Following the discovery of an analytical model to predict contraction in cheese chips cut by plastic scraping based on the rake angle used and the frictional coefficient between the cutting tool and the material, this study aims to investigate the generalization of such a model to homogeneous materials. The possibility of generating wavy flower-like chips and other three dimensional shapes by variation of rake angle and frictional behavior of a cutting tool is investigated and the consequences of this are discussed alongside of the accuracy of the model in predicting the experimentally estimated contraction ratio. An investigation is made into the use of different compositions of beeswax and vaseline and the mechanical properties of these compositions are estimated in various experiments. Finally, methods of estimating contraction and surface curvature are discussed in the context of understanding the morphogenesis of the chips obtained by scraping. The study shows qualitatively the possibility of generating three dimensional shapes by scraping by independent variation of both rake angle and frictional behavior in a single cutting tool. However, the model does not succeed in accurately predicting the measured contraction ratio, as done previously. The possible reasons for this are discussed, leading to further insight in the behavior of beeswax and possible improvements of the experiments and measurement techniques used.

Accident data shows that even though there is an overall reduction in moderate and more severe injuries, thoracic and lumbar spinal fractures still remain a significant concern of the reported injuries. Anthropomorphic Test Devices (ATDs), also known as crash test dummies, are currently used to evaluate occupant safety in automotive crash tests. However, due to the regulatory and test repeatability requirements, they are not omnidirectional in terms of their bio-fidelity, and are often sturdier than actual humans. With the rise in computational power, Finite Element (FE) Human Body Models (HBMs) prove a great alternative to ATDs and an important tool to simulate actual human body response in impacts at a full body or a system level. Additionally, these HBMs can be supplemented by injury risk prediction models based on physical tests, to provide a comprehensive solution in evaluating human occupant safety in different scenarios.As part of the development and continual improvement process of these HBMs they need to be validated under a wide range of operating conditions, especially in high load rate impact scenarios. The aim of this thesis was to investigate the bio-fidelity of the thoracic spine of the SAFER HBM in response to vertical impact loadings which can be seen in automotive run off the road accidents. This was done by replicating various tests done on human cadavers in FE simulations and comparing the response of the thoracic spine of the FE HBM against physical test data. Primarily, three different loading conditions were simulated: a quasi static test on a denuded thoracic spine with ribcage, a pilot seat vertical ejection test and an underbody blast (UBB) test representing high rate loading in military applications. The boundary conditions in these tests involving restraints, sleds and protective equipment were modelled, and the HBM positioned using pre-simulations. The seat vertical acceleration input pulses for these loadcases were classified as low and high rate tests based on their duration, and compared against measured vertical inputs in vehicle run off road accident scenarios. Additionally, the applicability of an existing Injury Risk Function (IRF) model for the lumbar spine was evaluated for the thoracic spine.The thoracic spine of the FE HBM was seen to be more compliant in flexion-extension and lateral bending, but was within the test corridor for torsion tests as compared to the experimental quasi static test data of the denuded thoracic spine with rib cage. In the pilot seat ejection test, the HBM FE time history response for seat force and vertical accelerations was found to correlate well to the experimental tests, which was evaluated using the ISO18571 correlation method. The model showed reasonable correlation to low rate UBB tests, but a reduced correlation to high rate UBB tests in terms of vertical accelerations, since the HBM was not modelled to account the high load rate effects present in such scenarios. A viscoelastic material model was investigated for the Nucleus Pulposus (NP) of the thoracic and lumbar spine, but was found not to show a significant improvement for a high rate UBB test, since the experimental NP test data did not cover the loading rates seen in the high rate UBB test. The IRF was found to be more sensitive and predicted higher risk of injury than seen in physical tests, but still predicted the location of risk of injury within the region of physically reported injuries.

The aim of this thesis is to study Nature-Based Solutions (NBS) as an emerging concept andexamine how the mitigating effects of NBS on urban flooding and urban heat island (UHI) areevaluated at the moment. This has been attempted through a structured literature review of 38peer-reviewed studies. The findings show that the assessments of NBS connected to urbanflooding and urban heat islands are dominated by models and simulations. Although NBS areassessed both in relation to urban flooding and urban heat islands, urban flooding appears tobe favoured among the authors and figures more often than urban heat islands in the reviewedarticles. The articles show that NBS are highly scale dependent and context dependent. NBSas an umbrella concept; with its rebranding treatment of already existing urban planningterms, runs the risk of oversimplifying complex issues.

Home care services are the largest employers in Sweden, operating throughout the whole country with different conditions in every municipality. Some operate in widespread geographical areas such as Värmdö Municipality in Sweden, which also contain a large number of islands where care receivers live. The home care services face substantial challenges related to daily route planning and scheduling. Traditionally done manually, the process is time-consuming, inefficient, and difficult to accommodate the complex mix of operational, social, and legal constraints that exist in real-world scenarios. This thesis addresses these challenges by developing an automated routing and scheduling model for the home care services that not only improves efficiency but also accounts for sustainability through the integration of electric vehicles (EVs).

The study introduces an advanced Vehicle Routing Problem with Time Windows (VRPTW) model implemented using Google OR-Tools and PostgreSQL with pgRouting extensions. The model incorporates a comprehensive range of constraints, including hard constraints such as lunch breaks and nurse qualifications, as well as soft constraints such as caregiver-patient continuity and language compatibility. Real world data provided by Djurö hemtjänst formed the basis for model development and validation.

A key component in this work is the integration of electric vehicle range constraints into the route planning process. This is achieved through a simple temperature-sensitive energy consumption model based on empirical data for the Nissan Leaf, which dynamically adjusts vehicle range according to ambient weather conditions. Additional EV related constraints include recharging and route adjustments based on limited range. The model also accounts for appointments located on the islands of the municipality by the inclusion of ferry travel. 

Results from the model demonstrate a significant reduction in total travel time and planning time compared to the current manual process. Travel time has been reduced with 37 % and travel distance by 38 %. Comparing to the real world example data, provided in this project, this leads to work time savings of 42 hours per day, which corresponds to 2 hours per person per day. Moreover, the optimized routes suggest improved caregiver continuity, by increasing the preferred caregiver-care receiver match by 61 %. The work also explores a sensitivity analysis, which reveals the effect of various operational parameters. Furthermore, with the reduced travel distance follows a reduction in fuel consumption discussed in a sustainability analysis, which confirms reduced carbon emissions and operational costs. Moreover, the integration of an EV fleet was found to not negatively impact the operational efficiency of the model.

This thesis shows that integrating automatic route planning, while still considering real world constraints, can greatly improve the operational, economic, and ecological efficiency of the home care services. The approach and methodology presented have rather broad applicability, offering valuable insights for other sectors that rely on complex scheduling and routing, such as school transport and delivery services.

In this dissertation, we study two related classes of problems within the field of systems theory: the analysis and optimization of networked dynamical systems, and the reconstruction of unknown cost functions in control and learning frameworks. These problems naturally arise in a wide range of applications, from engineering systems to natural phenomenon.

The first class of problems concerns control and optimization over large-scale networked systems. Achieving controllability efficiently is an essential topic. The problem of ensuring controllability with a minimal number of control inputs is investigated first. Moreover, optimal control placement with limited controls are studied to enhance energy efficiency, and reduce computational and implementation costs. The second class focuses on inverse problems of optimal control, which aim to reconstruct unknown cost functions from observed behaviour. These problems are valuable for uncovering the objectives underlying complex systems in nature and society. Both cases are considered: when the system dynamics are known a priori and when they are unknown.

Specifically, Paper A investigates the minimal control placement problem for networked systems derived from Turing's reaction-diffusion model, a classical framework for understanding self-organization and pattern formation in biological systems. The eigenstructure of the diffusion matrix is fully characterized, and by introducing  symmetric control sets, we establish the necessary and sufficient graph-theoretic condition that guarantees controllability of  diffusion systems over networks of arbitrary size and parameters. These results are further extended to the reaction-diffusion systems.

After studying network controllability, Paper B extends the analysis to energy-efficient control placement in networked systems. By classifying network symmetries and exploiting symmetric control combinations, we develop a method that enables efficient computation of the spectrum of the controllability Gramian through lower-dimensional representations. This approach is further generalized to non-symmetric cases, where upper and lower spectral bounds are derived. Moreover, by utilizing the trace of the controllability Gramian as the objective, we propose a closed-form algorithm for optimizing control placement under constraints of limited control inputs and system controllability, with simulations validating its effectiveness.

