This thesis was carried out on a concept vehicle equipped with four-wheel steering and bidirectionaldriving. The work was divided into two parts: vehicle dynamics modeling and pathplanning for autonomous navigation functionality.

For the vehicle dynamics study, two single-track (bicycle) models were developed: a conventionalmodel with a linear tire assumption and a modified model that included load transfer anda nonlinear Pacejka tire model. Their performance was evaluated for vehicle states such as yaw,yaw rate, and lateral velocity, using experimental data collected from specific maneuvers. Rootmean square error (RMSE) was used to assess accuracy, and three two-way ANOVA studies(one for each state) were conducted to analyze the effects of model type and maneuver type onestimation accuracy. The results show that the conventional model generally performs betterfor yaw and yaw rate estimation, while the modified model improves lateral velocity estimation, particularly in maneuvers where load transfer is important. The ANOVA analysis showedthat maneuver type has a significant influence on estimation accuracy across all vehicle states.

For autonomous navigation study, three path planning algorithms were developed: StandaloneA-star, standalone Rapidly exploring Random Tree (RRT) and hybrid A-star and RRT algorithms.Their performance was assessed using three different simulation environments: staticobstacles environment, dynamic obstacles environment and hybrid environment. These simulationswere developed using Gazebo software. The data collected from each trial was examinedthrough multiple case studies. The results show that the hybrid algorithm planner performswell in all simulation environments.

The aim of the thesis project was to explore the possibility of architectural interventions inheritage sites, looking at a disused mill for the production of olive oil in the village Olivetta SanMichele in Liguria. Using digital tools for the surveying of the site, the project also explored thepossibilities of digital fabrication of historical form.

In railway project appraisal, the train traffic and complementing and competing modes’ supplies are one of the most important factors for estimation of travel demand and hence project benefits. Today, the forecast timetables that the Swedish Transport Administration (Trafikverket) uses for passenger forecasts are based on past developments, supplemented by an expert assessment of expected future supply developments. There are some doubts as to whether the method is sufficiently objective and consistent between different objects to provide a fair planning basis. The aim of this study is to develop a data-driven method for the supply that can be used as a starting point for travel forecast generation.

In this study, a model, Multi-Task Heterogeneous Graph Attention Neural Network (MT-HGATNN), was developed and applied to forecast train supply across multiple train categories and line sections, or station pairs. By leveraging structured timetable data, socioeconomic inputs, and scenario-based forecasts, the model provides accurate and interpretable insights into both current and future rail demand.

The model performs well in both retrospective validation (2023 data) and prospective forecasting (base year 2045), with model training on and a case study of East Coast line (Ostkustbanan). It is offering a framework for strategic transport planning under uncertainty. The multi-task learning approach enables joint modelling of multiple train categories, improving parameter efficiency and allowing the model to exploit latent interdependencies between tasks. The ability to integrate temporal granularity (hourly slots), spatial structure (station pair-level analysis), and diverse exogenous variables (e.g., GDP, car ownership, fares) further strengthens its applicability in complex real-world settings.

There are several promising directions for extending and deepening the work presented in this study. One important step would be to evaluate the generalisability of the developed MT-HGATNN model by applying it to other rail corridors or potentially at the national network level. This would help assess the model’s robustness across different geographical contexts and service supplies.

In the life sciences, modern state-of-the-art super-resolution imaging systems are invaluable tools for observing and better understanding life at length scales previously unobservable. One such system, leveraging a parallelized REversible Saturable OpticaL Fluorescence Transitions imaging technique using reversibly switchable fluorescent proteins, called Molecular Nanoscale Live Imaging with Sectioning Ability, abbreviated as MoNaLISA, was developed by the Advanced Optical Bioimaging group at the Science for Life Laboratory in Stockholm, Sweden, to enable prolonged super-resolution imaging of samples over a large field of view.

During imaging with MoNaLISA, a stack of raw frame data is collected and reconstructed into a single super-resolved image using the Python-based software platform called ImSwitch. For time-lapse recordings, thousands of raw frames are collected across multiple stacks. To monitor the sample dynamics over time lapses, either manual reconstructions of each stack as they are acquired or a single reconstruction after all stacks have been acquired must be performed, both of which are time-consuming and tedious, and leave the user with no real-time feedback. The primary goal of this thesis work was to address these issues by developing software that can be integrated into the MoNaLISA system and perform real-time reconstructions on individual raw frames, thereby enabling real-time feedback. 

The software was developed with the help of existing software within the ImSwitch platform, based on MoNaLISA's operating procedure, and leverages the GPU-accelerated library CuPy to perform fast reconstructions. The resulting software can process individual raw frames in 0.4 ms to 1.1 ms, with an average time of approximately 0.7 ms, which is below the MoNaLISA system's approximate raw frame acquisition speed of 1.6 ms to 3.5 ms, and with reconstruction quality comparable to that of the ImSwitch reconstructions. This, in turn, implies that real-time reconstruction with satisfactory image quality is achievable.

This thesis consists of three papers, an individual summary of each paper, and an introduction. The papers are all related to existence, uniqueness, or regularity theory of local and nonlocal partial differential equations (PDEs).

In Paper A, we establish uniqueness for viscosity solutions of the inhomogeneous nonlocal infinity Laplace equation Lu = f, where the right hand side f is a bounded, continuous, and nonpositive function. Uniqueness is proven through a comparison principle.

In Paper B, we use Perron's method to construct viscosity solutions to the equation ∂u/∂t = L u in Ω, and u = g in the complement.

In Paper C we study regularity of a minimizer of the expression J(u) := ∫ F(∇u) dx, where F(x) is a strongly convex function whose second derivatives might jump at |x| = 1. The specific form of F gives rise to a free boundary Γ, and the resulting Euler-Lagrange equation varies over Γ. In this paper we only consider two-phase flat points. We show that under some regularity and non-degeneracy assumptions the asymptotic expansion of a minimizer u can be written as u(x) = a + ν · x + p(x) + q(x), where a ∈ R, ν ∈  R^n. The function p is a broken polynomial that is defined as a C^1 function consisting of one polynomial in the upper half space and another polynomial in the lower half space, and the function q is a rest term. We derive the PDEs that are satisfied by p and q, respectively, and show many regularity properties for the terms in the expansion. This paper is intended to be the first part of a project that aims at establishing regularity of the free boundary Γ.

Celler lagrar fett i lipiddroppar som utvecklas från strukturer i det endoplasmatiska retiklet för framtida behov. Mekanismerna genom vilka små reglerande proteiner påverkar transport och reaktioner i membran hos stressade tumörceller är i stor utsträckning okända. I denna studie undersöks hur det hypoxi-inducerbara, lipiddropps-associerade proteinet (HILPDA) samverkar med tre enzymer som styr riktningen på triacylglycerolflödet: adipos triglyceridlipas (ATGL) samt diacylglycerol O-acyltransferas 1 och 2 (DGAT1 och DGAT2). Vi använde beräkningsbaserad strukturprediktion tillsammans med inbäddning i membran och korta molekylära simuleringar för att skapa och pröva realistiska modeller i fosfatidylkolin-bilager och lysophosfatidylkolin-miceller. Modellerna placerar HILPDA vid ingången till den katalytiska klyftan hos ATGL, där åtkomsten till det katalytiska centret blockeras utan att centret ockuperas, förenligt med en ingångsblockerande, icke-kompetitiv verkan. För DGAT1 förkastades poser som tränger in i den intramembrana kammaren; i stället visar modellerna en perifer, cytosolisk/membranvänd bindning som bevarar kammaren och dess laterala tillträdesvägar, i linje med rapporterad aktivering av HILPDA. För DGAT2 framträder två återkommande laterala lägen: ett som tillfälligt blockerar en port till det katalytiska stället (inhiberande) och ett som ligger på avstånd från portarna och fungerar som en aktiverande “ställning”, med högre flexibilitet i gränssnittet än i de andra komplexen. Membranrekonstruktionerna återgav kända särdrag hos de tre enzymerna samt en N-terminal domän i HILPDA som förankras i membranet. Sammantaget etablerar resultaten en membrankänslig strukturell ram: proteinet stoppar ATGL, håller DGAT1 aktivt från periferin och påverkar DGAT2 på ett kontextberoende sätt. Arbetsflödet—strukturprediktion följt av membraninbäddning och kort simulering—utgör en generell metod för att identifiera gränssnitt mellan mikroproteiner och enzymer vid organellernas kontaktzoner.

