Neutral surfaces, on which the neutral density variable is constant, are widely used for understanding oceanic flow patterns and analysing hydrographic observations. The locally referenced normal vectors of a neutral surface are the so called diapycnal vectors. A neutral density variable for a given domain can be found through the diapycnal vectors and then used to construct neutral surfaces. Finding a realistic formulation of the neutral density variable and computing it for general hydrographic data present significant challenges in two key areas. Firstly, due to the non-zero helicity of the diapycnal vector field it is not possible to find neutral surfaces whose normals exactly align with the field. Secondly, evaluating neutral density is computationally heavy due to the size of global oceanic data. Given the current limitations, investigating alternative formulations of neutral density and developing computational methods for its evaluation remains an important area of ongoing research. This thesis is concerned with the construction of neutral density via a minimization problem that penalizes deviations from desirable properties of this variable. A finite difference scheme for two dimensional domains is suggested as an approach to solving the Euler-Lagrange equation corresponding to this optimization problem. The method is then investigated for two dimensional vector fields under varying rotational properties. The method’s performance is investigated for conservative and non-conservative test fields as well as a diapycnal vector field. In this analysis, convergence is shown in the 𝐿2-norm with respect to step-size for all fields and root mean square measures are analysed to evaluate consistency with theoretical results through numerical experiments. The results of this study indicate that the method works well when the vector field is aligned with the grid directions. As this is the case for the diapycnal vector field, the method shows promise for application in physical oceanography.

Cardiovascular disease is the leading cause of death globally, and mechanical circulatory support devices such as the intra-aortic balloon pump (IABP) play a vital role in treatment in acute care. This thesis develops and validates a lumped-parameter Simscape model of the pneumatic system in Getinge's CardioSave IABP. The model incorporates conservation of mass, momentum and energy for compressible gas flow, and includes the compressor, safety disk (SD), tubing, and balloon. Three balloon representations were compared: (i) linear spring-damper, (ii) nonlinear spring-damper, and (iii) a soft spring with check valves to capture inflation and deflation thresholds. Parameters were estimated using least-squares optimization and the model was validated against experiments using a 40 cc balloon under external pressures of 25, 50 and 100 mmHg (gauge). The check valve model best reproduced the observed threshold behavior and achieved the lowest average shuttle pressure error. A use case study across atmospheric pressures of 450-760 mmHg and heart rates 30-200 bpm showed 10-90 % inflation and deflation times < 0.12 s, with full inflation volumes within acceptance limits. The index-1 DAE system was solved with an NDF2-based implicit solver, yielding stable and efficient simulations with small dispersive errors. While the simplified compressor model does not capture measured LPM-RPM curves across altitudes, the system-level dynamics are reproduced, making it useful for evaluating control strategies and design verification.

Federated learning (FL) is one of the most popular way to collaboratively train models while preserving data privacy. Participants train their model locally and share only their gradients instead of their personal data. However, recent gradient attacks have shaken this guarantee of "privacy by design" by reconstructing the participants data from the shared gradients. Serious improvements have been achieved by first inferring the labels of the data, making it easier to then reconstruct the input data. Until now these attacks have been studied only in the context of classification models, leaving the regression case unaddressed. In this paper we develop a gradient-based attack on labels in the context of a regression model being trained under a FL framework. This attack relies on solving an approximated linear system of equations of gradients and labels, calibrated using auxiliary data. Our experiments show promising results about inferring labels considering a FL regression model.

This thesis investigates funding value adjustment (FVA), focusing on differences between symmetric and asymmetric FVA in derivatives pricing. Symmetric FVA assumes a single funding rate for both lending and borrowing, while asymmetric FVA accounts for two distinct rates, the more common scenario in practice. A challenge with asymmetric FVA is the reliance on computationally intensive Monte Carlo estimates, which require extensive sampling to achieve accuracy for portfolios with many derivatives. Therefore, it would be advantageous if symmetric FVA could effectively approximate asymmetric FVA.

Through analytical methods and numerical simulations, this study explores whether symmetric FVA can approximate asymmetric FVA under certain conditions. A simplified model is used to formulate the asymmetric FVA in closed form, used to derive an asymptotic result showing that asymmetric FVA converges to symmetric FVA as the expected portfolio value goes to infinity. Additionally, an approximation formula is derived for asymmetric FVA. Furthermore, Monte Carlo simulations within a foreign exchange model further evaluate the feasibility of using symmetric FVA as a substitute for asymmetric FVA. The findings suggest that symmetric FVA is a viable alternative when the expected portfolio value exceeds three times the portfolio's annual volatility.

The thesis also contributes by more clearly explaining the derivation of the value adjustments by Bugard and Kjaer.

Graph neural networks (GNNs) are a powerful framework for learning graph representations by recursively aggregating information from neighboring nodes. In this work, we propose alternative training methods for GNNs that avoid conventional stochastic gradient descent by employing random weight sampling. We introduce a novel alignment measure (ALI) to quantify the correspondence between graph embeddings and their labels, demonstrating that it correlates well with final model performance and can serve as an effective proxy during model selection. Based on ALI, we first present a random sampling approach that selects the best-performing model from a set of randomly initialized GNNs. This method offers a competitive alternative to traditional training. We further extend this approach by aggregating multiple small, independently sampled GNNs into a single, sparse ensemble model. This ensemble strategy not only enhances overall performance but also significantly reduces computational costs. Evaluations on benchmark datasets such as Mutagenicity, NCI1, and COLLAB demonstrate that our methods are competitive and time efficient, while in some cases outperforming conventional approaches for models with large hidden dimensions. Overall, our findings highlight the potential of random weight sampling and ensemble techniques as viable alternatives to standard training methods for GNNs, opening new perspectives for efficient graph representation learning.

This thesis investigates the application of Surrogate Modelling techniques to approximate the behavior of high-fidelity industrial process simulations developed in Dymola. The study, conducted in collaboration with Optimation AB, aims to reduce the computational cost of these simulations while maintaining acceptable accuracy. The focus is on steady-state simulations, with data-driven models trained using Machine Learning techniques such as Multiple Linear Regression, Polynomial Response Surface Model (PRSM), Support Vector Regression (SVR), and Neural Network (NN). Two models are developed and exported as Functional Mock-up Units (FMU) to generate training and test datasets. Various regression techniques are compared in terms of predictive accuracy, generalization ability, and computational efficiency. A classification model is also implemented to ensure the validity of input samples for regression models. The Surrogate Models are benchmarked against the original Dymola simulations to evaluate performance. The results indicate that Neural Network-based Surrogate Models outperform traditional regression methods in capturing complex system behavior. However, preprocessing strategies such as exploratory data analysis and outlier filtering significantly impact performance. The study demonstrates that Surrogate Modeling can effectively accelerate simulation workflows, by making rapid steady-state predictions. Future work includes investigating how Machine Learning models developed in Python can be exported as FMUs and extending the approach to dynamic system simulations.

Nodular cast iron is widely used at Scania in a variety of components. This thesis aims to assessthe low cycle fatigue properties of the nodular cast iron GJS-500-7, a widely used cast iron gradeat Scania. A total of 40 uni-axial test specimen were cut out of 5 serially produced Scania parts.Monotonic uni-axial tests were performed to verify the casting process and enabled the implemen-tation of a new material curve in Abaqus. A large number of uni-axial low cycle fatigue tests wereperformed for a variety of strain ratios. A combined Basquin and Coffin-Manson axial fatiguemodel was fitted and three separate kinematic hardening models were fitted by experimental dataand implemented into Abaqus. Low cycle fatigue testing in force-control was performed simulta-neously for verification of cyclic models. The performance of the three models were evaluated inthe ability to recreate the strain in force testing. All three kinematic models, linear, multi-linearand Chaboche, failed to recreate the mean strain from the ratcheting effect which occurred in theforce-controlled testing, but were sufficient in their ability to capture the amplitude in the strain-time signals. The multi-linear model was not stable for this analysis. The linear model performedthe best and is recommended for use in cyclic analysis. An acceptance criterion for total strainamplitude in GJS-500-7 can not be fully recommended, however recommendation is made forthe use of strain-controlled alternating points which have been used historically for dimensioningagainst low cycle fatigue.

