission into the cell. All of these steps are directly or indirectly affected by the composition, structure, and characteristics of the plasma membrane that forms the signal transmission barrier, shapes the sites for both extracellular sensing and intracellular signaling cascades. In immune cells, these two are connected by immune receptors that bridge the membrane.
Immune receptors were repeatedly reported to be localized in specific membrane domains, so called immune domains, that involve mutual interactions between the receptors, co-receptors and the lipid environment. Clearly, immune cell activation depends on and may affect this domain formation. In this project we envisage to characterize the architecture and interaction dynamics of these immune domains, and their coupling to receptor signaling.
To investigate this mutual interplay and decipher the driving forces underlying immune domain formation, we will use both coarse-grained and atomistic molecular dynamics (MD) simulations of receptors and their environment in membranes of increasing complexity, from symmetric simplified model membranes to asymmetric membranes with a composition close to the plasma membrane. We expect the simulations to yield a fingerprinting of the lipid nanometer environment of immune receptors depending on the sequence as well as on the orientation of their transmembrane (TM) region within the membrane. The composition of the immune domain composition may also influence membrane shape such as curvature, which is as well an essential factor for the efficiency of binding of large ligands or of association and clustering of receptors. The mutual influence of the composition of immune domains and signaling-required receptor assemblies on the membrane shape will be screened employing coarse-grained MD simulations of a recently developed bicelle setup which is particularly suitable to study protein- and lipid-induced membrane reshaping processes in an unbiased manner.
In the long run, combining in silico molecular scale insights of immune domain composition, structure, and dynamics with results from immune cell activation, receptor localization and membrane phase in superresolution microscopy, and proteo-lipidomics, as provided by a collaborative network of experimental labs, will yield a comprehensive view on immune cell activation and its modulating key factors.
Scientific field: 201-02 Biophysics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Resolving the Structure of mRNA-Vaccine Lipid Nanoparticles (Q4/2021–Q1/2022)
Lipid nanoparticles (LNPs) are very successfully employed as novel transport vehicles for mRNA vaccines. A major gap in our understanding and thus obstacle for future developments of nanoparticle-mRNA drugs, however, is the lack of a molecular picture and molecular insight into LNPs. In this project we aim to provide unique insight at the atomistic scale into the structure and mechanisms of these carriers.
Scientific field: 201-02 Biophysics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
A deep unsupervised Model for Protein Design (Q4/2021–Q1/2022)
The design of new functional proteins can tackle many of the problems humankind is facing today but so far has proven very challenging. Analogies between protein sequences and human languages have been long noted and a summary of their most prominent similarities is described. Given the tremendous success of Natural Language Processing (NLP) methods in recent years, its application to protein research opens a fresh perspective, shifting from the current energy-function centered paradigm to an unsupervised learning approach based entirely on sequences. To explore this opportunity further we have pre-trained a generative language model on the entire protein sequence space. We find that our language model, ProtGPT2, effectively speaks the protein language and can generate de-novo sequences with natural properties in a matter of seconds.
Scientific field: 201-02 Biophysics
University: University of Bayreuth
Target system: GPGPU cluster Alex
Dynamics of B2AR-Gs(GTP) (Q1/2022)
Scientific field: 201-02 Biophysics
University: Leipzig University
Target system: parallel computer Fritz
Medicine
MASCARA: Molecular Assessment of Signatures ChAracterizing the Remission of Arthritis (12/2022–01/2024)
The classification of chest x-rays for radiologists is a time-consuming task. Recently, neural networks have been shown to be quite useful for a first assessment of chest x-rays. For this, a large amount of annotated data is necessary. Even though there are strong efforts to anonymize the patient identity, it still may be possible to identify a patient. Thus, in this work, we will investigate the generation of synthetic chest x- rays using latent diffusion models. This way, we will be able to generate large amount of data necessary to train chest x-ray classification networks while preventing leakage of patient identity.
Scientific field: 205-01 Epidemiology, Medical Biometry, Medical Informatics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
Final Report (01/2024)
The project aimed to harness the power of advanced neural networks for the classification of MRI images in psoriatic arthritis (PsA), seronegative rheumatoid arthritis (RA), and seropositive RA. The scientific objectives were centered around the development and application of ResNet neural networks and a novel neural network for the differentiation of inflammatory arthritis types based on MRI scans and 3D articular bone shapes.
