[dipl] Perustieteiden korkeakoulu / SCI

Permanent URI for this collectionhttps://aaltodoc.aalto.fi/handle/123456789/21

Browse

Recent Submissions

Now showing 1 - 20 of 5612
  • Formal correctness proof of a mapping from a software specification language to the relational algebra
    (2025-05-24) Dargent, Stan
    School of Science | Master's thesis
    This thesis explores the connection between predicate logic and relational algebra, two foundational pillars of database theory. We introduce a mapping from logical formulae to relational algebra and prove the semantic equivalence between the two domains. This correspondence creates opportunities for query optimization and enables the use of database queries via a declarative approach, arguably more readable and less error-prone. The logical formulae introduced are part of our language, which is essentially the predicate logic but without function symbols. On the other hand, the relational algebra formalized in this project is the standard one. Additional features are also presented in this work. In particular, we extend our language, we formalize a type system for our relations and language – and the correspondence proven between these two domains can also be established under our type system. While still not widely used, theorem provers are getting increasingly important in computer science where software needs to be reliable. Hence, all results and proofs from this project are rigorously formalized using the Isabelle theorem prover, ensuring that our proofs are free from human error.
  • Internet delivered therapy for insomnia: clinicians experiences and views
    (2025-05-22) Leskinen, Alisa
    School of Science | Master's thesis
    Insomnia is a common condition affecting people’s life all over the world. Efficient treatment for insomnia helps to reduce the burden on healthcare systems. Recommended first-line treatment for insomnia is cognitive behavioral therapy for insomnia (CBT-I), that offer longer-lasting results than medication. However, limited number of trained professionals and treatment costs have accelerated the development of digital CBT-I. This thesis investigates the experiences and views of Finnish clinicians treating sleep problems about HUS Helsinki University Hospital iCBT-I program. This thesis was conducted as qualitative research using semi-structured interviews with specialists treating sleep problems as a primary data. Thematic analysis was used to obtain answers for the two research questions focusing on the role of iCBT-I in insomnia treatment and its benefits and drawbacks. The results emphasized that iCBT-I is considered as an effective treatment, particularly for individuals who are motivated for the treatment modality, self-sufficient and technologically proficient. Insomnia having variety of comorbidities, iCBT-I may not be suitable for all patients, especially those with severe mental health challenges or difficulties using digital tools. The most important benefits of iCBT-I that were found were accessibility, scalability, and cost-effectiveness for individuals. The study identified challenges such as limited therapist interaction and higher dropout rates should be considered when implementing this treatment model.
  • The role of user experience and cognitive efforts in ERP perfor- mance
    (2025-05-22) Pekkinen, Peter
    School of Science | Master's thesis
    To fill a gap found in the ERP literature, aspects of user experience and cognitive efforts were studied to assess their effects on ERP performance for users. This is done by using a hierarchical linear regression model to analyze survey results from a higher education organization. It was found that lower work difficulty for system usage both improved task performance and decreased stress for users. Physical effort and hours worked with the system per week were also found to help with task performance, but they were also found to increase stress. This implies that while work difficulty should be minimized for ERP systems, there was found to be a need for a balanced approach for physical efforts and time spent with the system. Thus, it was found that to increase performance ERP systems should be easy to use for users without trivializing and automating all everyday tasks.