The formation of metal slivers during cutting and stamping operations is a major defect that affects part quality and tool wear, especially in the automotive industry where throughput is high and tolerances are tight. This work aims to identify the geometric parameters influencing this phenomenon and to better understand its mechanisms. The study focuses on a 12-stage progressive tool used to produce ball-bearing rings from C45E steel for the automotive industry.Each stage performs a distinct forming or cutting operation along the strip. Sliver formation was primarily observed in three tool regions — stages 4–5, 8–9, and 11–12 — corresponding respectively to cutting the carrier outline (thin metal bands connecting each part to the strip and designed to bend during forming), center-hole trimming, and final ring separation.A Taguchi L16 experimental design was implemented with six factors: the punch and die edge radii for all cutting operations (2 factors) and the punch–die clearances for stages 1, 2, 3, and 8 (4 factors). Fifteen tests of 400 cycles each were performed, with systematic collection of sliver number, size, length, and type. Statistical analyses using Ellistat showed that sliver count was highly random and poorly correlated with the studied factors (R² between 0.155 and 0.251), whereas morphology depended strongly on local tool geometry and on the clearance of the preceding cutting stage.In the stage 8–9 area, larger punch radii increased flake width by about 0.302 mm per 0.1 mm rise in edge radius (p = 0.012), while in stage 11–12, a larger die radius increased overall flake size by 0.051 mm per 0.1 mm (p = 0.055). In stage 4–5, greater clearance at stage 3 produced longer flakes, with an effect of +0.874 mm per 0.1 mm increase (p = 0.050). Numerical simulations confirmed that slivers form when previously sheared edges re-contact the tool and detach more easily after local deformation.This study shows that while sliver frequency is difficult to predict, morphology can be controlled by adjusting punch–die clearance and tool radii. Combining experiments and simulation provides practical guidelines for tool optimization: limit punch and die radii to reduce flake size, finely tune early-stage clearances to minimize flake length, and design dies to prevent unnecessary contact between the die and the strip such as U-shaped grooves to prevent carrier contact.

This thesis examines how organizational culture shapes gender equality

practices in male-dominated, technology-intensive organizations in Sweden.

The study is situated within contemporary societal challenges characterized by

technological development, the climate crisis, and persistent gender inequality,

where organizations are often portrayed as central arenas for change. At the

same time, previous research demonstrates that many technology-intensive

organizations are characterized by norms, hierarchies, and knowledge ideals

that reproduce inequality.

Drawing on feminist organization studies, organizations are understood as

gendered, and gender equality practices are conceptualized as situated,

relational, and culturally embedded processes through which gender is done.

The thesis consists of four studies in two interconnected yet organizationally

distinct contexts within Sweden’s technology-intensive landscape: technical

higher education and the financial technology (fintech) industry. The first

context is a technical university, where future engineers study, and technical

knowledge is produced and disseminated. The second context is the Swedish

fintech industry, with a particular focus on rapidly growing scale-up

organizations in which engineers work and technological innovation is

commercialized. Both contexts are numerically male-dominated and subject to

expectations to engage in gender equality and diversity work, yet are

characterized by different organizational logics and understandings of change.

Methodologically, the thesis employs a qualitative research approach combining

interviews, job shadowing, and document analysis. The first two papers analyze

gender equality practices in engineering education and demonstrate howandrocentric cultures shape both the scope of such practices and the

organizational conditions that enable them. The latter two papers focus on the

fintech industry and analyze how understandings of diversity are constructed at

the industry and organizational levels, and how gender equality practices are

integrated into, and constrained by, homosocial cultures.

By synthesizing the findings from the four studies, the thesis makes three key

contributions. First, it demonstrates that gender equality practices in these

contexts are shaped more by cultural norms than by strategic problem analysis,

resulting in initiatives that signal progress while leaving deeper power structures

intact. Second, it advances understanding of homosociality by showing how

men’s engagement is both enabled and constrained by the masculine legitimacy

they embody, positioning them as legitimate actors of change, yet often without

disrupting underlying hierarchies. Third, it contributes to research on

organizational change by revealing the inherent ambivalence of gender equality

practices: practices aimed at transformation may simultaneously reproduce

gendered power relations.

Scarcity of battery raw materials leads to an increasing need to improve current battery recycling routes. Traditional hydrometallurgical recycling techniques use extensive amounts of chemicals to sufficiently recycle all elements of end-of-life lithium ion batteries (LIBs). To minimize chemical use and increase efficiencies, simultaneous precipitation has emerged as an alternative recycling technique. This thesis aims at creating and optimizing a closed-loop recycling process from a synthetic LIB recycling leachate. The first step is the simultaneous precipitation of nickel, manganese, and cobalt with oxalic acid as the precipitant, followed by subsequent precipitation of lithium carbonate with sodium carbonate as precipitant. The produced nickel manganese cobalt (NMC) oxalate is mixed with lithium carbonate and calcinated to obtain a NMC cathode, ready for use in battery applications. Optimization was carried out by varying the molar ratios, temperature, concentrations, and pH. 

The highest NMC removal while maintaining a low Li removal was reached at a 2.5 molar ratio for OA/Me²⁺, (where Me=Ni+Mn+Co) with a 1M concentration of oxalic acid added with a volumetric ratio of 1:1 to synthetic LIB leachate at room temperature. This yields a metal removal of 43.0% of manganese, 99.3% of cobalt, and 98.1% of nickel. In the second lithium recovery step, the highest Li removal was achieved at 90°C and a 1.05 molar ratio of Na/Li, where 76.1% of lithium precipitated. The last calcination step successfully converted the NMC oxalate to a NMC cathode. The process that gave the lowest level of cationic mixing was when the NMC oxalate was sintered at 500°C for 5h, then mixed with lithium carbonate and calcinated at 850°C for 12h. 

Electric and hybrid passenger vehicles offer the ability to recover braking energy and optimize power distribution, enabling significant improvements in efficiency compared to conventional vehicles. In this study, we quantify and compare the energy-saving potential of three optimization methods—mixedinteger linear programming (MILP), nonlinear programming (NLP), and dynamic programming (DP)—for a battery-electric vehicle following the Urban Dynamometer Driving Schedule (UDDS). A simplified longitudinal EV model with aerodynamic, rolling, and gravitational resistances is used to simulate vehicle dynamics, battery charging/discharging, and regenerative braking. The reference case enforces exact tracking of the UDDS speed profile, including late, hard braking at stops. Over a receding horizon of five seconds, MILP linearizes resistive forces and solves a linear program (≈16 ms per step), while NLP retains full nonlinear dynamics with a quadratic speed-deviation penalty (≈1.7 ms per step). DP serves as an offline global benchmark across a discretized state-of-charge and velocity grid. Results show that, relative to the baseline, NLP achieves the largest net-energy reduction (31.1 %), followed by DP (23.5 %) and MILP (18.2 %). NLP’s ability to model exact aerodynamic drag and smoothly coast into stops yields the highest savings but requires more computational effort. MILP offers a compromise between efficiency and real-time feasibility on automotive ECUs, although its hard speed bounds induce oscillatory speed errors. DP, while globally optimal under exact speed tracking, is best suited as an offline reference. These findings suggest that hybrid approaches—such as DP-trained lookup tables or variable-efficiency NLP surrogates—could approach global optimality while meeting real-time constraints. Future work will validate these methods on  additional driving cycles and incorporate variable motor/regenerative efficiency maps to better align with real-world performance. 

The use of Unmanned Surface Vehicles (USVs) in dynamic maritime environments demands robust real-time perception systems for safe obstacle avoidance. This thesis develops a LiDAR-based obstacle detection and tracking framework for hydrofoiling USVs. The proposed pipeline fuses 3D LiDAR point clouds from an Ouster OS1-128 with position and orientation estimates from an SBG Ellipse-D GNSS-aided INS. An important component of this work is a robust point cloud preprocessing step that filters water surface detections using attitude-informed RANSAC water surface estimation and dynamic radius outlier removal. The resulting clean data supports a Multi-Object Tracking pipeline that combines Euclidean clustering, Global Nearest Neighbor association, and Kalman Filter tracking. In parallel, a Rolling Local Occupancy Grid map is maintained to represent static obstacles, with a segmentation step separating static and dynamic clusters for each of the methods. All components operate in real-time and are integrated with ROS2. The pipeline was validated both in a Unity simulation environment, developed within this thesis, and with real-world datasets collected from the Evolo hydrofoiling USV. Results demonstrate strong tracking performance in sparse conditions (MOTA: 91.6%, MOTP: 1.99 m). However, in cluttered nearshore environments, local occupancy mapping proved significantly more robust than object-based tracking. Combined with the proposed preprocessing step, this map-based approach is recommended as the perception layer for future onboard obstacle avoidance on Evolo. The evaluation also confirms LiDAR as a suitable sensor for this application, proving capable of detecting even small obstacles with sufficient range (65 m) and resolution.