Extracellular vesicles, particularly exosomes, have attracted increasing attention in recent years due to their roles in intercellular communication in both healthy and diseased cells. They are increasingly recognized as mediators in cancer, as tumor-derived exosomes promote cancer progression and metastasis through the transport of oncogenic cargo to both healthy cells and other cancer cells. This discovery has led researchers to investigate the potential of exosomes as carriers of biomarkers for use in non-invasive cancer diagnosis and prognosis. Commonly diagnosed cancers that still rely on invasive procedure for diagnosis, such as colorectal cancer, could benefit from such approaches. Both exosome research and its clinical applications are currently hindered by the lack of standardization in exosome isolation methods. Therefore, the aim of this study was to establish a protocol for isolation and characterization of exosomes from the colon cancer cell line SW480 and mouse plasma to create a foundation for further studies of exosome content and its role in colorectal cancer. This was done using a commercial kit which isolates exosomes by polymer precipitation, followed by characterization using nanoparticle tracking analysis and western blotting. NTA results showed that the size distribution of SW480 and mouse plasma exosomes was consistent with previous studies, and western blot analysis confirmed the presence of exosomal surface markers in the isolates. However, the presence of cellular contaminants was also observed in the exosome preparations, reflecting the main disadvantage of precipitation-based methods. Additionally, RNA was extracted from the exosomes and analysis revealed an RNA profile consistent with previous studies of exosomal RNA, supporting successful isolation. Moreover, as estrogen receptor beta expression in the colon has been observed to be protective against colorectal cancer, its effect in SW480 cells on genes involved in exosome production and secretion was studied, as this may represent one of the mechanisms through which estrogen exerts its protective role. However, the results remain inconclusive and need further investigation.

In conclusion, this study led to the establishment of a protocol for successful isolation and characterization of exosomes from colorectal cancer cells and mouse plasma.

Urinary tract infections (UTIs) are one of the most common bacterial infections worldwide, with uropathogenic Escherichia coli (UPEC) as the leading cause. Key mechanisms of UPEC pathogenesis are its ability to invade urothelial cells, form intracellular bacterial communities (IBCs), and undergo filamentation, traits often considered exclusive hallmarks of virulence. In this project, an ex vivo mouse bladder infection model was established and optimized to investigate similarities and differences in these morphological adaptations between the pathogenic strain UTI89 and the non-pathogenic laboratory strain MG1655. Both strains were capable of invasion, IBC formation, and filamentation, although with different frequencies and morphologies. UTI89 generated compact, tightly packed IBCs and did so more efficiently than MG1655, which appeared more dispersed and often wrapped around the host nucleus. MG1655 also displayed a higher abundance of single adherent bacteria scattered across the urothelium. Western blot analysis showed reduced and altered FimH expression in MG1655, compared to UTI89, potentially explaining its strong adhesion capacity but limited ability to invade and develop mature IBCs. Both strains filamented after urine exposure. UTI89 filaments were generally shorter and reverted back to rod-shaped cells more efficiently, whereas MG1655 filaments were longer and tended to undergo lysis instead of successful reversion. These results suggest that phenotypes long considered UPEC-specific can also appear in non-pathogenic strains, although with less efficiency and lower viability. The findings demonstrate that the ex vivo bladder model is a useful tool for demonstrating infection dynamics, as it allows direct comparison of invasion, IBC formation, and filamentation between strains. These results also challenge narrow definitions of uropathogenicity by suggesting that morphological traits traditionally viewed as markers of pathogenicity may in fact represent broader adaptive strategies of E. coli.

Syftet med denna studie var att förbättra den termiska stabiliteten hos frystorkade enzymkulor som används i RT-qPCR-analyser, med fokus på att identifiera nya formuleringar som bibehåller enzymatisk aktivitet under accelererade stressförhållanden. Totalt testades 15 formuleringar, med varierande buffertsystem, pH, hjälpämnen och saltinnehåll. Stabiliteten utvärderades genom enzymaktivitetstester för DNA-polymeras och omvänt transkriptas, kompletterat med funktionell RT-qPCR-testning. 

Resultaten bekräftade att den etablerade referensformuleringen (Ref-1) har sämre stabilitet under stressförhållanden(55°C) , med funktionsbortfall efter cirka 32 dagar, medan de nyframtagningar formuleringarnaför samma enzym visade förbättrad stabilitet. Metionin och Prionex ökade stabiliteten med cirka 20-25 %, trehalos gav en ytterligare förbättring på omkring 11 %, och Brij 58 bevarade ungefär 20 % mer aktivitet jämfört med Tween 20. Buffertkompositionen visade sig också vara en viktig faktor, där formuleringar baserade på HEPES visade bättre stabilitet än referensformuleringen med Alt-Buffer. Dessutom visade resultaten att KCl påverkar enzymaktiviteten negativt och minskar aktiviteten. Jämförelsen mellan enzymaktivitetstesterna och funktionell RT-qPCR tester visade att aktivitetsanalyser är användbara för att följa gradvisa aktivitetsförluster, men att funktionella tester ger en mer rättvisande bedömning av hur analysen faktiskt fungerar. 

Sammanfattningsvis visar resultaten att val av buffertsystem och stabiliserande tillsatser har stor betydelse för stabiliteten hos frystorkade enzymkulor och att rätt kombination av komponenter kan bidra till förbättrad prestanda i RT-qPCR under stressade förhållanden. 

The human cerebral cortex works through an intricate organization of diverse neuronal and glial populations. These are arranged into layers, columns, and long-range circuits that collectively orchestrate perception, learning, and behaviour. Decoding this organization at cellular and molecular resolution is essential for understanding normal cortical function and for explaining how dysregulated gene and protein activities contribute to neurological and psychiatric disorders. Recent advances in single-cell and spatial omics now make it possible to interrogate the human cortex in situ, revealing how transcriptional programs, protein localization, and tissue architecture converge to define cellular identity and state within precise anatomical niches. This project leverages these advances to construct a high-resolution, spatially anchored, multimodal view of the human cortex. Here, we establish a multi-omics framework integrating high-resolution spatial transcriptomics with multiplexed spatial proteomics. The results demonstrate the feasibility of aligning transcriptomic and proteomic layers within the same cortical tissue, providing a proof-of-concept for multi-omics integration. This work establishes a workflow for future studies aiming to extend integrative approaches beyond microglia to other cortical cell types, thereby advancing our understanding of molecular and cellular organization. 

The phosphate fertilizers are a non-renewable resource and global reliance on mining capacities poses significant environmental threats, including landscape destruction, radioactive waste and water pollution through eutrophication. At the same time, Wastewater Treatment Plants aim to lower phosphorus effluent levels. Anaerobic digestion technology for sludge treatment re-solubilise phosphorus that was captured by Enhanced Biological Phosphorus Removal. This study investigates possible integrated solutions to optimize existing technology and implement phosphorus recovery as fertilizers for agriculture. Moreover, bioenergy potential of wastewater sludge could be reached.

Lab-scale, thermophilic, mesophilic and room temperature acidogenic digesters were used to ferment Thickened Waste Activated Sludge (TWAS), mixed with various organic and inorganic substrates, including whey permeate concentrate (WPC), pre-fermented WPC (FWPC), acetic acid (AA), sodium acetate (SA), and real food waste. The results demonstrate that a significant role in phosphorus solubilization plays no chemical acidification, nor biological activity. Acetic acid, which lowered pH to ~4.8, provided carbon source for the process, also achieving over 73% of P release within 6 hours, whereas SA, providing the same carbon source has not shown similar results (40% in 12 hours), keeping near-neutral pH 6.6. Pre-fermented substrates as FWPC and FUSFW consistently outperformed non-fermented ones WAS and USFW, achieving >70% P release by providing immediate VFAs and low pH environment, compared to >50% with untreated. This acidogenic stage also significantly increases soluble COD, indicating a high potential base for methane production. 