Paper C addresses inverse optimal control for continuous-time linear quadratic regulators over finite-time horizons, namely the reconstruction of unknown cost matrices R, Q, and F in the objective function  from observed optimal control trajectories. The underlying linear system is assumed to be known. Both problem settings where R is either unknown or given are investigated. Firstly, two methods are developed to reconstruct R: one that leverages the full trajectory of the optimal feedback matrix and provides the necessary and sufficient conditions for uniqueness, and another that relies only on selected time points to reduce computational burden, which is particularly effective if F is given as positive definite. Secondly, when R is given, we investigate the role of system controllability in determining the well-posedness of the inverse problems. This assumption is subsequently relaxed, and sufficient conditions are established to ensure well-posedness, along with explicit analytical expressions for Q and F. Finally, the structural equivalence between IOC problems with unknown and given R is characterized under certain circumstances.

Paper D investigates inverse reinforcement learning (IRL) to reconstruct the unknown cost function in a model-free setting, where system dynamics are also unknown. Conventional IRL algorithms often require on-policy data collection and bi-level optimization, which impose potential practical limitations. To overcome these challenges, we propose a direct and adaptive IRL algorithm that learns from off-policy data satisfying only a mild persistence of excitation condition. By employing Nesterov-Todd (NT) step primal-dual interior-point iterations, the cost parameter is updated through simple one-step recursions, avoiding repeated forward RL computations. Theoretical analysis quantifies the impact of system noise and establishes sublinear convergence of the proposed algorithm. This method is further generalized to nonlinear objective functions via differential dynamic programming, where gradients of the loss function are computed through a backward pass. Numerical simulations demonstrate the efficiency and effectiveness of the proposed approach.

Press-fits, also known as interference fits, are a widely used technique to achieve mechanical connection between components without the use of fasteners. In this thesis work, the influence of surface texture on press-fit behavior was investigated. The performance of the press-fits is influenced by several factors. One of the factors is surface texture and experimental investigations have shown that it is a key role in the press-fits performance. Further, to ensure the torsional strength of the press-fits, the general industrial practice is to compare the press force measured during assembly with design guidelines. The changes in surface texture causes variations in press force and introduce uncertainty in the torsional strength of the assembly. Hence understanding the influence of surface texture will facilitate better design and quality assurance during production. The first objective of the thesis is to identify appropriate surface roughness parameters for press-fit application. The second objective is to develop a predictive model for press force. Prototype shafts were manufactured with different surface roughness parameters, assembled with hubs and tested for torsional strength. The actual press forces were compared with predicted values. Two approaches were used to predict press force and compare with actual values. In the first approach, errors ranged from 16 to 48 % and in the second approach, errors were -5 % to 7 %. The torsional strength test showed that the selected roughness parameters for adhesively bonded assemblies meet the acceptance criteria.

The transition to a carbon-neutral Nordic power system, characterized by high shares of variable renewable generation, requires flexible resources that can respond within minutes. Aggregated fleets of electric vehicles (EVs) offer such flexibility but have yet to be fully integrated into balancing markets. This thesis develops a risk-aware bidding framework that enables an EV aggregator to participate in the Nordic 15-minute Manual Frequency Restoration Reserve (mFRR) Energy Activation Market introduced in March 2025.

A multi-stage mixed-integer linear stochastic optimization model is proposed to determine profitable and feasible bids, while respecting the operational constraints of the EV fleet. Two formulations are analyzed: (i) a fixed energy schedule, where the charging plan is determined separately and treated as a commitment, and (ii) a joint spot and reserve approach, where the energy schedule is optimized together with reserve bids. Conditional Value-at-Risk (CVaR) is incorporated to penalize outcomes with low profit, enabling the aggregator to trade-off expected gains against downside risk.

A case study of 1 000 Swedish EVs with typical home-charging patterns shows that participation in the mFRR market can greatly reduce charging costs when the spot schedule is fixed. When spot and reserve bids are instead optimized jointly, the aggregator can even profit from the charging process by fully leveraging the fleet’s flexibility. Introducing risk aversion reduces downside risk with only a minor impact on expected profit.

In several industries, components that are both lightweight and meet strict requirements need to be manufactured. Structural optimization is an approach, which follows these requirements to be formulated mathematically. 

This master´s thesis in concerned with density based topology optimization with the goal of minimizing structural compliance (maximizing stiffness) under both volume and stress constraints. Two types of stress constraints are studied. A global constraint that approximates the maximum von Mises stress in the structure, versus fully local constraints applied to each finite element, which often in topology optimization leads to problems with thousands to millions of constraints in contrast to the global stress constraint requiring only one. 

The method is based on solid isotropic material with penalization, to link densities to material properties, while the design variables are updated with the method of moving asymptotes. The study compares global and local stress constraints in terms of their ability to manage local stresses and overall structural compliance.

The results shows the global stress constraint performed effectively, while local stress constraints are preferred in safety-critical applications requiring strict stress control. However, the result also indicate that both approaches require further tailored implementations to solve specific problems. 

The European Union has new rules for sustainability reporting called CSRD and ESRS. These rules ask companies to give clear and comparable information that others can trust. This thesis looks at how the rules change the early design phase of large renovation projects in Sweden, how they affect the investment decisions of property owners, and how roles and responsibilities change across contract forms.We used a review of earlier literature and official rules, and interviews with key people in the value chain such as property owners, contractors, and consultants. We then compared the material to find common patterns and differences.The results show that the rules mainly give a stronger frame for management and control. They ask for clear goals, clear responsibility, and proof of results. The idea of double materiality helps firms choose what to measure for both the company and society. Sustainability moves into design earlier through life cycle assessment and life cycle cost, design for reuse and product traceability, stronger energy performance, and social requirements in the supply chain. Digital tools replace many spreadsheets. Shared data platforms and LCA tools make follow up easier, but they also demand high data quality and good assurance. Benefits include more transparency, better comparison, and stronger support for investors. Main challenges are more administration and the need for skills and solid data systems.In conclusion, to deliver projects well under CSRD and ESRS, teams need early and clear requirements in design, robust data systems, standard traceability, and clearly defined roles that are built into contractor processes, especially in design build. These steps improve decisions, reduce risk, and increase trust in reported results.

This thesis investigates a previously underexplored phenomenon in railway engineering, sleepertilting, a form of microrotation that can occur when sleepers are subjected to dynamic loads during train passage. The work is based on observations and technical indications suggesting that sleepers not only move vertically but may also rock side to side. This movement, which is not always accounted for in current computational models, can affect track stability over time and increase maintenance needs.The aim of the study has been to enhance the understanding of the sleeper’s dynamic behaviour and its impact on the ballast structure and the overall longevity of the track system. Through a literature review, knowledge has been gathered from scientific articles, technical reports, and manuals, with particular focus on load transfer, ballast properties, and stress distribution in the track superstructure.The results show that rail deflection under train load can force the sleeper to rotate slightly along its longitudinal axis. This tilting causes an uneven distribution of pressure within the ballast, potentially leading to local degradation, settlements, and changes in track geometry. Existing theoretical models tend to simplify this motion pattern, thereby overlooking critical localized effects. The study emphasizes the need for further understanding of this phenomenon and its consequences, especially in a future of increasing axle loads, higher train speeds, and more frequent traffic.

Software, programmability and connectivity are increasingly pervasive aspects of society. In the span of a lifetime, everything from vehicles to phones has evolved from fixed-function to generally programmable devices that are always online. As a result, their functionality is entirely determined by their software, which may change from update to update and depending on what applications you have installed. The software itself is increasingly written by outsourced consultants or even by AI, with vulnerabilities appearing and being fixed on a daily basis.

Where does one even begin to secure modern digital equipment? The highest levels of assurance certification demand formal verification - mathematical proofs that rule out vulnerabilities based on models of hardware or programming-language semantics. This is costly and time-consuming, and so is applied mainly to small, critical components. The way to gain scalability is to base your assurance statements on highly automated verification tools whose correctness depends only on minimal, auditable trusted code bases.

