22455 research outputs found
Sort by
Forward-modelling the hot circumgalactic medium in X-rays
Get PDFThe hot phase of the circumgalactic medium (CGM) offers a unique window into the gas flows that shape galaxy evolution, serving as both a reservoir and a conduit for baryons cycling in and out of galaxies. The hot CGM, as observed in X-rays, probes the diffuse baryonic gas, helping to constrain feedback mechanisms and galaxy formation models. Recent advances in observational capabilities, particularly from the X-ray telescope eROSITA, have opened new avenues to investigate the properties of the hot CGM around Milky Way (MW)-mass galaxies, yielding critical benchmarks for cosmological hydrodynamical simulations. Yet significant challenges remain in interpreting these measurements due to projection effects, contributions from unresolved sources, and the complex influence of a galaxy’s large-scale environment.
In this thesis, we build a fully self-consistent forward model of the hot CGM using the TNG300 cosmological hydrodynamical simulation. We construct a novel lightcone and generate mock X-ray observations based on intrinsic gas cell properties, so-called LC-TNGX, to enable direct comparisons with observations. Our analytical modelling captures intrinsic X-ray surface brightness profiles across stellar and halo mass bins. We find that higher stellar mass bins correspond to shallower slopes of the intrinsic galactocentric profiles, quantified via decreasing values of the profile surface brightness exponent β. Critically, we quantify the effect of satellite galaxies incorrectly identified as centrals in stacking experiments, which biases the derived hot CGM X-ray surface brightness profiles. For stellar mass bins similar to MW-masses, M31-masses, and twice M31-masses, we demonstrate that even modest contamination fractions (as low as 1%) can dominate the measured X-ray surface brightness profiles at large radii. Specifically, in the MW-mass bin, misclassified centrals contributing 30%, 10%, or 1% of a stacked sample dominate the measured surface brightness profile beyond radii of 0.11 × R₅₀₀, 0.24 × R₅₀₀, and 1.04 × R₅₀₀, respectively.
Building on this framework, we develop forward models tailored to the eROSITA stacked X-ray radial surface brightness profiles of MW-mass galaxies. Our model includes two key emission components: hot gas around both central and satellite galaxies, and point-source contributions from X-ray binaries (XRBs) and active galactic nuclei (AGN). We simulate mock observations using the TNG300-based gas profiles, carefully matching stellar mass and redshift distributions from the observations, and explore how variations in the underlying halo mass distribution affect the results. We show that for galaxy samples matched in stellar mass, increasing the mean halo mass by a factor of ~2 leads to a ~4× enhancement in the stacked X-ray luminosity of the hot CGM. By incorporating empirical constraints on AGN and XRB luminosities, we identify the model that best matches the eROSITA data. In the MW stellar mass bin, our model agrees well with prior literature. We find that within approximately 40 kpc from the galaxy centre, the hot CGM and point-source emission each contribute ~40–50% of the total stacked X-ray emission. Beyond ~40 kpc, the hot CGM emission from satellite galaxies, tracing more massive host halos with a mean M₂₀₀ ~ 10¹⁴ solar masses, dominates the stacked signal. This approach offers a novel method to jointly constrain the mean AGN X-ray luminosity and the radial hot CGM gas distribution in MW-mass galaxies, enabling tests of AGN feedback prescriptions in hydrodynamical simulations.
In addition to observational projection effects, we investigate how the large-scale cosmic environment shapes the hot CGM X-ray properties. Using the DisPerSE filament-finding algorithm on our TNG300-based lightcone spanning 0.03 ≤ z ≤ 0.3, LC-TNGX, we classify central galaxies into five distinct large-scale environment categories: clusters and massive groups, cluster outskirts, filaments, filament-void transition regions, and voids/walls. We find that the X-ray surface brightness profiles of central galaxies in filaments with M₂₀₀ > 10¹² solar masses are 20–45 % brighter in the radial range of (0.3–0.5) × R₂₀₀ compared to those in voids and walls. This excess arises from higher average gas densities, temperatures, and metallicities in filament galaxies, revealing a clear imprint of the cosmic web on hot CGM properties. Our findings highlight the importance of accounting for the cosmic environment in interpreting X-ray CGM measurements and suggest promising avenues for future studies exploring how the assembly history, gas accretion, and connectivity in the cosmic web shape the hot gas content around galaxies.
