OPUS Online Publikationen der Universität Stuttgart
Not a member yet
17148 research outputs found
Sort by
Consecutive SSCs increase the SSC effect in skinned rat muscle fibres
No full textMuscle function is essential for generating force and movement, with stretch-shortening cycles (SSCs) playing a fundamental role in the economy of everyday locomotion. Compared with pure shortening contractions, the SSC effect is characterised by increased force, work, and power output during the SSC shortening phase. Few studies have investigated whether SSC effects transfer across consecutive SSCs. Therefore, we investigated SSC effects over three consecutive SSCs in skinned rat muscle fibres by analysing the isometric force immediately before stretch onset (Fonset), the peak force at the end of stretching (Fpeak), and the minimum force at the end of shortening (Fmin), along with mechanical (WorkSSC) and shortening work (WorkSHO) at different activation levels (20%, 60%, and 100%). Each SSC was followed by an isometric hold phase, allowing force to return to a steady state. Results indicated an increase in both Fpeak (20.3%) and WorkSSC (60.9%) from SSC1 to SSC3 across all activation levels tested. At 20% and 60% activation, Fonset, Fmin, and WorkSHO increased (range: 4.5-28.5%) from SSC1 to SSC3. However, at 100% activation, Fonset and WorkSHO remained unchanged, while Fmin decreased (- 8.5%) from SSC1 to SSC3. These results suggest that the increase in SSC effects at submaximal activation may be primarily due to increased cross-bridge forces. The absence of increases in Fonset, Fmin, and WorkSHO at 100% activation suggests that increases in Fpeak and WorkSSC may not be attributed to increased cross-bridge force but could instead be caused by additional effects, possibly involving modulation of non-cross-bridge structures, likely titin, and their stiffness.Projekt DEALDeutsche Forschungsgemeinschaf
Simplified multireference coupled‐cluster methods : hybrid approaches with averaged coupled pair theories
No full textWe define an approximation to the internally contracted multireference coupled‐cluster method with single and double excitations by a hybrid approach. The rationale is to treat the external pair energy contributions by the coupled‐cluster method, which provides accurate results for a large part of the correlation energy while being tractable as the involved pair cluster operators commute. For the internal and semi‐internal contributions, for which the coupled‐cluster part becomes involved due to non‐commuting operators, a linearized approach based on the coupled‐electron pair approximation (CEPA) is used. For the latter, the CEPA(0) method, the averaged coupled pair functional (ACPF), the averaged quadratic coupled‐cluster (AQCC) method, and the averaged CEPA method are tested. We test the methods concerning size consistency, potential energy curves for C2, N2, CN, and O3 and for the singlet‐triplet splitting of ortho‐, meta‐, and para‐benzynes. Our results show that AQCC provides the most accurate results and stable performance. The main drawback of the method is that it shows small violations of size consistency.Deutsche ForschungsgemeinschaftProjekt DEA
Design piracy, design protection and conflict strategies in the «Barmer Artikel» industry (20s to 60s)
No full textDealing with design piracy is a blind spot in German business history, but studies on the importance of industrial property rights as an element of business strategies are also scarce in general. This article uses the example of the «Barmer Artikel», a specific segment of Wuppertal’s textile industry, during the period from the 1920s to the 1960s to examine how market strategies as well as legal disputes, conflict behaviour and juridical solutions were mutually dependent with regard to design protection. It shows that the enforcement of intellectual property rights was primarily a question of opportunity costs and that both plaintiffs and defendants carefully considered their actions, e.g. whether they appealed to the local arbitration court, took ordinary legal action or whether they avoided proceedings altogether. The findings reveal that property rights regimes are contingent and are equally determined by legal and economic considerations. As a result, the theoretical enforceability of property rights was factually more important than their practical enforcement
Fabrication and characterization of cryogenically grown granular aluminum films