The work in this thesis is focused on the creation of formal verification tools in the domains of machine code and networking. The common denominator is the use of an interactive theorem prover to gain assurance of the results. Specifically, the thesis presents (i) methods to decompose verification tasks for unstructured programs, especially as applied to machine code, and (ii) a formal model of the P4 domain-specific networking language, accompanied by (iii) a proof-producing symbolic execution tool and finally (iv) a complete toolchain to verifiably compile a verified P4 program down to a software switch.

Bird collisions can damage facilities such as airports and pose risks to human safety and infrastructure functionality. They also threaten bird populations, since such collisions often result in fatalities. Consequently, reliable bird detection systems are of growing interest. In reality, surveillance cameras typically capture birds that appear very small on the screen, making it difficult for traditional object detection algorithms to detect such targets. To address this challenge, this thesis explores the use of the YOLO (You Only Look Once) algorithm combined with various background subtraction methods to fully utilize the motion data from surveillance video. The detection pipeline consists of three components: preprocessing, background subtraction, and object detection, and the first two components are evaluated through experiments. In addition to existing background subtraction methods, this thesis proposes a Kalman filter–based background subtraction strategy to reliably mask input frames. The masked frames are then passed to Ultralytics YOLO11 (trained on the SOD4SB dataset) with the aid of Slicing Aided Hyper Inference (SAHI), aiming to achieve stable detection under different lighting conditions. The results showed that a preprocessing filter, such as a median filter or Gaussian filter, helps with tiny object detection. However, in rare cases it can cause significant information loss if the object is extremely small. Background subtraction improved tiny object detection, and the proposed Kalman filter– based methods were generally more stable than other existing methods (such as frame differencing and Mixture of Gaussians) at guiding the object detection algorithm to achieve more accurate detection of tiny moving objects. Overall, this thesis provides a possible approach to tiny object detection when motion information is available.

Smart contracts are immutable programs deployed on blockchain platforms to enable decentralized applications and financial systems without intermediaries. However, their immutability and public execution make them especially vulnerable to security flaws, which can lead to irreversible significant financial losses. A promising approach to improve security is the use of invariants, properties that must always hold during contract execution. FLAMES (Fine-tuned Large Language Model for Invariant Synthesis) is a tool that leverages fine-tuned large language models to automatically generate security-relevant invariants for Solidity smart contracts. Automatically generating and validating such invariants remains a key challenge. This thesis evaluates FLAMES ability to synthesize and validate security-relevant invariants for Solidity smart contracts. The goal is to determine whether these invariants are both syntactically correct and semantically effective. To this end, two automated evaluation pipelines were implemented. The first tests whether smart contracts with injected invariants still compile successfully. The second examines whether these invariants prevent real exploits while preserving benign functionality. Experimental results show that FLAMES20K and FLAMES100K achieved high compilability rates and successfully patched known vulnerabilities. This work bridges the gap between theoretical invariant generation and practical smart contract security. By automating the validation of synthesized invariants, the proposed framework supports the development of more secure and reliable blockchain applications.

This thesis presents the design, implementation, and evaluation of a multi-precision Single Instruction-Multiple Data (SIMD) floating-point unit (FPU) integrated into the Dynamic Reconfigurable Resource Array v2 (DRRA-2) coarse-grained reconfigurable architecture, developed as part of the SiLago framework. The work addresses the increasing demand for adaptable and energy-efficient hardware capable of supporting diverse arithmetic formats such as FP8, FP16, and FP32, commonly used in modern AI and Digital Signal Processing (DSP) workloads. Conventional arithmetic units are typically optimized for single precision, which limits their efficiency and flexibility when executing mixed-precision tasks. This limitation becomes particularly relevant in reconfigurable fabrics like DRRA-2, where architectural flexibility is essential to balance performance, area, and power. The challenge, therefore, lies in developing a reconfigurable SIMD FPU that can support multiple number representations while conforming to the architectural constraints of the SiLago design methodology. To this end, we designed a modular arithmetic unit capable of switching between formats through operand packing and shared datapaths. The unit was implemented in SystemVerilog, integrated into a complete DRRA-2 fabric, and synthesized using the TSMC 28 nm process technology. The design was evaluated in multiple configurations, highlighting trade-offs between area, power, and control complexity. The resulting design provides low-power, area-efficient execution of floating-point operations with multiple precisions. It supports flexible deployment scenarios where different accuracy and performance levels are required, without redesigning the underlying hardware. As the first multi-format SIMD FPU integrated into DRRA-2, this work extends the arithmetic capabilities of the SiLago framework and strengthens its suitability for applications requiring adaptable and precision-scalable computation.

Power consumption is one of the most important considerations during digital system development. As designs scale, it often takes days, weeks, or even months to estimate power via simulation after several verification steps. It would greatly improve power efficiency if engineers, especially designers, could know the power consumption early in the design stage and modify the design considering both power and performance. This thesis focuses on machine learning-based ASIC (Application-Specific Integrated Circuit) power prediction. Two approaches are developed, specifically targeting an ASIC block case in the communications industry:

• An architecture-based approach that spans multiple levels of abstraction across an ASIC block, enabling power analysis and prediction down to the most basic building components.

• A flow-based approach that accelerates the power simulation process.

Based on our studies, we provide a prediction of power distribution under a given stimulus in a simplified and accelerated manner. Additionally, to standardize the entire process of modeling, including power generation flow for data collection, automation of machine learning model training and optimal algorithm selection of the circuit block, scripts are developed to automate the workflow, improving the efficiency for future usage. Finally, we propose two directions for future investigation: (i) Generate more stimuli covering a wider power range to make the analysis more comprehensive. (ii) Upgrade the power prediction level, using small components to predict the power of larger ASIC blocks based on connection and inclusion relationships between modules, ultimately extending to the whole ASIC design.

As the use of generative AI tools grows across creative domains, so too does the need to reflect on their environmental impact. This thesis presents a research through design study that involves design and evaluation of a conversational sustainability reflection system, aimed at supporting creative practitioners who use AI considering sustainability trade-offs within their workflows. Grounded in principles from behavior change, explainable AI, and interactive reflection, a chatbot prototype that engages users in emotionally supportive and non-judgmental dialogue was developed. The design process followed an iterative, research-through-design approach, incorporating thematic analysis from surveys and user testing with creative AI practitioners. Based on survey insights, the design focused on embedding low-pressure, emotionally supportive prompts and summary-style sustainability feedback into creative workflows. Key findings from the qualitative user study further revealed that users appreciate empathetic prompting and low-pressure framing, which foster reflection without inducing moral fatigue. However, challenges remain in the timing, clarity, and personalization of sustainability feedback, as well as in maintaining conversational continuity. The study also found that users are more likely to adopt the tool if it is embedded into existing creative platforms, rather than used as a standalone application. The results suggest that conversational agents can serve as effective mediators of value-sensitive reflection, provided they are context-aware, emotionally attuned, and technically unobtrusive. Future research may explore personalization techniques, long-term engagement, and deeper integration with creative ecosystems.

Triple-negative breast cancer (TNBC) is an aggressive subtype of breast cancer that lacks targeted treatment options. Pathological complete response (pCR) following neoadjuvant therapy is a crucial prognostic indicator; however, both TNBC subtype classification and pCR prediction remain challenging using conventional methods. This thesis investigates whether deep learning models trained on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) can accurately classify TNBC and predict pCR. A flexible 3D convolutional neural network pipeline was developed and applied to the MAMA-MIA dataset, a large multicenter collection of expert-annotated breast DCE-MRI scans. The study systematically evaluated three design factors: temporal encoding strategies (e.g., delta-based subtraction), auxiliary input channels (e.g., segmentation masks and parametric enhancement maps), and model capacity (ResNet18, ResNet50, CBAM). All models were trained and evaluated using five-fold cross-validation. Results showed that temporal encodings improved classification performance over static input, and that adding parametric enhancement maps yielded further gains. However, segmentation masks did not yield improvement when used as additional input channels. Increasing model capacity and incorporating attention mechanisms (CBAM) achieved the highest performance, with an AUC of 0.81 for TNBC and 0.76 for pCR. These findings suggest that deep learning models can extract clinically relevant information from spatiotemporal DCE-MRI patterns, with implications for non-invasive tumor characterization and early treatment response prediction.