  • Fabless Business Model in the Battery Cell Industry
    (2025-05-20) Wärnman, Isak
    School of Science | Master's thesis
    Customers of battery cells are increasingly engaging with companies that innovate on cell components or their integration, rather than just cell makers, as they seek access to next-generation cell performance. These technology companies often have no established in-house cell manufacturing capability, which is costly, time-consuming, and risky to develop. This makes it challenging to deliver cells to prospective customers, slowing the diffusion of new technologies. A fabless business model, where technology companies outsource manufacturing to contract cell makers, offers a potential solution. Using a design science approach, this thesis investigates whether a fabless cell maker model is technically feasible in the battery cell industry, and under what conditions it is attractive to technology companies, cell makers, and customers. Based on a literature review, a set of hypotheses was developed to capture key feasibility conditions and stakeholder motivations. These were tested and iteratively refined through 25 semi-structured expert interviews with stakeholders across the battery value chain. The research identifies conditions under which the fabless model is feasible and the capabilities required to adopt a fabless model. The results suggest that the fabless model is feasible under four conditions: (1) the fabless company offers strong performance differentiation to offset added risk perceived by customers and cell makers; (2) the fabless company can develop manufacturable cell designs through in-house expertise or close collaboration with cell makers; (3) sufficient volume for a single cell design can be aggregated to enable cost-effective production; (4) the fabless company is prepared to accept product liability risk for imposed cell design, chemistry, and supply chain choices. The study contributes to theory by exploring how modular value chain literature extends to the battery industry. For practitioners, the findings offer a framework for evaluating when fabless arrangements are viable, and provide guidance on how firms might position themselves in relation to fabless companies. Future research could quantitatively validate the findings, conduct case studies to understand how fabless arrangements are implemented in practice, and explore other modular arrangements, such as separating electrode production from cell assembly and finishing.
  • The Diagnosis of Market Maturity in Disruptive Industries: A Case Market Study on the Direct Air Capture Market
    (2025-05-20) Juurakko, Kustaa
    School of Science | Master's thesis
    Abstract Accurately assessing market maturity is essential for organizations aiming to navigate or invest in disruptive industries, where traditional indicators are often missing or underdeveloped. This thesis explores the concept of market maturity in such contexts by developing and applying a stakeholder-based framework to evaluate the maturity of the emerging Direct Air Capture (DAC) market. While DAC serves as the empirical case, the primary contribution of this thesis lies in extending market maturity assessment as a strategic tool for decision-making under uncertainty. The framework developed in this thesis integrates insights from market timing theory, maturity models, and stakeholder analysis based on literature and 12 interviews with professionals working within disruptive industries. Framework defines five key dimensions of market maturity: (1) technological and ecosystem readiness as a base for maturity analysis, (2) profitability as a key maturity indicator, (3) competitive landscape and growth potential, (4) low-risk cash flow mechanisms, and (5) market creation through early customer engagement. Additionally, 17 expert interviews, with DAC technology providers, investors, policymakers, and corporate customers, were conducted to shape the understanding from DAC ecosystem. The findings indicate that the DAC ecosystem is still in an early stage of development, primarily limited by high capture costs and lack of regulatory recognition for “negative” emission credits. However, there are clear signs of progress. These include emerging long-term offtake contracts, a growing consensus around solid sorbent technologies, and increasing awareness of DAC’s role in decarbonizing hard-to-abate sectors. In the Finnish context, colder climates and the availability of BECCS create additional challenges for the breakthrough of DAC. The thesis offers managerial implications related to collaboration with technology providers and infrastructure operators, customer-aligned financial analysis, and strategic partnerships to accelerate market entry. Overall, the thesis contributes to the literature on market maturity in disruptive contexts and provides a practical framework for organizations seeking to evaluate and engage with immature markets in a structured, stakeholder-informed manner.