No full textThe thesis investigates the fabrication and superconducting properties of cryogenically grown granular aluminum thin films, with a focus on understanding how deposition parameters influence superconductivity. Using thermal evaporation techniques across a wide substrate temperature range, including an unprecedented 25 K regime, the study systematically correlates growth conditions (such as evaporation rate and oxygen flow) with film morphology and superconducting behavior. A dome-like dependence of the superconducting transition temperature (Tc) on resistivity is observed and refined, with 25 K films exhibiting the highest Tc and sharpest dome. Additionally, contrary to conventional assumptions, a clear dependence of Tc on film thickness is found, revealing enhanced superconductivity in thinner films. The work provides a detailed phase diagram, explores paraconductivity, and examines Kondo-like effects, establishing granular aluminum as both a platform for studying disorder-driven superconductivity and a promising material for quantum technologies
Electrical impedance-based tissue classification for bladder tumor differentiation
No full textIncluding sensor information in medical interventions aims to support surgeons to decide on subsequent action steps by characterizing tissue intraoperatively. With bladder cancer, an important issue is tumor recurrence because of failure to remove the entire tumor. Impedance measurements can help to classify bladder tissue and give the surgeons an indication on how much tissue to remove. Over the years of research, it became obvious that electrical impedance spectroscopy is a very promising tool for tissue differentiation, but also a very sensitive one. While differentiation in preliminary studies shows great potential, challenges arise when transferring this concept to real, intraoperative conditions, mainly due to the influence of preoperative radiotherapy, possibly different tumor types, and mechanical tissue deformations due to peristalsis or unsteady contact force of the sensor. This work proposes a patient-based classification approach that evaluates the distance of an unknown measurement to a healthy reference of the same patient, essentially a relative classification of the difference in impedance that is robust against inter-individual differences and systematic errors. A diversified dataset covering multiple disturbance scenarios is recorded. Two alternatives to define features from the impedance data are investigated, namely using measurement points and model-based parameters. Based on the distance of the feature vector of a unknown measurement to a healthy reference, a Gaussian process classifier is trained. The approach achieves a high classification accuracy of up to 100% on noise-free impedance data recorded under controlled conditions. Even when the differentiation is more ambiguous due to external disturbances, the presented approach still achieves a classification accuracy of 80%. These results are a starting point to tackle intraoperative bladder tissue characterization and decrease the recurrence rate.Projekt DEALUniversität Stuttgar
Empirical risk minimization in the interpolating regime with application to neural network learning
No full textA common strategy to train deep neural networks (DNNs) is to use very large architectures and to train them until they (almost) achieve zero training error. Empirically observed good generalization performance on test data, even in the presence of lots of label noise, corroborate such a procedure. On the other hand, in statistical learning theory it is known that over-fitting models may lead to poor generalization properties, occurring in e.g. empirical risk minimization (ERM) over too large hypotheses classes. Inspired by this contradictory behavior, so-called interpolation methods have recently received much attention, leading to consistent and optimally learning methods for, e.g., some local averaging schemes with zero training error. We extend this analysis to ERM-like methods for least squares regression and show that for certain, large hypotheses classes called inflated histograms , some interpolating empirical risk minimizers enjoy very good statistical guarantees while others fail in the worst sense. Moreover, we show that the same phenomenon occurs for DNNs with zero training error and sufficiently large architectures.Projekt DEA
Bridging cognitive and deep learning models of attention
No full textNeural attention mechanisms, drawing inspiration from the cognitive modeling of human attention, have led to significant advancements in deep learning models across the fields of computer vision (CV) and natural language processing (NLP) (Gupta et al., 2021). Despite these technological strides, AI models still fall short of human performance in tasks demanding nuanced comprehension (e.g., reading comprehension), as well as in out-of-the-box data domains and novel modalities (Sarker, 2021). The goal of this dissertation is to bridge human and data-driven models of attention to enhance the performance of neural systems for CV and NLP tasks.