The recovered phosphorus in the supernatant was successfully precipitated as struvite and brushite, achieving as high as 97% and 95% recovery, respectfully. Struvite precipitation was optimal at pH 10.0 with P:Mg ratio 1:1.2, while brushite formation was most effective at lower pH as of 8.0 and P:Ca ratio 1:1.2. This research validates a sustainable, circular economy model for WWTPs, where acidogenic co-digestion of WAS with organic wastes, can recover valuable nutrient and simultaneously pre-treat sludge for enhanced biogas production.

Lifespan varies by several magnitudes across the animal kingdom and has evolved many times independently, yet the exact mechanisms underlying this variation remain only partially understood. Previous studies investigating the mechanisms behind longevity have typically focused on pairwise comparisons between long-lived and short-lived animals, which has limited capacity for finding larger evolutionary patterns. With the development of new genomic tools, increased quality of genomes, and more powerful computational resources, the possibility for comprehensive studies across broad animal clades is greater than ever before. In this project, expanded analytical capacity was leveraged to perform a data-driven analysis of long-lived mammals, with the goal of identifying genetic patterns underlying both established ageing mechanisms and novel contributors to longevity, as well as other genomic signatures associated with extended lifespan. The study used genomes from 50  mammals which were chosen based on the top and bottom 20% of available mammals by lifespan. Single copy orthologs and gene families were inferred using Orthofinder and then analyzed for gene family size variation and differentiating evolutionary rates in amino acid sequence. The analysis identified 924 genes with significantly different evolutionary rates between the short-lived and the long-lived species, and the enrichment analysis of these genes showed that the results contained a multitude of genes related to the formation and function of neurons. This demonstrates the genetic mechanisms behind recent studies that have proposed that one of the strongest traits that correlates with lifespan is the number of cortical neurons, with a correlation more than twice as strong as body mass. Additionally, the analysis identified genes related to aging mechanisms such as DNA repair and the immune system. The analysis of the gene families identified 83 gene families that were significantly correlated with longevity, and the enrichment analysis revealed that many of them had functions related to toxic metal ion response. While there were also significant gene families related to more classic longevity mechanisms such as DNA repair and immune response, the metal ion detoxification represented a novel finding with possible theoretical backing for its potential importance to ageing mechanisms.

Urinary tract infections (UTIs) are among the most common infectious diseases in humans, with most cases caused by uropathogenic E. coli (UPEC). The increasing prevalence of antibiotic resistant UPEC has made UTIs progressively harder to treat, resulting in prolonged hospital stays and higher healthcare costs. Alternative therapeutics are therefore needed, which require a deeper understanding of UPEC behaviour during infection. Fluorescently labelled biosensors could be applied to this problem as they provide a means to study bacterial stress responses through production of a GFP signal when transcription is activated, which enables real-time monitoring using high-resolution fluorescence microscopy. In this project, two novel biosensors were engineered and transformed into a model UPEC strain, together with four biosensors already available, to evaluate their functionality during an in vitro infection model of UTIs. The bacteria were first grown under infection-like conditions to assess fluorescence in the absence of bladder cells. Based on these results one biosensor, pSEVA631(sp)-PhdeA-gfpASV (designed to detect acidic stress), was implemented in the infection model. The number of bacteria that emitted GFP increased significantly over time, suggesting elevated acidic stress as the infection progressed. However, whether the GFP signal was exclusively due to acidic stress could not be confirmed, as the hdeA gene is also transcribed during stationary growth. Therefore, further studies are needed to validate the specificity of the pSEVA631(sp)-PhdeA-gfpASV biosensor. Nevertheless, these findings provide a foundation for the future biosensor-based research that may contribute to novel approaches for UTI treatment.

The domestication of horses has profoundly influenced human history, yet genetic analyses of performance-related traits remain limited. Genetic variants downstream of the EDN3 gene have been linked to blood pressure regulation and athletic performance in horses. Here, we prepared lung tissue samples from Standardbred and Przewalski's horses for Hi-C sequencing to investigate three-dimensional chromatin interactions of this genomic region. Using mechanical tissue fragmentation, cell fixation, and a comprehensive Hi-C protocol, we generated sequencing libraries exhibiting appropriate DNA fragment sizes and concentrations for high-quality sequencing. Our results demonstrate successful sample preparation enabling detailed regulatory genomic studies of the EDN3 gene downstream module. These findings provide a foundation for further exploration of regulatory mechanisms influencing blood pressure and performance traits in domesticated horses.

This study examines the coordination of land-use and energy planning, aimingto promote sustainable regional development within the context of the greentransition. The thesis is based on a case study, which is connected to a projectinitiated by the Swedish Energy Agency, “Inter-Municipal Energy and SpatialPlanning in the Green Transition”, involving four municipalities, Boden,Gällivare, Jokkmokk, and Luleå, in Sweden’s Norrbotten region. Utilising aqualitative approach, which includes a literature study, document analysis,observations, and interviews with municipal representatives and the CountyAdministrative Board, the study examines inter-municipal collaborationmechanisms and shared challenges. Findings highlight the essential role ofland use in energy planning and the need for inter-municipal and state-levelcoordination. The study also identifies the complexity arising from multiplestakeholders, national interests, and escalating regional investments.Cooperation across jurisdictions emerges as a key factor in aligning local andregional priorities and presenting unified positions to investors andgovernment bodies. This study seeks to contribute to understanding regionalenergy transition governance and offer insights for improving future planningpractices.

The folding of the genome plays a large role in how cells differentiate. Through advances in chromosome conformation capture technology, new methods have lead to higher resolution results. HiCap makes it possible to analyse the promoter-enhancer interactions at a single interaction resolution, allowing for detailed analysis of promoter-enhancers interactions. Promoter-enhancer interactions are important for understanding certain diseases where enhancer mutations may cause changed chromatin topology or enhancer function. In this project, a metainteractome analysis was performed on HiCap data collected over a decade. The main goal of the study was to see how the promoter-enhancer interactions in cells of different type and in different states differ. 

Due to lack of computational resources, only 18 samples were analysed. It was possible to accurately divide these samples into different groups through a combination of PCA and K-Means Clustering based on their promoter-enhancer interactions. All samples had at least some interactions in common with other samples showing that there are a number of base interactions that all samples have.

 Based on degree distribution, the promoter-enhancer interactions showed that the interactions acted as a scale-free network and not as a normal random distribution.Due to the lack of samples analysed, it was not possible to reach any conclusions and a full scale version of the project is needed. The project also highlighted a need for specialised high memory and low core count HPCs.

Urban Green Infrastructure (UGI) plays a critical role in enhancing urban resilience, mitigating climate impacts, and improving environmental quality. This thesis presents a Deep Learning framework for large-scale UGI mapping using Sentinel-2 satellite imagery and auxiliary building footprint data. A transformer-based Swin UNETR model, enhanced with residual channel attention modules, is developed to perform semantic segmentation of UGI across 75 cities representing five Köppen-Geiger climate zones. Reference labels are generated using NDVI thresholding and refined with global building masks to reduce misclassification in built-up areas. The model achieves strong performance, with an average overall accuracy of 94.35%, precision of 91.72%, recall of 95.15%, F1-score of 93.36%, and IOU of 87.68%. Additional validation using high-resolution manually annotated ground truth in five cities reveals that NDVI threshold-based labels tend to generalise vegetation boundaries, masking segmentation errors in narrow or shaded green features. Visual analyses further highlight variations in performance across climatic contexts, with highest accuracy observed in continental and temperate cities. This study confirms the effectiveness of attention-augmented transformer models for scalable UGI mapping and underscores the importance of integrating contextual spatial data and fine-grained validation for reliable urban vegetation monitoring.