Ensuring operational safety and efficiency in industrial settings often depends on accurately monitoring water levels in storage tanks. In this project, we explore whether passive acoustic data can be used to predict liquid levels, leveraging a dataset collected from real-world pump stations. The dataset consists of real-time audio recordings paired with depth sensor measurements gathered under operational conditions. Two complementary approaches were applied: first, a series of neural network models, including convolutional neural networks and transformer-based architectures, that analyze spectrograms to capture key acoustic features; and second, a physics-based acoustic resonance model that computes the fundamental frequency from the audio signal and correlates it to water levels using established resonance equations for cylindrical tanks. Our results demonstrate that passive acoustic sensing can estimate water depth with surprising precision: a station-specific CNN achieved a mean absolute error of just 1.2 cm over a 1 m range, and multi-station models performed comparably when tested on familiar sites. However, no model generalized reliably to unseen stations, highlighting the sensitivity of acoustic signatures to tank geometry and local operating conditions. The resonance-based method failed to produce a consistent frequency-level mapping under real-world noise and flow variations, and thus proved unsuitable without further refinement. These results confirm that, with station-specific calibration or on-site fine-tuning, passive acoustic monitoring offers a non-invasive complement to traditional sensors.

Money laundering is a global challenge that undermines financial systems and supports criminal activity, making effective anti-money laundering measures necessary. However, like most other realistic anomaly detection problems, challenges such as label scarcity, data contamination, imbalance, noise, etc. complicate the effective application of machine learning in anti-money laundering systems. Although different model architectures, data synthetisation, and active learning techniques have been tried, selfsupervised learning has recently shown promise. This thesis examines whether self-supervised learning can enhance anomaly detection in an anti-money laundering setting with applied active human-in-the-loop active learning feedback. To address this, the thesis compared the performance of an anomaly detection model (PReNet) trained on raw data against the same model trained on contrastive self-supervised embeddings of the same data. The thesis evaluates the performance of the PReNet model under two active learning strategies, based on uncertainty- and certainty based sampling over 9 iterations of updates. The results indicate that self-supervised embeddings did not enhance model performance relative to raw data on our dataset, contradicting previous literature. After 9 iterations, the raw baseline models performed substantially better than their embedded counterparts based on domain expert feedback. Future work could explore alternative self-supervised learning methods for tabular data, the use of different anomaly detection models, and broader datasets to assess how these techniques may be developed further.

This thesis investigates the generalizability of a multimodal encoder trained on panoramic images and point cloud data using supervised and self-supervised learning approaches. The study is conducted within the context of Giraffe360, a company that utilizes visual and spatial data for automated real estate floor plan generation. Interest has been raised to explore self-supervised learning for visual-spatial encoder for transfer to downstream tasks. While self-supervised learning has achieved significant success in vision, its adapation to multimodal scenarios, specifically with point cloud data, remains underexplored. To address this, an existing vision-based self-supervised learning framework is adapted for multimodal pretraining, integrating image and point cloud data into a unified encoder architecture. Its performance is compared against several supervised learning baselines trained on classification and dense prediction tasks. Generalizability is assessed through tasks critical to Giraffe360 — room classification, door detection, and room layout estimation. Evaluation methods include linear probing, k-NN classification, and attention visualization. Experimental results show that while the vision-only model trained with self-supervised learning demonstrates promising transferability to downstream tasks, the extension of the self-supervised pretraining approach to the multimodal setting fails to produce an effective encoder. Encoders pretrained on supervised dense prediction tasks, such as object detection, demonstrate adequate generalization, including to global tasks like room classification. Multimodal object detection achieved the best overall generalizability across select tasks. Analysis of attention maps revealed that meaningful cross-modal spatial alignment is learned primarily in dense prediction pretrains, whereas the multimodal self-supervised model fails to align modalities. These findings suggest that direct application of single-modality selfsupervised learning methods to multimodal encoders, while promising, is nontrivial and that careful adaptation is necessary. While the hypothesis that self-supervised pretraining would universally outperform supervised learning was not confirmed, the results underscore the potential of self-supervised learning in multimodal learning.

The purpose of this thesis is to investigate the behavior of quantum magic in many-body systems. According to the Gottesman-Knill theorem, the particular class of quantum states referred to as stabilizer states can be efficiently simulated on a classical computer. The property of non-stabilizerness, or magic, can be broadly defined as the distance between a general state and a stabilizer state, serving as a measure of classical hardness as well as a resource for quantum computation. Numerical studies of this property, as a feature of many-body systems, are classically hard. This work uses a combination of quantum circuit simulations and tensor-network methods to study the spread and characteristic features of magic in quantum systems. We demonstrate that this can be done with relative efficiency, while addressing questions about the characterization of non-local magic and its relation to quantum computational complexity.

This thesis investigates thermal light sources for application in chaotic LiDAR, focusing ontheir optimization and the unexpected appearance of a structured interference or beating patternin the second-order coherence measurements. Chaotic LiDAR leverages the photon bunchingcharacteristics of chaotic light to measure distances with high temporal resolution, enablinglong-range and avoids ambiguity in range sensing.

The first part of the work explores how spectral filtering and light source configuration affectphoton correlations. Several filtration schemes at telecom wavelengths (1550 nm) were testedon various light sources, including a sub-threshold continuous-wave laser, halogen light andsunlight. Narrowband filtering using a notch filter combined with an etalon significantlyimproved the observed photon bunching for classical sources.

In the second part, the unexpected beating pattern in g^(2) when using the sub-threshold laserwas further studied. The pattern’s visibility and number of periods (oscillations) were foundto strongly depend on the laser bias current and its spectral alignment with etalon transmissionpeaks. The phenomenon was linked to the ratio of spontaneous emission to stimulated emissionnear the lasing threshold. Attempts were made to model the behavior using a two-modestatistical mixture framework, though no conclusive fit was obtained.

The findings suggest complex emission dynamics near the lasing threshold that are not capturedby standard models. This work advances the understanding of photon statistics in thresholdlasers and informs the design of improved bunched light sources for future chaotic LiDARsystems.

Biosensors are analytical devices that combine a biological recognition element with a physicochemical transducer to detect specific analytes and convert their presence into measurable electrical signals. Among various biosensing modalities - including electrochemical, optical, thermal, and surface plasmon resonance, the streaming current approach is examined in detail. This method utilizes electrokinetic principles rather than redox reactions, generating an electrical current as fluid flows through charged microchannels, driven by ion movement.This project centers on the development of a low-cost, printed circuit board (PCB)-based picoammeter (pA) for signal acquisition from streaming current biosensors. It also includes a data analysis platform and a user-friendly software interface. The PCB incorporates an analog multiplexer to enable signal capture from four sensor channels, facilitating multiplexed biosensing.Integrating electronics with biosensing technology enhances portability, cost-effectiveness, miniaturization, and ease of use—key attributes for point-of-care (POC) diagnostic applications.Keywords

The design of supersonic and hypersonic vehicles, such as reusable launchers and high-speed aircraft, must account for extreme aerothermodynamic loads that can trigger complex Fluid–Thermal–Structural Interaction (FSTI) phenomena. These include flutter and thermal buckling, potentially leading to structural fatigue or even catastrophic failure. To predict such interactions, this work develops an efficient aeroelastic coupling tool and methodology using DLR’s TAU flow solver along with Ansys structural and thermal solvers within the ATSI framework. The resulting tool aims to support the preparation and interpretation of future experimental campaigns.

The methodology was validated using three benchmark cases of increasing complexity based on the canonical problem of shock impingement on a thin elastic panel. A first verification was conducted using a static setup where the significance of thermal effects was highlighted. The second case, where dynamics were introduced, posed a stringent test to capture the expected stable limit cycle oscillation (LCO) under the complete absence of dissipative mechanisms. A quasi-steady formulation was used to improve efficiency which proved accurate, while structural velocity effects, initially neglected due to a limitation on the steady-state formulation of the Tau solver, proved to have significant effects when introduced through the Enriched Piston Theory (EPT) correction. Added structural damping showed similar improvements. The third case, based on an experimental configuration, exhibited higher uncertainty but maintained good agreement. Further enhancements are expected through a simplified implementation of EPT for viscous setups and additional sensitivity studies on timestep refinement and structural modelling improvements.

Despite challenges in the last configuration, the developed tool demonstrated strong predictive capabilities and good efficiency, establishing a robust foundation for future coupled studies.