  • The Role of a Procurement Sustainability Data Platform in Facilitating Supply Chain Emission Reduction Through Supplier Collaboration: a Case Study on Procurement Analytics Software Company
    (2025-05-22) Nyyssönen, Viliina
    School of Science | Master's thesis
    A significant portion of companies carbon footprint lie in the supply chain, and reducing scope three emissions is gaining importance. Companies aim to minimize their environmental impacts trough sustainable supply chain management practices and collaboration is recognized as a key component to drive emission reduction in the supply chain. The aim of this study is to understand how a software platform can facilitate collaboration for emission reduction and how buyer-supplier collaboration can support emission reduction in supply chains. Research is conducted as a qualitative case study for the collaborative partner. In total ten semi- structured interviews with procurement, sustainability and product management professionals were conducted including five interviews with the case company’s customers and five interviews within the case company. The results showed that importance of scope three emission reduction and supplier collaboration is recognized in procurement teams, however lack of primary emission data from suppliers hinders the adoption of collaboration practices. Collaboration opportunities start from supplier selection and qualification process, and extend to supplier relationship management practices, development programs for key suppliers and joint development projects. Communication, knowledge sharing and commitment are highlighted. While supply chain emissions are of key importance, this research showed that companies typically analyse suppliers’ sustainability more holistically from ESG perspective. Data platform should solve the complexities of emission data management and turn the data into clear practical insights and recommendations for procurement teams not equipped with broad sustainability knowledge. Functionalities of the software should support efficient data collection in vast quantities and multiple data sources, data analysis for different user types and provide clear explanations, insights and decision-making support as well as reporting capabilities on implemented emission reduction projects.
  • Levelling up continuous improvement through organizational change management
    (2025-05-26) Immonen, Veera
    School of Science | Master's thesis
    Continuous improvement is a crucial for the organizations to stay relevant. As new business models, markets, technologies, customer demands, and employee expectations increase, its importance for organizations grows. However, continuous improvement perceptions, culture, leadership and culture, and systems need to be updated to secure the fit to the modern organizations and environments. This thesis examines continuous improvement in a case company with complex and modern operations. The study is conducted as a qualitative single case study. I assess the current situation through the lenses of continuous improvement understanding, culture & leadership, and systems. I identify aspects that hinder the implementation and success of continuous improvement. These aspects form a basis for a roadmap. Based on further analysis on the aspects, I provide practical suggestions for the development roadmap. To level up the state of continuous improvement, organization should create and implement a collective understating on it. There should be one narrative, branded attractively. Organization must understand and commit to it. As a next step, the organization should strengthen the continuous improvement leadership and tune in the systems. There should be a sense of responsibility and clear responsibilities around continuous improvement. Systems should be clear, collaborative, pragmatic and fit to the purposes. They require promotion and training. Furthermore, organization must enhance culture around feedback. Cross-functional collaboration needs to be facilitated by improving understanding on other functions and getting people to know each other across the value chains. On the other hand, collaboration should be a process, not personal dependent phenomenon. The organization should also improve project and portfolio planning and management. Projects should be initialized systematically and prioritized on a structured way. Portfolio needs careful management, and personal workloads need to be in control. Finally, the organization should demand similar practices and aligned objectives. The study contributes to the continuous improvement and change management literatures, linking these themes together. It provides insights on how continuous improvement is perceived, and how it needs to be considered as an organizational transformation, focused on mindset and culture.
  • Value Perception and Pricing Strategies for B2B Generative AI Solutions
    (2025-05-25) Löfqvist, Oscar
    School of Science | Master's thesis
    The rapid adoption of generative AI is reshaping how B2B firms perceive, create, and seize value. Despite heavy investments, many companies struggle to translate generative AI solutions into sustainable revenue and profits. Despite the central role of pricing in commercial success, little guidance exists for monetising generative AI-driven offerings. This research examines how companies evaluate the value of generative AI solutions and translate that value into effective pricing strategies. Drawing on 31 semi-structured interviews with buyers and suppliers indicates that the highest perceived value of generative AI solutions comes from automating tasks and time savings. Although initial market enthusiasm may temporarily elevate willingness to pay, respondents anticipate that generative AI capabilities will ultimately become commoditised and become an expectation rather than a value-adding feature, reducing price premiums. Moreover, interviews reveal a shift from conventional software licensing toward value-based pricing. However, it is constrained by persistent challenges in measuring and quantifying generative AI, and respondents remain divergent in their perception of inference cost. These insights were synthesised into a double-dimensional pricing matrix. The framework outlines four pricing strategies: value-based, usage-based, traditional SaaS, and credit-based pricing strategies. It maps them against specific cost considerations of large-language-model powered generative AI solutions and customer-perceived value. This gives managers a practical tool to align pricing decisions with product characteristics, market expectations, and commercial viability. This thesis bridges a critical gap in modern pricing by integrating customer-perceived value and strategic pricing theory with the realities of generative AI. It contributes to existing research by expanding it into the generative AI area and proposes a practical framework for managers to price generative AI offerings. In doing so, it advances academic understanding of value dynamics in the emerging area of generative AI. It equips managers with a framework to choose and adapt pricing models that protect margins and capture the value of generative AI.