The automotive industry is rapidly evolving through the integration of Ad- vanced Driver-Assistance Systems (ADAS), which enhance safety and comfort by combining sensors such as radar, lidar, cameras, and ultrasonic devices. Radar is gaining momentum as a key sensing technology thanks to its ro- bustness, affordability, and reliable performance under diverse weather and lighting conditions, and it is being increasingly adopted by several Original Equipment Manufacturer (OEMs). As ADAS progresses toward higher reso- lution and wider coverage, antennas operating at millimeter-wave frequencies, particularly in the 77 GHz band, face increasing challenges in design and fab- rication. One of the main limitations is the appearance of ripple when the antenna’s field of view (FoV) becomes very large. This thesis, conducted in collaboration with KTH Royal Institute of Technology, Gapwaves AB, and Northern Waves AB within the SEMA project (6G Sustainable and Energy- Efficient Metallic AM Antennas and Microwave Components), addresses this issue by focusing on improving a waveguide slot array antenna for automotive radar applications. The main objective is to reduce ripple in the radiation pattern while maintaining a wide FoV of ±60◦, ensuring a low-profile design suitable for vehicle integration. The study employs CST Studio Suite for full-wave simulations, as well as COMSOL Multiphysics and a KTH-developed ray-tracing tool. These tools are used to design and optimize both the antenna and a dielectric radome with lensing capabilities. Coupling reduction was achieved through the implementation of soft sur- faces that suppressed surface currents between adjacent elements, resulting in coupling levels below –40 dB across the operating band and significantly reducing ripple in the radiation pattern. However, the introduction of these surfaces was observed to reduce the antenna’s field of view. To compensate for this effect, a dielectric lens was designed to extend the FoV close to its original value, targeting at least ±60◦. The lens was integrated with matching lay- ers to minimize reflections while providing mechanical support to the overall structure. However, further investigation is required to assess the interaction between two adjacent lenses when placed in close proximity, as preliminary results indicate that energy is redirected under such conditions.

Crosstalk in printed circuit boards (PCBs) poses challenges for signal integrity in high-frequency and compact systems, yet mitigation is often guided by empirical rules rather than systematic analysis. It is of particular concern in applications such as high-speed digital systems, multi-antenna wireless systems, and automotive radar, where unwanted coupling can lead to bit errors, degraded channel isolation, or false detections with direct impact on end-user performance and safety. To systematically study crosstalk in PCB transmission lines, this thesis adapts a reaction-theorem-based method originally developed for antenna placement optimization. Within this framework, two formulations are employed to evaluate coupling: the coupling density, denoted 𝜅, related to 𝑆₂₁, and the generalized impedance density, denoted 𝛿₂₁, related to 𝑍₂₁, which is proportional to 𝑆₂₁ under weak coupling. Simulations of microstrip and stripline geometries in CST show that both 𝛿₂₁ and 𝜅 visualizations identify the regions responsible for coupling. For countermeasure design, 𝜅 is particularly useful, since results demonstrate that placing shorting vias or continuous copper walls in regions of positive 𝜅 reduces crosstalk, whereas placement in negative 𝜅 zones increases it. In striplines, an iterative application of this method can suppress coupling almost completely in a narrowband case, achieving more than 50 dB reduction in |𝑆₂₁|. The method can be easily integrated into existing electromagnetic design workflows and was found to add only a small computational cost compared to the full-wave field simulations already required for such systems. The results establish reaction-theorem-based visualization as a systematic tool for identifying coupling paths and guiding countermeasure design in PCBs. Beyond the methodological contribution, the approach offers potential for more sustainable electronics design through reduced design iterations and material use, while at the same time supporting improved signal integrity, better reliability in high-frequency operation, and more efficient use of compact layouts.

Despite the increase of music generated by Artificial Intelligence (AI) tools such as Suno and Udio, little research has been done on the songs they create. Previous research has largely focused on techniques for AI music detection, while the potential biases and patterns in the vocals the models generate have been left unanalyzed. This thesis aims to develop a pipeline that allows for population analysis of singing voices present in music generated by Suno and Udio. In order to accomplish this, we investigate two types of methods. One approach uses Mel-Frequency Cepstrum Coefficients (MFCCs) for feature extraction together with Gaussian Mixture Models (GMMs) to model vocal characteristics. The other approach uses deep learning models to extract features directly, with two speaker recognition models and one singing voice representation model. We evaluate both methods through testing with real songs that we process using source separation and silence removal. Based on the initial test results we then apply the most reliable model — the singer representation model — to the dataset of AI singers and use K-medoids clustering as well as Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction to examine the data. The results from the singer representation model showed only limited quantitative success with the K-medoids clustering, while qualitative testing on real songs suggests the approach is somewhat successful. The UMAP projection showed quite distinct separation between the Suno and Udio songs, as well as between female and male vocals. Issues with distortion or incorrectly converted audio files in the datasets we used were discovered, which likely had significant impact on the clustering and visualization results. An implementation error in our UMAP usage of the MFCC approach initially showed it as being much less reliable than further testing seemed to indicate, making this method possibly interesting for more thorough testing. We motivate further research into how the voice characteristics of the generated vocals relate to their textual prompts and vocals of real artists.

Unmanned aerial vehicles (UAVs) are increasingly used in civilian and military domains, raising concerns regarding security, privacy, and airspace safety. Reliable detection and recognition of UAVs are therefore critical. Among various sensing modalities, machine learning with radio frequency (RF) signals oers a low-cost and robust solution, capable of operating eectively even under adverse environmental conditions. However, most existing RF-based recognition systems assume a closed-set scenario, where all UAV types are known during training. This assumption is unrealistic in practice, as novel and unknown UAVs frequently emerge. As a result, conventional classifiers suer from overconfident misclassifications, retraining from scratch is inecient, and incremental learning approaches face catastrophic forgetting when adapting to new classes. This thesis proposes a unified framework that addresses these challenges by combining open-set recognition, clustering-based novel class discovery, and incremental learning with memory replay. First, a signal-semantic open-set recognition method separates unknown RF signals from known classes. Next, novel UAV categories are discovered through clustering with automatic cluster number selection using K-Means and Gaussian Mixture Models (GMM), supported by loss functions that enforce intra-class compactness and inter- class separability. Finally, an incremental learning module incorporates these new categories into the model while mitigating catastrophic forgetting via a replay strategy with minimal storage overhead. The framework is validated on a real-world dataset comprising 18 known UAV classes and 6 unknown classes. Experiments demonstrate that the proposed approach achieves clear separation of unknown classes in semantic space, robust cluster validity under noisy conditions, and eective incremental learning with minimal sample replay. Overall, the results establish the framework as a practical and scalable solution for RF-based UAV recognition in dynamic and open-world environments.

Accurate characterization of grain size and distribution in metallic materials is extremely important for predicting their macroscopic properties and ensuring quality in industrial applications. Traditional metallographic analysis relies on time-consuming manual methods or automated image processing techniques with limited-precision. This thesis investigates the development and application of a deep learning pipeline for the automated segmentation of grain boundaries in Light Optical Microscopy (LOM) images of electrical steels, leveraging Electron Backscatter Diffraction (EBSD) data to generate annotations for model training. To answer the central research question of whether a deep learning model trained on EBSD-LOM correlative data could replicate expert analysis, a comprehensive evaluation of leading segmentation architectures was conducted. U-Net and U-Net++ architectures with ResNet encoders were optimized and rigorously assessed not only on pixel-level accuracy but, more critically, on their ability to reproduce key microstructural statistics that are vital for materials characterization. The study successfully demonstrates that an appropriately configured deep learning model can achieve results that are in very close agreement with manual annotations. The optimal model proved highly effective, accurately replicating grain count, size distributions, and standard industrial metrics like the ASTM Grain Size number. A subsequent investigation into post-processing with Generative Adversarial Networks (GANs) for post- processing indicated limited utility for global enhancement of already high- quality segmentations, unless the process was performed locally under user- supervision. This thesis validates that appropriately configured and optimized deep learning models, trained with EBSD-derived annotations, can provide a robust and accurate automated solution for quantitative microstructural analysis.