Housing association fees are an important factor in determining condominium prices. Despite their importance, it remains uncertain how much these costs are reflected in sale prices. We examine how deviations from expected monthly fees affect apartment prices in Stockholm. The data include more than 80,000 transactions from 2017, 2021 and 2023. We apply a two-step method. First, we model expected fees based on age, size, and location. Then we used the residuals in a hedonic price model with spatial controls. Higher-than-expected fees reduce sale prices. On average, each additional 1,000 SEK in monthly residual cost reduces the price by 2–3%. The elasticity remains stable over time but weakens in 2023. Our spatial analysis confirms stronger capitalisation in central areas compared to peripheral zones, and quantile regressions reveal small variation across the price spectrum. The effect does not appear driven by omitted variables or spatial autocorrelation. Our results help inform valuation and oversight in cooperative housing.

This thesis advances the understanding and application of synthetic data for manufacturing parts recognition. Vision-based inspection systems in manufacturing rely heavily on real image data, which are costly to collect, annotate, and adapt across products and environments. To address these challenges, this work presents a systematic investigation of how synthetic data can be effectively generated, evaluated, and applied for robust and scalable performance. The research introduces a series of new industrial benchmark datasets covering multiple manufacturing use cases and factory environments: SIP-17, SIP15-OD, and SIP2A-OD, to enable unified evaluation of sim-to-real transfer in classification and detection tasks. Building on these datasets, a domain randomization pipeline is developed to systematically explore the effects of rendering parameters, material variability, and illumination on model generalization. To further automate data generation, the thesis proposes Synthetic Active Learning (SAL), a closed-loop framework that identifies model weaknesses and adaptively refines synthetic data generation without requiring real samples or manual tuning. Experiments across the benchmark datasets show that the proposed method improves model robustness compared to existing approaches while reducing manual labeling requirements. Collectively, these contributions provide new insights into how synthetic data can be systematically leveraged to build data-efficient, automated, and reliable vision systems for manufacturing, aiming to support future development of intelligent and flexible production systems.

The COVID-19 pandemic influenced urban mobility worldwide, prompting travel restrictions, physical distancing, and lockdowns. Active transport modes, particularly cycling, emerged as socially distanced and adaptable alternatives. In response, many cities rapidly implemented cycling-friendly measures, such as repurposing streets, e-bike subsidies, and traffic-calming initiatives. Investigating these trends statistically, this study centers on three hypotheses: (1) pandemic conditions spurred a notable rise in cycling relative to pre-pandemic levels; (2) new biking infrastructure and lockdown policy interventions were drivers of this rise; and (3) an important motivation for biking under the pandemic was essential errands, like shopping for groceries. Drawing on publicly available weekly bicycle counts from Paris and Barcelona, we employed time series regression models to address our hypotheses. Results confirmed a substantial surge in cycling in both cities. Lockdown policy stringency was associated with increased cycling in both cities. Paris seemed to show a stronger sensitivity to infrastructure interventions. The findings underscore that this devastating health crisis also served as a catalyst for a shift toward more sustainable travel.

Increased cycling contributes to more sustainable mobility, and better insight into the drivers behind this mode is important for the design of policies that will effectively promote it. In this study we identified correlates with cycling in three European cities: Tallinn (Estonia), Braga (Portugal), and Istanbul (Turkey), using data from a six-month cycling incentive program.

Specifically, we investigated the impacts of wind speed, car ownership, and public transport use on weekly cycling frequency by estimating mixed ordinal probit models. An important early finding was that only 15% of participants accurately reported their observed cycling frequency. Further, across all the examined cities, wind speed had a significant (p-value <0.001), negative relationship, with estimates of -0.125 (Braga), -0.079 (Istanbul), and -0.058 (Tallinn). This suggest that higher wind speeds prevented people from cycling. Likewise, private car ownership generated significant (p-value < 0.001) negative associations of -0.423 (Braga), -0.419 (Istanbul), and -0.949 (Tallinn). Thus, car access seems to reduce cycling frequency. Lastly, public transport usage was positively correlated with cycling: 0.225 (p-value = 0.008) (Braga), 0.339 (p-value = 0.003) (Istanbul), and 0.714 (p-value < 0.001) (Tallinn). To promote cycling, the findings highlight the importance of infrastructure such as wind-protected bike lanes, as well as policies that reduce car ownership for instance car-sharing solutions. The positive relationship between public transport use and cycling underscored the potential benefits of better integrating these modes, for example through bike-sharing facilities near transit hubs. As the effects varied notably across cities, we recommend that future research address the drivers behind these differences, which may relate to extant infrastructure.

Failure-transparent programming models abstract from failures by fully masking them from the programmer. They are widely used for programming distributed systems, as failures otherwise are considered a core difficulty. The most widely used of its kind for processing data is stateful dataflow streaming, a model restricted to static, directed, acyclic graphs of stateful stream processors. However, its restrictions limit the applicability of the model, as it lacks support for compositional patterns and replicated data types, making it difficult to express certain applications. Moreover, there is a lack of formal foundations and proofs of failure transparency.

This thesis contributes a semantics-agnostic definition of failure transparency, and two proofs of failure transparency, one of which is for a model of a stateful dataflow streaming system. It additionally contributes two novel programming models based on stateful dataflow streaming. The first provides extensions for compositional patterns, allowing it to express use cases such as a shopping cart. The second provides extensions for windowed conflict-free replicated data types, implemented in a low-latency programming system for global aggregations.

This thesis demonstrates the utility of failure-transparent programming models for distributed systems by contributions to its formal foundations and by making it applicable to a wider range of applications.

This paper gives an in-depth theoretical analysis of the direction and speed selectivity properties of idealized models of the spatio-temporal receptive fields of simple cells and complex cells, based on the generalized Gaussian derivative model for visual receptive fields. According to this theory, the receptive fields are modelled as velocity-adapted affine Gaussian derivatives for different image velocities and different degrees of elongation.  By probing such idealized receptive field models of visual neurons to moving sine waves with different angular frequencies and image velocities, we characterize the computational models to a structurally similar probing method as is used for characterizing the direction and speed selective properties of biological neurons. It is shown that the direction selective properties become sharper with increasing order of spatial differentiation and increasing degree of elongation in the spatial components of the visual receptive fields. It is also shown that the speed selectivity properties are sharper for increasing order of spatial differentiation, while they are for the inclination angle $\theta = 0$ independent of the degree of elongation.

By comparison to results of neurophysiological measurements of direction and speed selectivity for biological neurons in the primary visual cortex, we find that our theoretical results are qualitatively consistent with (i) velocity-tuned visual neurons that are sensitive to particular motion directions and speeds, and (ii)~different visual neurons having broader {\em vs.\/}\ sharper direction and speed selective properties.  Our theoretical results in combination with results from neurophysiological characterizations of motion-sensitive visual neurons are also consistent with a previously formulated hypothesis that the simple cells in the primary visual cortex ought to be covariant under local Galilean transformations, so as to enable processing of visual stimuli with different motion directions and speeds.

This thesis investigates how to evaluate Retrieval-Augmented Generation (RAG) systems for technical maintenance without relying on ground-truth references using the LLM-as-a-Judge paradigm. Specifically, it addresses two research questions. First, how stable and interpretable are referencefree metrics across commonly used LLM-as-a-Judge evaluation frameworks? Second, to what extent can these metrics predict human-perceived usefulness? To answer this, we compare three open-source LLM-as-a-Judge evaluation frameworks (RAGAS, MLflow, and DeepEval) and their available referencefree metrics on real-world data in terms of consistency, discriminative ability, and predictive power. For the first research question, a balanced set of positive and negative interactions was evaluated across 10 repeated runs per framework. Stability was analyzed using average standard deviation and Intraclass Correlation Coefficient (ICC), while discriminative ability was assessed through score differences between human-labeled categories using Welch’s T-test. For the second research question, the evaluation metrics were used in a set of predictive models to predict human-labeled usefulness. Performance was measured using Accuracy, F1 score, and Receiver Operating Characteristic - Area Under the Curve (ROC AUC). The results highlight that certain metrics, especially those assessing Faithfulness, Answer Relevance, and Context Relevance, are not only consistent across runs but also effective predictors of usefulness. The bestperforming model achieved an above 0.7 Accuracy and F1 score, with a ROC AUC of 0.8, indicating that reference-free evaluations can approximate subjective human feedback to a degree exceeding random guessing. Overall, this work identifies a set of metrics, primarily from MLflow and RAGAS, that combine high consistency with strong discriminative power, demonstrating their practical utility for evaluating RAG outputs. We further show how these metrics can feed into predictive models to distinguish useful from non-useful responses and highlight qualities that make effective answers. Finally, we present a set of proposed use cases for integrating such predictive evaluators into live RAG systems.

