As climate change imposes rapid transformation toward net zero carbon emissions, the transformation of the energy sector poses a key role. With the rapid advancement of artificial intelligence (AI) and its seemingly endless possibilities, it raises the question whether synergies can be realized between AI and the renewable energy sector. The objective of this thesis is to investigate AI and data science applications within the renewable energy sector and provide an overview for energy engineers without previous knowledge of data science and AI. A systematic literature review was conducted to answer these questions. The thesis analyzed 171 peer-reviewed articles and review articles from 2022-2024 using the databases Scopus and Web of Science. Synthesizing the results, three prominent AI application areas were identified; Forecasting, Predictive maintenance and Integration to the grid. ML and ANN represent the most distinguished applications of AI with its speciality on time-series data which is abundant in renewable energy systems and forecasting in particular. Prognostic maintenance applies AI in many different ways ranging from image processing to assess the state of the machinery, to predicting failures with historical data to reduce downtime of power plants. The study further indicates that AI may be particularly well-suited to solve complex nonlinear relationships needed to increase the yield and integrate renewables in the grid. In addition to mapping the current technological landscape of AI, the study presents technical challenges, ethical and societal issues and trends.

Since the Brundtland Report, sustainable development has been established as a central norm within urban development practice, with sustainable urban development often treated as an almost self-evident objective for contemporary urbanisation, where environmental protection, economic growth, and social welfare are combined within a shared policy agenda. This self-evidence can, however, be understood as an expression of common sense—assumptions that appear universal and uncontroversial, but which are in fact historically and geographically situated and shaped by specific power and knowledge relations. In practice, this implies that sustainable urban development does not constitute a unified concept, but rather a complex, mutable, and often conflictual process that is continuously reconfigured across different socio-material contexts.Despite this acknowledged complexity, sustainable urban development has been criticised for being operationalised through a discourse of ecological modernisation, in which technological innovation and resource efficiency are expected to enable continued economic growth without compromising the needs of future generations. This pragmatic policy logic has contributed to a depoliticisation of the sustainability issue by downplaying conflicts between social, economic, and ecological justice and growth-oriented market logics. As a result, sustainability risks reproducing prevailing urbanisation trajectories rather than transforming them. In this sense, sustainable urban development projects can be understood as a sustainability fix, where an explicitly sustainable planning practice legitimises a continued urban development practice that does not substantially differ from dominant urbanisation practices.In the Swedish context, this practice has evolved through a long tradition of working with sustainable urban development, where Sweden has positioned itself as an international forerunner, with a large number of ambitious projects and several national strategies aimed at strengthening innovation, collaboration, and local capacity for sustainable development. A central component of this agenda has been the development of sustainable neighbourhoods, where projects function as experimental environments for testing new technical, social, and planning solutions. Several flagship projects in metropolitan regions have received considerable attention and have circulated within planning and policy discourses as best-practice examples of sustainable urban development.This dominance of metropolitan-based examples has, however, contributed to a knowledge gap regarding sustainable urban development in more peripheral and peri-urban environments. Peri-urbanisation describes a contemporary pattern of urbanisation in which central and peripheral urban characteristics are intertwined in heterogeneous and fragmented landscapes characterised by in-betweenness. These environments are not the outcome of prevailing planning ideals, but rather of unplanned processes emerging as spatial and functional structures expand. While peri-urban environments have often been marginalised within both research and planning practice, they today constitute central arenas for contemporary urbanisation and thereby potential sites for innovation and transformation.Against this background, the thesis aims to examine what happens when the concept of sustainable neighbourhoods is mobilised to peri-urban contexts. The focus is on the spatial, institutional, and conceptual reconfigurations that emerge when policies, practices, and ideas are translated across contexts. The thesis draws on theories of policy mobility, which emphasise that urban policies do not circulate as fixed entities but are transformed through processes of translation, mutation, and reconfiguration. This perspective enables an analysis of how sustainable urban development is constituted through a dialectical interplay between dominant discursive articulations and local conditions within the situated context.Theoretically, the study is grounded in a post-structuralist and relational tradition, where space is understood as assemblages—dynamic configurations of socio-material relations in continuous processes of becoming. In combination with critical urban theory, this highlights how the mobilisation of urban policy discourses can function as hegemonic interventions that reproduce unevendevelopment and prevailing power relations, while the contingency of spatial and discursive configurations simultaneously opens up possibilities for counter-hegemonic articulations and alternative development trajectories. The thesis also integrates literature on place-based development, social innovation, and transformative innovation policy to explore how sustainable transformations can be contextually grounded in local conditions rather than scaled from previous examples.Empirically, these questions are examined through an analysis of planning documents for sustainable neighbourhood development in Sweden, as well as an in-depth case study of Jakobsdalen in Borlänge—a project that represents the mobilisation of sustainable neighbourhood development in a peri-urban context. Borlänge, with its industrial history and experience of territorial stigmatisation, provides an illustrative case of how sustainability discourses are mobilised to reconfigure place identity in order to enhance attractiveness. The case study thus enables an analysis of both the possibilities and challenges associated with sustainable neighbourhood development in contexts characterised by peripheralisation and complex socio-spatial configurations.The overarching aim of the thesis is to problematise how sustainable neighbourhoods are conceptualised and mobilised within planning and policy discourses, and to identify the implications of this for peri-urban environments. The study also seeks to explore possibilities for the emergence of more context-sensitive and transformative practices by developing and testing a designerly approach in which sustainable urban development is treated as a boundary object. This approach aims to enable shared processes of interpretation between actors and contextual conditions, thereby organising and opening up for transformative and innovative solutions adapted to situated contexts.Through this approach, the thesis contributes to a deeper understanding of sustainable urban development as a relational and situated practice. It demonstrates how dominant urban articulations of sustainability risk reproducing prevailing development trajectories while simultaneously constraining the translation of policies across contexts. The thesis shows how this generates significant challenges for sustainable urban development in peri-urban environments. At the same time, it highlights the potential of these environments as sites for innovation and transformation, where contextual dynamics can be mobilised as resources rather than constraints. The thesis therefore argues for the need for a more pluralistic and context-sensitivemobilisation of sustainable urban development that recognises diversity and opens up for alternative urbanisation practices.

Semi-empiriska ”group-contribution” metoder (GC-metoder) erbjuder ett snabbt och enkelt sätt att förutsäga polymerers egenskaper med hjälp av regressionsanalys, antingen direkt utifrån experimentella data för den aktuella egenskapen eller indirekt via korrelerade strukturella parametrar. Befintliga GC-metoder har emellertid huvudsakligen optimerats för syntetiska polymerer. I detta examensarbete presenteras en ny parametrisk regressionsmodell anpassad för GC-analys av biopolymerer såsom kitosan, cellulosa, hemicellulosor och polynukleotider. Metoden bygger på den rotationsisomeriska tillståndsmodellen (RIS-modellen) och gitterteori, och tar hänsyn till intermolekylära interaktioner genom att relatera Tg till fundamentala termodynamiska storheter. Metoden är väl lämpad för integrering med maskininlärningstekniker och möjliggör både statistisk hypotesprövning och kontinuerlig, iterativ modellförbättring med hjälp av Bayesianska metoder. Arbetets huvudsakliga fokus ligger på metodutveckling, men som ett konceptbevis tillämpas modellen för att visa hur glasomvandlingstemperaturen (Tg) hos xylanliknande polymerer kan förutsägas.

The fourth industrial revolution envisions flexible manufacturing in which autonomous mobile robots collaborate over shared wireless networks. Fifth-generation (5G) cellular technology—and its successors Beyond 5G and 6G—has been positioned as the enabling infrastructure, promising network slicing, quality-of-service differentiation, and mobile edge computing that could untether industrial robots from cables without sacrificing control fidelity. The concurrent rise of physical AI and the AI-native radio access network paradigm, which converges AI workloads with network infrastructure, further amplify both the promise and the complexity of this integration. This thesis investigates whether current research is equipped to exploit that infrastructure—and identifies several structural misalignments that the field must address. 

The work proceeds through four phases. A systematic mapping study of edge computing for cyber-physical systems finds that industrial manufacturing is the single largest application driver, yet mobile edge computing in the 5G sense and systematic treatment of trustworthiness attributes are largely absent from the literature. A focused survey of what this thesis terms communication-aware motion planning reveals a field that is terminologically fragmented, methodologically confined to motion-level optimization with channel-centric metrics, and almost entirely disconnected from real telecommunications infrastructure. The survey yields a taxonomy and evaluation criteria that provide a structured vocabulary for characterizing and comparing approaches across the robotics and telecommunications communities. 

In response, a planning framework called RoboPlan5G demonstrates joint communication and mission planning by formulating multi-robot coordination in the Planning Domain Definition Language, treating 5G physical resource blocks as explicit planner decision variables alongside task allocation and sequencing. Incorporating network resource allocation into the mission planner reduces spectrum requirements by fifty percent while satisfying plan constraints within acceptable computation time. 

An empirical measurement campaign on a private 5G industrial testbed then challenges a central assumption shared by most existing communication-aware planning approaches: that favorable channel conditions translate reliably into end-to-end performance. A commercial ray-tracing simulator predicted signal-to-interference-plus-noise ratio with reasonable accuracy yet systematically over-predicted throughput, because multiple-input multiple-output spatial rank adaptation—the dominant source of prediction error—lies beyond the reach of channel-centric models. A data-driven Gaussian process regression model reduced prediction error by approximately two-thirds and eliminated systematic bias. The testbed also did not support the radio access network-level slicing that both the planning framework and the broader literature presuppose, revealing a gap between standardized interfaces and deployed capabilities. 