Taken together, the progress outlined in this thesis underscores that fully understanding the hot CGM requires an integrated approach, combining X-ray observations, careful modelling of projection and environmental effects, and robust comparisons across multiple hydrodynamical simulations. The methods and models developed here provide a critical framework for interpreting hot CGM observations enabled by current and future X-ray surveys, allowing us to disentangle the complex interplay of baryonic physics, AGN feedback, halo demographics, and the cosmic web in shaping the hot gas around galaxies. As new data arrives from next-generation missions like NewAthena and HUBS, alongside increasingly sophisticated simulations, we stand at the threshold of using the hot CGM as a precise cosmological and astrophysical probe. The challenges ahead in linking microphysical feedback processes to large-scale observables, quantifying environmental impacts, and constraining the diverse mechanisms governing baryon cycles offer an exciting frontier where theory and observation converge to refine our models of galaxy formation and evolution
Statistical methods leveraging uncertainties in machine learning
Get PDFIn today’s data-driven landscape, machine learning methods are increasingly applied in domains that demand high levels of safety, reliability, and interpretability. However, hidden influencing factors and limited data availability can significantly impair model performance and amplify predictive uncertainty. Recognizing, quantifying, and—where possible—reducing uncertainties such as aleatoric and epistemic uncertainty has therefore become a central concern. This is particularly true in critical fields like autonomous driving, medical diagnostics, finance, weather forecasting, and industrial production, where dependable predictions are not merely advantageous, but essential. Despite its relevance, the broader adoption of uncertainty quantification in practice is often hindered by high computational demands and growing model complexity. Furthermore, aleatoric uncertainty—stemming from noise and imperfections in the data itself—poses a fundamental challenge to the reliability of data-driven models. This dissertation explores multiple strategies for uncertainty quantification across four publications. These contributions examine both the necessity and the practical implementation of probabilistic techniques, while also introducing novel, less computationally intensive methods that reduce model complexity without sacrificing robustness. Publication 1: The first study addresses epistemic and aleatoric uncertainties in the preprocessing phase of industrial production modeling. Aleatoric uncertainties are predefined based on expert experience, setting the bounds for acceptable input variation. Given the limited spatial distribution of measurement data, uncertainty-aware interpolation is applied for data augmentation. Probabilistic Gaussian Process Regression is employed to model prediction intervals and serve as a basis for generating synthetic input data. Results using real production data from ams OSRAM show that even with sparse measurements, highly accurate models can be constructed. Publication 2: The second study introduces a novel modeling approach that combines two types of target variables: a continuous regression target and an ordinal classification target. A customized loss function, paired with fuzzy logic, enables the model to optimize regression estimates while simultaneously improving classification performance. The method shows particularly strong performance in imbalanced data scenarios, leading to a significant reduction in latent uncertainty. Applied to housing market data in the United States, the approach yields up to a 17.1% improvement in F1-score. Publication 3: The third publication investigates the integration of data-independent, expertderived knowledge into data-dependent learning models. Unmeasured or unquantified latent uncertainties can reduce model robustness. This approach trains models using both observed data and qualitative expert assessments—without requiring expert input at inference time. Results using synthetic data generated via variational autoencoders (VAEs), based on real-world use cases from ams OSRAM, demonstrate improved optimization even with a marginal increase in mean absolute error (MAE). Publication 4: The fourth study addresses the challenge of training with small datasets. A probabilistic modeling approach is presented that estimates the latent distribution of a target variable using ordinal class labels. This enables the generation of additional, reliable input data to support model training. With only 5–10% of the original training data, the method achieves notable improvements: up to 10% in mean squared error (MSE), 5–10% in coefficient of determination (R2), and approximately 8% in prediction coverage.In der heutigen datengetriebenen Welt werden Machine-Learning-Methoden zunehmend in sicherheitskritischen und hochzuverlässigen Anwendungsbereichen eingesetzt. Verdeckte