This study examines the relationship between customer satisfaction and individual-level bank revenue growth, drawing on data from 19,060 Swedish retail banking customers that combine survey responses with objective bank records. Furthermore, we investigate whether the impact of satisfaction on revenue growth is sustained over time, specifically one, two three and four years after the measurement of satisfaction, and whether this effect differs across customer satisfaction levels. The results show that higher satisfaction is associated with greater sustained revenue growth, with more pronounced effects for customers in the medium-high and highest satisfaction groups. By contrast, no significant sustained revenue growth is found for customers with low and low-medium satisfaction. The findings do not support the hypothesis of diminishing returns when moving from medium-high to the highest satisfaction levels, although weak indications suggest scope for further exploration. Overall, the findings demonstrate the long-term revenue growth of satisfied customers and emphasize the importance of targeting customers with lower to medium-low satisfaction to enhance overall revenue performance.

Autonomous driving for tractor-trailer vehicles presents significant challenges due to their large size and complex dynamics, especially in urban environments with narrow roads and sharp turns. While data-driven approaches such as imitation learning and reinforcement learning have achieved strong performance for passenger cars, their application to tractor-trailers remains limited by the scarcity of high-quality datasets and the sample inefficiency of reinforcement learning. This thesis addresses these challenges by proposing a novel three-stage framework for autonomous decision-making for tractortrailer vehicles. Firstly, an optimization-based trajectory conversion method transforms passenger car trajectories into feasible tractor-trailer trajectories, ensuring that the converted data respects tractor-trailer kinematics. Secondly, this converted dataset is used to train an imitation learning policy, providing a strong initialization for subsequent online reinforcement learning training. To further improve sample efficiency, a pre-trained passenger car model is also utilized. Thirdly, three approaches for integrating imitation learning and reinforcement learning are investigated: direct fine-tuning, fine-tuning with Kullback-Leibler divergence regularization to constrain policy updates, and simultaneous imitation learning and reinforcement learning training to mitigate catastrophic forgetting. The proposed methods are evaluated using the Waymo Open Motion Dataset and the Waymax simulator, focusing on left and right turning scenarios that are particularly challenging for tractortrailers. The results demonstrate that the trajectory optimization method effectively converts passenger car data into tractor-trailer-compatible data, and that integrating imitation learning and reinforcement learning improves both sample efficiency and policy robustness compared to baseline models. This work contributes a scalable framework for leveraging abundant passenger car data to train tractor-trailer policies, bridging the gap in dataset availability and enabling safer and more efficient autonomous heavy-duty vehicles.

The gaming and animation industries are becoming increasingly demanding and complex, as they strive to create more realistic and immersive media. Modern video games can feature thousands of assets, and producing them all requires significant time and advanced modeling skills. Recent advances in deep learning and generative AI have led to models capable of generating 3D assets from a single image, streamlining the content creation process. Designers can now produce preliminary 3D models from simple sketches or reference images. However, relying on a single view often leads to incomplete or inaccurate representations. Incorporating multiple views can significantly reduce hallucinations and better capture the user’s intent. This project aims to support designers by accelerating the asset creation pipeline, enabling faster ideation and providing solid 3D outputs that require minimal post-editing. While several models exist for multi-view 3D generation, few are opensource or readily usable. One of the most promising and publicly available models at the time of writing is TRELLIS [1]. However, as noted by its authors, the multi-view version of TRELLIS does not always perform optimally across all inputs. This project focuses on improving the performance of TRELLIS’s multi-view conditioning without retraining the model. To this end, we propose VALS (View-Aligned Latent Stitching), a novel inference-time technique for improving multi-view 3D asset generation through latent space manipulation. VALS operates by extracting and fusing view-specific latent representations to create more accurate 3D reconstructions. We present two variants: VALS-A, which assumes axisaligned views, and VALS-T, which generalizes to arbitrary camera viewpoints using transformation matrices. The first part of the work focuses on improving geometry, where inconsistencies are typically more noticeable, while the second part addresses texture generation and how manipulating texture latents can improve visual coherence. To compare the proposed methods with the original model, both quantitative and qualitative evaluations are conducted, the latter based on a user study. In addition to proposing a practical technique, this thesis offers new research insights into how TRELLIS processes multi-view information and how latent representations can be leveraged for more robust 3D generation. Our results are especially promising when conditioning on real photographs or complex assets, highlighting the method’s potential to enhance the asset creation pipeline and support rapid ideation in production environments.

Condition monitoring of industrial systems, particularly rotating machinery, is a critical component of predictive maintenance strategies, enabling reduced downtime, lower operational costs, and improved system reliability. Within the ITER project, crucial remote handling systems need continuous monitoring and evaluation. This thesis investigates the application of machine learning models for anomaly detection in multivariate time series sensor data, with a focus on unsupervised learning methods to reflect real-world scenarios where labelled fault data are often unavailable. The study evaluates the performance of one-class support vector machines (OC-SVMs) and autoencoders (AEs), with particular attention given to model accuracy, scalability, and computational efficiency. OC-SVMs with an RBF kernel and fixed parameters (ν = 0.01,γ = 0.8) globally outperforms the competition with low computational overhead. The best AE configuration — a 4-layer encoder and decoder network with L2 regularisation, weight decay, and dropout — also shows strong performance, though with increased training and inference complexity. The evaluation uses a custom weighted geometric mean to balance specificity and recall, reflecting the industrial need to avoid both false alarms and missed detections. The study also discusses scalability, noting that the primary computational bottleneck lies in feature extraction rather than inference. Overall, the findings support the deployment of lightweight, unsupervised models for effective and interpretable condition monitoring in real-world industrial applications. Additionally, ensemble methods using classifier aggregation across vibration channels demonstrate improved detection stability and overall performance, particularly when aggregating decision functions. Finally, limitations of unsupervised learning methods are addressed by emphasising their inability to dig into root cause classification of anomalies, leaving area for future work.

Language integrated query (LINQ) is a powerful extension to C# that enables integration of database queries into the programming language. However, not all LINQ queries can be translated to SQL, and the programmer is not always aware of the performance of the queries against a database, particularly when using indexes. This thesis addresses these challenges through theoretical and practical contributions. We develop a novel object-based core language based on C# , incorporating query capabilities similar to LINQ, and present a type-and-effect system for analyzing query expressions. The core language has two features, it can guarantee that the query expressions can be translated into SQL and that queries that do not utilize any database search index can be detected. The work builds on existing research in the area of language integration. After LINQ was created, a method to guarantee translation to SQL was developed. We extend this work to include the concept of indexes and how they can be utilized in query expressions. Based on the core language, we have created a C# analyzer that can detect some problematic queries. The analyzer can be integrated into an existing build pipline and will generate compiler errors and warnings if problematic queries are found. The analyzer was validated using a set of example quries as well as running on Challengermode’s codebase. The analyzer successfully identified several performance-critical issues in production code, including 11 non-indexed queries and 4 potentially problematic large queries. This shows that this type of analyzer can give insight into performance issues before they are discovered in production. This thesis has shown that a type and effect system can be used to detect if an object-based language with query capabilities similar to LINQ can be translated to SQL and if it can be translated to SQL it can be detected if the query utilizes any database search index.

This thesis investigates the challenges of concurrent schema evolution in Delta Lake, a widely adopted storage layer for large-scale data processing. While Delta Lake provides ACID properties for data consistency through optimistic concurrency control, its handling of concurrent schema modifications remains conservative, often rejecting operations that could otherwise be executed safely. To address this limitation, an analysis of its concurrency control algorithm and a series of exploratory experiments were conducted to study Delta Lake’s behaviour under different workloads, including concurrent appends, schema changes, and conflicting schema evolution scenarios. These experiments confirmed that while concurrent data appends are safely supported, metadata updates are systematically blocked, even in non-conflicting cases. This motivated the design of a novel optimization for Delta Lake’s concurrency control algorithm, enabling safe schema evolution without false conflicts. The proposed solution was implemented and evaluated through a custom benchmarking tool, designed to simulate concurrent readers, writers, and schema-changing operations. A comprehensive set of validation experiments compared the official Delta Lake release with the modified version across scenarios, ranging from conflict-free appends to large-scale real-world workloads. Results demonstrate that the optimized version preserves consistency while significantly reducing unnecessary transaction failures, therefore improving throughput and enabling native support for concurrent transactions using dynamic schemas. Overall, this work contributes both an in-depth empirical characterization of Delta Lake’s concurrency model and a practical enhancement that facilitates more flexible and efficient schema evolution in modern data lake deployments.