Localization is critical for autonomous vehicles to reliably estimate their position and orientation within complex and dynamic environments. This thesis develops a lightweight, computationally efficient localization algorithm with guaranteed convergence, specifically tailored for intersection scenarios. Intersections pose significant challenges due to vehicles approaching from multiple directions and frequent occlusions from surrounding vehicles, complicating observability and localization accuracy. The proposed bearing-based Riccati localization method integrates bearing measurements from known static landmarks and vehicle velocity measurements (linear and angular). The method is validated experimentally using a scaled vehicle platform (SVEA) equipped with a camera and ArUco markers as landmarks, simulating intersection conditions at scale. Scenarios examined include nominal driving conditions, aggressive maneuvers near operational limits, and degraded sensing conditions characterized by sensor inaccuracies and measurement noise. The approach demonstrates high accuracy and robustness, maintaining reliable localization even during dynamic maneuvers and sensor disturbances. This thesis underscores the potential of Riccati localization as a practical and scalable solution for autonomous vehicle localization at intersections. Future work should involve formal analysis of tolerable delays to maintain system stability, exploration of cooperative multi-agent scenarios, and integration with advanced communication infrastructures such as 5G to further enhance real-time localization performance.

We present a framework that enables autonomous systems to collect rich, informative data from partially unknown nonlinear plants while strictly respecting stability and hard state–input constraints. Starting from a stabilisable linear approximation, the unmodelled residual dynamics are learned on-line with Gaussian-process (GP) regression, which delivers both a mean estimate and high-confidence variance envelopes. These uncertainty bounds are embedded in a probabilistic control-invariant set (PCIS) that contracts the state space to a region where every state is guaranteed to remain safe with probability 1 − 𝛿. At each sampling instant a small quadratic programme selects the control input that (i) keeps the state inside the PCIS, (ii) tracks the nominal stabilising input as closely as possible, and (iii) actively excites poorly modelled directions so that the GP posterior variance shrinks over time. As learning progresses the PCIS expands automatically, allowing progressively more aggressive exploration without sacrificing guarantees. The framework is validated on two benchmarks of increasing complexity: (a) a two-state, unstable polynomial system and (b) a laboratory three-tank process with multi-input actuation and water-level constraints.

This thesis investigates autonomous waypoint navigation for the single-strut electric hydrofoil vessel FoilCart II through the integration of modeling, planning, and control. A nonlinear six-degree-of-freedom first-principle model was derived to capture vessel dynamics and used as the baseline for controller design. A waypoint-based global planner was implemented using Bézier-spline trajectories with explicit acceleration minimization to satisfy seakeeping comfort limits. At the control level, a hierarchical architecture was developed, combining a nonlinear model predictive controller (MPC) at the coordination layer with the existing linear quadratic regulator (LQR) at the execution layer. The integrated MPC-LQR framework achieved substantial improvements in trajectory tracking compared to LQR alone. In the tested waypoint mission, position and heading root-mean-square errors were reduced by an order of magnitude, while accelerations remained within the ±0.25𝑔 comfort band. Real-time feasibility was demonstrated with an average MPC solution time of approximately 9.5ms at a 10Hz update rate. In parallel, this thesis also explored model augmentation from resid- ual dynamics using the Wide Array Nonlinear Dynamics Approximation (WyNDA) algorithm, which is a data-driven dynamics approach that augments a known baseline model by treating it as the fixed part of the dynamics and discovering the missing physics as residual terms. Although it did not succeed in identifying a usable discrepancy model under the available data, WyNDA exhibited accurate state estimation and demonstrated real-time computational performance, highlighting its potential for future augmentation with higher- frequency sensor data. The main limitations of the study include reliance on a non–real–time simulator, the assumption of fully observable states, and validation restricted to cruise speeds of around 8 m/s. Nevertheless, the results confirm that combining trajectory planning with a two-layer MPC–LQR structure provides a solid foundation for autonomous hydrofoil navigation and serves as a baseline for further model augmentation and real-world implementation.

The continuous advancement of artificial intelligence and big data in the current era has driven increasing demands for computing power. However, traditional von Neumann architectures are constrained by processing limitations, making further enhancements in computing power prohibitively expensive. Memristors, as devices that integrate storage and computing capabilities, have emerged as a promising research field. Nanofluidic memristors, in particular, have garnered significant attention due to their ease of fabrication and biocompatibility. This project developed a funnel- shaped nanofluidic memristor with an asymmetric structure. Using an ionic liquid (BMIMTFSI) and an electrolyte solution (KCl), the device achieved a memristive effect based on asymmetric ion transfer. Distinct hysteresis curves were obtained across seven sweep rates, ranging from 50 mV/s to 200 mV/s. Furthermore, to account for variations in the taper angle within the asymmetric structure, 18 devices with taper angles ranging from 0° to 90° were fabricated and tested at various angles. Distinct patterns in the hysteresis loop opening area were observed. The relationships among the hysteresis loop opening area, sweep rate, and taper angle were qualitatively analyzed. In addition, to explore the neuromorphic computing capabilities of memristors, pulse tests were conducted on the fabricated devices. By applying a specific number of read and write pulses, the short-term plasticity (STP) and short-term depression (STD) functions were measured. Finally, the basic functionalities of the devices, such as handwritten digit recognition, tic-tac- toe game and a simplified super mario game were evaluated through online simulations.

This thesis presents a comprehensive analysis and evaluation of generation shift key (GSK) approaches within flow-based market coupling (FBMC) simulation for power systems. As European electricity markets transition toward full liberalization through regional market coupling, FBMC has emerged as a critical methodology for optimizing cross-border electricity trading while accounting for transmission network constraints. However, the strategies and methods for calculating GSKs remain underdetermined and require further investigation. This study develops a mathematical framework that presents the transmis- sion model in the FBMC, encompassing nodal optimal power flow, FBMC optimization, and redispatch modeling. In this research, five GSK strategies are examined, including two novel approaches. The proposed strategies include one with non-negativity constraints and another that allows negative values for enhanced flexibility. These are compared against existing GSK strategies through comprehensive simulations. The methodology employs a system with three zones, examining four seasonal cases under both perfect and imperfect forecasting scenarios. Each case incorporates realistic load and renewable energy variations. Results demonstrate that the proposed GSK strategies offer significant adaptability advantages over traditional approaches. The total operational costs analysis reveals that the proposed GSK strategies can achieve cost reductions compared to existing approaches. The methods also show a better performance in system marginal cost optimization and reduced redispatch requirements under forecast uncertainty. This research contributes to sustainable power system development by enhancing FBMC efficiency, supporting renewable energy integration, and advancing economic welfare in electricity markets. The adaptive GSK framework provides transmission system operators with improved tools for managing cross-border electricity flows while maintaining system security and economic efficiency.

Typical electrical insulation such as that used in power transformers consists of some form of cellulose such as paper or pressboard, impregnated with an insulating liquid such as a mineral or ester oil. One of the most significant factors affecting insulation degradation is moisture, which accelerates aging and increases the risk of dielectric failure. In recent years, ester-based oils have emerged as environmentally friendly alternatives to traditional mineral oils, raising questions about their dielectric behavior under various moisture conditions. This Master’s thesis investigates the effect of moisture content on the dielectric properties of oil-impregnated pressboard insulation using Dielectric Frequency Response (DFR) measurements. The study examines three widely used insulating oils: mineral oil (Nytro 10XN), synthetic ester (Midel 7131), and natural ester (FR3), across four moisture content levels: <0.5 %, 1 %, 2 %, and 4 %. Pressboards were initially dried, then impregnated with oil and conditioned in an enclosed environment using saturated salt solutions to achieve specific moisture levels. Karl Fischer titration was used to determine moisture content, while DFR measurements were conducted using the Megger IDAX 300 system over a frequency range of 1 mHz to 1 kHz. Results showed that the dielectric losses tan δ increased with moisture content for all oil types. In general, ester oils showed higher dielectric losses than mineral oil at all moisture levels. In the cases with a higher ε′′ but lower tan δ, ester oils indicated that they may have higher polarizability, particularly in the low frequency regions.