The thesis culminates not in a single superior framework but in a harder-won insight: the field's algorithmic sophistication obscures a disconnect from industrial reality operating at multiple levels simultaneously—conceptual, modeling, infrastructure, and evaluation. At the conceptual level, motion planners are extended with signal-level metrics when mission planners should be extended with service-level network abstractions. At the modeling level, planners optimize for channel quality while the phenomena that determine throughput in modern cellular systems lie beyond channel-centric models. At the infrastructure level, the network capabilities assumed in the literature remain incompletely realized in commercial equipment. The contribution of this work lies in exposing these disconnects systematically and in offering both analytical tools and a constructive artifact that together chart a more grounded path forward. Achieving the smart factory vision requires not merely better algorithms but a reorientation of assumptions, formalisms, and evaluation practices—from motion to mission planning, from channel metrics to network-service abstractions, and from simulated validation to measurement-grounded evaluation.

Rapid changes and increasing climatic variability across the widely varied Köppen-Geiger regions of northern Europe generate significant needs for adaptation. Regional planning needs high-resolution projected temperatures. This work presents an integrative statistical bias correction framework that incorporates Vision Transformer (ViT), Convolutional Long Short-Term Memory (ConvLSTM), and Geospatial Spatiotemporal Transformer with Attention and Imbalance-Aware Network (GeoStaNet) models. The framework is evaluated with a multicriteria decision system, Deep Learning-TOPSIS (DL-TOPSIS), for ten strategically chosen meteorological stations encompassing the temperate oceanic (Cfb), subpolar oceanic (Cfc), warm-summer continental (Dfb), and subarctic (Dfc) climate regions. Norwegian Earth System Model (NorESM2-LM) Coupled Model Intercomparison Project Phase 6 (CMIP6) outputs were bias-corrected during the 1951–2014 period and subsequently validated against independent historical observations (1951–2014) of day-to-day temperature metrics, extreme value distributions (99th percentile), and thermodynamic coupling (Diurnal Temperature Range). The ViT showed improved performance (Root Mean Squared Error (RMSE): 1.01 °C; R² : 0.92), allowing for production of credible bias-corrected projections. Under the SSP5-8.5 scenario, the Dfc and Dfb climate zones are projected to warm by 4.8 °C and 3.9 °C (Summer T m a x ), respectively, by 2100, with expansion in the diurnal temperature range by more than 1.5 °C. The Time of Emergence signal first appears in subarctic winter seasons (Dfc: ~ 2032), signifying an urgent need for adaptation measures. The presented framework offers station-based, high-resolution estimates of uncertainties and extremes, with direct uses for adaptation policy over high-latitude regions with fast environmental change.

The world stands before the challenge of making the switch from fossilfuels to green energy in order to fight climate change. To achieve this, the availablegreen energy sources need to be broadened and optimized, to cover for thesubstantial gap in production as fossil fuels produce more than twice the amount ofenergy as wind, solar and hydropower combined. The latter is considered the mostflexible out of the three mentioned and therefore is a very attractive resource tooptimize. This project attempts to take a further step in this process of makinghydropower more effective by analyzing how different regression algorithms can beused to predict the inflow into hydro reservoirs by using open source meteorologicaldata, which in this case is the flow of tributary rivers. It concludes that there is nosingle algorithm that safely outperforms the rest in the general case, or deliverssatisfying results. Additional work is needed to determine which further modificationsto the algorithms and input data, such as regional precipitation, seasonal snowmeltand local climate could help explain the variability of inflow.

Den snabba integreringen av artificiell intelligens (AI) inom utbildning har möjliggjort ett paradigmskifte från att se datorer som statiska verktyg till att behandla dem som symbiotiska partners inom en samverkande “människa-AI-allians”. Detta examensarbete undersöker tillämpningen av generativ artificiell intelligens (GenAI) inom matematikundervisning, med specifikt fokus på konstruktion och utvärdering av en AI agent utformad för att ge återkoppling enligt de fyra nivåer som definieras av Hattie och Timperley: uppgift, process, självreglering och person. Genom en iterativ designprocess innefattade studien utveckling och testning av flera versioner av agenten med hjälp av modellen Llama 3.3 70B, huserad på en lokal server på en svensk gymnasieskola. Prompt engineering delen vägleddes av IDEA ramverket och tillämpade principerna PARTS (Persona, Aim, Recipients, Theme, Structure) och CLEAR (Concise, Logical, Explicit, Adaptive, Restrictive) för att förfina agentens pedagogiska beteende. Empiriska data samlades in från ungefär 60 konversationer med svenska gymnasieelever och analyserades genom tematisk analys för att kartlägga interaktioner mot de teoretiska återkopplingsnivåerna. Resultaten bekräftar att genAI agenter är kapabla att ge återkoppling på alla fyra nivåer, även om återkoppling på uppgiftsnivå förblev mest förekommande på grund av stora språkmodellers grundläggande tendens att ge omedelbar verifiering. Tematiska fynd belyste en framgångsrik förskjutning mot återkoppling på processnivå och användningen av ledtrådar av typen “Vart ska jag härnäst?” för att aktivera elevernas självreglering. En kritisk “upplösning av självreglering” observerades dock i komplexa textuppgifter, där agenterna ofta föll tillbaka på direkt instruktion, vilket potentiellt kan framkalla en “expertise reversal effect”. Även om genAI erbjuder en kraftfull potential för omedelbar kognitiv stöttning (scaffolding), lägger den en betydande utvärderingsbörda på de lärande, som måste besitta tillräckliga förkunskaper för att upptäcka sakfel eller “hallucinationer”. Studien drar slutsatsen att en framgångsrik integrering av AI agenter kräver ett “humans in the loop” ramverk för att mildra tekniskt beroende och säkerställa att teknologin förstärker, snarare än ersätter, elevens egen kognitiva ansträngning.

This deliverable describes the main developments carried out in WP6, with focus on models and algorithms to improve long-term and short-term timetabling of the railway network. The activities in WP6 aim at increasing infrastructure and transport utilisation capacity through optimised and robust timetables, synchronized with rolling stock planning. These objectives have been targeted through

a. The development of advanced algorithms for the generation and adjustment of timetables and rolling stock planning, which will be further developed and completed in WP7

b. The definition a suitable family of use cases, which will be demonstrated in WP7

c. The implementation of specific technical enablers.

Addressed technical enabler and the defined use cases and demonstrators are synthetically reported in the background Section 3 of this document.

All the timetabling and rolling stock planning problems tackled in WP6 require to find good quality solutions fulfilling various physical and logical requirements, and the business rules of the railway infrastructure managers and railway undertakings. As such, they can be viewed as optimization problems, which can be modelled and solved by means of mathematical optimization, an AI discipline which concerns the making of optimal decisions. With few exceptions, the models developed in WP6 are based on Mixed Integer Linear Programming or Constraint Programming, solved then by means of specialized commercial solvers (as CPLEX, or GUROBI), or by ad-hoc heuristic algorithms, such as local search, genetic algorithms, simulated annealing. Mathematical decomposition and graph theory are also exploited to model and solve various problems.

Although the algorithms developed in WP6 are not yet fully completed – they will be in the first year of WP7 – still some interesting and promising conclusions can be drawn. In fact, tests on realistic instances have been performed. It turns out that the developed methods work well for the size and the type of instances for which they are designed. In turn, this implies that we can expect they will tackle the instances arising in the demonstrations of the planned use-cases. Ultimately, this means that in general the approaches will be able to support human planners in their activities, and to automatize segments of the current planning process. Preliminary results show that solutions of high quality can be produced in short computing time. One limit is that, since the algorithms will be completed and demonstrated in WP7, these conclusions are still very preliminary. Also, the solution of full integrated problems appears to be still out of reach, and we need to content ourselves with tackling suitable subproblems. For instance, we can possibly compute an optimal or quasi-optimal timetable for the entire Norwegian network, and subsequently calculate an associated optimal rolling stock rotation, and solve the stabling problem, but we are still far from being able to solve to optimality the three problems jointly. 

In this study, we explore the feasibility of single-trial predictions of odor registration in the brain using olfactory bio-signals. We focus on two main aspects: input data modality and the processing model. For the first time, we assess the predictability of odor registration from novel electrobulbogram (EBG) recordings, both in sensor and source space, and compare these with commonly used electroencephalogram (EEG) signals. Despite having fewer data channels, EBG shows comparable performance to EEG. We also examine whether breathing patterns contain relevant information for this task. By comparing a logistic regression classifier, which requires hand-crafted features, with an end-to-end convolutional deep neural network, we find that end-to-end approaches can be as effective as classic methods. However, due to the high dimensionality of the data, the current dataset is insufficient for either classifier to robustly differentiate odor and non-odor trials. Finally, we identify key challenges in olfactory BCIs and suggest future directions for improving odor detection systems.

Decoding visual images from brain activity has significant potential for advancing brain-computer interaction and enhancing the understanding of human perception. Recent approaches align the representation spaces of images and brain activity to enable visual decoding. In this paper, we introduce the use of human-aligned image encoders to map brain signals to images. We hypothesize that these models more effectively capture perceptual attributes associated with the rapid visual stimuli presentations commonly used in visual brain data recording experiments. Our empirical results support this hypothesis, demonstrating that this simple modification improves image retrieval accuracy by up to 21% compared to state-of-the-art methods. Comprehensive experiments confirm consistent performance improvements across diverse EEG architectures, image encoders, alignment methods, participants, and brain imaging modalities.