  • From Obligation to Advantage: Approaching Regulatory Compliance as a Strategic Driver in EU Medical Software Companies
    (2025-05-24) Hänninen, Marcus
    School of Science | Master's thesis
    The complexity of the European Union’s Medical Device Regulation (MDR, 2017/745) and In Vitro Diagnostic Regulation (IVDR, 2017/746) poses consequential strategic and operational challenges for companies developing software-based medical devices. This thesis investigates how such companies navigate, adapt to, and strategically leverage regulatory compliance prerequisites within the European market. Based on ten semi-structured interviews with managers operating at the intersection of regulatory compliance and strategic decision-making, the study employs an inductive analysis using the Gioia method to explore the relationship between regulatory obligations, organisational strategy, and product development practices. The research develops an integrative framework capturing four core dimensions of regulatory adaptation: Dynamic Compliance Foundations, Embedding Compliance in Development & Organisation, Acceleration and Modularity in Compliance, and Strategic Shaping of Product and Market by Regulation. The findings indicate that compliance is not merely a legal hurdle but a strategic force, influencing product architecture, internal processes, and market positioning. Companies that proactively align regulatory foresight with development cycles, modularise their compliance scope, and embed cross-functional responsibility for quality management are better positioned to maintain agility and leverage certification as a competitive advantage. In addition to its theoretical contribution, the thesis presents a framework for practitioners that translates the theoretical insights into a more actionable model for managing regulatory strategy across the product lifecycle.
  • How to measure value in the emergency department? Developing an outcome-based ED performance measurement system
    (2025-05-13) Hagelstam, Julius
    School of Science | Master's thesis
    The landscape of emergency department (ED) care is evolving as the number of patients treated in EDs and the associated cost continue to rise. In this context, making informed decisions about how emergency care is organized requires data about the effects of care in relation to costs. However, current ED performance measurement systems focus on inputs and process efficiency rather than care effectiveness. This thesis aims to address this shortcoming by focusing on what ultimately should guide the decisions of both clinicians and policymakers: value, defined as patient-relevant health outcomes achieved in relation to their costs. To meet the challenges of measuring value in the unique context of ED care, this thesis employs a qualitative interview study posed to elaborate existing theory in value-based healthcare (VBHC) and ED performance measurement. The data for the study was collected from 16 expert interviews and structurally analyzed using the Gioia method of analysis, complemented by a thematic analysis to examine how patient-relevant outcome measures may be implemented in this setting. The findings suggest that implementing VBHC in the ED requires an enhanced focus on the acutely and severely ill, which in turn necessitates a clearer definition of the role of ED services. The shift from process efficiency toward value also requires advancements in three key areas: 1) data collection, 2) the culture of performance improvement, and 3) supporting IT infrastructure. Additionally, the study suggests that measuring patient-relevant outcomes in the ED setting requires segmentation of patients based on two key variables: 1) health problem severity or urgency and 2) the patient’s outset service need or functional ability. Based on this proposed segmentation logic and a synthesis of expert perspectives, the findings include an outcome-based performance measurement system tailored to the ED setting.