The growing demand for efficient and reliable Electric Vehicle (EV) powertrains has made accurate battery monitoring and thermal management essential for safety and performance. A key challenge is estimating internal cell temperature, which cannot be measured directly and is usually inferred from temperature sensors mounted on the surface of the cell using complex thermal models. Thermal behavior at the surface of the cell can vary significantly to that which is occurring internally and can take longer to notify the system of a potential thermal event. This thesis explores electrochemical impedance spectroscopy (EIS) as a method for internal temperature estimation using prototype Battery Management System (BMS) hardware with built-in EIS capability. A Python-based framework was developed to au- tomate frequency sweeps and collect data across a controlled temperature range from −10 ^◦C to 40 ^◦C to build a dataset of impedance vs. temperature measurements. The hardware was tested on a commercial lithium-ion cell in a thermal chamber and vali- dated against a laboratory-grade frequency response analyzer (FRA). Results show that the hardware reliably measured EIS data with the expected inverse relationship between impedance and temperature. From literature it was found that a frequency range from 10 to 100 Hz offered a method for temperature estimation using the impedance measurements which has minimal influence from the battery state-of- charge or state-of-health. These findings confirm the feasibility of EIS-based temper- ature estimation in EVs, though further refinement is needed to improve robustness and reduce data requirements.

This thesis focuses on the field of HVDC transmission systems, specifically exploring the modular multilevel converter (MMC) topology. A new variation, the Flying Capacitor Hybrid Modular Multilevel Converter (FC-HMMC), is proposed. While the original topology utilizes IGBT devices, these components are prone to fail due to high voltage and current during turn-on, leading to latch-up or avalanche breakdown. In contrast, thyristors typically short due to physical or electrical abuse, with very few failing open. However, integrating thyristors into the FC-HMMC topology presents new challenges that require a thorough investigation into the commutation process. The significance of this work lies in its contribution to analyzing and proposing solutions to the problems. The project involves building a detailed simulation model of the FC-HMMC from the ground up using PSCAD, analyzing its operating principles, and proposing a modified control algorithm to enable the replacement of IGBTs with thyristors. The proposed method is implemented and validated through simulation, with results demonstrating its effectiveness. This work provides new insights into converter design and control, paving the way for future research and practical applications in high-voltage power electronics.

Early identification of cerebral palsy (CP) is vital for timely intervention but hampered by scarce, highly imbalanced datasets. This thesis investigates whether a single, domain-motivated transformation— left–right reflection of eight joint-angle time-series filmed with smartphone videos can improve CP- risk classification. A cohort of 723 recordings from more than 400 at-risk infants was processed into movement clips and split infant-wise into an 80 pool and a 20 Two pipelines were evaluated: (i) TodyNet, a graph-convolutional multivariate time-series classification network, and (ii) a logistic-regression baseline trained on handcrafted kinematic features. Both models were trained with and without reflection using 5 × 5 stratified cross-validation. Reflection increased TodyNet’s mean sensitivity by 8.6 percentage points in cross-validation (𝑝 < 0.001) and by 18.7 pp on the test set without reducing specificity, while logistic regression gained 3.8 pp in specificity and 3.4 pp in overall accuracy. The results show that a free, biologically plausible augmentation can regularise both deep and classical classifiers, narrowing the gap between general-purpose Multi variate Time Series Classificatio (MTSC) models and clinically acceptable screening performance.

The rising demand of electricity from renewable energy sources calls for an improvement in the efficiency of power generation and transmission. Modular multilevel converters with their merits of modularity, scalability, and reduced harmonic distortion offer several advantages for medium and high voltage applications, making them interesting topologies for alternating current/direct current conversion. However, their employment comes at the expense of increased control complexity. Converter prototypes are therefore commonly adopted to validate the performance of novel control concepts as part of research activities in the field. This project is concerned with the development of a heat run test setup, which is used to electrically and thermally stress submodules prior to their commissioning to a modular multilevel converter prototype. More specifically, the functionality of the setup is extended to accommodate testing of submodules designated for a modular multilevel matrix converter prototype. Test conditions reflecting the operation of the converter are attained by adapting the voltage and current control loops, consisting of proportional-integral and proportional-resonant controllers respectively. This enables tracking of alternating current and voltage quantities. During the control development of the heat run test, an existing control hardware- in-the-loop setup was used. To ensure that the real-time simulations are accurate with respect to reality, aspects related to power electronic modelling such as sub-cycle averaging and choice of discretisation time are considered. Stable operation and robustness of the control implementation are guaranteed by varying parameter values of the simulated electrical model. Finally, a fidelity analysis is performed. The results confirm the high performance attainable by employing sub-cycle averaging in the modelling of power electronic circuits, as steady-state measurements from the real-time simulation platform and experimental setup show a high degree of overlapping in the time and frequency domains. Certain discrepancies were observed in the values of maximum current ripple and current transient response between the two systems, suggesting that a better estimation of inductance would have been needed to more accurately simulate the dynamics of the system. The results of the analysis can be used to direct future efforts aiming to further improve the fidelity of the hardware-in-the-loop system, while the developed control implementation can be used in the heat run test setup to accommodate testing of submodules for a modular multilevel matrix converter prototype.

This article develops the improved scalable analytical model (iSAM) for predicting the performance of the surface-mounted permanent-magnet (SPM) motor with the arbitrary magnet shape and magnetization direction. It extends the ability of the conventional analytical models. The database of the analytical model can store the electromagnetic information for the SPM motors, which enriches the scaling laws for electrical machines. As for the magnets with arbitrary shapes, they are divided into finite rectangles or fan-shaped segments based on their magnetization directions, which are then replaced by equivalent magnet currents along the magnet surfaces. Then, based on the iSAM, the air-gap field of the SPM motor can be obtained considering the iron nonlinearity. This method is beneficial for accelerating the optimization process aimed at reducing torque ripple in servo motor applications. The finite-element method (FEM) for eccentric magnetic pole servo motors is executed to verify the effectiveness of the proposed method. Additionally, a prototype with parallel magnetized PM is tested to demonstrate the accuracy of the iSAM. The proposed method is particularly useful for the modeling techniques and fault diagnosis of electrical machines.

Per- and polyfluorinated alkyl substances (PFAS) are a group of more than 4700 man-made chemicals. The commercial production of PFAS began in 1960 and continues today. They are used in, among other things, fire-fighting foams, surface coatings and packaging materials. Their persistent properties and resistance to degradation have led to the spread of PFAS in nature, which in several cases has caused water contamination. This report aims to investigate which PFAS occurs in drinking water, health risks with a focus on cancer linked to exposure, and treatment techniques for treating water contaminated by PFAS.

The PFASs most commonly detected in nature, including drinking water, are PFOA and PFOS. PFAS emissions in Sweden come mainly from fire-fighting foams that are often used during fire drills by the Swedish military. Other sources of spread can be emissions from industries that use or produce PFAS. Regulations and limit values for PFAS have become stricter in both the EU and Sweden, which has strengthened the need for effective treatment methods. The report examined three separation techniques for the purification of PFAS, which are adsorption with activated carbon, ion exchange and membranes. The results of the literature study show that the degree of purification was highest for membranes, followed by ion exchange. Activated carbon shows a lower efficiency, especially for short-chain PFAS, but is on the other hand more established and cheaper to purchase.