The A100 HPC Cluster provided the computational backbone for this project, enabling the training of complex ResNet models and the novel neural network on large datasets. The first part of the project used ResNet neural networks to analyze MRI scans from 649 patients, encompassing various sequences such as T1 coronal, T2 coronal, and fat-suppressed contrast-enhanced sequences. The models were tasked with distinguishing between seropositive RA, seronegative RA, and PsA based on these scans.
In the second part, a novel neural network was trained on 3D shapes of hand joints derived from high-resolution peripheral-computed-tomography (HR-pQCT) scans of 617 patients. This innovative approach focused on identifying joint shape patterns indicative of RA, PsA, or healthy controls (HC).
Achievements and Results
Classification Performance: The ResNet models achieved AUROCs of 75%, 74%, and 67% for seropositive RA vs PsA, seronegative RA vs PsA, and seropositive vs seronegative RA classifications, respectively. The novel neural network differentiated hand bone shapes with AUROCs of 82% for HC, 75% for RA, and 68% for PsA, highlighting its potential for clinical application.
Insights into Disease Patterns: The project revealed that all MRI sequences were relevant for classification, but the exclusion of contrast agent–based sequences only marginally affected performance. Additionally, psoriasis patients without clinical arthritis were often classified as PsA by the neural networks, suggesting early PsA-like MRI patterns in psoriatic disease.
Heat Maps for Clinical Application: The novel network provided heat maps identifying critical anatomical regions prone to arthritis-related changes, offering valuable insights for clinical assessments and potential early detection of disease.
The A100 HPC Cluster proved to be an invaluable resource, providing the necessary computational power to process large imaging datasets and complex neural network architectures. The computational efficiency and capacity of the cluster were sufficient to meet the project’s objectives, enabling rapid training and iteration of the models.
A significant highlight of the project was the development of the novel neural network capable of analyzing 3D joint shapes, a pioneering approach in the field of rheumatology. This advancement opens new avenues for the non-invasive and early detection of inflammatory arthritis, potentially improving patient outcomes through earlier intervention.
Publications
L. Folle, S. Bayat, A. Kleyer, F. Fagni, L. A Kapsner, M. Schlereth, T. Meinderink, K. Breininger, K. Tascilar, G. Krönke, M. Uder, M. Sticherling, S. Bickelhaupt, G. Schett, A. Maier, F. Roemer and D. Simon: Advanced neural networks for classification of MRI in psoriatic arthritis, seronegative, and seropositive rheumatoid arthritis. Rheumatology 61(12), 2022. DOI: 10.1093/rheumatology/keac197
L. Folle, D. Simon, K. Tascilar, G. Krönke, A.-M. Liphardt, A. Maier, G. Schett, and A. Kleyer: Deep Learning-Based Classification of Inflammatory Arthritis by Identification of Joint Shape Patterns—How Neural Networks Can Tell Us Where to “Deep Dive”. Clinically. Front. Med. 9(850552), 2022. DOI: 10.3389/fmed.2022.850552
L. Folle, P. Fenzl, F. Fagni, M. Thies, V. Christlein, C. Meder, D. Simon, I. Minopoulou, M. Sticherling, G. Schett, A. Maier, and A. Kleyer: DeepNAPSI multi‑reader nail psoriasis prediction using deep learning. Front. Med. 9, 2022. DOI: 10.3389/fmed.2022.850552
GastroDigitalShirt: Development and test of deep neural network models for the automatic detection of body sounds to monitor digestion in control group and patients with intestinal disorders (09/2022–11/2023)
The goal of this project is to develop an unobtrusive wearable technology for long-term digestion monitoring, termed GastroDigitalShirt. We investigate low-amplitude bowel sounds (BS) as indicators of digestive disorders, including chronic inflammatory bowel diseases (IBD). We analyze methods to: (1) acquire BS in unconstrained environments (i.e. daily life of patients), (2) determine the benefit of source separation techniques to interpret BS sources, (3) spot BS patterns in continuous audio data streams, and (4) investigate how BS analysis can help patients and clinicians in the diagnosis and treatment of gut diseases.