We hypothesize that the human–machine performance gap is due to a lack of adequate human-like attention functionalities in AI systems, given the relationship between attention functionality and task performance in humans (Pashler et al., 2001). To address this gap, we focus on three aspects that currently hamper the performance of attention-based deep neural networks (DNNs) (Kotseruba et al., 2016). First, the lack of interpretability, obscuring our knowledge of how these models process and prioritize information. Second, the challenge of generalizability across datasets and domains. Third, the substantial data dependency, hindering the development and scalability of the models for certain tasks. We explore if we can mitigate these issues by integrating DNNs with cognitive models of attention, especially for the tasks of reading and scene perception, where human attention has been widely studied and where DNNs fall short of human capabilities (Das et al., 2017; Mathias et al., 2021).
Accordingly, the manuscript develops along three research questions. The first is: What is the relationship between neural and human attention? Focusing on reading comprehension tasks, we uncover correlations between models and human-like attention on reading comprehension tasks. Our findings demonstrate that: a closer alignment with human attention patterns can in fact significantly improve DNNs task performance in both mono- and multimodal settings; that there is a trade-off between model complexity and attention-based interpretability; and that specifically text attention is significantly correlated to model accuracy.
Second, we ask: How does incorporating cognitive theories of attention into DNNs enhance model generalizability? We illustrate that using cognitive simulations as an inductive bias, along with specialized training, effectively compensates for the absence of human ground truth attention data in novel domains. We pioneer a method (known as deep saliency prediction (Wang et al., 2021)) to initiate training a DNN for visual saliency prediction by using cognitive model simulations as an inductive bias. Our text and image saliency models, informed by generalized eye movement behaviors simulated from cognitive models, are further refined with limited eye-tracking data, achieving significant performance improvements comparable to the state of the art across various domains and datasets.
Lastly, our third research question is: Can methods informed by cognitive models of attention effectively mitigate data dependency requirements? We apply our saliency prediction model in mono- and multimodal NLP tasks using a novel joint semi-supervised training method: we generate task-specific human-like attention by training our downstream task models and allowing for gradient flow in the saliency prediction model. Hence, we supervise neural attention layers of different downstream DNNs with different saliency predictions from the same model. This way, by supervising neural attention mechanisms with human-like attention, and jointly training both models for a given task end-to-end, we circumvent the need for task-specific human data.
Put together, our studies set forth a structured approach towards addressing key limitations of current data-driven deep learning models of attention. This thesis demonstrates that integrating them with cognitive science frameworks of human attention opens up new research possibilities, allowing to obtain models that are more efficient, more aligned with human cognitive processes, and that better perceive and understand the world in a human-like manner.Neuronale Aufmerksamkeits - Attention - Mechanismen, die sich an der kognitiven Modellierung der menschlichen Aufmerksamkeit orientieren, haben zu erheblichen Fortschritten bei Deep-Learning-Modellen in den Bereichen Computer Vision (CV) und Natural Language Processing (NLP) geführt (Gupta et al., 2021). Trotz dieser technologischen Fortschritte bleiben KI-Modelle bei Aufgaben, die ein nuanciertes Verständnis erfordern (z. B. Leseverständnis), sowie bei unkonventionellen Datendomänen und neuartigen Modalitäten immer noch hinter der menschlichen Leistung zurück (Sarker, 2021). Das Ziel dieser Dissertation ist es, menschliche und datengesteuerte Modelle der Aufmerksamkeit zu verbinden, um die Leistung neuronaler Systeme für CV- und NLP- Aufgaben zu verbessern. Wir stellen die Hypothese auf, dass die Leistungslücke zwischen Mensch und Maschine auf das Fehlen adäquater, menschenähnlicher Aufmerksamkeitsfunktionen in KI-Systemen zurückzuführen ist, wenn man die Beziehung zwischen Aufmerksamkeitsfunktionalität und Aufgabenleistungperformanz beim Menschen betrachtet (Pashler et al., 2001). Um diese Lücke zu schließen, konzentrieren wir uns auf drei Aspekte, die derzeit die Leistung von aufmerksamkeitsbasierten tiefen neuronalen Netzen (Deep Neural Networks, DNNs) behindern (Kotseruba et al., 2016). Erstens, die mangelnde Interpretierbarkeit, die unser Wissen darüber, wie diese Modelle Informationen verarbeiten und priorisieren, verschleiert. Zweitens, die Herausforderung der Verallgemeinerbarkeit über Datensätze und Domänen hinweg. Drittens, die erhebliche Datenabhängigkeit, die die Entwicklung und Skalierbarkeit der Modelle für bestimmte Aufgaben behindert. Wir untersuchen, ob wir diese Probleme durch die Integration von DNNs mit kognitiven Modellen der Aufmerksamkeit entschärfen können, insbesondere für die Aufgaben des Lesens und der Szenenwahrnehmung, bei denen die menschliche Aufmerksamkeit umfassend untersucht wurde und bei denen DNNs hinter den menschlichen Fähigkeiten zurückbleiben (Das et al., 2017; Mathias et al., 2021). Dementsprechend entwickelt sich die Arbeit entlang dreier Forschungsfragen.