Studies on PFAS and cancer show that PFOA is the most studied PFAS and has the strongest evidence for an increased risk of certain cancers. These cancers are mainly kidney and testicular cancer. More variable results have been identified for PFOS, and for newer PFAS (such as GenX and PFBS) there is a lack of sufficient research to draw clear conclusions about health effects in humans. In addition, the mechanism of how PFAS potentially causes cancer has not been established, however, there are theories based on non-genotoxic mechanisms (does not damage DNA directly)

The political tensions of the Cold War influenced both sport andscience. There is a long-standing tradition within both liberaldemocracies and authoritarian regimes of emphasizing the pur-portedly positive connections between sport, training, and mili-tary capability. The Swedish total defense is an interesting casein this context, given its social and political significance, the linksbetween the military and sport, and the extensive investments inscientific research. However, the application of sport and trainingscience in Swedish defense research has remained largely unex-plored. This article examines the use of sport psychology in theSwedish Armed Forces, with particular attention to the “psycholo-gization” of Swedish fighter pilots in the late 1970s. It analyzes thediscussions surrounding, as well as attempts to apply, knowledgeof mental training—including relaxation, meditation, and hyp-nosis—to enhance pilots’ training and performance. The articlecenters on the Swedish Defense Research Establishment (FOA)and the Division of Human Sciences, where Lars-Eric Uneståhl’smethods were debated and implemented.

This study explores various 3D printing techniques (dry and wet extrusion) as a tool to obtain biobased, porous absorbents with tailored porosity and shapes. The techniques explored herein included extrusion, 3D pellet printing, and filament making (dry extrusion), as well as 3D gel printing (wet extrusion). Furthermore, blends of biopolymers were explored to determine the best combination that allows for reproducible porous shapes. The results demonstrated that, with the current state-of-the-art, the 3D pellet and the filament maker printer require further optimization and formulation development to produce a printed porous shape. On the contrary, the use of a mini extruder enabled the successful extrusion of all formulations tested. Here, samples containing blends of proteins from various sources, lignocellulosic biomass, and sodium bicarbonate as a foaming agent exhibited a lower density (774 kg/m³) than those extruded without SBC. The extruded materials were highly porous, which allowed for a better free swelling capacity (FSC) in saline solution. Moreover, some formulations exhibited better blood absorption and more efficient liquid penetration than previous reports, which can be attributed to a more permeable porous network.

For the wet extrusion route, the combination of proteins, polysaccharides, and CaCl₂ was key to allowing for printable shapes both by hand and using a 3D gel printer at room temperature. However, the printed shapes had lower swelling capacities compared to those extruded, due to a very solid structure observed in the microstructural analysis. Overall, the study confirms the potential of creating formulations based on blends of biopolymers towards 3D printable shapes with high absorption properties. The high absorption property is relevant to creating customizable shapes used as alternatives in sustainable absorbent applications. This study highlights key challenges and opportunities in formulation development and 3D printing as an alternative technique to conventional reactive extrusion forming highly porous materials. 

This thesis investigates how artificial intelligence and computer vision can be utilized to automate the identification of installation symbols in digital electrical drawings for the company mbiz AB. In the electrical installation industry, planning and selfmonitoring are currently performed largely manually, which is both time-consuming and labor-intensive. A key technical challenge is that the symbols are extremely small relative to the overall size and high resolution of the drawings, complicating automated analysis. The study evaluates a methodology based on the object detection model YOLOv11 in combination with SAHI (Slicing Aided Hyper Inference), a technique that divides the drawing into smaller segments to preserve detail. The results show that the model achieved a Mean Average Precision (mAP) of 0.97, confirming that patch-based inference is an effective method for handling technical documents. The study resulted in a functional prototype that integrates visual detection with automated reading of drawing numbers and coordinate translation to PDF format. The conclusion is that the technology has significant potential to increase efficiency and quality assurance in future installation projects by reducing manual tasks.

This work presents a numerical investigation of the thermo-hydraulic performance of a tube-bundle cavity (TB)receiver for parabolic trough collectors (PTCs). The proposed receiver replaces the conventional single absorbertube with multiple smaller tubes arranged in a circular bundle, forming a linear cavity that improves solar absorptionand reduces temperature non-uniformities on the absorber surface. A three-dimensional CFD model isdeveloped under real-scale operating conditions to assess several TB configurations through a two-stage optimizationprocedure. The designs are evaluated using thermal efficiency, pressure drop, and overall efficiencymetrics. The results indicate that, despite higher flow resistance, TB receivers significantly enhance thermalperformance compared to the conventional design. Hotspot temperature increases are reduced by up to 77%,while temperature uniformity increases by approximately 23%. Among the investigated configurations, the 12-tube design provides the best thermo-hydraulic compromise, achieving a maximum overall efficiency of 0.76 atan inlet temperature of 450 K, corresponding to a 7% improvement over the standard receiver. Additional analysesover inlet temperatures ranging from 400 to 550 K confirm the robustness of the optimized TB configurationin mitigating hotspot formation while maintaining superior overall performance.

Improving energy efficiency in industrial manufacturing remains challenging despite substantial technical potential. This has resulted in a persistent energy efficiency gap, which is increasingly understood as a socio-technical issue driven by not only technology limitations but also organizational and informational barriers. This study investigates how energy waste is targeted in practice through an in-depth single case study of an automotive company. Fifteen energy efficiency measures (EEMs) were analyzed and classified by type of energy waste addressed, digital technologies applied, and organizational knowledge required. The results show that industrial efforts primarily focus on reducing idling energy losses, while fewer measures address more complex forms of energy waste, such as over-processing losses. Digital technologies are mainly applied and rolled out at lower maturity levels, emphasizing energy monitoring and visualization. Further, different types of organizational knowledge are associated with targeting energy waste: technical knowledge dominates isolated interventions, process knowledge supports standardized technology diffusion, and leadership knowledge is required for cross-functional coordination. The findings highlight that bridging the energy efficiency gap requires the alignment of technological solutions with organizational knowledge and routines. This study contributes empirical insights into how manufacturing companies can structure and prioritize energy efficiency efforts and provides a framework to support the implementation of energy efficiency measures in practice.

In the context of increasing energy security pressures and national decarbonization targets, energy communities are increasingly discussed as mechanisms for supporting the energy transition by combining local energy production and consumption with new forms of participation and governance. In Sweden, however energy communities have developed in the absence of a formal legal definition and under relatively weak institutional support. This raises questions about how such initiatives emerge and operate in practice.

Against this backdrop, this thesis examines how regulatory, institutional and social conditions shape the development of energy communities in Sweden. Drawing upon a case study of two energy communities in Sweden including qualitative interviews with actors involved in Swedish energy communities, this study analyzes how regulatory ambiguity, incumbent actors, and social dynamics influence the operation and perceived value of energy communities.

The analysis shows that energy communities persist through flexible interpretations of existing rules, allowing energy communities to develop through informal arrangements and workarounds. This enables operation, but still carries uncertainty as well as introduces unintended consequences, particularly with regard to community visibility. While knowledge of and active participation in energy communities among the general public is limited, the findings show that formal participation is not a prerequisite for participants to derive material value from residing in an energy community.

This thesis presents the design and validation of a 15 kW three-phase AC/DC converter using Sinusoidal-Triangular Current Mode (S-TCM) modulation and Zero Voltage Switching (ZVS) using SiC MOSFETs. To implement S- TCM, Autonomous Gate Drivers (AGDs) are used in place of conventional drivers to allow the converter to achieve soft switching at all operating points, significantly reducing switching losses and improving thermal performance. These new drivers, while allowing operation at very high switching frequencies, impose an unconventional hysteretic current control, which was analyzed. The use of S-TCM results in a Variable Switching Frequency (VSF), which spreads the harmonic spectrum and reduces conducted Electromagnetic Interference (EMI) compared to the sharp harmonic peaks of fixed-frequency converters. Firstly, a complete analytical model of the AGDs was developed to predict the ZVS transitions and determine the required filter inductance to ensure soft switching while limiting switching frequency and minimizing losses. Additionally, the harmonics spread spectrum was calculated to be able to optimize the design of the grid-connected LCL filter. Secondly, this work includes the analysis of the grid current control in the αβ frame required for the plant, and the implementation of a switching frequency control for the AGDs hardware. The complete system was simulated in PLECS, achieving an efficiency of 98.51 % with THD of 0.8746 %, being able to operate with Power Factor Correction (PFC) and under 4-Quadrant if necessary. The results confirm that it is possible to implement, with AGDs and SiC switches, a high-frequency, soft-switching three-phase converter with low harmonic distortion and high efficiency, validated by simulations and optimized to fully leverage the advantages of S-TCM modulation. As a continuation, future work should focus on building a hardware prototype to validate the analytical design, refining the control strategy, and extending the approach to incorporate high level control.