Non-Newtonian fluids are omnipresent in all natural and synthetic processes surrounding us. Due to our literal submergence in water and air, Newtonian fluids as such have a prominent vitality in many aspects of the industrial advancements, yet from a technical perspective, Newtonian fluids form a sub-set of the whole picture - the physics of flowing matter - and the expanded, general description, without any assumptions on the (i) constant viscosity, and (ii) single-source stress (i.e., viscosity), belongs to the non-Newtonian fluids. Under this taxonomy, an ever increasing level of complexity can be found, which deviates from the norm and intuitively familiar behavior of Newtonian fluids, and - with its many open problems at present - requires in depth, fundamental level investigations. Non-Newtonian fluids of different classes can have shear dependent varying viscosities, store elastic energy and release it out-of-phase with the local flow field and at multi-mode time-scales, undergo continuous and bi-directional micro-structural break-down and recovery, exhibit memory effects, age, and be made of living constituents. 

In this thesis, a variety of phenomena relevant to time-dependent flows and state transitions of non-Newtonian fluids are investigated experimentally using a wide range of flow measurement and rheological techniques. Fluids of viscoelastic (VE), elastoviscoplastic (EVP), and thixo-elastoviscoplastic (TEVP) natures were studied in regards to wall-slippage, pressure-driven duct flows, thixotropy and memory effects, elastic- and shear-banding driven instabilities, inertia-dominated instabilities and transition to turbulence, sub-yielding dynamics, and flow around obstacles and ducts with varying cross-sections. To that end, flow velocimetry techniques such as Particle Image Velocimetry (PIV), Laser Doppler Velocimetry (LDV), Doppler-Optical Coherence Tomography (D-OCT), medium structure measurements with intensity recordings of Optical Coherence Tomography (OCT) and Polarized Light Imaging (PLI), and rheometric assessments with steady- and oscillatory shear-, as well as parallel superposition rheology, have been used. The results regarding the development and usage of the measurement techniques, as well as physical interpretations of the studied instationary flow phenomena, are reported in the appended papers, and summarized in the upcoming chapters. 

The aim of this thesis is to deepen understanding of how sustainability is integrated into urban development through different governance instruments.The thesis focuses on two central instruments in contemporary urban development in Sweden: certification systems for sustainable urban areas and urban experiments. Previous research on certification systems has primarily examined the content, indicators, and structures of existing systems. This thesis instead focuses on how certification systems for urban areas are developed and shaped through trade-offs between competing principles and requirements that guide their design. By analysing a development process, the thesis contributes knowledge about the rationales and considerations that underpin certification as a governance instrument. In the case of urban experiments, it is often assumed that experimental projects generate learning that can be transferred beyond individual cases. However, questions of how such learning is organised and scaled within municipal organisations have received less attention. This thesis therefore examines how learning from urban experiments can be identified, organised, and scaled in municipal practice. The research draws primarily on transdisciplinary and design-oriented approaches, in which knowledge is developed through close engagement with practice. A central empirical component is the development of the certification system Citylab Post-Construction, in which the researcher was actively involved. This provided an in-depth basis for analysing the trade-offs and considerations that shape certification system design. In addition, the thesis includes two empirical studies of urban experiments in municipal contexts, examining both the testing of new approaches to structuring upscaling and municipal practitioners’ perspectives on learning and organisational conditions. The findings show that the development of certification systems for sustainable urban areas is characterised by recurring trade-offs between competing principles, such as simplicity, comprehensiveness, and methodological credibility. Making these trade-offs explicit enhances understanding of why certification systems take the forms they do and what limitations follow from different design choices. The studies of urban experiments show that while experimentation can create opportunities for learning and innovation, such learning often remains project-bound unless supported by organisational structures. Upscaling learning depends primarily on organisational conditions within municipalities, including mandates, resources, and recipient capacity, rather than on the content or scope of individual experiments. Clarifying what should be scaled, how it can be scaled, and under what conditions is therefore a central part of making experimental knowledge actionable. Overall, the thesis shows that certification systems and urban experiments shape sustainable urban development in fundamentally different ways. Certification systems tend to reinforce established standards and ways of working, while their reliance on measurable and influenceable indicators means that sustainability dimensions that are difficult to quantify or attribute to specific actors risk receiving less attention. Urban experiments, in contrast, often stabilise project-based approaches to sustainability work, where continuous cycles of pilots and innovation can overshadow the upscaling of already tested solutions. At the same time, both instruments can play important roles in supporting the integration of sustainability in urban development, provided that their governing effects and limitations are recognised and managed in practice.

Urbanising landscapes are undergoing rapid transformation as densification, sprawl, land-use competition, and climate change reshape urban green and blue spaces. These dynamics influence not only ecological functions, but also everyday opportunities for people to encounter nature. In particular, children’s access to nearby nature is increasingly recognised as important for wellbeing, environmental awareness, and the development of nature stewardship. However, the conditions that enable children’s meaningful and lasting connections with nature remain unevenly studied. This licentiate thesis examines how formal land use planning and management of urban green-blue infrastructure, and practices of civic organisations, shape children’s opportunities to meaningfully connect with nature in urbanising landscapes. The research is situated within sustainability science and social-ecological systems research and adopts stewardship as an analytical lens for examining how responsibility for human-nature relations is distributed across society. The thesis consists of two qualitative case studies conducted in the Stockholm region. Paper 1 analyses formal land use planning and management processes shaping multifunctional urban green-blue infrastructure, based primarily on semi-structured interviews with municipal and regional officials. Paper 2 investigates how Swedish civic outdoor organisations foster children’s nature connection through pedagogical practices, recurring outdoor activities, and intergenerational learning. The findings show that formal land use planning and management processes play a decisive role in shaping structural conditions for nature access, yet outdoor recreation and children’s everyday contact with nature are often weakly prioritised in the formal governance of land use characterised by sectoral fragmentation and competing land-use interests. At the same time, civic outdoor organisations provide important relational infrastructures that enable children’s repeated, playful, and meaningful engagement with nature. Taken together, the studies demonstrate that children’s opportunities to develop lasting relationships with nature depend on the interaction between structural and relational conditions. Urban land use planning and management determines whether nature is physically available and accessible in a broad sense, while practices of civic organisations contribute to the experiences, knowledge, and meanings through which children learn to relate to landscapes. By linking formal land use planning and management with lived practices of nature engagement, the thesis contributes to sustainability science by highlighting how stewardship emerges across public and civic sectors and by bringing children’s experiences more explicitly into discussions of sustainable urban development.

Digital media operates on a curious boundary between storage and loss. While each new storage format promises a permanent solution to our exponentially expanding media libraries, they inevitably fail or otherwise become unusable. This paper reflects on a long-term design process that attempts to bring a different paradigm to the experience of personal digital media: destruction. We present an annotated portfolio of a set of sound listening devices, critically unpacking the particular temporal, perceptual, and experiential qualities that emerge when designing for the loss of personal media. These annotations show how destruction comes to matter in designing against the traditional bias towards growth and accumulation in HCI.

In recent years, emerging technologies have led to an increase in the number of safety-critical services and applications that require reliable communication services. To accommodate performance requirements, mobile network operators offer service level agreements (SLAs) that specify guaranteed quality of service (QoS) targets. However, maintaining these targets is challenging in highly mobile scenarios, where changing propagation conditions, network load, and interference can alter the statistical properties of wireless performance metrics. To address the issues with varying statistical properties, ML-based predictive QoS (pQoS) has been proposed to help the network proactively detect and prevent future network degradations.

This thesis formulates problems of data-driven QoS prediction for wireless networks, with a focus on the wireless channel, throughput, and latency. The main contribution is a set of prediction methods that address both dynamic environments and practical deployment constraints. A comprehensive empirical comparison of neural-network architectures for channel prediction provides guidance for selecting suitable models in mobile wireless settings. To reduce the effects of concept drift in pQoS models, the thesis introduces a distributed joint clustering and prediction framework that groups network cells and trains cluster-level predictors while keeping the number of models manageable. For user-side prediction in high-mobility scenarios, the thesis proposes geographical clustering combined with federated learning, enabling local adaptation while respecting privacy and communication constraints. These prediction frameworks are developed by iterative approximate solvers with convergence guarantees to improve pQoS accuracy. The first algorithm is evaluated using a network digital twin (NDT) simulation tool presented in this thesis. The thesis also presents an NDT framework that predicts the current achievable user throughput based on the network state.

Overall, the results show that combining clustering, distributed learning, and realistic system modeling can substantially improve the robustness of QoS prediction in challenging wireless environments. More generally, the thesis provides methods that can support dependable communication for future wireless networks. Future research should develop a theoretical foundation for integrating uncertainty-aware and SLA-driven objectives into predictive models. This will enable research to create pQoS frameworks that jointly optimize prediction accuracy, risk mitigation for critical services, and adaptability in non-stationary wireless environments.

This project proposes the development of an AI-enhanced Facility Management Control System (FMCS) for industrial energy optimization, specifically targeting resource management in reflow ovens. Traditional FMCS solutions provide real-time monitoring of key utilities such as cooling water, electricity, and compressed air. The advancement in machine learning technologies enables the transition from passive monitoring to predictive control, facilitating proactive resource dispatching and cost reduction. The key challenge lies in building an accurate prediction model using a limited dataset. To address this, a deep learning-based FMCS was designed, utilizing a Long Short-Term Memory (LSTM) architecture tailored for time-series data. The model comprises 4-6 LSTM layers followed by a two-layer neural network with ReLU activation functions. The system specifically focuses on monitoring nitrogen consumption while maintaining oxygen concentration within acceptable limits, with the aim of identifying inefficiencies and reducing operational waste. The empirical results demonstrated that the model could help save approximately 60% of the nitrogen cost, demonstrating the feasibility of the solution in production.