Scientific field: 205-01 Epidemiology, Medical Biometry, Medical Informatics; 407-06 Biomedical Systems Technology; 409-07 Computer Architecture and Embedded Systems
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Natural Sciences
Chemistry
CatalAcetylen: Acetylene selective hydrogenation to ethylene catalyzed by bi- and trimetallic alloys: AI search for new catalysts (07/2022–12/2022)
Ethylene (C2H4) is a building block for the production of polymers. The employed catalysts used for enabling the polymerization process are sensitive to impurities. C2H4 is usually produced by cracking light alkanes and contains acetylene (C2H2) which deactivates the catalyst. Therefore, it is important to reduce the amount of it before starting the polymerization reaction. This important issue can be alleviated by selective hydrogenation of C2H2 to C2H4. In order to achieve high control over selectivity, we can employ binary and ternary transition-metal alloy surfaces as catalysts. However, there are few known alloys for such reactions. Applying only extensive computational search and/or trial and error experimental methods hinders the catalytic search. We will apply an AI tool based on compressed sensing, in conjunction with DFT calculations to identify descriptors for predicting catalytic activity and selectivity from a relatively small set of training DFT data.
Scientific field: 303 Physical and Theoretical Chemistry; 307 Condensed Matter Physics
University: Berlin
Target system: parallel computer Fritz
Physics
HESSML: Advanced Machine Learning Analysis for the H.E.S.S. Telescopes (08/2022–01/2024)
The High Energy Stereoscopic System (H.E.S.S.) telescopes in Namibia observe the very-high-energy gamma-ray sky from 20 GeV to 10s of TeV. H.E.S.S. operates by observing the Cherenkov light emitted when such a gamma-ray creates a particle cascade in the atmosphere; the nature, origin and energy of the incident particle can be determined using image processing techniques. State-of-the-art deep learning methods allow for high-sensitivity image analysis at high-speed, making them an attractive analysis prospect for H.E.S.S.. With this project, we aim to develop a variety of new analysis methods for H.E.S.S. data, including improved identification of muonic events, exploration of geometric deep learning techniques, fast simulation approaches and new data augmentation strategies.
Scientific field: 311-01 Astrophysics and Astronomy; 309-01 Particles, Nuclei and Fields
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Intermediate Report (01/2024)
Imaging Air Cherenkov Telescopes (IACTs) are fundamental to ground-based observations of the very high-energy sky. To perform ground-based gamma-ray astronomy, an effective rejection of the hadronic background is pivotal. Within the current research project, we are investigating algorithms based on state-of-the-art machine-learning techniques to improve instrument resolution and sensitivity.
We proposed a new deep-learning-based algorithm for classifying images by interpreting the detected images as sets of triggered sensors, which are then represented as graphs and analyzed using graph convolutional networks. This innovative method, particularly effective for images cleaned of night sky light, circumvents challenges associated with sparse images encountered in traditional deep learning techniques like convolutional neural networks. Our exploration of various graph network architectures has yielded promising results, showcasing improvements over existing machine-learning and deep-learning-based methods. This memory-efficient approach has promising potential to be adapted to IACT arrays with a large number of telescopes.
For precise observations using IACTs the instrument response has to be derived, requiring large simulation libraries. These simulations, characterized by computationally and memory-intensive calculations, are repeatedly performed for different observation intervals to account for variations in the optical sensitivity of the telescopes. Leveraging generative models based on deep neural networks, specifically Wasserstein Generative Adversarial Networks, we propose an efficient method for storing extensive simulation libraries in a memory-efficient manner and rapidly generating simulations. In our study, we applied this approach to a state-of-the-art IACT camera featuring over 1,500 pixels, and demonstrated the successful generation of high-quality images with high fidelity with respect to important physical parameters. Notably, the substantial increase in simulation speed, on the order of 100,000, holds promising implications for efficient and fast simulations for IACTs.