Die erste lautet: Wie ist die Beziehung zwischen neuronaler und menschlicher Aufmerksamkeit? Indem wir uns auf Leseverständnisaufgaben konzentrieren, decken wir Korrelationen zwischen Modellen und menschlicher Aufmerksamkeit bei Leseverständnisaufgaben auf. Unsere Ergebnisse zeigen, dass eine engere Angleichung an menschliche Aufmerksamkeitsmuster die Leistung von DNNs sowohl in mono- als auch in multimodalen Umgebungen erheblich verbessern kann, dass es einen Kompromiss zwischen Modellkomplexität und aufmerksamkeitsbasierter Interpretierbarkeit gibt und dass insbesondere die Textaufmerksamkeit signifikant mit der Modellgenauigkeit korreliert ist.
Zweitens fragen wir: Wie kann die Einbeziehung kognitiver Theorien über Aufmerksamkeit in DNNs die Verallgemeinerbarkeit von Modellen verbessern? Wir zeigen, dass
die Verwendung kognitiver Simulationen als induktiver Bias zusammen mit spezialisiertem Training das Fehlen menschlicher Ground-Truth-Daten zur Aufmerksamkeit
in neuartigen Domänen wirksam kompensiert. Wir führen eine Methode ein (bekannt als Deep Saliency Prediction (Wang et al., 2021), um ein DNN für die visuelle Salienzvorhersage zu trainieren, indem wir kognitive Modellsimulationen als induktiven Bias verwenden. Unsere Text- und Bildsalienzmodelle, die durch verallgemeinerte Augenbewegungsverhaltensweisen, die von kognitiven Modellen simuliert werden, informiert werden, werden mit begrenzten Eye-Tracking-Daten weiter verfeinert und erreichen signifikante Leistungsverbesserungen, die mit dem Stand der Technik in verschiedenen Domänen und Datensätzen vergleichbar sind. Unsere dritte Forschungsfrage lautet schließlich: Können Methoden, die auf kognitiven Aufmerksamkeitsmodellen beruhen, die Anforderungen an die Datenabhängigkeit wirksam abschwächen? Wir wenden unser Salienzvorhersagemodell in mono- und multimodalen NLP-Aufgaben an, indem wir eine neuartige Joint Semisupervised Trainingsmethode verwenden: Wir erzeugen aufgabenspezifische, menschenähnliche Aufmerksamkeit, indem wir unsere nachgelagerten Aufgabenmodelle trainieren und einen Gradient Flow im Salienzvorhersagemodell zulassen. Daher überwachen wir die neuronalen Aufmerksamkeitsschichten verschiedener nachgeschalteter DNNs mit unterschiedlichen Salienzvorhersagen aus demselben Modell.