Application-Specific Integrated Circuits (ASICs) offer superior performance and energy efficiency compared to traditional processors, but at significant manufacturing costs. The SiLago framework addresses that through a new design methodology, enabling simple composition by abutment of pre-designed blocks. Within this framework, Dynamically Reconfigurable Resource Array version 2 (DRRA-2) provides a Coarse Grain Reconfigurable Architecture (CGRA) for streaming applications. It defines a structure of cells, each containing resources, that are abutted to form arbitrary systems. This thesis explores the application of Design For Testing (DFT) techniques and Built-In Self-Test (BIST) for that hierarchical architecture, with a focus on the Data Path Unit (DPU) resource. DFT techniques are integrated during the design process to strengthen the manufacturability of large designs. BIST is particularly useful to provide testing capacities once a chip is installed on a circuit. This work presents the implementation of a BIST architecture for the DPU using standard Electronic Design Automation (EDA) tools. The BIST builds upon scan chain insertion and is based on the STUMPS architecture (Self- Testing Using MISR and Parallel SRSG). We evaluate both an automatically generated and a custom BIST to explore different design choices. Then we investigate architectural considerations such as scan compression, reusing existing pins, and sharing BIST hardware across identical blocks. Finally, design optimisations are applied to improve the BIST’s fault coverage within the DPU. Experimental results show that the best BIST configuration for the DPU obtains a fault coverage ranging from 71% to 78%, depending on the design. For this, the total area overhead is 22% to 35% of the DPU’s area. However, the implemented BIST requires special hardware to control the primary ports of the DPU. It degrades the maximum frequency from 975 to 852 MHz. To avoid this frequency reduction cost, alternative approaches to control the primary ports should be studied. For tests without BIST, scan compression mixed with fullscan provides a 20% reduction of test time for a similar fault coverage than fullscan only.

Understanding and modelling user enjoyment in Virtual Reality (VR) is an important challenge for the design of adaptive and engaging interactive systems. Existing enjoyment computational methods are often limited to using text and audio data. These methods seldom capture the fleeting emotional nuances inherent in immersive VR interactions. To overcome this limitation, this work explores whether VR-native behavioural signals can be integrated into an existing enjoyment detection framework to improve predictive performance. Using the built-in sensors of the Meta Quest Pro, this study collected real-time multimodal behavioural data, including gaze trajectories, upper-body posture, and facial expression signals. The study investigates how incorporating facial expression signals improves multimodal fusion prediction performance. Using the Meta Quest Pro’s built-in sensors, the study collected synchronised gaze, facial expression, and upper-body posture data during gameplay. To validate the framework’s efficacy, a multiplayer collaborative VR version of the board game Pandemic was implemented. This successfully established a high-quality dataset comprising synchronized behavioral data and self-reported enjoyment scores from 20 participants. To address the challenges of frequency and density heterogeneity in multimodal data, this study systematically evaluated three preprocessing strategies: raw data baseline, time-window, and peak-interval filtering. Experimental results revealed significant differences in modality sensitivity to preprocessing strategies: facial expression data demonstrated the strongest predictive power after extracting high-intensity emotional bursts with peak-interval filtering, whereas audio and text modalities performed optimally under the contextpreserving time-window strategy. Furthermore, a comparative analysis of fusion architectures revealed that a prediction-layer-based late fusion strategy substantially outperformed early fusion and single-modal baseline models. This approach enabled the integration of complementary information from different modalities at the prediction level, demonstrating strong statistical significance. These results suggest that facial expression signals provide complementary information to speech and text, and can contribute to more accurate enjoyment prediction in collaborative VR environments.

The rapid growth of network complexity and the increasing demand for flexible infrastructure present significant challenges for automated resource allocation in industrial network systems. Traditional rule-based methods or single-solution optimizers often struggle to manage the combinatorial explosion of device configurations and lack support for dynamic changes or multiple feasible options. This study aims to model the network equipment combination problem as a constraint satisfaction problem (CSP) and address it using state-of-the-art constraint programming techniques. We concentrate on generating all valid, resource-optimal combinations of potential network nodes while minimizing the utilization of additional shared devices. We implement a CSP model utilizing Google OR-Tools’ constraint programming SAT solver (CP-SAT) solver and propose an incremental graph decomposition strategy that partitions the global problem into subgraphs based on connectivity. Each subproblem is solved independently, with their solutions subsequently merged through an efficient deduplication mechanism. Additionally, a mixed integer programming (MIP) version is implemented for comparative analysis. Experimental results indicate that our CP-SAT approach significantly outperforms the MIP model in terms of scalability and solution diversity, particularly in large-scale network scenarios. The method can efficiently enumerate hundreds of optimal configurations within seconds. This work contributes a practical, scalable, and dynamic CSP-based optimization framework for network configuration, which can be extended to other domains involving complex resource allocation and combinatorial design.

Most progress in humanoid control is driven by experiments on proprietary, closed-source systems. However, closed-source systems hamper communitydriven development, make it difficult to modify or extend the hardware, and often lead to costly repairs that can cause critical delays. To address these issues, this thesis presents an open-source humanoid robot and demonstrates its capabilities on motion tracking and locomotion tasks. Our humanoid robot is compact, low-cost, and its 3D-printed components can be modified or extended with additional sensors or manipulation tools. We create an accurate simulation model of the humanoid robot, including identifying actuator parameters, and use it to train policies for motion tracking and locomotion in simulation. For motion tracking, the robot follows a reference motion, either from a prerecorded motion sequence or in real time from pose estimates derived from webcam frames, whereas for locomotion, it follows a velocity command. For both tasks, the trained policies are transferred to the real robot without further fine-tuning, demonstrating the accuracy of the simulation model. We evaluate variants of motion-tracking policies and compare their performance. Furthermore, we quantitatively evaluate sim-to-real transfer for locomotion and observe a sim-to-real gap; yet the trained policy remains robust to this gap and successfully tracks velocity commands on the real robot.

Object detection is a research area within computer vision that includes both localizing and classifying objects in images. Its utility spans various domains, ranging from autonomous driving to finding anomalies in medical radiology. However, a challenge across these fields that has intensified as datasets grow in scale is the requirement for accurately labeled datasets. Traditionally, the identification and correction of labeling errors have relied on labor-intensive manual review by human annotators, a method which is expensive and difficult to scale. The goal of this thesis is to develop a framework that leverages a Vision-Language Model (VLM) to automate the task of finding and correcting object detection labeling errors present in real-world datasets, thereby improving model performance. This research addresses a gap in the field, as previous studies have focused on cleaning synthetic datasets with VLMs, overlooking the complexities of noisy labels found in real-world data. In this work a two-stage cleaning pipeline is applied to a real Oriented Bounding Box (OBB) dataset focused on identifying license plates and company logos on vehicles. First, a task model flags potential error proposals. Secondly, these proposals are validated by a VLM (Qwen 2.5 VL). The study evaluates multiple strategies, comparing simple binary classification to Chainof- Thought (CoT) reasoning and the effects of fine-tuning. The results demonstrate that while binary prompting and classification performed best, fine-tuning increased cleaning performance for all settings. Consequently, the fine-tuned binary VLM was selected to produce the final cleaned training datasets. The results show that an object detector trained on the uncleaned and Zero-Shot cleaned dataset performs similarly, while the fine-tuned VLM-cleaned dataset significantly improved performance. These findings suggest that fine-tuned VLMs can be effectively used to clean real object detection datasets, providing a scalable option to improve model accuracy.

Metal artifacts (MA) in cone beam computed tomography (CBCT) images are caused by X-Rays utilized in the imaging process being irregularly absorbed by metallic objects in a patient’s body. The artifacts can resemble dark areas or light streaks in the image, obscuring the underlying tissue required for accurate analysis or treatment of a patient. Metal artifact reduction (MAR) is the process of removing these artifacts through inpainting or interpolation. In this thesis, we implement two deep learning methods based on denoising diffusion probabilistic models (DDPMs) for the purposes of MAR: RePaint, which utilizes an unconditional DDPM, and Palette, which is a conditional DDPM trained specifically for inpainting. In addition to these two DDPMs, we also test a simple U-Net architecture. While previous research has been performed testing DDPMs for the purposes of MAR to promising results, little has been done to investigate whether these methods are time-efficient enough to be implemented in near real-time settings where computational times are an important factor. We compare the aforementioned deep learning methods with state of the art linear interpolation MAR methods to determine which, if any, of the deep learning methods could feasibly be implemented in real pipelines without a considerable increase in computation times as compared to interpolation methods. We compare the methods with respect to both quantitative metrics, such as structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and execution times, as well as perform a qualitative comparison of the resulting processed images. We observe that both the simpler U-Net-based method and Palette show significant improvement in MAR quality compared to the linear methods while performing comparably in terms of execution times, with U-Net performing slightly better than Palette in both aspects. We did not get satisfactory results for the RePaint method, likely due to needing a deeper network to fully train an unconditional DDPM. We can reject the method due to its considerably higher computation time. These results suggest that relatively simple U-Nets can be trained and implemented into reconstruction pipelines for processing CBCT images in order to improve MAR quality while maintaining similar processing times as linear methods.

