[diss] Perustieteiden korkeakoulu / SCI

Permanent URI for this collectionhttps://aaltodoc.aalto.fi/handle/123456789/52

Browse

Recent Submissions

Now showing 1 - 20 of 1688
  • The effect of magnetic configuration on fast ions in stellarators
    (2026) Kontula, Joona
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-04-17
    Using nuclear fusion as a baseline electricity source in the future would greatly benefit from steadystate operation. Stellarator fusion reactors are suitable for this task as, unlike tokamaks, they forgo the need for plasma current drive, instead relying on the magnetic field coils alone for shaping the plasma equilibrium. The complex three-dimensional magnetic structure of stellarators makes charged particle orbits complex compared to axially symmetric devices. Assumptions of toroidal symmetry and related invariants cannot be applied and the magnetic field has to be computationally optimized for improved particle confinement. Computational optimization is also an opportunity to craft a wide operational space of magnetic field topologies, each of them with unique properties. Exploring this operational space nevertheless requires a sophisticated design and simulations for both performance and operational safety. One important factor for both safety and nuclear fusion capabilities in stellarators is the behavior of fast particles. Such ions are created with either neutral beam injection (NBI) or ion cyclotron resonance heating (ICRH). In this thesis, the behavior of fast ions in different magnetic configurations in stellarators was assessed using the Monte Carlo orbit-following code ASCOT. Two stellarators, Wendelstein 7-X and TJ-II, were included in the study. NBI power loads to the ICRH antenna in Wendelstein 7-X were predicted to be within safe operational limits. The magnetic configuration was predicted to determine both the power load magnitude and its distribution on the antenna. We also developed theoretical models and methods for simulating neutral beam current drive in stellarators using ASCOT. The model was validated against experimental results in TJ-II and found to accurately predict the current drive under stable operating conditions, although charge-exchange reactions were not considered in the assessment. Finally, we simulated neutron production during deuterium NBI operation in Wendelstein 7-X to study the feasibility of installing a scintillating fiber neutron detector. The fusion reactivity and neutron production were simulated with ASCOT and found to be mostly independent of the magnetic configuration and instead determined by the plasma temperature and density profiles. During operating conditions planned for deuterium operation, the neutron flux to the detector was predicted to be sufficient for time-resolved measurements, allowing for observations of the fusion reactivity during deuterium NBI operation. The methods developed in this thesis improve the accuracy of fast ion simulations and measurements in stellarators, allowing for better estimations of phenomena such as neutral beam current drive and NBI power loads.
  • Massively parallel algorithms for sparse graphs
    (2026) Latypov, Rustam
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-04-10
    Most real-life graphs are large and sparse, consisting of billions of nodes connected by billions of edges. Single-machine computation for such graphs is often slow and inefficient. The modern approach is to deploy a fleet of machines that work on a given graph problem in parallel. The Massively Parallel Computation (MPC) model captures the strengths and limitations of such parallel systems and is a widely accepted theoretical framework for analyzing modern parallel algorithms. This thesis considers fundamental graph problems for uniformly sparse graphs in the MPC model, focusing on memory-optimal solutions. Computation is performed on arbitrarily many machines simultaneously while using roughly the same amount of memory that is needed to store the input itself. We resolve open problems and advance the state of the art by developing novel deterministic algorithms for graph problems such as connectivity, vertex coloring, maximal matching, and maximal independent set. Our focus is on uniformly sparse graphs, technically known as lowarboricity graphs. These graph families form the setting for most known lower bounds in MPC, and designing memory-optimal algorithms for them is notoriously difficult. Algorithm complexity is measured in communication rounds and is typically expressed as a function of the total number of nodes n or the diameter of the graph D. We establish the first O(log D)-round algorithm for connectivity, as well as the first O(log log n)- and O(log D)-round algorithms for 3-coloring, maximal matching, and maximal independent set on forest graphs. We also show that strengthening the MPC model, by allowing adaptive queries to a distributed data store, enables a constant-round O(alpha)-coloring algorithm for graphs with bounded arboricity alpha. One major technical contribution of this thesis is a novel graph exploration technique called balanced exponentiation, on which many of our results rely. This technique has been developed into a parameterized, standalone tool that allows nodes to explore surrounding sparse regions of the graph without prior knowledge of their location, while incurring minimal memory overhead.