Durch die Überwachung neuronaler Aufmerksamkeitsmechanismen mit menschenähnlicher Aufmerksamkeit und das gemeinsame Training beider Modelle für eine gegebene Aufgabe von Anfang bis Ende umgehen wir so die Notwendigkeit aufgabenspezifischer menschlicher Daten. Zusammengenommen stellen unsere Studien einen strukturierten Ansatz zur Überwindung der wichtigsten Einschränkungen aktueller datengesteuerter Deep-Learning-Modelle der Aufmerksamkeit dar. Diese Arbeit zeigt, dass die Integration dieser Modelle mit kognitionswissenschaftlichen Modellen der menschlichen Aufmerksamkeit neue Forschungsmöglichkeiten eröffnet, die es ermöglichen, Modelle zu erhalten, die effizienter sind, besser auf menschliche kognitive Prozesse abgestimmt sind und die Welt auf eine menschenähnliche Weise besser wahrnehmen und verstehen
Service-oriented integration of blockchain smart contracts
No full textBlockchains are distributed systems that enable autonomous entities to collaborate without mutual trust or trusted third parties. To enable this, a blockchain maintains its state in a replicated, append-only ledger that only allows changes via agreed-upon transactions approved by means of a consensus mechanism. Although early blockchains aimed to enable self-governing digital money at a global scale, modern blockchains have expanded their applicability to other domains, such as identity management, international trade, and supply chains. This is facilitated by the introduction of smart contracts, which are immutable programs deployed directly on blockchains and executed by the network nodes in a provable manner. They enable enterprises to execute the logic that governs their collaborations reliably without depending on intermediaries.
The emerging blockchain business opportunities have motivated enterprises to integrate them with their processes and systems. However, specific technical challenges complicate this task. For example, the interactions between client applications and blockchains have unclear quality of service. Further, many business scenarios require integrating with services hosted across multiple blockchains. However, blockchains lack standardization, are heterogeneous, and do not readily support cross-chain transactions.
To facilitate the integration of blockchains into enterprise applications, the presented thesis proposes to interpret blockchain-based services as regular services in a service-oriented architecture, which allows the adoption of existing Web services concepts to tackle the integration challenges at hand. In particular, this work presents specifications to enable the uniform description, addressing, and invocation of smart contracts. Furthermore, it analyzes the reliable messaging guarantees of blockchains and their ability to run atomic transactions that can be executed in parallel safely, and introduces a uniform approach to handle the problem of blockchain forking. Moreover, the thesis analyzes the characteristics of existing approaches that support distributed business transactions across multiple blockchains and introduces a novel approach that enables executing such transactions in an atomic and serializable manner. To enable the integration of blockchains into business processes this work introduces an approach that supports designing blockchain-aware process models and executing them using standard-compliant engines.
Finally, the SOSI method combines the aforementioned contributions to enable service-oriented integration of smart contracts and other blockchain-based services. The technological support for using the SOSI method is facilitated with a comprehensive toolchain, which is implemented prototypically to prove the practical feasibility of the introduced concepts
Hydrogeodesy : a Bayesian perspective
No full textWhile historically focused on local scales, modern hydrologic studies have increasingly adopted a global perspective, recognizing water as a finite resource and the interconnection between regions. This global perspective puts hydrology within the water cycle framework, offering a comprehensive view of water dynamics across regions and scales. Despite this framework’s conceptual clarity, accurately quantifying the global water cycle remains challenging due to the complexity of capturing localized and large-scale patterns, variations in topography, climate, and land use, as well as temporal variability. These complexities hinder comprehensive measurements, resulting in knowledge gaps around key water cycle components, including river discharge, surface water storage, soil moisture dynamics, and subsurface water storage and flow.
Inspired by the existing knowledge gaps in the water cycle, an emerging field known as Hydrogeodesy comes to the forefront. Hydrogeodesy is the discipline that uses terrestrial and primarily spaceborne geodetic data, both geometric and gravimetric, to support global water cycle quantification. Utilizing technologies such as satellite altimetry, gravimetry, imaging, InSAR, GNSS, and GNSS-Reflectometry, hydrogeodesy offers direct or indirect measurements of key water cycle components, including terrestrial water storage, and river discharge, significantly advancing our understanding of water dynamics. Despite advancements in spaceborne geodetic sensors, hydrogeodesy faces challenges such as limitations in the spatiotemporal resolution of satellite measurements, measurement uncertainties, unobserved variables, inconsistencies in background models, and the difficulty of separating aggregated measurements. Possible solutions to these challenges involve combining different data types, including satellite, ground-based observations, and model outputs, to benefit from their complementary strengths. However, this presents its own challenges, as it requires reconciling datasets with varying resolutions, accuracies, and temporal scales.