Phishing is one of the most prevalent forms of cybercrime and continues to pose a growing threat as digital devices and services become more widespread and cyberattacks become increasingly sophisticated. Although organizations often address phishing risks through structured training among their employees, similar educational efforts aimed at the public are less common. This creates a gap in cyberthreat awareness and digital resilience. The purpose of this thesis is to investigate how phishing awareness can be strengthened among Swedish citizens by evaluating the feasibility and perceived usability of a game-based educational solution. A comparative study was conducted with 29 participants divided into two groups, who completed a pre-test and a post-test to measure phishing detection performance in terms of accuracy and recall. User perceptions were also collected through a survey with Likert-scale items and open-ended questions. The results show that participants who used the game-based method demonstrated a higher development of recall across all age categories, with the strongest effect observed among participants aged 18-29. The gamebased approach was also associated with higher engagement, motivation, and perceived learning. However, the text-based method was still perceived as more time-efficient. In conclusion, the findings suggest that the developed prototype can serve as an alternative to existing phishing awareness initiatives. However, further development of the game design are required to improve usability and achieve greater acceptance.

While Large Language Models (LLMs) have shown exceptional performance on natural-language tasks, they struggle with domain-specialized queries. This thesis investigates the effectiveness of Domain-Adaptive Continuous Pretraining (DAP) for enhancing cybersecurity awareness of three opensource pretrained LLMs—Llama-3.1-8B, DeepSeek-R1-Distill-Qwen-14B, and Llama-3.3-70B—on a relatively small domain-specific corpus (1M, 50M, 118.8M tokens). The adapted models are evaluated against their base counterparts and a cybersecurity LLM baseline, Llama-Primus-Base (8B parameters, 2.77B tokens). Across three benchmarks—CTI-MCQ, CyberMetric, and SecEval—the DAP models outperformed base models and Llama-Primus-Base, with the 70B model demonstrating better results than the open-source baseline models. These results indicate that DAP can enhance LLMs’ cybersecurity understanding with a small dataset size and no Supervised Fine-Tuning (SFT)/Reinforcement Learning with Human Feedback (RLHF).

This master’s thesis investigates and compares control strategies for the stabilisation and reference tracking of a two-wheeled inverted pendulum robot. The study focuses on three model-based approaches: linear quadratic regulator (LQR), linear model predictive control (MPC), and constraint tightening MPC (CTMPC), alongside a data-driven neural network-based controller (NNC) designed for realtime implementation. The evaluation is divided into simulation and experimental phases. The simulation phase consists of two parts: trajectory simulations under nominal initial conditions and estimation of the region of attraction (RoA) using a Monte Carlo approach. To this end, we employ a certified terminal set computed directly from the nonlinear dynamics, which guarantees invariance for the closed-loop system and allows stopping simulations once the set is reached. Validation on hardware is carried out for the LQR and NNC controllers through stabilisation and reference tracking tasks, while MPC and CTMPC are not feasible for real-time implementation on the embedded platform. This limitation motivates the adoption of the NNC, which approximates the behaviour of CTMPC while remaining suitable for deployment on hardware. In simulation, all controllers perform comparably under nominal conditions. The estimated RoA is identical for the model-based controllers, whereas the NNC exhibits a smaller RoA. In hardware experiments under uncertainties, however, the NNC outperforms the LQR, benefiting from being trained on the robust CTMPC policy. The study demonstrates that while conventional model-based controllers achieve strong nominal performance, robustness to uncertainties can be improved through robust formulations and their learned approximations. The NNC offers a computationally efficient way to imitate CTMPC control behaviour on embedded hardware, enabling improved performance over the LQR under real-world conditions without requiring online optimisation.

Photomasks are essential templates that enable accurate pattern transfer during photolithography. With the continuous downscaling of device geometries, photomask fabrication demands strict control over etching processes to minimize critical dimension (CD) variation. Dry etching, particularly ICP-RIE, provides anisotropy, precision, and process flexibility, but is sensitive to pattern density and plasma conditions. This work investigates dry-etch optimization for various photomask pattern densities, on chromium photomasks by varying Cl2/O2/He composition and O2 partial pressure. Results demonstrate strong correlations between gas flow composition, plasma density, and local pattern loading, offering insights into recipe development for improved uniformity and CD control in photomask manufacturing. Notably, lower-density patterns showed the largest improvement under the optimized settings.

Efficient wood defect detection is increasingly critical for the sustainable forestry industry to ensure material quality and maximize resource utilization. However, traditional manual inspection is time-consuming and subjective, while state-of-the-art deep learning models are often too computationally intensive to operate on the resource-constrained edge devices required for field applications. This thesis addresses the challenge of enabling accurate, realtime wood defect detection on lightweight hardware by optimizing the tradeoff between model complexity and detection performance. This problem is significant because existing lightweight solutions often degrade accuracy unacceptably when applied to fine-grained industrial tasks, leaving a gap between theoretical architecture design and practical deployment needs. The project tackles the difficulty of reducing computational cost (FLOPs) without sacrificing the ability to detect subtle defects like cracks and knots. The research methodology followed a two-phase experimental design based on the YOLOv8n baseline. Phase I focused on structural lightweighting, systematically evaluating convolutional variants such as GhostConv, GSConv, and RepConv to identify the optimal backbone-neck configuration. Phase II introduced accuracy-enhancement modules, including SE, ECA, and SimAM, to recover representational power lost during compression. The study identified a hybrid ”GhostGS” architecture combined with the parameter-free SimAM attention module as the optimal solution. This final configuration achieved approximately a 45% reduction in parameters and a 42% reduction in FLOPs compared to the YOLOv8n baseline, while retaining over 94% of its detection accuracy. These results demonstrate that a targeted combination of architectural simplification and attention mechanisms can yield a highly competitive performance-efficiency balance. This work enables the deployment of reliable automated inspection systems on low-power embedded devices, facilitating scalable and sustainable wood processing workflows that were previously hindered by hardware limitations.

In military and industrial operations, Life Cycle Management (LCM) plays a critical role in ensuring that deployed systems maintain operational effectiveness throughout their entire lifespan. In cases where systems such as aircraft and vehicles must operate in remote locations away from the normal supply chain, maintenance and repair capabilities depend entirely on a preconfigured set of resources, known as a ”fly-away kit”, that is brought along during deployment. Traditionally, Systecon tackles the optimization of fly-away kits using the simulation tool SIMLOX. The tool Opus Evo iteratively searches for the optimal kit configuration and evaluates them using SIMLOX. The objective is to reduce computational cost by finding suitable surrogate models to approximate SIMLOX outputs, while optimizing fly-away kit configurations to maximize the quality of these outputs. Four surrogates with lower computational cost: Gaussian Process (GP), Multi-Layer Perceptron (MLP), Random Forest (RF) and Support Vector Regression (SVR), were evaluated for their ability to approximate SIMLOX and support the optimization of kit configurations. The surrogates were integrated into the EVO environment, and trained through online active learning. SIMLOX was not replaced completely and was required during the active learning training cycle. The cases studied were two kinds of training data sizes defined by sliding window sizes (fixed datasets from the most recent iterations) and three varying numbers of kit items, resulting in six distinct test-case scenarios. All surrogates reduced runtime cost compared to SIMLOX, except for GP for test case with 218 items and sliding window size 10. SVR showed overall best runtime for all test cases. For 50 item count, SIMLOX performed the best for the optimization performance metric (MoE), and SVR among the second best. The overall best surrogate for higher item counts was GP for optimization performance. To understand how well the surrogates perform in SIMLOX, two approximation performance metrics (MSE and R²) were used. Although RF achieved the best approximation performance for both metrics among the surrogates, this did not translate to better optimization performance. This reveals that approximation accuracy does not necessarily reflect better kit configuration optimization. Despite bad R², the surrogates performed well in MoE, and supported kit configuration optimization.

Large Language Models (LLMs) such as ChatGPT have rapidly become part of students’ everyday learning practices in higher education, particularly within STEM disciplines. While these tools can support programming, problem-solving, and writing, their convenience also introduces risks of overreliance, uncritical acceptance of automatically generated outputs, and cognitive offloading that may undermine self-regulated learning (SRL) — the ability to plan, monitor, and reflect on one’s learning. Despite growing awareness of these challenges, few existing AI tools are explicitly designed to foster SRL during students’ everyday interactions with commercial LLMs. This thesis addresses this gap by introducing ChatWise, a browser extension developed to strengthen students’ SRL when using generative AI tools. Instead of restricting the AI system itself, ChatWise focuses on helping students engage with ChatGPT more reflectively and strategically. The tool provides adaptive, metacognitive feedback in real time, classifying each prompt using AI-based real-time classification grounded in SRL theory to the SRL strategies elaboration, effort regulation, organisation, rehearsal, help seeking, and critical thinking. Through immediate visual and textual feedback, ChatWise encourages students to reflect on their prompting behaviour and rephrase prompts in ways that better support learning. Following a Design-Based Research (DBR) approach, ChatWise was iteratively designed and evaluated through a within-subject study conducted in an authentic learning setting, involving 17 STEM undergraduates, of whom 9 completed both study phases. Quantitative analyses of prompt logs and survey data showed an increase in high-quality, SRL-aligned prompts when students used ChatWise, with substantial variation in how individual students’ prompting strategies developed over time. Qualitative findings revealed increased reflection, greater awareness of productive prompting strategies, and improved engagement with learning tasks. Participants described the adaptive SRL feedback as motivating and helpful for increased understanding of how to interact with ChatGPT more intentionally for improved learning in STEM higher education. These first results suggest that adaptive metacognitive feedback can foster more reflective and self-regulated engagement when interacting with generative AI during studies. This work contributes both practical and theoretical insights into how educational AI tools can promote SRL and metacognitive awareness. The thesis concludes with design recommendations for supporting reflective, sustainable, student-centred use of generative AI in higher education.