Einflussfaktoren sowie begrenzte Datenverfügbarkeit können jedoch die Modellgüte erheblich beeinträchtigen und die Vorhersageunsicherheit erhöhen. Die Erkennung, Quantifizierung und – wo möglich – Reduktion von Unsicherheiten, insbesondere aleatorischer und epistemischer Art, hat daher zentrale Bedeutung gewonnen. Dies gilt insbesondere für Anwendungsfelder wie autonomes Fahren, medizinische Diagnostik, Finanzwesen, Wetterprognose und industrielle Produktion, in denen verlässliche Vorhersagen nicht nur wünschenswert, sondern essenziell sind. Die breite Anwendung von Unsicherheitsquantifizierung scheitert jedoch häufig an hohen Rechenaufwänden und zunehmender Modellkomplexität. Zudem stellt die aleatorische Unsicherheit – verursacht durch zufällige Messfehler und Datenrauschen – eine grundlegende Herausforderung für die Verlässlichkeit datenbasierter Modelle dar. Diese Dissertation untersucht verschiedene Strategien zur Quantifizierung von Unsicherheiten anhand von vier begutachteten Publikationen. Die Beiträge beleuchten sowohl die Notwendigkeit als auch die praktische Umsetzung probabilistischer Verfahren und schlagen darüber hinaus neuartige, rechenökonomische Alternativen vor, die ohne signifikante Steigerung der Modellkomplexität eine robuste Modellierung ermöglichen. Publikation 1: Die erste Studie thematisiert epistemische und aleatorische Unsicherheiten in der Vorverarbeitung industrieller Produktionsdaten. Aleatorische Messunsicherheiten werden auf Basis von Erfahrungswerten vordefiniert, um einen Rahmen für die Eingabedatenunsicherheit zu schaffen. Aufgrund der räumlich begrenzten Messpunktverteilung werden Interpolationsverfahren mit Unsicherheitsberücksichtigung zur Datenanreicherung eingesetzt. Eine probabilistische Gaussian-Process-Regression dient zur Modellierung von Prognoseintervallen und zur Generierung zusätzlicher Eingabedaten. Die Anwendung auf Produktionsdaten der Firma ams OSRAM zeigt, dass auch mit wenigen Messwerten präzise Modelle realisierbar sind. Publikation 2: Die zweite Studie stellt einen innovativen Modellierungsansatz vor, der zwei Zielgrößen kombiniert: eine kontinuierliche Regressionsgröße und eine ordinale Klassifikationsgröße. Eine angepasste Verlustfunktion in Kombination mit Fuzzy Logic ermöglicht eine gleichzeitige Optimierung beider Zielgrößen. Der Ansatz zeigt insbesondere bei unausgeglichenen Klassenverteilungen signifikante Verbesserungen und reduziert latente Unsicherheiten durch die Kombination beider Ziele. Die Anwendung auf US-Immobiliendaten zeigt eine Verbesserung des F1-Scores um bis zu 17,1 %. Publikation 3: Die dritte Veröffentlichung befasst sich mit der Integration von datenneutralem Expertenwissen in datenabhängige Lernmodelle. Nicht gemessene oder nicht quantifizierbare Unsicherheiten können die Modellstabilität gefährden. Der vorgestellte Ansatz kombiniert trainingsseitig beobachtete Daten mit qualitativen Experteneinschätzungen – ohne dass Expertenwissen zur Vorhersagezeit erforderlich ist. Ergebnisse mit synthetischen Daten, erzeugt mittels Variational Autoencoders (VAE) auf Basis realer ams-OSRAM-Anwendungsfälle, zeigen eine verbesserte Modelloptimierung trotz eines geringen Anstiegs des mittleren absoluten Fehlers (MAE). Publikation 4: Die vierte Studie widmet sich der Problematik kleiner Stichprobenumfänge. Ein probabilistischer Modellierungsansatz wird vorgestellt, bei dem die Verteilung der Zielgröße über ordinale Klassen abgeschätzt wird. Dadurch lassen sich zusätzliche, zuverlässige Eingabedaten für das Modelltraining erzeugen. Bereits mit 5–10% der ursprünglichen Trainingsdaten lassen sich deutliche Verbesserungen erzielen: eine Reduktion des mittleren quadratischen Fehlers (MSE) um bis zu 10 %, eine Steigerung des Bestimmtheitsmaßes (R2) um 5–10% sowie eine um rund 8% verbesserte Abdeckung der Prognoseintervalle
Calcium carbonate biomaterials
Get PDFCalcium carbonate is a chemical substance that is not only considered the most important compound for biomineralisation in marine organisms and invertebrates, but also has significant implications for industry, the carbon cycle, and biomedicine. Through evolving and adapting to ecological challenges, organisms have developed the ability to biomineralise complex hierarchical structures comprising biopolymers and inorganic minerals. Due to their outstanding material properties, such as high strength and toughness, while retaining a low density compared to the pure inorganic mineral, biomineralised materials have long been a source of inspiration for fabricating and optimising man-made materials. Accordingly, it is of great importance to understand the hierarchical design and biomineralisation principles of calcium carbonate hard tissues. This dissertation presents and discusses in great detail findings related to the microstructure, crystallographic texture, and nanomechanical properties of various calcium carbonate biomaterials. This work focuses mainly on analysing marine invertebrate hard tissues, such as bivalve molluscs shells, rhynchonellate brachiopod shells and sea urchin skeletal elements.