AHPO4TauID: Automatic hyperparameter optimization of Graph Neural Networks for tau neutrino identification in KM3NeT/ORCA (05/2023–07/2023)
KM3NeT/ORCA is a neutrino telescope currently under construction in the Mediterranean deep sea. It detects neutrinos of all flavors produced in the atmosphere indirectly by measuring the Cherenkov light emitted when fast charged particles are produced during neutrino-nucleon interactions. One of the scientific goals of the KM3NeT Collaboration is to further constrain the least-well-known elements of the three-flavor leptonic mixing matrix. To this end, we will in particular seek to identify tau neutrino interactions in the detector volume, a task known as tau neutrino event identification. A new approach to this problem is studied in this project, where state-of-the-art Graph Neural Networks are trained to directly identify tau neutrino events in the detector data. The promising results achieved in first trainings will be further improved in this NHR project by conducting a systematic search for the optimal hyperparameters of the networks, a task known as automatic hyperparameter optimization.
Scientific field: 309-01 Nuclear and Elementary Particle Physics, Quantum Mechanics, Relativity, Fields
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex
Final Report (11/2023)
Different optimization steps for training Graph Neural Networks (GNNs) on the task of tau neutrino event identification in KM3NeT/ORCA were performed. The last optimization step consisted of an automatic hyperparameter optimization (AHPO) that systematically searched through the hyperparameter space by stepwise parameter selection and subsequent automated training of the resulting GNNs. The chosen hyperparameters, various metrics that were monitored during training, and the final performance metric of the GNNs were logged. In total, about 2000 GNNs could be trained with the 46000 GPU-hours of computing time granted for this project. The data from this AHPO was used to identify the most significant hyperparameters, i.e., the hyperparameters whose values have the greatest influence on the resulting performance of the GNNs.
The most significant hyperparameter was identified to be the learning rate. The boxplot on the left shows the distribution of the value of the performance metrics over many trained GNNs for learning rates in different ranges. It indicates that learning rates between 10-3 and 10-2 result in the best-performing GNNs. Analogous investigations were done for many other hyperparameters in the GNN architecture, demonstrating which of these are essential and which are of less significance for the performance in the studied classification task. Importantly, this systematic analysis a-posteriori confirmed the general, best-guess hyperparameter choices that were made in previous KM3NeT studies, published in two dissertations and several conference proceedings, using very similar GNN architectures for classification and regression tasks. Due to a lack of corresponding resources within the international KM3NeT Collaboration, this important systematic study could not be done earlier. Another goal of the AHPO was to achieve a performance gain in classification compared to the first-step manual optimizations done earlier. However, the AHPO did not show a significant performance gain over the manual optimizations. Presumably, the reason is that the GNN architecture was already well-optimized due to its use in previous studies. Finally, using the insights, strategies, and methods developed during this project for future architectures and applications in the analysis of neutrino telescope data, more efficient, faster and more trustable optimizations will be possible.
The resources granted to this project have been sufficient to achieve its goals. The baseline target was to train between 1000-1500 different GNNs, but the resources turned out to be even enough to train about 2000 GNNs. This number of GNNs allowed us to compute a large enough sample to draw statistically relevant conclusions on the significance of most of the probed hyperparameters. The AHPO ran for most of the time in parallel on 64 GPUs, i.e., on eight computing nodes with eight GPUs per node.
Publications
The accomplishments of this project are documented in the master’s thesis of Lukas Hennig.
MultiDyn: Dynamics of complex networks of oscillators (01/2023–12/2023; LARGE SCALE)
This project is designed to provide a wide-ranging stability phase diagrams classifying the distribution of self-organized periodicities and complex motions of various sorts in families of coupled nonlinear oscillators when more than one control parameter is varied simultaneously.
Scientific field: 310 Statistical Physics, Soft Matter, Biological Physics, Nonlinear Dynamics
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: parallel computer Fritz
CLSfiniteV: Finite volume study of 2 + 1f QCD from lattice simulations (10/2022–09/2023; LARGE SCALE)
High precision calculations from lattice QCD are needed nowadays in many respects. In order to obtain such results it is mandatory to have good control over all systematic effects involved in such calculations. In this study we plan to investigate the systematic effects introduced by the finite volume of the system, which is important in particular for quantities like low energy constants, pseudoscalar masses and decay constants, or the axial charge of the nucleon.