Artificial intelligence (AI) has seen a big increase in popularity due to the improvement and general availability of large language models (LLMs) such as ChatGPT and Claude. LLMs have granted people without programming experience the ability to generate complete web applications with a single prompt. When an LLM creates an entire web application for a person who cannot understand the code generated, it is critical that the person knows what to expect of the application’s state of security. This study attempts to provide that knowledge by analysing AI-generated web applications in terms of common web application security issues, assessing the prevalence and severity of discovered vulnerabilities. Due to the hasty advancement of AI, there is currently a deficit of studies analysing the security of AI-generated web applications, a deficit that this study attempts to reduce. The LLMs ChatGPT-5 and Claude Sonnet 4.5 were queried for complete web applications in several isolated conversations. A standardised ”prompting script” was created and used for each conversation, ensuring reproducibility across the web application generations. Furthermore, it ensured that the prompts were written as if by a person without any programming experience asking for a complete web application solution. For each generated web application, a vulnerability assessment was made using a custom suite of automated scanners focusing on the top three vulnerabilities of the OWASP Top 10 (2021), summarising the prevalence and severity of discovered vulnerabilities. The results showed that web applications created by non-technical prompting of LLMs exhibit multiple recurring vulnerabilities categorised as A01- Broken Access Control and A03-Injection. Furthermore, the majority of the vulnerabilities have CVSS scores of the Medium or High category. The reliability of the results would be improved by analysing additional web applications and performing manual penetration testing for verifying the output of the automated scanners, making the study a good stepping stone to further research. This study provides a basis for advising people without programming experience to be wary of the demonstrated risks of having an LLM create complete web application solutions.

Industrial vacuum conveying systems are critical for bulk material handling, yet current monitoring relies on threshold-based rules that cannot detect subtle anomalies. This thesis proposes a two-stage pipeline combining unsupervised deep learning for anomaly detection with interpretable machine learning for root cause characterization. In Stage 1, two reconstruction-based deep learning models (a State Space Model (SSM) (KambaAD) and an attention-based transformer (Anomaly Transformer)) are evaluated on real vacuum conveying data. Results show that including operational context (machine settings) significantly improves detection, with both models achieving Area Under Receiver Operating Characteristics (Curve) (AUROC) above 0.96 when context is available. In Stage 2, fault-specific signatures derived from SHapley Additive exPlanations (SHAP) analysis enable a rank-based characterization approach that achieves 96% top-3 accuracy, allowing operators to quickly identify likely fault types. Key contributions include: (i) a curated industrial dataset covering 30 configurations and 6 fault categories, (ii) an empirical comparison of transformer and SSM architectures for industrial Time Series Anomaly Detection (TSAD), and (iii) a complete two-stage pipeline designed for industrial deployment where interpretability is essential.

The introduction of the European Union’s second Payment Services Directive (PSD2), which mandates secure, customer-consented access to bank account and transaction data, has enabled access to detailed, transaction-level financial data. This development creates new opportunities for behavioural customer segmentation based on observed financial activity rather than static demographic attributes. However, such data are high-dimensional, heterogeneous, and largely unlabeled, posing significant challenges for traditional clustering methods in terms of robustness, interpretability, and stability. This thesis evaluates the suitability of topology-preserving and deep representation learning approaches for behavioural customer segmentation in an open banking context. Using anonymised and categorised transactional data from nearly 10,000 individuals, four unsupervised segmentation pipelines are compared: direct constrained K-means clustering, a Self-Organizing Map (SOM)-based approach, a sequential autoencoder (AE) followed by SOM, and a topologyregularised AE–SOM architecture inspired by the Deep Embedded Self- Organizing Map (DESOM) framework. All pipelines are evaluated under identical conditions using internal clustering metrics, stability analysis, and qualitative interpretability through visualisation and cluster profiling. The results demonstrate a clear performance hierarchy, with the DESOMbased representation learning approach consistently achieving the most compact, well-separated, and stable clusters. While standalone SOMs provide strong visual interpretability in the original feature space, deep representation learning significantly improves structural cluster quality, albeit with some loss of feature-level transparency. Overall, the findings indicate that hybrid models combining autoencoders and Self-Organizing Maps are viable and effective methodologies for behavioural segmentation of PSD2-enabled transactional data, with trade-offs between interpretability and representational power that should be carefully considered in practical applications.

Technology production is a complex process that requires the accumulation of resources and in which very little is done individually. Given the fragmentation of tech development and the limits of individual action, how much can tech workers affect the tech they produce? Do they care about the tech they are building and its implications for the world? Who should act to ensure that technology is “ethical”? What are the possibilities for individual and collective action that exist in tech workplaces? These are the questions I aim to answer in my dissertation. 

This dissertation is an interdisciplinary project connecting four studies with the methods from design research, ethnomethodology and sociology. The papers are published within the fields of human-computer interaction (HCI) and computer supported cooperative work (CSCW) and positioned within the tradition of studying ethics in practice, which approaches ethics by studying the work practices in specific tech organisations. Theoretically, my work is grounded in the philosophical tradition of ethics of care with its focus on concrete situated actions instead of omnirelevant rules, in the ethnomethodological tradition of respecifying theoretical phenomena as matters of practice, and in moral anthropology with its focus on ethics as shaped by communities.   

The four papers combined in the thesis explore the ethical reasoning of tech practitioners in academic settings, government officials working with AI in Sweden, and technology practitioners working in commercial companies in Europe. I rely on a variety of qualitative data: (1) a video-ethnography of design ideation sessions at the university; (2) an interview-study of ethical responsibility by tech practitioners in academia, governmental sector and the industry; (3) a a workshop-based study of voicing discomfort (critique) in a design workshop with academic; (4) an interview-based study of ethics discourses in tech companies in Europe. The studies contribute to design and HCI with an empirical research of ethical reasoning in everyday design work.  

The thesis aims to advance an anthropological, human-centred take on ethics in HCI. I argue in favour of including affect and emotion of tech practitioners into the discussion on ethics in HCI and emphasise the limitations of an individualistic take on ethics. Affect is important for connecting the level of everyday practices with structural power relations. Understanding the limitations of individual action and focusing on the potential for collective action instead is important for balancing responsibility with power. This thesis also draws attention to the inequality within the tech sector as a factor that hinders the possibility for forming a shared agenda and engaging in collective action. 

With the advancement of automotive intelligence and autonomous driving technologies, seat comfort and interactivity have become critical factors shaping the driving and riding experience. However, current seat design mainly focuses on static comfort, with limited attention paid to dynamic driving conditions—particularly during turns, where centrifugal forces cause physical displacement, muscle fatigue, and motion sickness. This study investigates how dynamic, interactive seat design can improve comfort during vehicle turning through a Soma Design approach that uses body awareness as the foundation for interaction design. Rather than relying solely on quantitative ergonomic metrics, the design process draws on first-person somatic exploration to identify and refine experiential qualities that emerge from temporal coordination between seat dynamics and bodily perception. An initial exploration using a manually controlled inflatable seat prototype tested bodily responses across multiple driving scenarios—acceleration, braking, cornering, and autonomous mode transitions. Turning scenarios exhibited the most prominent discomfort and demonstrated the clearest improvement potential through dynamic support. Based on these insights, the hardware was upgraded to a digitally programmable pneumatic system. The seat contains four inflatable silicone actuators on both left and right sides—one in the backrest and one in the seat cushion. A series of interaction workshops tested variations in inflation timing, airflow rate, vehicle speed, and airbag positioning across right-angle turns and roundabouts. Results show that temporal coordination between support delivery and bodily anticipation significantly improved comfort. Specifically, at turning speeds between 15–25 km/h, participants consistently reported high comfort when inflation began 1.5–2 seconds before turn entry, compared to without support. The study found that seat interaction effectiveness lies not in static support strength but in what we term “somaesthetic coordination”—temporal alignment with the body’s changing states during turns. This finding establishes a foundation for intelligent seat control systems that prioritize temporal synchronization over force magnitude. This research contributes a Soma-based methodology for optimizing dynamic seat interaction pattern during vehicle turns, extending Soma Design principles to automotive contexts. The findings provide design implications for intelligent seating systems in both conventional and autonomous vehicles, offering new directions to enhance comfort under dynamic driving conditions.

large language models (LLMs) are increasingly adopted across industries due to their strong performance on a variety of natural language processing (NLP) tasks. However, these models require significant computational resources during inference, resulting in high energy consumption and increased environmental Impact. As the demand for sustainable AI grows, optimizing the runtime efficiency of LLMs without compromising performance has become a key concern. In this work, we investigate the Impact of optimization techniques, such as Quantization and Pruning, on the energy efficiency and performance of LLMs during inference. We evaluate four models: GPT-2 1.5B, DeepSeek- R1 1.5B, DeepSeek-R1 7B, and Mistral 7B, across three NLP tasks: text completion (LAMBADA), sentiment classification, and binary question answering (BoolQ). We compare performance across four precision formats (FP32, FP16, INT8, and INT4) and apply structured Pruning to the 7B models. Our results show that FP16 achieves more than four times the energy saving of full-precision without compromising accuracy. INT8 and INT4 experienced small performance drops, though implementation limitations likely prevented full utilization of their efficiency potential. Pruning improved efficiency in Mistral 7B but degraded DeepSeek 7B, due to architectural differences. We also performed statistical tests to verify the significance of observed differences in accuracy and energy consumption. Our findings showcased how effective model optimization strategies can significantly reduce energy consumption without compromising performance, and suggest future work to further explore additional methods like fine-tuning or prompt design for further gains.