Molluscs, particularly bivalves, are globally widespread and present in highly diverse marine environments. Bivalves are renowned for their high diversity in lifestyles and for generating hard tissues with hierarchical structures. About 15 main calcium carbonate microstructures are recognised today in modern bivalve shells. An important part of the bivalve composite hard tissue is the organic matrix, which regulates and organises the crystallographic orientation and morphology of the inorganic minerals. Bivalve molluscs can secrete four main carbonate polymorphs, namely calcite, aragonite, amorphous calcium carbonate or (rarely) vaterite and use the biomineralised hard tissues mainly for protection of the soft body, but also to pursue diverse lifestyles. To gain insight into the hierarchy, microstructure, and nanomechanical properties of bivalves, I characterised differently cut shells of 30 different species from 11 orders by EBSD analysis, thermogravimetric analysis, electron and confocal laser microscopy imaging, and nanomechanical testing.
The structural results provide a detailed illustration of the microstructure and changeovers between the different layers. If the microstructures feature the same calcium carbonate phase, the changeover is generally smooth, with the initial crystallographic texture being transmitted from the shell portion of the adjacent layer. This biomineralisation mechanism is particularly fascinating for the aragonitic myostraca, the bivalve adductor muscle attachment sites, as their microstructure generation is largely influenced by competitive growth determinants rather than biological control. Bivalve myostraca are not only interesting because they enable the organism to form a strong connection to the adductor muscle fibres, but also because of their outstanding nanomechanical properties, such as a significantly enhanced hardness, compared to other shell layers or geological aragonite. The microstructure and nanomechanical properties of myostraca from different bivalve orders and bivalves following different lifestyles (such as burrowing, swimming, or attaching to substrates) are characterised. Previous studies have indicated that bivalve myostraca are strictly conservative in microstructure and texture, consistently forming large, prismatic units. However, the measurements presented in this work reveal a broad diversity in myostracal microstructures, demonstrating how the different crystal arrangements, textures, and twinning modes of the myostraca and other shell layers can influence material properties.
Like bivalve molluscs, rhynchonellate brachiopods also use specialised epithelial cells to mediate crystal growth at muscle attachment sites of their calcium carbonate shells. Although rhynchonellates and bivalves share similar living environments, lifestyles and modes of biomineralisation, they are biologically distant organisms. Thus, comparing the microstructural similarities and differences between rhynchonellates and bivalves provides insight into the biological convergence of muscle attachment sites and the different approaches for forming these important organic-inorganic interface structures. High-resolution EBSD measurements for the shells of modern two- and three-layered rhynchonellates show that the general microstructure of brachiopod muscle attachment sites differs strongly from bivalve shells in both calcium carbonate phase and crystal morphology. However, the muscle attachment sites of both invertebrate classes share some characteristics that may be important for a strong muscle-shell attachment. This includes the adoption of crystallographic texture from adjacent shell layers with the same calcium carbonate phase, as well as the co-orientation of crystallographic c-axes parallel to the muscle bundles at muscle attachment sites.