Scientific field: 309 Particles, Nuclei and Fields
University: Universität Regensburg
Target system: parallel computer Fritz
Intermediate Report (01/2024)
Within the past NHR project we have made another great step forward in our large scale project of computing important matrix elements related to Beyond-the-Standard-Model (BSM) interactions (which are not accessible experimentally) and (lower) moments of the parton distribution functions. These computations are of great interest for current and upcoming experiments (like the Electron Ion Collider) with respect to nucleon structure. Utilizing the granted NHR resources we could improve on our statistical and systematic uncertainties in a particular relevant region in the parameter space that enters our fits from which we extract the final (physical) matrix elements.
CatalAcetylen: Acetylene selective hydrogenation to ethylene catalyzed by bi- and trimetallic alloys: AI search for new catalysts (07/2022–12/2022)
Functional renormalization group calculations (Q1/2022)
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: RWTH Aachen
Target system: parallel computer Fritz
Massively parallel simulation via Markov Chain Monte Carlo techniques (Q1/2022)
Scientific field: 307-02 Theoretical Condensed Matter Physics
University: Julius-Maximilians-Universität Würzburg
Target system: parallel computer Fritz
Engineering Sciences
Thermal Engineering/ Process Engineering
Flow around a Wind Turbine Blade at Reynolds Number 1 Million (Q1/2022)
The cost of energy produced by wind turbines has been undergoing a steady reduction. Wind energy supplied 15% of the electricity demand of the European Union in 2019. Since rotor blades are the determining component for both performance and loads, they are the objective of further optimizations. To obtain high efficiencies, an increased use of special aerodynamic profiles is observed possessing large areas of low-resistance, which means laminar flow is maintained. In order to design such profiles, it is necessary to include the laminar-turbulent transition in CFD simulations of wind turbine blades. Thus, the objective of the project is to carry out high-fidelity numerical simulations of the flow around a wind turbine blade at a realistic Reynolds number to get a deeper insight into this phenomenon and especially the transition process under different levels of the turbulence intensities of the approaching flow.
Scientific field: 404-03 Fluid mechanics
University: Helmut Schmidt University Hamburg
Target system: parallel computer Fritz
Evolution of drops in homogeneous isotropic turbulence (Q1/2022)
Scientific field: 404-03 Fluid mechanics
University: University of Bremen
Target system: GPGPU cluster Alex
Computer Science, Electrical and System Engineering
HEISSRISSE: Massively Parallel Simulation of the Melt Pool Area during Laser Beam Welding using the Lattice Boltzmann Method (10/2022–11/2023)
HEISSRISSE: As a flexible and contact-free joining technology, laser beam welding has increasingly gained importance. Processing of alloys with large melting range poses a challenge due to their solidification cracking tendency. Solidification cracks form due to critical stress and strain states of the dendritic microstructure with interdendritic melt. Despite the high industrial relevance, there are only approaches addressing single aspects of the problem, metallurgically or structurally oriented. The research unit “Solidification Cracking during Laser Beam Welding – High Performance Computing for High Performance Processes” aims at developing quantitative process understanding of the mechanisms of solidification cracking and their relation to process parameters. The sub-project aims at simulating the dynamics of the melt pool with a resolution of about one micron. This is achieved using a numerical model consisting of approximately 109 computational cells that have to be computed for more than 105 time steps. The simulation will model the phase change at melting and solidification, the energy input from the laser, the expansion and contraction, and the heat and mass transport in the melt pool. Simulations of such complexity require the computational power of parallel high-performance systems and can only be realized using optimized parallel algorithms and modern software technologies. In this sub-project, the lattice Boltzmann method (LBM) will be used and extended to correctly capture the various physical effects in the melt pool. Compared to other numerical schemes, the LBM is well-suited for parallel computing and modern computer architectures that include hardware accelerators. The implementation will be based on the HPC framework waLBerla specifically designed for implementing complex multi- physics applications. Using abstraction layers and code generation concepts, software developed with waLBerla is sustainable, i.e., portable to future computer architectures. One of the significant aspects of this project lies in the validation of newly implemented models and algorithms, and in the interoperability with models from the partners in the research unit. This will be achieved via the close cooperation with neighboring sub-projects.
Scientific field: 409-08 Massively Parallel and Data-Intensive Systems
University: Friedrich-Alexander-Universität Erlangen-Nürnberg
Target system: GPGPU cluster Alex