To address some of the challenges listed above, Bayesian approaches offer viable solutions by providing probabilistic interpretations and uncertainty quantification. Bayesian approaches offer a robust framework for updating prior knowledge with new data to yield a posterior distribution, enabling a probabilistic interpretation and explicit uncertainty estimation of parameters. This is especially valuable in hydrogeodesy, where parameters like river discharge, soil moisture, and groundwater storage are often estimated indirectly and carry substantial uncertainties. This habilitation thesis provides a foundational discussion on Bayesian modeling and statistics and demonstrates the versatility and power of Bayesian methods in enhancing our understanding of water cycle components by presenting three distinct Bayesian applications in hydrogeodesy.
The first study applies a Bayesian approach, specifically the Kalman filter, to estimate river discharge using spaceborne geodetic measurements. In hydrogeodetic studies, the Kalman filter and dynamic systems are especially valuable, as they enable the integration of multiple data sources and the continuous updating of estimates with incoming measurements. This is particularly beneficial for river systems, which inherently function as a dynamic system. To assess this potential, a method is introduced that uses the cyclostationary properties of discharge as prior information, while observed altimetric discharge data provide the likelihood. Together, these yield a posterior providing an unbiased daily discharge estimate. The method is applied to the Niger River basin and its main tributaries and validated against in situ data from 18 gauges. Results show a high average Correlation Coefficient (CC) of 0.9 and an average relative Root Mean Squared Error (RMSE) and bias of 15%. This method effectively estimates daily river discharge across entire basins and shows promise for global application, especially in data-scarce regions. With satellite altimetry data from multiple virtual stations and historical discharge data, daily discharge estimates with an error under 20% could be attainable in many river basins worldwide. The growing availability of spaceborne geodetic data, such as that provided by SWOT, further enhances this potential by delivering comprehensive measurements of river height and width, along with global discharge estimates.
In most real-world applications, including hydrogeodesy, the Gaussianity assumption required by the Kalman filter does not hold, limiting its applicability. Inspired by this challenge, and motivated by the need to overcome the limitations of the poor spatial resolution of the GRACE and GRACE-FO missions, the second study proposes a Bayesian method to downscale GRACE data, proposing a nonparametric method to infer the posterior distribution directly, without any assumption for the likelihood or posterior. The prior distribution is obtained based on GRACE data values using the monthly variation of GRACE data. To model the likelihood functions, copulas are employed to capture dependencies among multivariate distributions. Monthly empirical copulas are constructed and fitted to analytical copulas, conditioned on specific quantile values, reflecting the dependency between GRACE and fine-scale data. A key advantage of this copula-supported Bayesian approach is its capacity to represent uncertainties in both data and models, even with variable input quality.
The proposed downscaling approach is applied to the Amazon Basin, utilizing four different fine-scale datasets: WGHM, PCR-GLOBWB, SURFEX-TRIP, and the ensemble of flux data and soil moisture data from GLEAM and ASCAT. Validation is conducted against two independent datasets: space-based Surface Water Storage Change (SWSC) and GPS-observed Vertical Crustal Displacement Change (VCDR). In SWSC validation, downscaled results capture spatial variations in river storage with high CC and a relative RMSE of 26%. VCDR validation involves two analyses: comparing GPS-VCDR with TWSF-based VCDR using Green’s function convolution, where downscaled products yield RMSE values between 2.27 and 5.65 mm/month, outperforming input fine-scale data with 14 mm/month RMSE. In terms of CC, downscaled results achieve an average value of -0.81 versus -0.73 for the input. The proposed Bayesian framework effectively downscales GRACE data, with performance highly dependent on input data quality. The copula-supported Bayesian approach offers valuable uncertainty quantification even with inconsistent input data. This method aids in understanding water storage variations in small catchments, supporting local hydrological studies, and can be applied to other water cycle parameters as an alternative to traditional methods.