  • Examining the causal relationship between the neglected post-acquisition integration process and constrained capability in exploring strategic options
    (2025-05-26) Merilampi, Jesse
    School of Science | Master's thesis
    This thesis explores how a neglected post-acquisition integration process can raise negative emotional reactions and resistance towards the acquiring organization’s practices inside the acquired company and affect the acquired organization’s capability of exploring and exploiting its strategic options. A qualitative case study approach is applied to examine the phenomena in a Finnish subsidiary of a global technology corporation. Personnel are interviewed from all levels of the acquired organization and the interview results are coded employing the Gioia methodology. The results show that the neglected post-acquisition integration process is critically important to leverage the benefits of the acquisition. In the examined case, the acquiring corporation coerced the acquired organization to implement practices, which conflicted its existing practices and organizational culture. The conflict between these practices and the decreased autonomy created tensions inside the case organization, which materialized as a negative emotional reaction and resistance towards adopting the new ways of working. Consequently, resistance towards the new practices have eroded the case organization’s capability of renewing its portfolio and processes.
  • The role of analyst sentiment in excess return forecasting, and its implications for investors
    (2025-04-25) Kohvakka, Eetu
    School of Science | Master's thesis
    This study evaluates the role of analyst sentiment in trading, and its implications for traders. The research problem of this study is to gauge if active investing with machine learning is worth the effort, how analyst sentiment features in the data affects neural network machine learning model accuracy and portfolio performance, how it compares to passive index investing, and to whom it is suited for. Neural network machine learning models are trained with and without features that are based on the consensus of subjective financial analyst estimates. The neural networks are built similarly to those presented in risk premium forecasting literature from the past few years, which have been proven to offer significant economic gains to investors employing them. To highlight the importance of data quality, the experiment is repeated with two different datasets. The top decile by predicted excess returns makes the long portfolio, and the performance of the long portfolio is compared to the S&P500 index. As the focus is on alpha values, the returns used in the data, in predictions, and for the S&P500 are excess returns over the 10-year constant maturity bond yield. The data for this study is sourced from Wharton Research Data Services (WRDS), using tables from Compustat, CRSP and I/B/E/S. The long horizon excess returns of the best machine learning portfolios can double the excess returns of the S&P500 index. When controlling for the presence of analyst features, significant changes can be seen in the portfolio performances. Removing the analyst features from the data generally improves the out of sample performance of the portfolio. The required break-even trading costs are 0.15 USD at the highest, which limits the feasibility of this strategy. In theory with fractional shares, a retail investor can accomplish this. Even in perfect conditions, an individual investor is better of investing in the S&P500 index. For those investors with better capabilities of managing large portfolios and a more frictionless access to the market, these strategies are sound.
  • Molecular dynamics simulations of dynamic crack propagation in brittle materials
    (2025-05-07) Karhulahti, Gabriella
    School of Science | Master's thesis
    In this thesis, molecular dynamics is used to simulate crack propagation in brittle materials. The purpose of the study is to investigate the suitability of classical interatomic potentials for describing atomic fracture mechanisms. Three materials (silicon, iron, and nickel) have been studied that have different crystal structures, allowing for a comparative analysis. In addition, two potentials per material were analyzed to assess potential discrepancies. The study is limited to materials exposed to tensile stress caused by dynamic loading. Pristine crystal structures containing a predefined seed crack were used for the simulations. Griffith's theory has been used to calculate the theoretical values for the critical load for each material. The theory is limited to elastic materials that fracture in a brittle manner, meaning that no plastic deformation will occur before the fracture. The theory is formulated using continuum mechanics which does not take certain atomistic fracture mechanisms into account. Two of these mechanisms are lattice trapping and dislocation emissions, which will prevent propagation to some extent and cause a higher critical load. Previous fracture studies on the same materials and crystal orientations have exhibited brittle fractures. Most of the simulations in this study showed typical characteristics of ductile fractures. The simulation of a crack in iron along the (110) crystal axis, in particular, exhibited dislocation emission. Although most of the materials did not undergo ideal brittle fractures, plastic deformation was limited and the simulated critical loads were close to the theoretical values. Additionally, a fracture study on hexagonal ice was attempted, but was not feasible due to long simulation times and incompatibility of the ice structure and potential. This study has shown that classical potentials are capable, to some extent, of capturing crack propagation in brittle materials. Although the simulated critical load values are consistently higher than theoretical predictions. The fracture behaviour is consistent with previous experimental and computational studies.