  • New stochastic and computational approaches in atom probe tomography
    (2026) Shaikh, Aslam
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-04-09
    Atom probe tomography (APT) provides quantitative three-dimensional compositional mapping of materials with near-atomic spatial resolution and parts-per-million chemical sensitivity, thereby enabling fundamental insights into material nanostructures and properties. However, despite these advantages, the physical processes governing field evaporation remain incompletely understood. Conventional reconstruction methods rely on geometric projection models that assume independent and deterministic evaporation events, overlooking the stochastic and correlated nature of ion emission that gives rise to detector and reconstruction artifacts. This thesis establishes a physics-based framework for understanding and modeling field evaporation as a stochastic, field-driven process whose dynamics are governed by the geometry and composition of atomic neighborhoods, local field fluctuations, and surface evolution. In this framework, field evaporation is treated not as an idealized geometric projection but as a coupled physical process shaped by stochastic perturbations, correlated ion emission, and surface dynamics such as diffusion and the roll-up mechanism. It begins with the quantitative study of stochastic perturbation mechanisms in single-element crystalline systems, advances to the characterization of correlated, avalanche-like emission dynamics in multicomponent materials, and culminates in the development of a probabilistic, Gaussian process regression (GPR)-based reconstruction methodology that can systematically capture the stochasticity in ion emission behavior. The three publications included in this thesis form a coherent progression toward this framework. In Publication I, stochastic perturbation mechanisms — initial velocity perturbations and roll-upmotion — were incorporated into the Robin–Rolland electrostatic model, reproducing experimentally observed detector maps and quantitatively linking perturbation energy distributions to the discrepancies observed between simulated and experimental detector maps with the help of the structural similarity index measure (SSIM) for single-crystalline systems such as Al and Ni. Publication II demonstrates that field evaporation in multicomponent systems is not a sequence of independent, Poissonian ion emissions, but a correlated, avalanche-like process with long-lived spatial and temporal memory. These cascades, within a single pulse, follow a Poisson process compounded with a subcritical branching process, and also exhibit Omori–Utsu-type waiting-time correlations over an extended pulse range (corresponding to 103 evaporation events), revealing the correlated nature of field evaporation in APT. These processes can be understood by examining the surface evolution in the simulations, where the evaporation front behaves as an interface moving through a heterogeneous medium and may become locally pinned by atoms with a high evaporation field. In Publication III, a GPR-based reconstruction framework was developed that is capable of learning the hidden correlations encoded in the detector data due to the stochastic nature of field evaporation.The new framework yielded significantly improved spatial resolution compared to the conventional wide-field-of-view geometric approach.