The field of speech compression is actively changing with the currently evolving field of Artificial Intelligence (AI), with multiple AI speech compression models being developed at a fast pace. With the goal to raise end-to-end quality of speech transmission and storage while maintaining low bitrate, low model size and low computational complexity, it is important that many different optimizations are explored in order to develop an efficient model. In this Master’s thesis project, a state-of-the-art speech synthesis model Framewise Autoregressive Generative Adversarial Network (FARGAN) was adjusted and evaluated, in order to explore new versions of the model and thereby advancing research in the area. A number of adjustments were tested and the three most interesting ones were evaluated using three objective evaluation models: Perceptual Evaluation of Speech Quality (PESQ), WARP-Q and Perceptual Objective Listening Quality Analysis (POLQA) as well as one subjective evaluation method: Multiple Stimuli with Hidden Reference and Anchor (MUSHRA). The three adjustments made were changes to the size of the subframes which the FARGAN model synthesizes speech from, which in turn affected the number of weights of the model. This kind of adjustment does not seem to have been explored in previous work. The results showed that we were able to recreate the original FARGAN model in terms of quality, and that our new, adjusted models produced lower quality speech than the original one. However, depending on how an application might weigh importance of quality, model size and computational complexity, it could be worth exploring two of our new, adjusted models, one of them being smaller in model size and the other being less computationally expensive during inference.

Application-Level Synthesis (ALS) frameworks enable the automatic generation of hardware from high-level dataflow specifications. Sylva is one such framework that synthesises Synchronous Dataflow Graph (SDF) into complete hardware systems through binding, placement, and routing. Validating the correctness of these generated systems is a crucial step; however, existing simulation approaches either operate at an overly abstract level, omitting timing details, or at the Register Transfer Level (RTL) level, where simulation is costly and time-consuming. This thesis introduces a modular simulator, GLICsim, designed to validate Sylva-generated systems at the dataflow abstraction level. The simulator employs Python for orchestration, Protobuf for schema-based communication, and file-based interaction for enhanced debuggability and reproducibility. Its architecture models address translation, process execution, and data transfer explicitly, ensuring accurate timing and buffer behaviour. Two representative case studies are used to evaluate the simulator: the Sobel edge detection algorithm and the LeNet-5 convolutional neural network. Outputs are compared against global software reference models to ensure functional correctness, while timing, throughput, and buffer behaviour are validated against Sylva’s synthesis predictions. The results demonstrate that GLICsim provides reliable validation of functional and timing properties, while offering modularity, extensibility, and transparency that are not available in existing simulation frameworks.

Accurate prediction of chemical reaction yields is a critical task in synthetic chemistry, directly impacting the design and optimization of reaction pathways. This thesis focuses on yield prediction for Pd-catalyzed Buchwald– Hartwig amination reactions—a widely used method for forming C–N bonds. While deep learning models, particularly artificial neural networks, have shown promise in modeling such reactions, they often struggle with practical limitations: they are sensitive to noisy descriptors, vulnerable to skewed target distributions, and typically require large, high-quality datasets to perform well. This thesis proposes a more robust and interpretable alternative to deep learning approaches by combining gradient boosted tree models with a modular, chemically informed feature engineering pipeline. Initially, benchmark models based on artificial neural networks using quantum mechanical descriptors were reproduced. These were then systematically compared to gradient boosted tree models trained on both quantum mechanical and alternative descriptor sets, including Morgan fingerprints, Differential Reaction Fingerprints, and RDKit molecular descriptors. To address the resulting high-dimensional feature space, a two-stage feature pruning strategy was implemented using correlation-based and model-aware genetic algorithms. The results show that gradient boosted tree models consistently outperformed their artificial neural networks counterparts in terms of robustness and predictive accuracy, particularly on skewed and noisy datasets. Moreover, models trained on pruned, multimodal descriptors achieved similar or superior performance to those using computationally expensive quantum mechanical features, while greatly improving training efficiency and interpretability. However, external validation on proprietary Chemify reactions highlight the need to include broader experimental conditions such as reaction time, temperature, and molecular quantities in the modeling process to allow for generalizability. By demonstrating that simpler, tree-based models with tailored feature selection can match or surpass deep learning in this context, this work offers a scalable and transparent framework for reaction yield prediction. These insights contribute to the development of more efficient and accessible tools for cheminformatics and automated synthesis planning.

Signal Level Attenuation Characterization (SLAC) is a protocol to automatically match and associate devices through the use of the HomePlug Green PHY (HPGP) specification, a technology to transceive signals over a live power wire. SLAC is at the heart of the Vehicle to Grid (V2G) movement, but the research into using this technology for industrial purposes is lacking. The work is done at the request of EpSpot, a company that develops solutions for a fossil free future. The findings of this work is meant to aid future development of products utilizing SLAC for Plug and Charge (PnC) solutions. To do this, ISO 15118-3, a specification that outlines the requirements for SLAC is used as a base for reference. This thesis demonstrates and analyses the implementing of SLAC through a physical layer without auxiliary wiring. To achieve this, architecture and sequence diagrams of ISO 15118-3, are used for reference. Through a comparative analysis between the developed solution and ISO15118-3, the conclusion is that SLAC is feasible to implement through a physical layer without auxiliary wiring. However, this comes at the loss of features aimed towards functionality and safety.

Autonomous navigation in planetary-like environments presents unique challenges due to the absence of GPS signals, limited semantic structure, and visual ambiguity caused by repetitive textures or harsh lighting conditions. Traditional place recognition and localization methods either rely on dense maps and structured environments or only provide coarse retrieval without estimating full 6-DoF (Degrees of Freedom) poses. This limits their applicability in the context of real-time Simultaneous Localization and Mapping (SLAM) for field robotics and planetary exploration. This thesis addresses the problem by developing a multi-modal system that performs both place recognition and relative pose estimation in unstructured, GNSS-denied environments. The proposed approach fuses visual features extracted from a transformer-based encoder (DINOv2) with 3D geometric descriptors from a LiDAR-based backbone (SONATA). These features are projected and aligned in 3D space to produce interpretable correspondences, from which the system estimates full 6D poses. On the retrieval side, DINOv2 descriptors are aggregated using SALAD, a learned VLAD-style module, and searched efficiently using FAISS indexing. The system is evaluated on the Etna volcano dataset, representative of planetary terrains. The results show that the proposed model outperforms established retrieval methods like NetVLAD and TransVPR and achieves more stable pose estimation than handcrafted or regression-based alternatives. The fusion of LiDAR and vision improved robustness in scenes with low texture or poor illumination, validating the hypothesis that multi-modality can bridge the gap between accuracy and generalization. Importantly, the system produces interpretable outputs and operates within real-time constraints for retrieval, although further optimization is needed for pose estimation. This thesis demonstrates that it is feasible to move beyond retrieval-only frameworks and provide full, explainable 6D poses suitable for SLAM. Future work should focus on improving runtime efficiency in the pose estimation module, incorporating more diverse datasets, and testing deployment on real robotic platforms. These developments could contribute to more autonomous and trustworthy robotic systems for exploration, disaster response, and agriculture in extreme environments.

In-the-wild automotive user studies are costly, time consuming and difficult to scale, yet early design demands fast and iterative feedback. This thesis investigates whether generative agents, built on large language models, can simulate participants responses and produce synthetic data for such studies. Two real-world cases were used to compare results: an interruptibility assessment study and an emotion recognition study, with agents embodying synthetic personas and exposed to multimodal driving context. Alignment with human data was evaluated through majority-vote consensus and a machine-learning generalisation test. An ablation further examined three architectural features and their impact on results: personas, contextual templating, and drive–goal conditioning. Results show that synthetic agents approximate human behaviour in structured tasks such as interruptibility, but perform poorly in affective tasks like emotion recognition. Architectural additions yielded inconsistent or negative effects. These findings indicate that performance is strongly task dependent. Generative agents may therefore serve as a complement rather than a substitute for participants: they are most effective in narrowly defined studies that rely on filtered and task-relevant input data, while final validation must always be based on human responses.