  • Incident Cause Classification in Insurance claims using Generative AI
    (2025-05-20) Uzair, Mohammad
    School of Science | Master's thesis
    Automation plays a critical role in modern insurance operations, enabling companies like OP-Pohjola to process claims more rapidly, reduce manual workloads, and improve customer satisfaction. However, claim automation rates are often reduced when essential categorical data are left unspecified during claim notification. Customers frequently struggle to choose from predefined incident categories, submitting instead free-text descriptions of their incidents. These unstructured inputs disrupt downstream automation workflows, forcing claims into manual handling, even when they would otherwise qualify for automated processing. This thesis explores the potential of large-scale generative language models, particularly GPT-4o, to address this challenge. The study introduces a solution that classifies missing incident cause categories by analyzing free-text descriptions and other structured claim data. The system is planned to be integrated as an API within OP-Pohjola’s internal claims handling platform, where it will be automatically triggered when claims lack the necessary categorization. By mapping user-inputted text to predefined category labels, the solution restores automation eligibility to such claims. The experimental results presented here are promising. GPT-4o demonstrates the ability to reliably and accurately infer incident categories from a wide range of diverse and informal claim descriptions without the need for domain-specific training or large annotated datasets. By fine-tuning prompts and restructuring the classification taxonomy, the system achieved remarkable improvements in classification accuracy, while maintaining impressively low false positive rates. The results also show a direct and measurable increase in simulated automation rates, streamlining claim processing and reducing manual intervention. This approach offers a timely solution to the growing demand for efficient and scalable automation in insurance. It bridges the gap between unstructured user input and the structured data requirements of automated workflows, providing a lightweight alternative to traditional machine learning methods. The findings not only demonstrate the viability of generative models in practical insurance applications but also underscore their broader potential to improve automation rates in various industries.
  • Temperature dependence of the superfluid weight and order parameters within mean-field and dynamical mean-field theory in flat band systems
    (2025-05-19) Aaltio, Anton
    School of Science | Master's thesis
    Superconductivity manifests in many different forms and its origin in an experimentally studied sample is not always clear. A method helping to distinguishing the nature of the superconductivity in a sample relates to the studying temperature dependencies of relevant quantities. In accordance with this method, the goal of this thesis is to study how the temperature dependencies of the superfluid weight and the order parameters in multiband systems with s-wave pairing are described by considered analytical models. The models used in this thesis include power law models, exponential models and extension of known single-band Bardeen-Cooper-Schrieffer theory results. We evaluate the models on numerical data obtained by using both mean-field and dynamical mean-field theory (DMFT). The considered lattices are the Lieb lattice alongside its modifications and the dice strained lattice. Other prominent aspects of the thesis are comparing the mean-field and DMFT results and studying the effects brought by including long-range hoppings to the lattices. In mean-field theory, the results show that for arbitrary flat band systems with s-wave pairing, the superfluid weight and order parameters do not have universal temperature dependencies among the considered models. The order parameters of the orbitals where the flat band states reside exhibit the closest we have to a universal description as they are almost always predicted accurately by the power law models with the optimal exponents between 3.15 and 3.24. For the superfluid weight and the order parameters of the orbitals where the flat band states vanish, increasing the band gap from a band touching is seen to increase the optimal exponents of the power law models often describing these quantities. The accuracies of the power law models for the superfluid weight also simultaneously decrease. When the band gap is at least moderate, for each of the two quantities, there exist very similar temperature dependencies in multiple lattices. As opposed to the mean-field results, DMFT shows all quantities at least roughly following the power law models and the exponents are clearly higher. Thoroughly considering the DMFT results is made difficult by the error related to determining the critical temperatures. We still find the band gap having a similar effect on the temperature dependencies as in mean-field theory.