In contrast to the biomineralised hard tissues of bivalved invertebrates, nucleation of calcitic skeletal elements in echinoids involves intracellular processes. The latter usually comprises an intricate arrangement of highly co-oriented trabeculae into various porous, yet robust stereom architectures, historically described as single-crystalline. Applying the unprecedented EBSD data analysis method of pattern matching to different stereom architectures in the tests and spines of Cidaris cidaris and Paracentrotus lividus, this work reveals the presence and distribution of internal small-angle misorientations. As these misorientations are mostly located at the trabecular junctions, the quality of single-crystallinity appears to be influenced by the respective stereom architecture; however, sea urchin calcite cannot be considered single-crystalline invariably. Furthermore, I demonstrate the presence of poorly co-oriented, polycrystalline areas at both muscle attachment sites of echinoids (at test tubercles and the base of the spines) and the cortex, which encases the primary and secondary spines of C. cidaris. The microstructure generation of the cortex is determined by two factors: (i) The cortex crystals nucleate with c-axes oriented perpendicular to the stereom, which functions as a nucleation template. Thus, the smoothness of the outer stereom surface determines the texture of the initial cortex portion. (ii) A competitive growth mechanism determines the further microstructure and texture of the cortex.
Furthermore, this study demonstrates how the EBSD pattern matching method can improve the angular precision and reveal small-angle misorientations within highly co-oriented structures, such as the sea urchin stereom. By employing this method to other calcium carbonate biomaterials, such as eggshell grains, myostracal prisms or large columns in bivalve shells, I highlight in great detail their crystallographic texture and microstructure, demonstrating the large potential of this data evaluation technique for future studies
Structure and properties of layered intermetallic compounds and supertetrahedral phosphidosilicates
Get PDFDetecting gender discrimination in natural language processing
Get PDFTechnology is an essential part of human life. It helps us be more productive and shapes the way we live. Modern technology is composed of algorithms, but modern algorithms not only benefit society, they can also harm it. By misrepresenting certain genders or reinforcing stereotypes, algorithms can contribute to discrimination. The field of natural language processing (NLP) is particularly prone to such issues. This doctoral thesis analyses gender discrimination in NLP and
proposes a way to make it, specifically large language models (LLM), less gender discriminatory by improving their training data.
The first part of this thesis defines what algorithmic gender fairness means. This definition is then applied to analyse the results of information retrieval methods and the results of search algorithms. This analysis reveals that the representation of different genders in algorithmic output remains insufficient, underscoring the need to improve current approaches.
Building on the foundation of fairness in information retrieval and search, the focus then shifts to LLMs, a rapidly evolving technology that increasingly shapes everyday life. The model GPT-3 (Brown et al., 2020), as implemented in the system ChatGPT (OpenAI, 2022), is analysed with regard to how it responds to prompts in English and German from a female, male, or neutral perspective. The analysis of the prompt results shows that attempts to reduce gender discrimination after training can introduce new problems, for example, an over-representation of female personas in response to neutral prompts or an exaggerated emphasis on diversity in gendered prompts. These findings suggest that “downstream” mitigation, after model training, is not the right approach. Instead, mitigation should be done “upstream”, before training, by improving the quality of the training data itself.
This is addressed in the third part of the thesis. The first publication of this part introduces a modular, language-agnostic pipeline designed to detect discrimination in English newspaper texts. This pipeline combines linguistic discourse analysis with computational techniques. Using information extraction methods, it identifies the actors mentioned in a text, how they are referred to (nomination), and how they are described (predication). Therefore, it is possible to analyse quantitative metrics for each gender in the text and, additionally, qualitative metrics like the sentiment towards actors of each gender. The pipeline is scaled up in a second publication to process an entire corpus of German newspaper articles. This work also publishes the most significant German newspaper corpus to date, spanning four decades and comprising 1.8 million texts. A third publication further extends the pipeline and utilises it to generate a gender-balanced corpus, drawing on the German newspaper corpus from the second publication.Technologie ist ein zentraler Bestandteil des modernen Lebens. Sie steigert unsere Produktivität und beeinflusst maßgeblich unsere Lebensweise. Gleichzeitig bergen algorithmische Systeme nicht nur Potenziale, sondern auch Risiken für die Gesellschaft. So können sie etwa Geschlechterrepresentationen verzerren oder bestehende Stereotype verstärken und dadurch Diskriminierung begünstigen. Besonders im Bereich der Verarbeitung natürlicher Sprache (Natural Language Processing, NLP) treten solche Problematiken verstärkt auf. Die vorliegende Dissertation befasst sich mit unterschiedlichen Teilbereichen des NLP und untersucht, wie geschlechtsspezifische Diskriminierung in diesen entstehen kann. Darüber hinaus wird ein Ansatz vorgestellt, um einen konkreten Bereich, Large Language Models (LLM), durch eine Analyse der Trainingsdaten weniger diskriminierend zu gestalten.