This thesis addresses the problem of detecting and locating snug in the advanced tightening technique during industrial assembly. It proposes a deep learning model based on CNN and a Hybrid Model combining CNN with a rule-based algorithm. Firstly, the torque-angle traces are uniformly resampled to a fixed length input, and the standardized torque, angle, and their first-order differences are extracted as features. Then, 1D-CNN is used to categorize whether the trace contains a snug, and then to predict the coordinates of the snug, with the predicted point mapped back to the original traces through Euclidean distance. To further utilize the existing rule-based algorithm from Atlas Copco, this thesis designs three Hybrid Models, among which the model using an additional fully connected layer to combine the CNN and the rule-based algorithm performs the best. The experimental results show that the pure CNN classification model achieves an F1 score of 0.9764 on the test set; in the coordinate prediction, the MAE of the pure CNN is reduced by 44% compared to the rule-based algorithm, and the MAE of the Hybrid Model is further reduced by approximately 2.7% compared to the pure CNN. This thesis demonstrates that data-driven methods can significantly improve the accuracy of snug detection, and integrating physical or rule prior information can further enhance model performance under limited sample conditions. This method provides an efficient and feasible technical path for industrial tightening quality control and fault diagnosis.

The introduction of Generative AI (GAI) into marketing workflows marks not only a technical but also an ontological change to what it means to be a marketer. Using a Value Sensitive Design (VSD) lens to engage with 12 marketing professionals from Sweden and Vietnam, this study investigates the values and tensions that arise when marketing professionals integrate this technology into their daily work. We discovered how marketers from Sweden and Vietnam expressed, negotiated and occasionally sacrificed personal values in GAI-powered workflows. The findings show that marketers feel more empowered than ever to achieve and learn, thanks to the use of GAI. At the same time, marketers are grappling with the professional identity shift of becoming a supervisor of GAI while managing the appearance of using GAI for work. This work contributes to existing empirical research on the use of GAI in professional workplace by foregrounding the values at stake for marketers and articulating the changing role of marketers in the age of GAI as shared and differed in two different contexts.

The doorway effect predicts crossing a doorway impairs memory performance in both physical and virtual environments. While environmental features and the way people perceive information from their surroundings appear to influence memory performance, findings on the doorway effect in virtual environments remain inconsistent, particularly regarding the integration of multimodal sensory cues. This study examines the influence of environmental features on memory performance of object memory and object–location memory in a walking-based immersive virtual reality environment replicating a realworld layout to maintain ecological validity. Memory performance was assessed through recognition of object and reconstruction of the spatial layout. Two variables were tested: sensory modality (unimodal vs. multimodal environmental cues) and environmental functional change (constant living-room scenario vs. living-room-to-office transition). Results showed no significant effects of doorway crossing, multimodal environmental cues, or functional changes on memory performance, suggesting the link between environmental features and the doorway effect remains inconclusive. Nonetheless, evidence of the influence of environmental features on memory encoding and retrieval was identified, positioning the findings as an active area for investigation into the mechanisms that trigger the doorway effect.

The increasing demand for accurate traffic monitoring in urban areas has prompted design offices to explore artificial intelligence (AI) for vehicle detection and classification. This thesis investigates the application of computer vision and deep learning to automate traffic analysis at intersections using real-world video footage provided by design offices. The dataset consists of multiple video recordings from different urban intersections under varying conditions (e.g., weather, lighting, camera angles). Using a YOLO11-based object detection model, combined with BoT-SORT for tracking, vehicles were detected, classified, and counted across multiple scenarios. Performance was evaluated using standard metrics (counting accuracy, classification accuracy, tracking consistency), and results were compared against both manual counting and traditional traffic analysis tools. The proposed pipeline reaches 94% of human accuracy in ideal conditions and matches human-level performance in most real-world scenarios, while also aligning with state-of-the-art solutions with broader applicability. Key challenges include occlusion, low resolution, and night-time conditions. These findings show that AI-based methods already provide human-level accuracy in traffic analysis and, given their scalability and consistency, are likely to become the preferred solution for large-scale monitoring.

This thesis explores the application of clustering algorithms to identify consumer segments based on spending behaviors and demographic characteristics. The thesis is conducted in partnership with Svenska Handelsbanken. The bank's current tools for financial overview lack the capability to contextualize and give feedback on customers' spending relative to relevant peers. This study aims to identify data-driven customer groups within the bank’s user base from customer spending data and to evaluate the spending groups against each other and the clustering algorithms utilized, as a pre-study to later be able to create the comparison groups necessary to develop individualized products, such as relativized spending feedback. The study employs two clustering algorithms, k-means and HDBSCAN, to analyze and group customers based on their spending profiles derived from transactional data. The efficacy of these algorithms is compared to identify distinct consumer groups. Logarithmized spending data and dimensionality reduction with t-SNE and PCA is used to prepare the data for clustering. The analysis successfully identified distinct consumer groups with demographic and consumption differences, and the best results were achieved with HDBSCAN based on reduced data via t-SNE. The clusters were validated with new data inputs, ensuring stability and robustness in the customer segments. This study lays the groundwork for banks to offer personalized financial insights and services, with the potential to enhance customer satisfaction and decision-making, as well as targeted initiatives from the bank. Future research can build on these findings to create even more granular clusters or incorporate external demographic data and analyze household-level spending, further refining customer segmentation and personalization efforts.

This thesis presents the design and development of a fabrication-ready Photonic Integrated Circuit (PIC) using open-source automated design tools, GDSFactory and Nazca Design. The circuit is based on a hexagonal waveguide mesh architecture and aims to explore the capabilities, limitations, and practical challenges of using these tools for complex PIC design. A modular and iterative design approach was adopted, leveraging each tool’s strengths: GDSFactory for modular and hierarchical layout assembly and Nazca Design for parametric and curved structure construction. Key challenges included routing complexities, tool interoperability issues, and precision limitations of the GDSII file format, which introduced sub-nanometer misalignment. These were addressed through a combination of custom scripting and manual design strategies. The final layout was validated through design rule checks and deemed ready for fabrication. This work demonstrates the feasibility of a fully open-source PIC design workflow and provides a methodology for future programmable photonic circuit development.

Smart contracts are self-executing programs on the blockchain, representing the core of Decentralized Finance (DeFi). They can carry billions of dollars, making their correctness and security essential. However, like any other software application, smart contracts may contain vulnerabilities that can be exploited by malicious adversaries — a concern that is further exacerbated by their immutability and transparent nature. Fuzzing is a widely adopted automated testing approach for assessing and ensuring smart contract quality. Although powerful, its effectiveness heavily depends on the quality of the initial inputs, also referred to as seeds. This thesis identifies a key limitation in traditional smart contract fuzzers, such as Echidna, which often rely on randomly generated seeds. It investigates how these fuzzers can be improved in exploring deep contract states, reaching hard-to-trigger conditional paths, and increasing the number of executed instructions by providing higher-quality initial seeds. This thesis proposes AutumnEchidna, a smart contract pre-processing tool that leverages static analysis to generate optimized initial seeds for fuzzing. The methodology involves generating transaction sequences based on state dependencies and producing arguments through constraint solving, aiming to guide execution toward critical contract states. Experiments are conducted on two datasets: a Motivation Dataset and a Maze Dataset, designed to simulate complex input constraints and deep state transitions. Performance is evaluated based on instruction coverage and execution time under consistent configurations for both baseline Echidna (with random seeds) and AutumnEchidna (with optimized seeds). The experiment results show that AutumnEchidna improves instruction coverage by 1.26% on the aggregated Motivation Dataset and by 4.45% on the aggregated Maze Dataset. Additionally, it also reduces the execution time to achieve comparable or higher coverage. These findings demonstrate that optimized seed generation can enhance both the effectiveness and efficiency of smart contract fuzzing. This thesis concludes that incorporating static analysis to generate high-quality initial seeds is a promising approach for enhancing the performance of fuzzing strategies. 

This thesis focuses on energy savings in the downlink operation of Cell- Free Massive MIMO (CF-mMIMO) networks under real-time dynamic traffic conditions. We propose a Multi-Agent Reinforcement Learning (MARL) algorithm that allows each Access Point (AP) to autonomously manage its antenna switching and Advanced Sleep Mode (ASM) transitions. After the training process, the framework functions in a fully distributed manner, removing the need for centralized control and enabling each AP to adjust dynamically to real-time traffic fluctuations. Simulation results demonstrate that the proposed approach achieves up to 64.4% energy savings compared to systems without an energy-saving scheme, and 53.2% savings over systems that use only the active or the lightest sleep mode.