Although a direct posterior is obtained for each grid cell in the downscaling study, spatial dependencies among neighboring grid cells are not considered. Graphical models are particularly well-suited for capturing such spatial dependencies. To address this limitation - and inspired by the challenge of noisy water level estimates from satellite altimetry over inland water bodies - the third study presents a Bayesian approach that formulates a probabilistic graphical model known as a Markov Random Field (MRF), with a Maximum A Posteriori estimation of the MRF (MRF-MAP) as the objective. There to improve inland altimetry, a retracking method is proposed. Unlike conventional retracking methods that target a single waveform point, a holistic approach by identifying retracking lines within 2D radargrams, treating the radargram as a segmented image. This segmentation divides the radargram into Front and Back segments, resembling a binary image segmentation task. The proposed MRF-MAP framework uses spatial dependencies as prior information, with the likelihood based on the temporal evolution of pixel labels across groundtrack cycles. Two temporal energy functions are applied: 1D, based solely on pixel intensity, and 2D, which includes both intensity and bin values, with the posterior probability maximized using the maxflow algorithm. The maxflow algorithm is then applied to obtain MAP solution, yielding a segmented radargram where the retracking line is defined as the boundary between segments.
The proposed retracker method is applied to both pulse-limited and SAR altimetry datasets across nine U.S. lakes and reservoirs with varying altimetry characteristics. Validated against in situ data, the proposed method improves RMSE by approximately 0.25 m with the 1D temporal energy function and 0.51 m with the 2D function. The main advantage of the proposed method is its robustness against unexpected waveform variations, making it especially valuable for complex radargrams where conventional retrackers often deliver outliers. By integrating both spatial and temporal information, this method offers a more comprehensive understanding of the data and has broad applicability, such as improving the classification of SWOT pixel cloud points by incorporating spatiotemporal detail.
Through these case studies, the thesis illustrates the advantages of Bayesian approaches in improving the accuracy and reliability of hydrological estimates - such as river discharge, terrestrial water storage, and water level measurements - derived from spaceborne geodetic sensors. By integrating theoretical insights with practical applications, the thesis demonstrates how Bayesian methods can effectively improve spatiotemporal resolution, obtain uncertainties, enhance data fusion, and accommodate the complexities inherent in hydrological systems. This combination of foundational knowledge and real-world examples shall establish a base for advancing the use of Bayesian approaches in hydrogeodetic research and beyond. Moreover, by highlighting the challenges in hydrogeodesy, this thesis provides a clear direction for future research and development in the field. It emphasizes critical areas requiring attention, such as improving the spatial and temporal resolution of hydrological estimates, addressing inherent uncertainties in geodetic observations, and developing more effective methods for assimilating diverse data sources. The thesis encourages the refinement of geodetic data processing techniques and the adoption of probabilistic frameworks, such as Bayesian modeling, in future work. Building upon the work presented here, future studies can ultimately achieve more accurate and reliable insights into the Earth's systems
Lossfunction für Physics-Informed Machine Learning in Untergrundströmungen
No full textPhysics-Informed Machine Learning (PIML) methods for solving complex nonlinear partial differential equations (PDEs) have recently gained popularity in the simulation sciences. Unlike purely data-driven approaches, PIML integrates prior knowledge about the underlying physical system - expressed through the governing PDEs - into the learning process. This is done via a loss function that includes the PDE residual, thereby penalizing model outputs that do not fulfill the PDE. Hence, PIML methods are particularly appealing for scenarios with limited data availability, where extensive measurements or numerical simulation costs are prohibitive.
However, purely physics-informed models often suffer from various pathologies, rendering purely physics-informed learning ineffective. One major challenge is the complex loss landscape introduced by the PDE residual. This work examines how simplifying the governing PDE can enhance training performance or enable physics-informed learning in cases where it would otherwise be infeasible.