Im ersten Teil wird das Konzept der algorithmischen Gender-Gerechtigkeit definiert. Diese Definition
dient anschließend als Grundlage für die Analyse von Ergebnissen von Information-Retrieval- Systemen und Suchmaschinenergebnissen. Die Analyse zeigt, dass die algorithmische Repräsentation verschiedener Geschlechter nach wie vor unzureichend ist, was auf die Notwenigkeit der Verbesserung von gängigen Vorgehen hinweist.
Auf Grundlage dieser Erkenntnisse richtet sich der Fokus im zweiten Teil der Arbeit auf LLMs, eine Technologie, die rasant in allen Lebensbereichen adaptiert wird. Konkret wird das LLM GPT- 3 (Brown et al., 2020), in seiner Implementierung im System ChatGPT (OpenAI, 2022) betrachtet. Es wird analysiert, wie sich das System bei Prompts verhält, die aus weiblicher, männlicher oder neutraler Perspektive formuliert sind, sowohl auf Deutsch als auch auf Englisch. Die Analyse zeigt, dass der Versuch, geschlechtsspezifische Diskriminierung nachträglich aus dem System zu entfernen, problematisch sein kann. Die Antworten des Systems neigten zu Überkorrekturen: Bei neutralen Prompts wurden vermehrt weibliche Personen generiert und Prompts, die eine weibliche oder männliche Sichtweise einnahmen, führten zur Überbetonung von Diversität. Diese Beobachtungen machen deutlich, dass eine Korrektur nach dem Modelltraining (“downstream”)
nicht ausreicht. Stattdessen sollte bereits vor dem Training (“upstream”) angesetzt werden, durch eine gezielte Aufbereitung und Verbesserung der Trainingsdaten.
Um das Problem der geschlechtsspezifischen Diskriminierung bereits vor dem Modelltraining anzugehen, befasst sich der dritte Teil der Dissertation mit sogenannten „upstream“-Mitigation- Ansätzen. In einer ersten Publikation wird eine modulare, sprach-agnostische Pipeline zur Erkennung von Diskriminierung in Zeitungstexten entwickelt. Ziel ist es, diskriminierende Muster frühzeitig in Trainingsdaten aufzudecken. Die Pipeline kombiniert Methoden der linguistischen Diskursanalyse mit informatischen Verfahren und nutzt Ansätze aus der Information Extraction,
um die zentralen Akteur:innen eines Textes zu identifizieren, ihre Benennung (Nomination) sowie ihre Darstellung im Text (Prädikation) zu erfassen. Dadurch lassen sich sowohl quantitative Metriken im Bezug auf die Geschlechterverteilung im Text, als auch qualitative Aspekte wie das Sentiment gegenüber im Text genannten Geschlechtern analysieren. In einer anschließenden Publikation wird die Pipeline auf einen Korpus deutschsprachiger Zeitungstexte angewendet. Im Zuge dieser Publikation wird der bislang größte, öffentlich zugängliche, deutschsprachige
Zeitungskorpus veröffentlicht. Dieser umspannt vier Jahrzehnte und besteht aus rund 1,8 Millionen Texten. Eine abschließende Publikation erweitert wiederum die Pipline der vorhergehenden Publikation und nutzt diese, um auf Grundlage des deutschprachigen Zeitungskorpuses einen Gender-ausgeglichenen Korpus zu generieren