Domain-specific languages (DSLs) are programming languages designed for a specific domain, as opposed to general-purpose languages, which can be used for a variety of tasks. Examples of DSLs include HTML, TeX, or Verilog. The Miking framework is a meta-language system with the goal of facilitating creation of external domain-specific and general-purpose languages. The framework functions by compiling a minimal functional language called MCore to one of the several supported backends – OCaml, C, JavaScript or JVM, or by interpreting MCore through a self-hosted interpreter. Miking provides a foreign function interface (FFI), called the externals system, which lets MCore call backend functions. The current system only has support for compiled targets and requires a full recompilation of the compiler when new externals are defined. This thesis presents a redesign and a proof-of-concept implementation of a new externals system targeting the OCaml backend. The system provides a unified interface for definition of externals for both compiled and interpreted modes, with native support for basic types. It acts as a stepping stone towards a fully redesigned externals system supporting all the existing backends and language features.

The Domain Name System (DNS) is a vital part of the Internet infrastructure. It acts as a distributed database, providing address resolution of humanreadable domain names to machine-readable addresses. One of the most common external attacks against the Domain Name System is Distributed Denial of Service (DDOS) attacks, which attempt to overload the system by flooding it with requests. One of the ways to provide resilience against DDOS attacks is to ensure the server can handle requests at line rate. To serve every Internet user and withstand attacks, performance is vital. It mainly handles scaling by having many nodes serving traffic, but single-node performance is still important. Virtualization provides increased security and utilization of hardware resources, but it also increases the length of the Input/Output data path. DNS traffic primarily consists of small packets using the User Datagram Protocol, which causes significant challenges for the data path. To test the performance, I split the data path into two sections: host (hypervisor to virtual machine) and guest (virtual network interface card to application). The two metrics I looked at were the Zero-Loss Rate and Latency. To evaluate the performance of the host data path, three solutions were tested: Vhostnet, Vhost-user/OVS-DPDK, and Single-Root I/O Virtualization (SR-IOV). To test the guest data path, two DNS servers were created, one using the standard Linux Networking Stack and one bypassing it using the Data Plane Development Kit (DPDK). The results show that all host data paths could reach the maximum throughput of the Linux Networking Stack, with interrupt placement being a key factor. However, only the SR-IOV host data path could reach the maximum throughput of the DPDK version. The performance of the DPDK version is highly dependent on the host data path; however, with the proper configuration, it was capable of about 19x the performance of the Linux version.

Underwater robots offer a promising solution for exploring areas deep beneath the ocean surface, both for collecting data to study the marine environment and for conducting rescue operations in locations inaccessible to humans due to extreme underwater conditions. Bioinspired underwater robots have gained significant attention for their ability to mimic the locomotion of marine animals. One of the main reasons for adopting fish-like movement is that traditional commercial underwater robots, which typically rely on propellers, can potentially harm marine life through their mechanical movement and the noise they generate. The RoboIguana is a bioinspired robot modeled after the Galápagos marine iguana, the world’s only marine lizard. This unique animal uses its tail for swimming and its limbs for locomotion along the seafloor, demonstrating both streamlined propulsion and agile underwater movement. The aim of this thesis is to identify the control parameters required for the RoboIguana’s tail to generate optimized forward propulsion. These parameters are tested on the tail using a custom-built, rigid, and stable experimental setup designed to hold the tail underwater, while supporting the necessary sensors outside the water to analyze the tail’s movement. The forward propulsion force generated by the tail is influenced by four control parameters: yaw amplitude, sway amplitude, frequency, and the phase shift applied to the sway motion. Experimental results indicate that the propulsion force increases with higher movement frequencies and larger amplitude values. However, these increases also result in greater electrical current consumption by the tail.

Recently, Internet of Things (IoT) devices have experienced a massive increase in popularity and are being introduced into households to a greater extent. However, as with most adaptations of relatively new technology, there has also been an increase in attacks against such devices. Additionally, with devices constantly monitoring users’ activity and potentially harvesting data, the question of social sustainability becomes quite pressing. This thesis will explore the issue of IoT device security by comparing the security of a well-known, branded IoT camera with that of a simpler, low-cost IoT camera. This includes issues such as how vulnerable the devices are to attacks and potential data privacy violations. Although IoT device security is a commonly discussed hot topic, this could be quite relevant since the differences in security correlated to pricing are not as commonly discussed. The study was carried out by compiling a set of tests consisting of hardware, firmware, network, and software vulnerability exploitation, to then compare the results. The tests, including data extraction, file system extraction, port scanning, packet sniffing, reverse shell access, live feed access, and control hijacking, were performed in a controlled test environment. Results showed, contrary to the hypothesis, that when presented with this set of methods and resources, the branded camera was more easily exploited than the low-cost camera. The low-cost camera proved to be quite difficult to penetrate due to unforeseen operational shutdowns when probed, but still allowed another user to take control of it. The branded camera, although lacking open critical ports and insecure services, exposed unencrypted files on its SD card and allowed live feed access via Real-Time Streaming Protocol (RTSP) to be gained through brute-forcing. The results highlight the fact that there is no guarantee of better security in more expensive devices from reputable brands compared to simple, low-cost devices. However, it is important to mention that both devices lack in security and privacy issues, and that with more time and resources, both cameras are likely to be found to have even more critical flaws.

This thesis evaluates the effectiveness of five multimodal large language models, GPT-4o, GPT-4.1, Gemini 2.0 Flash, Gemini 2.5 Flash, and Mistral Medium, for extracting structured information from Swedish housing association reports. The models are benchmarked on their ability to retrieve information from 19 predefined fields using two inference strategies: page-bypage and multi-page. The evaluation focuses on three dimensions: extraction performance, cost-efficiency, and strategy-related tradeoffs. The results reveal statistically significant differences in both cost and precision, with 95% confidence intervals confirming that the multi-page approach consistently outperforms page-by-page processing on both metrics. In contrast, differences in recall and F1 score were not statistically significant, with overlapping confidence intervals suggesting these metrics are more sensitive to model-specific behavior than strategy choice. GPT‑4.1 with multipage processing achieved the highest precision of 0.70 (95% CI: 0.56, 0.84), while Gemini 2.0 Flash offered the best cost-efficiency, being over 40 times more cost-effective than GPT-4.1. These findings support the viability of multimodal large language models for structured information extraction from domain-specific, image-based documents in non-English contexts. They also underscore the importance of choosing the right model-strategy combination based on performance, cost, and application requirements. The study concludes with a guide for real-world deployment of such systems in production environments.

Micro-mobility services, such as shared electric scooters and bikes, have revolutionized urban transportation systems worldwide. While these services offer convenient and sustainable mobility, their rapid adoption has raised safety concerns, particularly regarding user accidents. Reliable accident detection is crucial to address this concern and improve user safety. This thesis examines the use of deep learning–based anomaly detection techniques to identify accidents in micro-mobility data, with a particular emphasis on autoencoders. It explores whether incorporating the temporal characteristics of data collected from Inertial Measurement Units (IMUs), which monitor motion metrics such as acceleration and angular velocity, can enhance the performance of accident detection methods. Furthermore, the thesis compares the effectiveness of these advanced methods to a traditional threshold-based approach. The research findings demonstrate that autoencoder-based methods outperform the simpler threshold-based method, highlighting the potential of deep learning methods over traditional, rule-based methods. However, the study found no significant improvement in accident detection when considering the temporal aspect of the data. This may be due to the limited testing of different hyperparameters and the small size of the accident data set, which constrained the analysis. The results suggest that further exploration of the architecture of the LSTM (Long Short-Term Memory) autoencoderbased method and the TCN (Temporal Convolutional Network) autoencoderbased method is required to determine whether these methods fail to capture temporal dependencies or if the data inherently lack such dependencies.