Risk identification play a significant role in Power grid security management. However, due to the dynamic and complex operating environment, they are vulnerable to various risks. Nowadays, most of the traditional and popular machine learning methods for identifying these risks rely heavily on supervised learning techniques, which require extensive labeled datasets. Nevertheless, in the real-world power grid industry, labeled fault data is rare and often inadequate due to the difficulties in data collection and the dynamic operating environment. This situation make the existing supervised techniques insufficient for comprehensive risk analysis in a lot of case. This thesis explores a innovative approach for capacitor, a criticial component in power grids system, risk identification using Deep Deterministic Policy Gradient (DDPG), which is a reinforcement learning (RL) algorithm suitable for continuous action spaces. To be more specific, the task of predicting capacitor health, which is represented by the Equivalent Series Resistance (ESR), is formulated as a Markov Decision Process (MDP). In this study, we utilize historical multivariate time-series data provided by Hitachi Energy Sweden AB to train the DDPG reinforcement learning model. The aim is to investigate the possibility of a new way of risk prediction without replying on substantial labeled data. The results of the experiment demonstrate that although the DDPG-based approach can perform ESR data forecasting, its predictive accuracy and robustness remain inferior compared to mainstream deep learning methods in this domain, such as Transformers and Modern recurrent neural network. These results indicate that while DDPG offers an innovative perspective, its practical effectiveness in static time-series data prediction scenario is limited. This study helps the field of reinforcement learning in power system risk identification by pointing out the limitations and difficulties of applying it in real-world industry. Furthermore, it also gives useful insights to companies like Hitachi Energy Sweden AB, which is that they need to evaluate carefully before using AI technologies to make power grid management more reliable and efficient.

Beamforming using Phase shifter (PS) is widely used in high frequency applications, such as 5G communication systems, satellite communications and medical imaging. Beamforming techniques and phased-array systems are implemented to Multiple Input Multiple Output (MIMO) systems to reduce the effects of high propagation path loss. PS plays a critical role in steering the beam generated by an antenna array. This thesis presents a D-band passive PS application-specific integrated circuit designed in 22-nm Fully Depleted Silicon On Insulator (FD-SOI) technology for 140 GHz application, which has high linear phase shift, and high resolution. The proposed PS combines of two topologies - a Switched Coupled Inductor (SCI) PS for fine tuning and a Switched Transmission Line (STL) PS for coarse tuning. In the SCI PS design, by switching the inductance with two shunt capacitors, small phase shifts are obtained. Two parallel transmission lines combined with a switch network are designed as a unit cell. By changing the state of the switch network, the unit cell can be configured in three different modes. By applying these three modes, larger and more accurate phase shifts can be obtained. Combining these two types of PS, simulation results show that the proposed PS provides a phase tuning range from 0° to 90° with a resolution of 5.625° and the Integral Non-Linearity (INL) and Differential Non-Linearity (DNL) are within 2°. The loss of the proposed design is below 20 dB at 140 GHz. These specifications meet the requirements for high-frequency D-band applications, where fine phase resolution, low non-linearity, and minimal insertion loss are critical for accurate beam steering at 140 GHz.

The growing demand for scalable model training in modern MLOps pipelines, together with the rise of large language models (LLMs), highlights the need for robust integration of distributed computing frameworks with reliable storage systems. This thesis presents a prototype system that integrates the Ray framework into Hopsworks to support scalable model training on heterogeneous compute resources, while also providing periodic data backups in the persistent storage, HopsFS. The system is containerized by Docker and orchestrated by Kubernetes. To evaluate the system, two experiments were conducted: the first benchmarked distributed training performance on ResNet152; the second finetuned the Llama-3.1-8B-Instruct model, leveraging LoRA for parameter-efficient adaptation and DeepSpeed for memory-efficient distributed training. The results of the experiments provide several key findings. In small clusters, inter-node overhead was negligible, implying that communication and coordination between nodes did not significantly affect training performance. Scaling resources improved training speed, although efficiency gains were limited, with doubling resources yielding only an approximate improvement of 16%, measured in seconds per epoch. The experiments also showed that dataset loading throughput and checkpoint saving throughput remained consistent across different cluster formations, indicating stable system I/O performance. Furthermore, the system successfully completed multiple training runs without failures, demonstrating its reliability under prolonged, resource-intensive workloads. Testing in different workloads and environments also confirmed system stability and flexibility. Overall, the results show that the integration of Ray with Hopsworks enables scalable model training while meeting the expected functionality, reliability, and stability. This prototype system extends Hopsworks’ capabilities to support advanced training pipelines for modern machine learning, especially LLM finetuning.

As the world increasingly turns towards greener technologies, the demand for minerals has skyrocketed. This increase in demand has led to higher demands on the mining industry, and, therefore increasing requirements on the machines. To be able to meet these requirements new control strategies are needed that can provide more optimal utilization of these machines. The goal of this thesis is to investigate the possibility of using an Extremum Seeking Controller (ESC) to improve the drilling performance of one of Epiroc’s rotary drills. Extremum Seeking Control is a model-free, real-time, adaptive algorithm that can solve static optimization problems for a given objective function. To solve the problem, a literature study was conducted, after which a case study was done in which a provided simulation of Epiroc’s Pit Viper 351 drill rig was used. Before the algorithm could be implemented, a Model Predictive Controller (MPC) had to be created with which the drill system could track a given reference signal. To evaluate the ESC, experiments were done using the provided simulation which tested the algorithm for constant and varying rock conditions. The optimization problem was set up both as a single and multiobjective optimization problem. Based on the results, it was concluded that while the extremum seeking algorithm can find optimal solutions, it suffers from slow convergence times and difficulty with tuning.

Hospital readmission within 30 days is a persistent challenge that strains resources and signals sub-optimal continuity of care. Data-driven Machine Learning approaches have become the prevailing strategy for mitigating this problem. Although these models built on electronic health records (EHRs) have boosted predictive performance, their reliance on purely correlational patterns limits clinical trust and the design of actionable interventions. This thesis aims to bridge this gap by marrying Machine Learning algorithms with state-of-the-art causal discovery methods to produce both accurate and interpretable predictions. In this thesis, we first construct a cleaned, harmonized EHR cohort of MIMIC-IV dataset and develop predictive models that estimate each patient’s likelihood of returning within 30 days, and analyze the most influential clinical factors. Furthermore, we apply complementary causal-discovery techniques on the most important variables, to uncover directed relationships among demographic, clinical and utilization variables. We find that this hybrid framework sustains strong predictive accuracy and uncovers stable, clinically plausible causal relationships. We hope that these insights will enable more targeted discharge planning and pre-emptive interventions, and ultimately reduce avoidable readmissions and improve healthcare efficiency.

This work investigates suitable polyhydroxyalkanoate (PHA) blends to use in Novo Nordisk’s injection pen application. Five PHA blends were analysed, each with varying ratios of PHAs, including poly-3-hydroxybutyrate-co-3-hydroxyvalerate (PHBV), poly-3-hydroxybutyrate-co-3-hydroxyhexanoate (PHBH350), and poly-3-hydroxybutyrate-co-4-hydroxybutyrate (P3HB4HB), with compositions ranging from 70:30 to 55:45 with or without talc. The study examined crystallisation kinetics, secondary crystallisation and mechanical properties through differential scanning calorimetry (DSC) and tensile testing. A life cycle analysis (LCA) was performed to evaluate carbon emissions, and the influence of different crystallisation rates on processing cycle time was studied.

The Avrami model was applied to isothermal crystallisation of the blend and revealed that an equilibrated balance between semi-crystalline and amorphous PHAs, particularly with shorter hydrocarbon chains and a small talc addition, accelerated crystallisation. Tensile tests indicated a higher tensile strength and low creep behaviour for blends have a ratio close to 1:1. In addition, blends containing PHBH350 exhibited lower tensile strains and toughness but a larger Young’s Modulus. LCA results demonstrated a carbon footprint reduction of 3% to 5% with talc addition and optimised crystallisation rates.