Building on the approach proposed by Piller for extending predictions of heat plumes generated by Groundwater Heat Pumps (GWHP), this study verifies the effectiveness of such PDE simplifications. Concretely, this work finds a moderate monotonic correlation (Spearman: 0.59) between the simplified PDE residual and the data loss, indicating that the simplification of the governing PDEs preserves enough of the physics to be useful while making training more tractable. To this end, Physics-Informed Neural Networks for Heat Plume Extension (HPE-PINN) and Physics-Informed-Neural Operators for Heat Plume Extension (HPE-PINO) are developed and compared against Piller's model, which makes use of Singular Value Decomposition (SVD) to reduce the dimensionality of the solution space. Throughout this process, several common PIML pathologies are encountered. A suite of techniques to mitigate their negative effects on training is presented, implemented, and validated to improve training stability and model performance. Comprehensive ablation studies highlight the effectiveness of normalization and hard enforcement of initial and boundary conditions in enhancing convergence, as well as the importance of Fourier feature embeddings to reduce spectral bias.Die Disziplin des Physics-Informed Machine Learning (PIML) zur Lösung nichtlinearer partieller Differenzialgleichungen (PDE) hat in den Simulationswissenschaften in jüngster Zeit an Bedeutung gewonnen. Im Gegensatz zu rein datengetriebenen Ansätzen integriert PIML Vorwissen über das zugrundeliegende physikalische System - ausgedrückt durch die maßgebenden PDEs - in den Lernprozess. Dies geschieht über eine Lossfunction, die das PDE-Residuum beinhaltet und somit Modellausgaben bestraft, die die PDE nicht erfüllen. Daher sind PIML-Verfahren besonders gut geeignet, wenn die Akquisition von Daten durch umfangreiche Messungen oder numerische Simulationen zu teuer ist.
Allerdings leidet rein physikgetriebenes Lernen oft unter verschiedenen Pathologien, die den Lernprozess verlangsamen oder ganz verhindern. Ursächlich hierfür ist unter anderem die komplexe Topologie der Lossfunction, die durch das PDE-Residuum bedingt ist. Diese Arbeit untersucht, wie eine Vereinfachung der maßgeblichen PDE den Lernprozess beschleunigen oder überhaupt erst ermöglichen kann.
Eine solche Vereinfachung wurde zuletzt von Piller für die Fortsetzung von Vorhersagen von Wärmefahnen, die von Grundwasser-Wärmepumpen (GWHP) erzeugt werden, vorgeschlagen und auch erfolgreich implementiert. Diese Studie bestätigt die Wirksamkeit solcher PDE-Vereinfachungen. Konkret wird eine monotone Korrelation (Spearman: 0,59) zwischen dem Vereinfachten Residuum und dem Datenverlust festgestellt. Dies deutet darauf hin, dass die Vereinfachung der maßgeblichen PDEs genug der Physik bewahrt, um gute Approximationen zu liefern und gleichzeitig den Lernprozess zu beschleunigen. Für die Verifikation der PDE-Vereinfachung werden Physics-Informed Neural Networks for Heat Plume Extension (HPE-PINN) und Physics-Informed-Neural Operators for Heat Plume Extension (HPE-PINO) entwickelt und mit Pillers singulärwertbasiertem Modell verglichen.
Während der Entwicklung von HPE-PINN und HPE-PINO werden häufige PIML-Pathologien festgestellt. Um die Trainingsstabilität und Approximationsgüte der Modelle zu verbessern, wird eine Reihe von Techniken zur Reduzierung der negativen Auswirkungen der Pathologien auf den Lernprozess vorgestellt, implementiert und validiert. Umfassende Ablationsstudien unterstreichen die Wirksamkeit der Normalisierung und des Hard Enforcement von Anfangs- und Randbedingungen zur Verbesserung der Konvergenz sowie die Bedeutung von Fourier Feature Embeddings zur Reduzierung des Spectral Bias