The rise of Autonomous Vehicles (AVs) presents major opportunities for sustainable and efficient logistics. Yet, scalable deployment is hindered by current 1:1 remote supervision models, where each vehicle requires a dedicated Remote Operator (RO). This thesis explores how Remote Operation Interfaces (ROIs) must evolve to support more scalable supervision frameworks, specifically parallel (1:n) and multi-user (m:n) operations. A qualitative, exploratory approach was employed, combining semi-structured interviews, archival research, and participatory work sessions with internal stakeholders at the company and external domain experts. Thematic analysis revealed three central challenges: the need for redesigned supervision interfaces, the evolving role of the RO, and human and organisational barriers to scalability. Findings show that current systems lack support for fleet-level oversight, workload balancing, and collaborative workflows. Key design requirements identified include integrated fleet overviews, predictive scheduling tools, context-aware alerts, and seamless role transitions. Multi-user operations emerged as essential in scenarios involving simultaneous interventions or geographically distributed sites. The thesis concludes with design recommendations and suggests future work, including iterative prototyping and the exploration of design frameworks such as Ecological Interface Design.

The latency and resiliency benefits of edge computing have driven a shift toward distributing computation closer to end users. Serverless frameworks facilitate the development of highly elastic applications, with serverless functions being able to scale from completely dormant to as many replicas as needed, and fitting well into an edge setting. However, as the explicit lack of integrated state management can be restrictive, stateful serverless approaches aim to extend the serverless paradigm with persistent state and improve the developer experience of scalable development. This thesis explores the use of Conflict-free Replicated Data Types (CRDTs) as a state persistence foundation for environments like edge and stateful serverless computing, where servers need to scale to dynamic loads. CRDTs offer strong eventual consistency with minimal coordination, making them promising candidates for replication in dynamic environments such as edge and serverless deployment models. However, high replica churn and the absence of stable peer discovery mechanisms pose new challenges for CRDT systems, particularly around metadata accumulation and system scalability. While solutions have been proposed for dynamic replica groups and metadata garbage collection (GC) separately, removing metadata is both especially useful and more challenging to perform when replicas can leave permanently and the total number of peers is constantly in flux. Our contribution is a prototype system that extends delta-based CRDTs (- CRDT) with a group management protocol and GC mechanism. Replication is achieved through peer-to-peer synchronization, while we introduce object storage as the sole shared resource to enable fully elastic deployment. GC in CRDTs requires stronger coordination between replicas, but our proposed mechanism can make opportunistic use of dynamic load patterns to remove stale metadata during replica shutdowns, maintaining consistency while minimizing impacts on availability. Evaluation shows that while metadata accumulation cannot be fully prevented, our system reduces it in scenarios with periodic loads. As the buildup of metadata can increase the computational load of CRDT functions, reducing it improves system performance. This work contributes an initial validation of scalable CRDT deployments in dynamic networks, and outlines future improvements in handling high loads, fault tolerance, and integration with serverless frameworks or web applications.

As hardware complexity increases, modular design methodologies like SiLago are gaining traction for enabling scalable, reusable, and predictable Very Large-Scale Integration (VLSI) integration. The SiLago framework builds on the concept of composition by abutment, assembling hardened, grid-aligned building blocks with pre-integrated logic, interconnect, and power infrastructure. Within this context, Dynamically Reconfigurable Resource Array v.2 (DRRA2) is a reconfigurable, SiLago-based coarsegrained architecture designed for streaming applications, such as dense linear algebra and neural networks. This thesis focuses on automating two critical backend tasks for DRRA2: floorplanning and power planning. Floorplanning requires precise placement of SiLago blocks based on architectural descriptions and synthesized netlists, while ensuring correct alignment, pin orientation, and abutment rules. Power planning involves generating a hierarchical power grid, including rings, stripes, and rails, that maintains electrical continuity across block boundaries and supports robust current delivery. The proposed automation flow is implemented in Python and generates TCL scripts compatible with standard EDA tools like Cadence Innovus. It includes JSON and netlist parsing, coordinates computation, pin constraint generation, and abutment-aware power structure creation. The flow has been validated across multiple DRRA2 configurations, demonstrating correct block alignment, reliable power connectivity, and clean Design Rule Checking (DRC) compliance. Limitations remain as the work is part of a research project, but the flow is built to be flexible and easily adaptable to future revisions. Overall, this work enables scalable backend integration of DRRA2 within the SiLago methodology, significantly reducing manual effort and supporting continued research on automated, composable chip design.

The necessity to make energy distribution more efficient is important for both ethical and environmental reasons: it ensures that consumers receive their required energy, and it lowers the risk of having electricity losses. Predicting energy consumption can provide some aid to plan for a period ahead, and today there are multiple models used for time series predictions. Two areas of these models are statistical and neural network. While both have their advantages and disadvantages, neural networks usually produce better prediction results. However, exogenous variables can be added to the testing, and can improve predictions, but there are few studies involving them. This study will therefore see how much improvement the use of exogenous variables to statistical methods can be made compared to neural network models without exogenous variables. The models investigated in this thesis is AutoRegressive Integrated Moving Average with Exogenous variables (ARIMAX), Seasonal ARIMAX (SARIMAX), NeuralProphet, MultiLayered Perceptron (MLP) and Gated Recurrent Unit (GRU). The first two are statistical, the last two are neural networks, and NeuralProphet is a combination as it uses both statistical and neural network methods in its implementation. These five models were set to predict 24 hours into the future, after training on data containing hourly energy consumption. The exogenous variables were the temperature of the area, as well as whether it was a weekday or weekend. The resulting optimal model turned out to be NeuralProphet, with SARIMAX performing almost as well, based on the error metrics. The two neural network models came thereafter, whose performances were almost equal, and MLP even surpassing GRU in some cases. The least optimal model was ARIMAX, which is not too surprising as it did not take seasonal patterns into consideration. However, doing statistical hypothesis testing of their performances revealed that only half of the pairwise comparisons rejected the null hypothesis, stating that the model choice made an impact on the prediction results. All comparisons relating to ARIMAX were rejected, and an additional one between NeuralProphet and GRU. This means that ARIMAX is proven to be the worst performing model when doing time series prediction, and NeuralProphet is proven to be better than GRU. Otherwise, it is concluded that more testing is required to provide results regarding the remaining models.

The global transition to clean energy has intensified the need for efficient solar power generation. This study investigates how to maximize energy output from photovoltaic power plants by comparing three system configurations: Fixed-tilt, Horizontal Single-Axis Tracking (HSAT), and Vertical Single-Axis Tracking (VSAT). The primary research question focuses on identifying which tracking system yields the highest energy production across diverse European locations, accounting for inter-row shading effects. A custom simulation framework was developed in Unity 3D, utilizing solar irradiance data from the PVGIS-SARAH3 database and implementing shading loss calculations based on the bypass diode model by Martínez-Moreno et al. Simulations were conducted for five cities with varying climates, and production was evaluated on monthly, daily, and hourly scales. Results show that VSAT consistently achieves the highest total energy yield, particularly excelling in spring and autumn months and in early morning and late afternoon hours. However, it also suffers from the highest shading losses, highlighting a tradeoff between performance and row spacing. Fixed-tilt systems demonstrated the lowest variability and shading loss but also the lowest yield. Welch’s ANOVA and Games–Howell post-hoc tests confirmed that the performance differences between the systems are statistically significant. The findings support the adoption of VSAT systems in locations with low land costs or dual land use potential, such as AgriPV, where increased energy yield justifies greater row spacing.

Sentiment analysis is a valuable tool for understanding and evaluating the tone and content of documents, articles and text. It can be used in conjunction with social media, to explore how public opinion can change over time, with a focus on political events. In this thesis, sentiment analysis together with clustering has been used to gain insight into that subject, using the 2024 Romanian presidential election as a relevant case study. By tracking changes in people´s perceptions of events over time, the effect social media can have on political events is analyzed in more detail. To achieve this, two sentiment analysis models were developed using the BERT and FastText libraries, and these models were trained to support both Romanian and English. After sentiment analysis, clustering was performed to further divide the results into distinct topics. This was achieved by using the spectral clustering method together with K-means. With a total of 8 494 tweets collected from the social media platform X before and after the first round of the election, a distinct negative shift could be observed, indicating a general dissatisfaction and a shift of the discussion regarding the election. Sentiment analysis was successfully performed, resulting in observable changes between before and after the election. Clustering helped provide more precise insight into which topics the changes in sentiment occurred. However, due to limited data before the election, especially in English, only the Romanian tweets could be clustered and the reliability is difficult to confirm. Further work is needed to be able to draw more general statements about political events, including collecting more data from various social media platforms and increasing the number of case studies to see if similar patterns can be observed.