This study investigates how visual cue design affects task performance and perceived presence in a virtual reality (VR) environment. Two cue types (particle and highlight), three colors (red, blue, yellow), and two opacity levels (α = 1.0, α = 0.5) were investigated in a visual search task involving 20 participants. The results show that perceptual color contrast relative to the surrounding environment was the primary factor affecting both task efficiency and perceived presence. Fully opaque cues with a strong contrast resulted in faster completion times, higher accuracy, and stronger presence ratings, whereas low-contrast cues blended into the environment and reduced performance. Qualitative findings further revealed that participants prioritized functionality and visual clarity over realism in cue design, and that highly visible cues did not reduce their sense of presence. Together, these results suggest that perceptual contrast is a key factor for effective attention guidance in VR and can support both performance and presence in visual search tasks. The study is limited by the restricted set of cue variations, use of a single environment, and a relatively small sample size of participants. Future work should explore additional environments and cue designs to further explore how perceptual contrast affects performance and presence in VR.

In the field of education, enhancing student engagement remains a challenge. Research indicates that personalisation offers one solution, though this can require significant manual effort. Large language models (LLMs) provide a viable way to reduce this workload. At the same time, employing lightweight models is crucial for local deployability. Rewriting structured problems, e.g., coding problems, into a specified theme while preserving solvability remains an open problem, and this challenge is amplified when using lightweight models, which struggle to maintain the original constraints. We propose a pipeline that combines an 8B LLM with Retrieval-Augmented Generation (RAG). By leveraging contextual retrieval together with a retry mechanism, the pipeline reliably produces rewritten problems with valid formatting. Using prompt engineering and placeholders to protect LaTeX and code snippets, the structural constraints of the original tasks are preserved. Experiments demonstrate that, even with an 8B model, the rewritten problems retain high solvability while successfully shifting to the target theme. The resulting pipeline supports personalised themed problems for learners, reduces manual adaptation effort for educators, and provides researchers with a reproducible framework for studying structured rewriting using lightweight LLMs.

Dynamic contrast-enhanced MRI (DCE-MRI) is widely used for breast cancer diagnosis and treatment monitoring, but repeated injections of gadoliniumbased contrast agents (GBCAs) raise safety, and environmental concerns. This thesis investigates whether contrast-free, multi-sequence breast MRI can be used to synthesize multi-phase DCE-like images under realistic data constraints. We develop a conditional generative adversarial network (cGAN) based on Pix2PixHD that maps pre-contrast T1-weighted images, diffusionweighted imaging (DWI) at multiple b-values, and apparent diffusion coefficient (ADC) maps to several post-contrast phases in a single forward pass. To emphasise enhancement dynamics rather than absolute intensities, we introduce a multi-scale subtraction consistency (MSSC) loss that compares pre-/post-contrast subtraction maps across multiple image scales.

Experiments on public breast MRI datasets show that the proposed MSSC model with full multi-sequence input improves standard reconstruction metrics (PSNR, SSIM, LPIPS, lesion-wise RMSE) compared with strong Pix2PixHD baselines. Lesion-level enhancement curves and qualitative assessments further indicate better preservation of tumour morphology and parenchymal texture with fewer artifacts. These results suggest that, given limited public data, a contrast-free, multi-sequence cGAN with scale-aware subtraction supervision can synthesize multi-phase DCE-like breast MRI that is competitive with real DCE-MRI while more faithfully capturing lesion-level enhancement behaviour.

Over the past few decades, code reuse attacks have shown how malicious actors can alter a program’s intended execution flow by taking advantage of benign code already present in the application. Class Pollution in the Python programming language is a novel variant of a code reuse attack, which can enable a malicious party to surgically mutate a variable in any part of the application in order to trigger a change in its execution flow. However, until now, little to no research has explored class pollution in detail, and no tool is readily-available to detect it. For this reason, as part of this degree project, a literature review on the causes and consequences of class pollution has been conducted, in addition to the methodical development of a tool capable of detecting class pollution, Classa. Additionally, an empirical study on the prevalence of class pollution in realworld Python code has been performed by running Classa against a dataset of 3000 Python projects, revealing, most notably, a critical vulnerability in a popular PyPI package with more than 30 million downloads. This vulnerability allowed for Denial of Service and Remote Code Execution, having since been responsibly disclosed and patched. Altogether, the results revealed that while not many real-world Python projects are susceptible to class pollution, it is a vulnerability that must be accounted for when building a secure application due to the serious consequences it can lead to.

Future wireless networks face growing energy demands due to dense infrastructure and computationally intensive processing. Cell-Free Massive Multiple-Input Multiple-Output (CF-mMIMO) has been recognized as a promising architecture to enhance user fairness and spectral efficiency through joint transmission among distributed Access Pointss (APs). However, largescale deployment of CF-MIMO with optical fronthaul introduces high endto- end power consumption across radio, fronthaul, and cloud domains. More recently, Open Radio Access Network (O-RAN) allows controlling radio and cloud resources jointly, allowing potential energy savings. To reduce the energy consumption of CF-mMIMO, this thesis considers CF-mMIMO on top of O-RAN architecture, and proposes a Reinforcement Learning (RL)-based framework for energy-efficient joint radio, fronthaul, and processing resource allocation in centralized CF-MIMO networks. This study first models the end-to-end power consumption of centralized CF-mMIMO, demonstrating that the power consumption in both cloud and radio dominantly depends on active antenna ports and the number of active radio units. Then, Proximal Policy Optimization (PPO) algorithm is designed to learn near-real-time control policies at the near-real-time RAN Intelligent Controller (near-RT RIC), dynamically adjusting active antennas based on long-term channel characteristics to minimize total network power while maintaining users’ Spectral Efficiency (SE) requirements. To enhance the performance, a greedy post-processing stage is developed to further prune redundant activations without compromising service quality. Simulation results show that the proposed RL-based controller achieves over 50% total power reduction compared to the full activation scheme, and around 15% reduction over the heuristic baseline. Specifically, by integrating a greedy strategy with the RL-based control, the total power consumption can be further reduced by an additional 50%. Moreover, the trained model performs well to different user densities and SE targets, demonstrating high robustness and scalability. Its inference time is at the millisecond (ms) level, offering a significant speed advantage over traditional optimization methods. The joint orchestration of radio and processing resources reduces the O-Cloud power consumption by 60%, demonstrating the significant energy-saving potential of the proposed end-to-end control framework.

Platform engineering teams in large enterprises face growing operational challenges due to the increasing scale and complexity of cloud-native systems. A significant portion of their workload consists of repetitive workflows, such as repeatedly searching logs, consulting documentation, and updating incident tickets. These tasks consume valuable engineering time, fragment knowledge across multiple systems, and delay efficient problem resolution. This thesis addresses this problem by designing and implementing an AI-driven chatbot that automates repetitive operational workflows through a conversational interface. The system is built on a LangGraph-based orchestration framework that coordinates a master-agent workflow, routing queries to specialized tools: Elasticsearch for log retrieval, Qdrant as a vector database for semantic documentation search, and ServiceNow for incident management. A continuous embedding pipeline, based on HuggingFace models, ensures that the knowledge base is regularly updated with documentation, logs, and historical incident data. The contribution of this work lies in providing a unified, context-aware interface for platform engineers that not only retrieves information but also executes repetitive workflows automatically. By integrating knowledge management with workflow automation, the system reduces the need for manual effort in operational tasks and supports more efficient platform operations. The solution is evaluated through deployment in an enterprise case study at Volvo Cars, demonstrating its potential to improve workflow efficiency, reduce repetitive work, and enhance the overall effectiveness of platform engineering teams.

The decarbonisation of European industry coincides with rising shares of variable renewable generation and increasing electricity prices, making industrial load flexibility a critical resource for cost-efficient, low-carbon operation. Yet many planning tools still treat demandside management in coarse, system-level terms and cannot represent the plant-specific operating rules, inter dependencies, and validity windows that determine what flexibility is actually realizable in factories. This thesis proposes a simulation-based planning framework that embeds a fine-grained model of Energy Flexibility Measures (EFMs) into an opensource environment. The formulation captures bidirectional load changes, hold and clean-up phases, recovery and minimum activation windows, cooldown dynamics, one-to-one pairings, and group-level exclusivity at 15-minute resolution, and couples them with dynamic tariffs, peak-power charges, and on-site assets such as photovoltaic (PV) system and battery storage. The framework is implemented as a modular Mixed Integer Linear Programming (MILP) component and evaluated for two real industrial systems: an oven-dominated diecasting plant and a site-infrastructure-oriented system with ventilation, electric vehicle (EV) charging station, PV system, and battery storage. For each system, multiyear simulations (2019–2024) compare baseline operation with and without EFMs under four tariff designs: static prices, static prices with peak shaving, day-ahead prices, and day-ahead prices with peak shaving. The results show that explicit EFM modeling systematically reshapes load toward low-price periods, reduces grid peak demand, and yields consistent grid-cost savings and, in most years, CO₂ reductions while respecting process and staffing constraints. Batteries are found to complement rather than replace demand response, primarily shifting midday PV surplus into higher-value evening hours, and peak-shaving charges can be reduced without increasing annual consumption. Overall, the thesis contributes a transparent, Energy Flexibility Data Model (EFDM) inspired EFM module and data layer, alongside a reproducible 15-minute planning workflow. It calls on industrial planners and tool developers to integrate structured, plant-level demand-side flexibility into energy system planning and tariff design so that future factories can co-optimize production and energy use rather than treating flexibility as an afterthought.