  • Adaptive Preference Learning from Multi-User Feedback
    (2025-05-07) Le, Duong
    School of Science | Master's thesis
    We consider the problem of multi-user preference learning, where the model learns from multiple human users and uses the learned information to quickly adapt to new users. Previous attempts at developing such a system assumed that users who are connected on platforms like social media tend to like similar products. This assumption is unlikely to hold in practice because users usually know only a small fraction of their connections. Other works imposed a structure on user preferences using shared latent functions. This method is limited by the representation capabilities of the chosen latent functions. In this work, we proposed a novel probabilistic approach to multi-user preference learning by partitioning user preferences into two components: an (optional) shared knowledge between users and a set of individual parameters. We trained a model on each user in the training user base and aggregated these trained models to extract both the shared and individual components simultaneously. The extracted information is stored using a Bayesian Neural Network, which facilitates efficient adaptation to new users. During learning and adaptation, we employed active learning to query items to users for labeling. We tested this framework on both synthetic and real preference datasets, and also applied the method to an existing retrosynthesis scoring system. The results show that our model outperformed even personalized models trained on the users in terms of recommending high-rating items to users.
  • Bayesian Meta-Learning for Task Adaptation Using Expert-Inferred Task Similarities
    (2025-04-28) Mäkinen, Lotta
    School of Science | Master's thesis
    Meta-learning, the concept of learning-to-learn, refers to machine learning methods that leverage prior knowledge across tasks to enable faster and more efficient adaptation to new tasks. However, the success of meta-learning typically depends on the similarity between the source and target tasks, and its performance can deteriorate under distribution shift. This thesis proposes a novel adaptation pipeline for Bayesian meta-learning, focusing on enabling efficient task adaptation through expert-inferred task similarities. We build upon a hierarchical Bayesian meta-learning framework and introduce a human-in-the-loop approach, where expert feedback is used to infer task similarities and construct informative priors for new tasks. The proposed pipeline is evaluated on synthetic datasets, assessing both the quality of task similarity inference and the downstream adaptation performance. Our experiments show that task embeddings inferred from noisy expert feedback improve adaptation to new tasks, with performance approaching oracle-level adaptation under low expert noise. Furthermore, active querying strategies based on Bayesian experimental design achieve better inference quality with fewer expert queries compared to random querying. These results demonstrate the effectiveness of expert-driven prior construction for improving task adaptation in Bayesian meta-learning.
  • Cyber security training in office and warehouse — Gamifying the cyber security training of the entire workforce of a large international logistics company
    (2025-05-26) Oksanen, Laura
    School of Science | Master's thesis
    Gamification is a process of applying game elements in non-game contexts. In comparison to traditional trainings, gamified cyber security trainings are more engaging and motivating, and may also improve learning outcomes. This master thesis studies gamification of cyber security awareness training of employees from a perspective of large international logistics company. Prior studies have focused on cyber security trainings of office employees. This thesis attempts to fill this gap by including production employees in the target group and evaluation test participants. The goal of the thesis is to identify how differently office and production employees react to gamification of cyber security trainings, and whether same design can be used to fill needs of both groups. The research included a literature review, implementation of two gamified cyber security training solutions, and collection and analysis of player experiences. The first solution is a cyber security escape room, which is designed to suit both office and production employees. The escape room got positive feedback from two test groups. Results indicate that same gamified cyber security training implementation can be used to train production and office employees, even for mixed groups, assuming both groups have been considered during the design phase. The second implemented solution is gamification of annual training material. Production and office employees have their own versions of the game. Results of an end of the training survey are compared to results of the previous non-gamified version of the training. Additionally, opinions on the gamified elements are identified with a survey form. Survey form results of production and office employees were compared, but no statistically significant differences were identified. Although prior literature focuses on gamified cyber security trainings of office employees, the results of the thesis indicate that the literature may be also used as a base when designing gamified cyber security trainings for production employees. All experimental results of the thesis are based on small sample sizes and are thus only indicative. Further research with larger sample sizes is required to confirm the results.