Finally, the blend containing 55 wt% PHBV and 45 wt% PHBH350 with a low amount of talc exhibited the best results in all measured parameters and was successfully moulded into a high precision component.

Textile waste, particularly polyester-based materials, represents a growing environmental challenge due to tow recycling rates and persistence in the environment. In this Master’s degree project, a simple one-pot transesterification method was developed to synthesize poly(butylene adipate-co-ethylene terephthalate) (PBAET) copolyesters from waste PET fabric, PET granules, and hydroxyl-terminated poly(butylene adipate) (PBA). Five copolyesters with varying PET/PBA weight ratios (50/50, 65/35, 35/65) were prepared and characterized. Molecular weights (Mn) ranged from 10 000 to 17 300 g/mol. The chemical structure of the copolymers was confirmed by NMR and FT-IR spectroscopy. Flexible films were fabricated via solvent casting and hot pressing and the thermal, mechanical, and biodegradation properties were systematically evaluated. Increasing PET content improved thermal and mechanical stability: elastic moduli ranged from 20 to 150 MPa, tensile strain from 9.4% to 27%, and tensile stress from 1.2 to 10.4 MPa. Biodegradation studies showed that the 35 wt% PET sample exhibited the highest mass loss after five weeks of soil burial — 8.4% at room temperature and ≥ 28% at 55 °C. Enzymatic degradation with Candida rugosa lipase resulted in 1.4–5.6% mass loss over nine days, while control samples remained stable. These results demonstrate that waste PET textiles can be upcycled into PBAET copolyesters with properties comparable to commercial PBAT. This provides a promising route toward more sustainable materials. Future work should focus on optimizing degradation rates and further improving mechanical performance.

Neutrophils are the immune system’s rapid responders, forming the largest population of white blood cells in humans. They express a diverse array of cell surface receptors, including formyl peptide receptor 2 (FPR2), a G protein-coupled receptor (GPCR) involved in initiating and regulating inflammatory responses. Given the importance of FPR2 in neutrophil activation, a novel FPR2 agonist was characterized, based on its effects on primary human neutrophils. Receptor specificity was confirmed through antagonist and desensitization experiments, demonstrating that the novel agonist signals selectively through FPR2. While the agonist triggered key neutrophil effector functions such as production of reactive oxygen species, degranulation, and chemotaxis, these responses were significantly weaker than those induced by the well-characterized FPR2 agonist WKYMVM, even at fivefold higher concentrations. This indicates a low activation profile not previously observed among FPR2 agonists, suggesting a distinct interaction mode. Notably, this novel agonist failed to induce intracellular calcium mobilization, a response traditionally considered essential for these effector functions and regarded as the earliest intracellular signaling event downstream an activated GPCR. This challenges the established view that these responses are strictly calcium-dependent, suggesting that the agonist may act as a biased FPR2 agonist, selectively activating specific downstream pathways without relying on conventional intracellular calcium signaling, making it a unique FPR2 agonist with an atypical signaling profile.

In the context of a growing interest in completely passive safety systems in the nuclearindustry, this study aims at a deep comprehension and improvement of the CATHARE3 condensationmodel for passive systems applications. In particular, two experimental facilities willbe analysed, COTURNE and PKL, allowing the study of condensation modelisation both inturbulent and laminar conditions under natural circulation. The developments produced inthis study will be considered by Commissariat á l’énergie atomique et aux énergies alternatives(CEA) for future versions of the CATHARE code. The study focuses on the comparison of theexperimental readings from the facilities with the simulation outputs of the same facilities onthe CATHARE code. In fact, regarding the PKL facility, a considerable difference between theexperimental data and simulation outputs is observed in terms of a sudden subcooling of liquidtemperature when condensation starts. This is a reason for concern, due to the fact that thePKL facility is the one studying the new passive Safety Condenser (SACO). Five parametersof the CATHARE code have been analysed, four of which introduced specifically for this work,in order to understand the impact of each on the heat exchange in the condensation scenario,and one has been identified as possible cause of the discrepancy: the division of the thermalresistance between the two interfaces gas-liquid and liquid-wall. This finding can contribute tothe improvement of the CATHARE code and to having a better understanding of condensationunder natural circulation at high Reynolds number.

Micro-electromechanical systems (MEMS) act as a cornerstone in the current technology industry, providing miniature mechatronic devices for creating compact, low-power, and efficient products. This master’s thesis project revolves around a compact MEMS-based loudspeaker, aimed to function optimally within the audio frequency range, i.e. 20 Hz to 20000 Hz. There exist some challenges with the existing piezoelectric-based MEMS loudspeakers in terms of sound output quality, which calls for identifying the cause for distortion in the existing range of MEMS loudspeakers. As a result, this master’s thesis delves into analyzing distortion in MEMS loudspeakers, examining its sources and their impact on the overall system output response. Understanding and investigating the sources of the distortion in a specific loudspeaker design contains many complications, some of which can be considered comprehensible at the master’s thesis level. Although conventional loudspeakers have well-understood sources of distortion, MEMS loudspeakers that use a novel driving method lack such a comprehensive understanding. Hence, a limited pool of information is available for the MEMS loudspeakers. This thesis identifies potential causes for distortion in a MEMS loudspeaker by experimenting and trying to understand its output response using a simulation model containing loudspeaker nonlinearities. The nonlinearity data was obtained from the existing background work performed on the MEMS loudspeakers. This comparative study between the experimental and simulation response states which nonlinearity plays what role in the overall system distortion. The results of this thesis can be considered for further research in MEMS loudspeakers, knowing the sources of distortion and their impact on the distortion of the system.

The ongoing proliferation of intelligent systems, driven by artificial intelligence (AI) and 6G, is leading to a surge in closed-loop inference tasks performed on distributed compute nodes.These systems operate under strict latency and energy constraints, extending the challenge beyond achieving high accuracy to enabling timely and energy-efficient inference.This thesis examines how distributed inference can be optimised through two key decisions: when to sample the environment and when to offload computation to a more accurate remote model.These decisions are guided by the semantics of the underlying environment and its associated costs.The semantics are kept abstract, and pre-trained inference models are employed, ensuring a platform-independent formulation adaptable to the rapid evolution of distributed intelligence and wireless technologies.

Regarding sampling, we studied the trade-off between sampling cost and detection delay in event-detection systems without sufficient local inference capabilities. The problem was posed as an optimisation over sampling instants under a stochastic event sequence and analysed at different levels of modelling complexity, ranging from periodic to aperiodic sampling. Closed-form, algorithmic, and approximate solutions were developed, with some results of independent mathematical interest.Simulations in realistic settings showed marked gains in efficiency over systems that neglect event semantics. In particular, aperiodic sampling achieved a stable improvement of ~10% over optimised periodic policies across parameter variations.

Regarding offloading, we introduced a novel Hierarchical Inference (HI) framework, which makes sequential offload decisions between a low-latency, energy-efficient local model and a high-accuracy remote model using locally available confidence measures. We proposed HI algorithms based on thresholds and ambiguity regions learned online by suitably extending the Prediction with Expert Advice (PEA) approaches to continuous expert spaces and partial feedback. HI algorithms minimise the expected cost across inference rounds, combining offloading and misclassification costs, and are shown to achieve a uniformly sublinear regret of O(T^2/3).The proposed algorithms are agnostic to model architecture and communication systems, do not alter model training, and support model updates during operation. Benchmarks on standard classification tasks using the softmax output as a confidence measure showed that HI adaptively distributes inference based on offloading costs, achieving results close to the offline optimum. HI is shown to add resilience to distribution changes and model mismatches, especially when asymmetric misclassification costs are present.

In summary, this thesis presents efficient approaches for sampling and offloading of inference tasks, where various performance metrics are combined into a single cost structure. The work extends beyond conventional inference problems to areas with similar trade-offs, advancing toward efficient distributed intelligence that infers at the right time and in the right place. Future work includes conceptual extensions like joint sampling-offloading design, and integration with collaborative model-training architectures.