The goal of the project presented in this report was to implement the photons as aprojectile in the Liège Intranuclear Cascade model (INCL), elaborated to simulate nuclearreactions. The energy of interest ranges from a few MeV up to 300 MeV. For this purpose,it is essential to understand the structure of the model, how it works and all its finedetails. The complexity is to describe the interactions that take place when the incidentparticle interacts with the nucleus. Two main mechanisms need to be taken into accountfor the realisation of the project. At low energies (below 30 MeV), the Giant DipoleResonance is supposed to be dominant. Going up in energy, a smooth transition towardsthe Quasi-Deuteron model must be implemented. This report is also dedicated to thevalidation of the photo-nuclear reaction model through comparisons with experimentaldata available on databases. Globally, the model predictions are consistent with theexperimental data. Looking at specific data, some discrepancies appear. One of themencouraged the implementation of the Delta production mechanism for photons withenergies higher than 150 MeV while others are probably due to missing mechanisms inthe photo-nuclear reaction model. The INCL predictions are also analysed in light ofexperimental data on 209Bi obtained via laser-based accelerators by collaborators fromthe University of Valencia, Spain.

Denna rapport presenterar resultaten från forskningsprojektet "Sårbarheter och hotanalys från sidokanaler med maskininlärning i fokus", finansierat av Myndigheten för civilt försvar (MCF) Projektet har genomförts i samarbete mellan KTH Royal Institute of Technology och Lunds universitet.

Projektets övergripande ambition har varit att öka kunskapen om fysiska attacker mot digitala enheter, med särskilt fokus på sidokanalsattacker förstärkta med maskininlärningstekniker. Genom att systematiskt studera nya attackvektorer, utveckla metoder för sårbarhetsanalys och utvärdera motåtgärder har projektet syftat till att bidra till ett mer proaktivt arbete med hårdvarusäkerhet.MCF:s stöd har varit avgörande för att möjliggöra långsiktig och explorativ forskning inom detta strategiskt viktiga område. Vi vill särskilt tacka Erik Sundström, som varit projektets kontaktperson och handledare vid MCF.

Vi hoppas att resultaten som presenteras i denna rapport ska vara till nytta inte bara för forskarsamhället utan även för beslutsfattare, systemutvecklare, utvärderare och andra aktörer som arbetar med att skydda samhällsviktiga funktioner.

Detta examensarbete omfattar 15 hp och har genomförts inom programmet Industriell teknik och produktionsunderhåll vid KTH Södertälje i samarbete med Coca-Cola Europacific Partners i Jordbro under höstterminen 2025. Studien fokuserar på utmatningen i produktionen, där färdig dryck grupperas och plastas in innan de skickas in till lagret för att sedan levereras till kund. De undersökta delarna av anläggningen omfattar inplastningsrobotarna samt pallbanan mellan inmatningen och palletiketterarna.

Syftet med arbetet var att analysera orsaker till driftstörningar i utmatningen samt att ta fram förbättringsförslag som kan bidra till ökad driftsäkerhet och minskat antal oplanerade driftstopp. För att kartlägga nuläget genomfördes semistrukturerade intervjuer och Gemba Walks. Drift och haveridata samlades in från SAP, medan trenddata och information om maskinfel hämtades från LineView. Det insamlade materialet analyserades med hjälp av metoder såsom Paretoanalys, Ishikawa, 5 Varför, FMEA och en förenklad RCM.

Resultaten visade att utmatningen är ett område där underhållsarbetet är mindre utvecklat än i andra delar av produktionen, trots att återkommande störningar och maskinfel påverkar tillgängligheten. Arbetet resulterade i förslag på förbättringsåtgärder och en rekommenderad underhållsstrategi med fokus på att förebygga återkommande fel, förbättra uppföljningen av störningar och skapa bättre förutsättningar för ett mer systematiskt underhållsarbete.

Today, there are several uses for satellites, one of them is communication. Having the ability send audio, video, or other types of data through electromagnetic waves, people all around the globe can communicate with each other. This kind of technology enables a lot of possibilities, and there exist both stationary and mobile communication. Ovzon AB is a Satellite Communication company which works with both these solutions. This Master thesis will be focusing on their SATCOM on the Move, SOTM, solution. It uses sensor fusion to combine multiple sensor data in order to track the antenna mounted on the vehicle. This technology has been around for a while, but the sensors in the system are still susceptible to disturbances. The scope of the project is to compare different sensor combinations through Unscented Kalman filter against the common combination INS and GNSS. The results indicate that there are differences between the sensors. However, the result also shows that there were some outliers depending on which sensor combination that was used. Overall, the conclusion of the thesis is that there were two groups which improved the antenna performance, the GNSS + Gyroscope and GNSS + Gyroscope + Accelerometer.

Moreover, this project has given more insight into design choices as well as optimization alternatives.

The increase in the number of simultaneous proton-proton collisions, pile-up interactions, will be a major challenge for the future High-Luminosity phase of the Large Hadron Collider (LHC) at CERN, making pile-up mitigation an important issue for the ATLAS experiment. The High-Granularity Timing Detector (HGTD), to be installed in the upcoming years, is one of the detector upgrades contributing to this pile-up mitigation. To report the current status of the HGTD development, an HGTD test-beam analysis was performed using data taken during two weeks in October 2025 at the CERN SPS test-beam facility. Synchronization between different detector systems in the test-beam environment was achieved. The full data reconstruction chain was demonstrated to be feasible although the measured hit efficiency and timing resolution were limited by non-optimal chip tuning including a very high threshold setting, as well as by low statistics. To study the effect of pile-up, the impact of jet pile-up contamination on the jet flavor-tagger was investigated under current LHC running conditions. A substantial degradation in flavor-tagging performance was seen, highlighting the importance of understanding pile-up effects in jet reconstruction. A supplementary timing-based pile-up mitigation study was also performed.

Railway infrastructure requires a more circular use of materials, but much research still focuses on single metrics and assumes that higher recycling rates are always better. This paper constructs a simple and consistent framework for evaluating recycling strategies for three railway assets (sleepers, rails, and overhead contact lines). The framework covers four dimensions: material utilization, greenhouse gas emissions, energy consumption, and financial costs. It follows the EN 15804 life cycle structure (Modules A-D) and links all metrics to the basic material flow. Recycling rate (R) and reuse or maintenance rate (U) are used as the primary decision variables.To avoid unrealistic results at extremely high recycling or reuse rates, the model introduces penalty functions. These penalties represent additional effort and losses, such as more complex dismantling, lower waste quality, or higher processing energy consumption. All metrics are standardized and can be combined through a simple weighted scheme, allowing for comparison of different strategies on the same scale.Case studies show that the optimal range for R and U largely depends on the asset type. For railway sleepers, moderate steel recycling and low reuse rates already offer significant benefits in terms of material utilization and emissions, while pushing recycling rates close to their maximum leads to higher costs and fewer additional benefits. Rails can more easily achieve higher recycling rates due to well-established scrap and rerolled steel routes. Overhead contact wires are primarily made of copper, a material that is highly recyclable in principle. However, in practice their effective recycling rate remains narrow because dismantling, contamination, and alloy degradation significantly constrain recoverability. As a result, low recycling rates fail to capture substitution benefits, while excessively high rates increase energy use and costs.Overall, the results indicate that higher recycling rates are not always better in practice. Recycling targets should be set based on asset type, rather than a single value for the entire railway system. This framework provides a practical approach to testing different recycling and reuse levels and supports procurement and maintenance decisions that balance material utilization, climate impact, energy demand, and costs.

This study presents the development and validation of a fast, iterative design workflow for a new class of fully electric long-range delivery drones that combine simultaneous vertical take-off and landing (VTOL) capability with high cruise efficiency in a fixed-wing configuration. With a growing presence of unmanned aerial vehicles (UAVs) in today's world, efficient and reliable design methodologies become increasingly important to decrease costs and accelerate development, especially in startup environments. The proposed workflow integrates multi-objective optimization of the drone's geometry with an aerodynamic analysis of increased fidelity level through computational tools. Rapid potential flow solvers are employed for preliminary design, followed by computational fluid dynamics (CFD) simulations of greater accuracy for further refinement, advanced analysis and validation. To test the speed and modularity of the workflow, two different UAV geometries are designed and optimized with the potential solver VSPAERO, then further analyzed with advanced CFD models in Ansys Fluent. The results obtained demonstrate that this approach significantly reduces design time and usage of computational resources compared to CFD-only methods, while maintaining accuracy and scalability for further iterations.

A major challenge in developing connected and automated vehicles~(CAVs) for urban environments is achieving a comprehensive understanding of the surrounding traffic scene. This relies on situational awareness, defined as the ability to perceive, interpret, and anticipate the behavior of surrounding road-users, which is essential to ensure safety. In particular, unprotected road-users, such as pedestrians and cyclists, are often occluded or located in sensor blind-spots of the CAV, which remains a critical challenge. This thesis aims to improve the situational awareness of the ego-vehicle, the CAV of primary interest, in urban environments by leveraging vehicle-to-everything (V2X) communication to incorporate information from connected road-users. A framework using set-based methods is developed to systematically handle uncertainties in measurements and initial conditions of detected pedestrians.

The objective is to address several key challenges that arise in real-world scenarios, including data inconsistency, data association, pedestrian motion prediction, and efficient reduction of redundant information. The thesis first proposes a shared situational awareness framework for occluded pedestrian-crossing scenario to compute an estimated set for the pedestrian. The framework is extended to handle measurements from V2X units that may be inconsistent with the ground truth of the detected pedestrian. To address scenarios involving multiple occluded pedestrians, a data association method based on intersection-over-union heuristics is introduced. Pedestrian motion prediction is further studied using both a data-driven approach and a bounded velocity–acceleration model applied to the estimated set obtained from the framework. An occlusion-aware extension is also developed to handle situations where occlusions affect both the ego-vehicle and V2X units by exploiting previously observed measurements. Finally, a method for selecting and filtering relevant information from multiple V2X units is proposed to reduce the computational load while maintaining effectiveness. The proposed methods are validated through numerical simulations and real-world experiments using Scania prototype automated vehicles.