  • Estimating wood pulp regional imports using AIS data
    (2025-05-08) Uusikumpu, Usva
    School of Science | Master's thesis
    This thesis explores the application of Automatic Identification System (AIS) data in estimating regional wood pulp imports, contributing to the growing field of AIS data analytics in maritime trade. While AIS data has been widely used in navigation, vessel tracking, and environmental monitoring, its potential for wood pulp trade estimation remains underexplored. Currently, the information business practitioners get from the imports suffers from delays and is aggregated on a high level, limiting its usefulness for real-time decision-making. This study proposes an alternative approach using AIS-based models to track vessel movements, identify wood pulp carriers, estimate cargo volumes, and predict arrival times. The methodologies used in this exploratory research include gathering potential methods from the previous literature, building two models to test how wood pulp imports could be estimated and using a correlation analysis to evaluate their performance against official imports. The models built focus on estimating eucalyptus pulp (BEKP) imports from Latin America to the European Union. The results indicate that AIS-based models have the potential to generate reliable import estimates and thus help in business decision-making. To unlock the full potential, the limitations of the current models should be reduced, for example, by making more effective use of vessel coordinate data. The findings also highlight how official import statistics cannot be used as a ground truth, making the evaluation of the model's accuracy difficult. This study advances the use of AIS data in import volume estimation by presenting a dedicated framework developed for this purpose. In future research, the framework and the methods discussed in this study could be applied to other bulk commodities or pulp grades, further strengthening the role of AIS data as a tool for business decision-making.
  • LLM-based code generation for optimized quantum circuits
    (2025-05-24) Jern, Linus
    School of Science | Master's thesis
    Modern Large Language Models (LLMs) routinely exceed their natural-language origins, matching or surpassing human performance in code, mathematics and multi-modal generation. Despite this, the potential of LLMs in quantum computing remains largely unexplored. An open challenge is generating quantum circuits on a large scale using LLMs and training LLMs on quantum-specific knowledge. This thesis explores whether the capabilities of LLMs extend to quantum circuit generation: can a fine-tuned LLM read a high-level description of an optimization problem and generate a high-quality parameterized quantum circuit? We introduce an end-to-end pipeline that (i) creates a novel supervision dataset, (ii) fine-tunes the LLM, and (iii) rigorously evaluates the resulting model. We present a data-generation framework that produces 14.000 optimized circuits (QAOA, VQE and Adaptive-VQE) for 12 graph-based optimization problems, expressed in platform-agnostic OpenQASM 3.0. A 3 billion parameter pre-trained base model is then fine-tuned on this dataset. We evaluate the model using three metrics: (i) the syntactic correctness of the generated QASM 3.0 code, (ii) the generated circuit’s expectation value compared to that of the optimized circuit in the dataset, and (iii) the Kullback-Leibler divergence between the generated circuit’s output probability distribution and that of the optimized circuits. Our evaluations show that our fine-tuned model, trained on our dataset, outperforms the state-of-the-art open-weight models on generating syntactically correct quantum circuits as well as generating parameters that are closer to the ground truth. These generated circuits could potentially be used as an initial point, called warm-start, for quantum optimization algorithms, such as QAOA and VQE, or as benchmarks for quantum hardware. Parts of this work appeared in a shorter collaborative paper; the author of this thesis implemented the data-generation framework, the fine-tuning pipeline and the evaluation suite.