Machine learning-based pricing strategies typically rely on either supervised learning (SL) for predictive modelling or reinforcement learning (RL) for sequential decision-making. However, both approaches face limitations in data-constrained environments like small and medium-sized enterprise (SME) platforms, where user interaction data are sparse, static, or subject to drift over time. This thesis addresses the gap between static pricing methods and fully adaptive dynamic pricing by proposing a sequential hybrid SL–RL framework tailored to SME conditions. The framework uses SL to approximate user behaviour and reward functions from historical data, which are then used to initialize a tabular RL agent that adaptively refines pricing policies. Unlike parallel hybrid approaches, this sequential design enables rapid early adaptation to reduce cold-start problems and continues to adapt to handle data drift scenarios. The method was evaluated in a simulation environment combining synthetic and real publisher data to reflect evolving user and market dynamics. Empirical results show that, with aligned predictive start, the hybrid agent achieves on average between –2 and 6 percentage points in cumulative reward relative to a traditional RL, while requiring approximately 30% fewer price adjustments. In edge-case scenarios without assumptions about user interactions, the framework outperformed an exploratory baseline by 7–24%, effectively shortening the cold-start phase. Overall, the hybrid SL–RL strategy demonstrated benefits to behavioural shifts, offering a strong trade-off between performance, stability, and data efficiency. These findings suggest that the proposed framework provides a viable and resilient pathway for digital publishers to adopt dynamic pricing, even in settings with limited and drifting data.

Cochlear implant (CI) users often face difficulty in listening compared to normal hearing. Research has shown that hearing training improves the ability to appreciate music, however, spectral elements such as pitch and timbre are especially hard to understand. This study aims to explore the use of a serious game to increase the motivation and engagement for CI users to train their hearing. The study used a “User Centered Design” approach, and based on the input from an exploratory interview with a single mixed hearing (CI and hearing aid) individual, a prototype game was developed. Input from the participant was gathered during a testing session of the game along with observations from the testing. The participant found the balance between challenge and usability to be good and the auditory and visual help mechanics were very appreciated. Necessary improvements were highlighted however, involving a robust visual effects integration to make feedback clearer and more intuitive. The results show that a serious game has potential to reduce cognitive load during hearing training with a CI. Future research should involve a larger group of users and apply longitudinal studies to validate the effect on both improving music perception as well as its effect on motivation and engagement in hearing rehabilitation.

The 5G technology has significantly extended the challenge of containerised network function configuration, management, and scaling on Kubernetesbased clusters. Scripted approaches to deploying and managing Kubernetesnative systems that are still widely used in telecom industry scenarios don’t scale, not offering full automation or efficient operations consistency. This thesis aims to evaluate how a GitOps-based lifecycle management system for 5G Kubernetes infrastructure, when combined with Infrastructure as Code and Configuration as Code, can help streamline the deployment and lifecycle management of containerised network functions. The industry gap the project aimed to fill pertained to the adoption of advanced automation practices such as GitOps in telecom-grade environments. A proof-of-concept was developed to automate the lifecycle of CNFs on Kubernetes, leveraging GitOps tools such as Flux CD and Infrastructure and Configuration as Code. The proposed system was validated against a legacy semi-automated system that functioned as a baseline. Evaluation of the two systems was based on reconciliation time, MTTA, scalability, and reliability, with the results showing a 50% improvement in the proposed system for both reconciliation time and deployment latency. Additional evaluation around effort and repeatability for reconciliations indicated minimal manual intervention with increased repeatability. The evaluation framework was also extended to accommodate scalability validation with an increase in CNF replicas, as well as to test for self-healing in the presence of configuration drift. The results of this work and internal discussions after the live demo indicate a readiness to evaluate this for adoption on a production system with further improvement strategies.

Every day, our brain performs impressive tasks, but what is truly remarkable is its energy efficiency. This has led to the development of spiking neural networks (SNNs) and event cameras, which are energy-efficient, biologyinspired models and imaging sensors. The high temporal resolution of event cameras also make them particularly well-suited for capturing rapid motion. Independently, drone detection is an increasingly important concern in both civil and military related contexts, often under energy-constrained conditions. When captured with an event camera, drones produce a characteristic spatiotemporal pattern due to their high-frequency spinning propellers. The here developed algorithm ’PropellerSpikes’ is an explainable SNN that detects drones because of this pattern by first detecting motion directions and then rotational motion. With appropriate parameter settings, the network was able to detect and localize rotating propellers in scenarios with various drone types, translational speed and altitudes. Estimations show that the network has very low-power capabilities if implemented on dedicated neuromorphic hardware.

Beta oscillations in the human brain appear as transient, high-amplitude events known as beta bursts, which underpin sensorimotor and cognitive processes. Their disruption is a hallmark of Parkinson’s disease (PD), yet the large-scale spatiotemporal organization of beta bursts in the resting brain remains poorly understood. In this single-subject feasibility study, we employ high-density optically pumped magnetoencephalography (OPM-MEG) to map beta burst-derived functional networks, capturing transient, event-locked connectivity across 26 cortical regions in a healthy adult at rest. Resting-state data were recorded with a 128-channel OPM array and source-reconstructed; beta bursts (13–30 Hz) were identified via thresholding and spatiotemporal clustering. Directed coburst interactions were inferred through lag-resolved cross-correlation, with statistical significance assessed by circular time-shift surrogates and FDR correction. Graph-theoretic analysis of the burst network suggested a lateralized, frequency-specific architecture: sensorimotor regions (postcentral gyri) functioned as high out-degree broadcasters, whereas frontal areas (orbitofrontal and rostral middle frontal cortices) emerged as convergence hubs. These findings suggests that even at rest, beta bursts form structured, propagating networks with distinct directional roles. Importantly, this work introduces a novel, burst-centric pipeline combining OPM-MEG, unsupervised clustering, time-lagged connectivity, and network theory. Although inherently tentative due to the single-subject design, it provides proof of concept that beta bursts can delineate dynamic functional networks. Extending this framework to larger cohorts and clinical populations—especially in PD, where burst dynamics correlate with symptomatology—may yield physiologically grounded biomarkers of network dysfunction.

AI music generation platforms allow users to generate music by just providing textual prompts. However, how the sentiment of the generated music aligns with the sentiment of the prompts has yet to be explored. This thesis examines the alignment between the valence (pleasantness) and arousal (intensity) of text prompts to the generated music. To achieve this, we fine-tune a pre-trained language model to infer valence and arousal (V-A) from text prompts, and train a deep neural network to infer V-A from music. We apply these models to a dataset of 6,086 text prompt-music pairs collected from two commercial AI music platforms, Suno and Udio, and analyze the relationships of the inferred scores. We conduct a human listening test on 24 text prompt-music pairs selected based on the model inferences, and manual inspections on specific examples to validate the model results. Our results indicate a significant but weak correlation between the V-A of text prompts and the resulting music. We include illustrative examples to support interpretation and discussions.

Accurate unemployment forecasting is crucial for economic policy and planning, especially for institutions such as central banks. Traditional econometric models often struggle to capture non-linear patterns in economic data. This study investigates whether machine learning (ML) methods, specifically Random Forest, XGBoost, Neural Networks, and LASSO regression, can improve forecast accuracy for the unemployment rate in Sweden, compared to benchmarks such as Random Walk and Autoregression, as well as the Riksbank’s current forecasts. Using a rolling window approach across 1-, 3-, 6-, 9-, and 12-month horizons and a dataset of economic and survey-based indicators, this study finds that Random Forest and XGBoost consistently outperform traditional models throughout the test period, which spans from 2018 to March 2025. LASSO performs particularly well at the 1-month horizon, but struggles at the longer horizons. The Neural Network models consistently underperform, likely due to insufficient tuning and adaptation. The analysis also highlights which variables are most important at different horizons: current labor market indicators dominate short-term forecasts, while macro-financial and international signals become more relevant for longer horizons. Compared to the Riksbank’s forecasts, the top performing ML models perform better at all horizons except at 9 months. However, this comparison is not entirely fair, as the Riksbank incorporates expert judgment and uses seasonally adjusted data, while this study does not. Overall, the findings suggest that machine learning models can enhance unemployment forecasting and serve as valuable complements to traditional methods, particularly when combined with expert knowledge in policy settings. These findings align broadly with previous research, where the main contribution lies in applying these methods to the Swedish context, offering both empirical insight and practical relevance for policymakers.