The rapid growth of the Internet of Things (IoT) presents challenges for low-power devices. The ESP32-C6 System-on-Chip offers versatile connectivity options, supporting both Wireless Fidelity (Wi-Fi) 6 and Thread, making protocol selection crucial. This thesis aimed to investigate the trade-offs in Round Trip Time (RTT), power consumption, security, and scalability when using different network layer protocols (Wi-Fi 6 with and without Target Wake Time (TWT), and Thread) and application layer protocols (Message Queuing Telemetry Transport (MQTT), Hypertext Transfer Protocol (HTTP), and Constrained Application Protocol (CoAP)) on ESP32-C6 based IoT devices. An experimental methodology was used for measuring RTT and power consumption on an ESP32-C6 client communicating with a local backend server. Security and scalability were assessed qualitatively. The results indicated that Wi-Fi 6 with TWT, especially when paired with CoAP, yielded the lowest RTT and the lowest power consumption. Thread exhibited higher RTTs but offers mesh networking capabilities. CoAP generally outperformed MQTT and HTTP in RTT across network types, although the difference was minor when using Wi-Fi. Robust security options exist for all protocol stacks, and a defense-in-depth approach is recommended. Scalability analysis highlighted that Wi-Fi 6 suits high-bandwidth needs while Thread is better for large-scale low-bandwidth deployments, with CoAP being advantageous for the latter. These findings provide developers with empirical data to make informed protocol choices for ESP32-C6 based IoT solutions, emphasizing the significant benefits of Wi-Fi 6 TWT for low-power applications, though support for TWT is still maturing.

This work addresses the challenges of underwater localization in confined environments where an Autonomous Underwater Vehicle (AUV) needs to perform aggressive and unconventional maneuvers in order to move around. Such scenarios include close proximity operations such as operating near divers, which requires high localization performance due to safety concerns. We formulate the localization as an inference problem with a probabilistic factor graph, that fuses the proprioceptive sensors on the AUV, namely a Doppler Velocity Log (DVL), barometer, Inertial Measurement Unit (IMU), Attitude Heading Reference System (AHRS) and a Global Positioning System (GPS). We expand on this by introducing a motion model factor that captures the AUVs unique actuator dynamics, in order to constrain the estimate of the AUV. In order to assess the impact of the proposed localization framework and the motion model factor, a dedicated case study was conducted to benchmark three inference strategies: filtering, fixed lag smoothing, and iSAM2, using both simulated and real world datasets, with and without the factor. The results showcase that it is possible to integrate a motion model in the factor graph in order to constrain the velocity and pose of the AUV. But due to the model being inconsistent with the dynamics, we see a degrading estimate in all three inference strategies when including the model. However, without the motion model factor all three different incremental inference strategies perform comparably, with the fixed lag smoother offering a slight advantage.

As wireless communications systems are heading towards their sixth generation (6G), the operating band of radiating systems is expected to shift to higher frequencies to accommodate the increasing transmission rates per link, in the order of several Gbps up to a few Tbps. In this regime, antennas can be coupled with microwave dielectric lenses to increase their directivity exploiting the same working principles of optical lenses in order to focus the radiated beam. An important concern in the adoption of lens antennas is to reduce the mismatch losses and other undesirable effects caused by internal reflections, which must be assessed during the design stage. However, the computer simulation of dielectric lens antennas can be very time consuming with conventional software based on the Finite Element Method (FEM) or Finite-Difference Time-Domain (FDTD), because of the typical large electric size of lenses. In this thesis, an efficient approximate method based on Geometrical Optics (GO) ray tracing is implemented for the calculation of the reflection coefficient of a single feed and mutual coupling of multiple feeds in two-dimensional dielectric lens antennas, for arbitrary shapes and number of reflections. The rays are traced inside the lens and the Fresnel coefficients are applied for each reflection. From geometrical considerations, the fields are approximated on surfaces in proximity of the feeds and the mutual and self-impedances are calculated with the Generalized Reaction Theorem (GRT). The method is validated for lenses inside a parallel plate waveguide (PPW), used to reduce the problem to 2-D. The Python codes estimate the couplings in terms of S-parameters, which are compared to full-wave simulations in various lenses, showing the improvement in the required computational time.

WebAssembly (Wasm) is a promising technology advancement for serverless platforms because it can provide faster cold-start times than Linux containers due to its efficient binary format. A relatively unexplored aspect of Wasm in the context of serverless environments is observability, i.e., the ability to measure the internal state of an application workload by examining its output. Similarly, any application scenario, observability for serverless Wasmbased applications can be achieved by instrumenting the application, e.g., with OpenTelemetry based tracing code. Although such instrumentation can be used for effective tracing for workloads at the application layer, we propose here system-level observability to compensate for, e.g., missing telemetry from application developers. To support this, we leverage Extended Berkeley Packet Filter (eBPF) to enable observability for Wasm-based serverless applications. eBPF is an efficient kernel-level tracing technology that allows user-defined code to be executed within the Linux kernel, capturing event data without modifying the application code. It provides an instrumentation-free way to achieve runtime observability of applications, making it particularly suitable for serverless applications running in Kubernetes environments which are always based on Linux. We implement a prototype of an observability framework for serverless Wasm workloads running in Kubernetes using SpinKube framework. Our key contribution is an eBPF agent that collects data from the Wasm workloads. The performance evaluations of the prototype system show that our solution introduces a latency average overhead of 17.8% and a maximal CPU overhead of 8%.

This thesis investigates Federated Physics-Informed Learning, a framework that integrates physical laws into federated machine learning to improve model performance across decentralized datasets. In federated learning, multiple clients collaboratively train a model without sharing raw data. Incorporating physics into this setting offers a promising solution by enforcing constraints derived from known physical principles. The central research question addressed is: how can physics-informed modeling enhance learning in a federated setup? This problem is both timely and relevant for engineering applications, as federated approaches reduce the need for expensive or centralized data collection while maintaining privacy. Despite growing interest in physics-informed machine learning and federated learning separately, combining these approaches remains a relatively unexplored area, particularly for systems governed by partial differential equations. To address this, a series of experiments were conducted using a combination of centralized, federated, and ensemble neural network architectures. The models were trained on distributed datasets with varying overlap, and performance was evaluated in terms of accuracy, stability, and adherence to physical laws. Special attention was given to the enforcement of physical constraints, such as initial and boundary conditions as well as the overall residual. The results indicate that explicitly enforcing initial and boundary conditions over the partial differential equation residual significantly improves performance compared to alternative strategies. This approach enhances both predictive accuracy and physical consistency across decentralized clients, demonstrating the potential of federated physics-informed learning for practical engineering applications. These findings provide a foundation for future research into scalable, privacy-preserving, and physically consistent machine learning systems, offering a new approach to addressing complex problems where data is distributed and costly to acquire.

Electricity consumption has increased over the last decade, driven by electrification and the adoption of new technologies. While this development contributes to the United Nations’ Sustainable Development Goals, it also places growing stress on power grid infrastructure. To ensure reliability and optimal utilization, grid operators require accurate short-term load forecasts to support proactive decision-making. While system-level forecasting using machine learning is common, few studies explore granular forecasts, as they demand large datasets and significant computational resources. This thesis investigates whether short-term electricity consumption in the Stockholm region can be forecast more accurately using granular customer metering point data compared to aggregated data from primary substations at the system level. Approximately 2.5 years of hourly consumption data from 500000 metering points were collected and clustered to create more homogeneous groups. XGBoost models were trained on both granular and system-level datasets using GPU-accelerated distributed computing to handle the scale. The results show that, despite the availability of more detailed features, models trained at the customer metering point level performed worse than models trained on system-level data across all prediction weeks. This suggests that increased granularity does not automatically improve forecasting accuracy, particularly when zero consumption values and limited feature engineering are present. The findings highlight both the potential and the limitations of bottom-up forecasting. While large scale GPU-accelerated training is technically feasible, future work should focus on improved feature engineering, handling zero values, refining clustering strategies, and testing advanced models such as transformers. This study provides practical insights for researchers and grid operators considering granular approaches to short-term load forecasting.

In modern industrial and computing environments, maintaining the optimal performance of critical components—such as fans, power supply units, and other hardware is essential to ensure system reliability and operational efficiency. This Bachelor’s thesis project presents a monitoring system that integrates sensors, real-time data analytics, and visual alerts to track the operational status of the fans, helping prevent failures, reduce maintenance costs, and improve overall system stability. The monitoring system consists of sensors that transmit data to a web application updated every 20 minutes. Users can access the interface from both mobile devices and computers by connecting to the local Wi-Fi network and entering the correct IP address. The system provides a clear visual status: green indicators show that fans are operating normally, while red signals identify fans requiring attention. The monitoring system is developed for CAE, a global leader in flight simulator manufacturing and pilot training. It is customized to monitor the numerous fans inside the simulators. For example, monitoring the fans for cooling the high-voltage hardware, like the 33kV projectors used in the visual systems. Malfunctioning fans can cause these projectors to shut down, interrupting training sessions and potentially leading to an “Aircraft on Ground” (AOG) status. Since simulator training is costly and involves pilots traveling from across the globe, any downtime results in significant financial and logistical consequences. By providing real-time insight into hardware conditions, the system enables CAE’s maintenance team to detect issues early, such as fans rotating too slowly, and take prompt action. The primary goal of this project is to minimize simulator downtime and help maintain uninterrupted, high-quality training operations.

As the digitalisation of society continues, the use of Large Language Models (LLMs) to aid in software development has become valuable. Despite their potential, LLMs often struggle with hallucinations, underperform on domain-specific and complex tasks, and cannot handle tasks that depend on classified or proprietary information. This thesis aims to explore whether coding knowledge extracted from human coding sessions, termed insights, can be used to improve code generation. Methods for extracting and applying these insights were explored, alongside experiments with fine-tuning an LLM using Group Relative Policy Optimisation (GRPO). Different prompts and reward functions were tested, and performance was compared with a pre-trained reference model. The findings show that GRPO improved code generation accuracy from 32.6% to 37.9% with only ≈1000 training samples. GRPO was also effective for teaching the LLM to follow a specific format while not requiring a carefully curated dataset. While insights are effective for error correction, further work is needed to establish their broader impact on code generation.

Modern data centers are facing an imbalance caused by the increasing network line-rate (now reaching 400 Gbps and moving towards 800 Gbps), creating a performance bottleneck in the capability of conventional CPU-based in-server systems. The industry solution involves deploying the DPU (Data Processing Unit) to offload network-intensive tasks, which offers many advantages: data takes the shortest path, the host CPU is freed for application logic, and a physically isolated data layer. This thesis presents the design and evaluation of NVIDIA’s BlueField-3 DPU, focusing on the inside processor: DPA (Data Path Accelerator), which manages the shortest data path within the DPU. We select the key-value storage function as a representative network-intensive task to thoroughly test the hardware’s performance boundaries. This work investigates the architectural trade-offs between two DPU storage modes: the self-contained Stand-alone Storage and the collaborative Host Cache Storage. To optimize performance within these paradigms, we designed and implemented distinct underlying data structures to leverage and benchmark the DPA’s cache and memory hierarchy. Our results confirm that high performance is achieved through a hybrid memory strategy and using data structures and algorithms with a minimal memory footprint for optimal cache residency. Using these optimized approaches, the BlueField-3 DPA can achieve line-rate performance for innetwork Key-value stores, reaching a peak bandwidth of up to 90 Mops and achieving an effective offload rate of 70 − 85% with a cache-conscious DPA cache structure design.

The 5G core network serves as the central component of the 5G system, responsible for managing connectivity, mobility, authentication, and service delivery across diverse access networks. Built on a service-based architecture, it enables flexible interaction between modular network functions through standardized APIs. Within this architecture, the network exposure function (NEF) plays a key role in securely exposing selected core network capabilities to external application functions, facilitating controlled access, policy enforcement, and service innovation. As 5G services expand, improving the efficiency and scalability of network function exposure has become a critical research focus in the design of flexible and programmable service exposure architectures. This thesis investigates the state of the art of current 5G core network exposure solutions and identifies several limitations, including architectural complexity, high deployment overhead for exposing new services, and poor scalability. To address these issues, it proposes a generic network function exposure strategy based on an enhanced service communication proxy (SCP) and predefined configuration files. The proposed approach is evaluated using two representative services from existing 5G NEF functionalities: authentication and key management for applications (AKMA) and UE_identifier. The results demonstrate that, while maintaining equivalent functionality and security levels, the solution significantly improves the flexibility of service exposure. This work contributes to the ongoing efforts to design lightweight, scalable service exposure frameworks for 5G and beyond networks.

Manual test case generation is a major bottleneck in the automotive industry, consuming significant engineering resources and slowing software verification cycles. This thesis addresses this challenge by developing a Hybrid Retrieval- Augmented Generation (RAG) pipeline that leverages Large Language Models (LLMs) to automate test case creation for Electronic Control Units (ECUs) within a custom testing framework. The pipeline integrates vector search, BM25 keyword retrieval, and Reciprocal Rank Fusion (RRF) to improve retrieval relevance and code generation accuracy. Three configurations were evaluated: baseline vector retrieval, Hybrid RAG, and Hybrid RAG with RRF across 100 test cases. The results show that the hybrid methods improve generation accuracy and functional correctness by up to 15%, while reducing test creation time from 15 minutes to approximately 15 seconds, with minimal runtime overhead. However, RRF’s benefits varied depending on query types, indicating that retrieval fusion effectiveness is contextdependent. Although further refinement is needed to enhance retrieval precision, contextual grounding, and consistency across diverse testing scenarios, the proposed RAG pipeline demonstrates overall that LLM-assisted automation can substantially accelerate ECU verification while maintaining code quality.

This thesis examines how adaptive user experience strategies can improve transparency around generative AI in journalism, with the dual goals of strengthening reader trust and preserving editorial integrity. Conducted in collaboration with Schibsted ASA, the study followed a three-phase, mixed-methods design that included speculative design exercises to explore future transparency scenarios. Current transparency practices in AI-driven journalism are largely declarative and static, relying on simple “AI-generated” labels that do not clarify how or why AI contributed to the final output. This limitation weakens accountability and reduces user trust, which motivates the research presented in this thesis. Semi-structured interviews with newsroom editors, UX designers, and AI specialists identified core challenges and user requirements, which informed the creation of three interactive prototypes: a CMS disclosure builder, an AI “show thinking” chatbot, and a customizable content settings panel. Each prototype was evaluated through expert heuristic review, think-aloud testing with editors, and a value-alignment assessment to examine usability, workflow fit, and trust effects. Findings show that progressive disclosure controls and provenance indicators help balance cognitive load with editorial oversight, while chain-of-thought explanations enhance verifiability but require personalization safeguards. In contrast, a generic settings panel without provenance cues reduced both trust and perceived control. Contributions include empirical evidence of transparency–usability trade-offs, a replicable multimethod evaluation protocol, and three design strategies for newsroom integration. The thesis concludes with practical guidelines—on-demand “how it was made” expanders that clarify AI contributions to both text and images, and plain-language disclaimers that emphasize continued human editorial oversight and outlines future work on large-scale reader testing, live CMS deployment, operational impact metrics, and context-sensitive disclosure triggers.

The technological capability to perform rendezvous between two controlled vehicles in space has witnessed a growing interest since the years of the space race between the U.S. and the Soviet Union in the 1960s as well as a progressive shift from automated and manned approaches to increasingly more autonomous systems. Moreover, in recent years, the growing number of debris comprising rocket bodies and defunct satellites in orbit around the Earth is shifting the focus of rendezvous to Active Debris Removal, which requires a servicing satellite to safely approach a non-cooperative target to begin proximity operations. The work presented in this thesis, developed in collaboration with the AOCS & GNC Design & Software department of OHB, aims to address such need by developing an integrated guidance software which a manoeuvring satellite may use to autonomously approach and inspect an orbiting target in a safe way. This is obtained by first developing closed-form impulsive manoeuvres for arbitrarily eccentric targets in the space of Relative Orbit Elements and then designing high-level guidance strategies which adopt such schemes to accomplish mission objectives. Ultimately, the implemented guidance law is validated in a high-fidelity MATLAB/Simulink simulator; the results demonstrate the effectiveness of the proposed architecture and its potential reusability across different mission scenarios.