  • Causality, identifiability, and representation learning for machine learning with non-i.i.d. data
    (2026) Hızlı, Çağlar
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-04-02
    Deep learning-based systems have achieved striking milestones over the past decade, such as defeating the world champion in Go, predicting protein structures with high accuracy, and demonstrating advanced natural language reasoning. Their capabilities keep improving as compute, data and model size scale, captured by empirical scaling laws. However, these capabilities are typically evaluated on held-out test sets assumed independent and identically distributed (i.i.d.) with the training data. Hence, while scaling laws suggest continued improvements, they also reveal a potential limitation: progress mostly targets in-distribution generalization and relies on large i.i.d. datasets. This contrasts with animal intelligence, which excels at out-of-distribution generalization and adapts quickly to changing environments that yield highly non-i.i.d. data. To address these challenges, this thesis focuses on machine learning with non-i.i.d. data, both learning from it and generalizing to it. Within this theme, we develop systems that (i) handle distribution shifts and low-data regimes, (ii) learn stable processes across changing environments, and (iii) transfer skills between environments and tasks. These directions are pursued through the lenses of causality, identifiability, and object-centric representation learning, respectively. First, to handle distribution shifts and low-data regimes, we investigate how causality can reason about counterfactual futures and pasts under such shifts. We learn from and make predictions for non-i.i.d. data. In two healthcare applications, our goal is to estimate how a change in treatment policy affects patient outcome trajectories. We observe patient histories as time series under different treatment policies, while the policies themselves are unobserved. To address this, we jointly model treatments and outcomes in continuous time. On real-world surgery-meal-blood glucose time-series data, the model learns clinically meaningful predictions, providing insights into how surgery affects glucose profiles through the causal pathways of meal intake and other metabolic processes. Second, to learn stable processes across changing environments, we investigate how identifiability can exploit departures from the i.i.d. assumption. Our goal is to identify latent dynamical systems from non-i.i.d. data, in the form of high-dimensional observations across non-stationary environments. To the best of our knowledge, we propose the first theoretical framework that identifies the unknown latent transition function, in addition to the latent states, and we show that identifying latent dynamics improves future predictions. Third, to transfer skills between environments and tasks, we study how object-centric representation learning can yield transferable representations that are useful under shifts in data-generating mechanisms. We propose an object-centric self-distillation algorithm that uses object annotations as weak supervision. We pretrain our models on i.i.d. data such as ImageNet and evaluate transfer learning for classification and segmentation. Our method consistently improves performance on both in-distribution ImageNet test set and held-out transfer benchmarks, compared to strong baselines. Taken together, these findings highlight the importance of combining higher-level cognitive abilities with lower-level pattern-recognition skills in order to handle changing environments and non-i.i.d. data as animals do. Future intelligent systems will need to integrate the best of both worlds, ideally guided by a few high-level principles since such principles are more likely to generalize across diverse tasks and learning setups.
  • Fractional rheology and data-driven approaches to characterize viscoelastic materials
    (2026) Miranda Valdez, Isaac Y.
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-04-02
    Viscoelastic materials play a crucial role in various aspects of life, impacting everything from the mechanical properties of biological tissues to the textures of food products. Accurately characterizing these materials is essential for understanding their complex behavior, although it poses significant challenges due to their inherent time-dependent properties. This dissertation explores advanced methods for characterizing the rheological behavior of viscoelastic materials by integrating fractional calculus and Bayesian Optimization. By employing fractional calculus, this thesis presents mathematical frameworks that capture the intricate dynamics of viscoelasticity, thereby moving beyond traditional integer order models. Additionally, Bayesian Optimization is utilized to model complex materials, simplifying the parameter estimation process in viscoelastic modeling and advancing material design. The five articles included in this thesis demonstrate the practical application of these methodologies across various soft matter systems, highlighting their potential to accelerate material characterization processes and improve predictive accuracy. Publication I examines colloidal hydrogels through the perspective of fractional order viscoelasticity, establishing a connection between the material's microstructure and its macroscopic rheological response. Publication II identifies a research gap in fractional-order viscoelasticity: the absence of intuitive computational methods, which has hindered the widespread adoption of these models. This publication addresses the computational complexity of fractional order models and provides a Python package that enables easy and efficient computation. Publication III introduces the use of Gaussian Processes as a regressive method for modeling complex rheological behaviors. Furthermore, it demonstrates the predictive capability of Gaussian Processes, which allows for the generation of machine learning-based rheograms that align with experimentally measured ones. Publication IV presents Gaussian Processes within the framework of Bayesian Optimization, demonstrating how this approach can be used to infer viscoelastic parameters in curve-fitting routines, thereby enabling the fitting of models from Publication II with quantified uncertainty. Lastly, Publication V applies Bayesian Optimization to accelerate material research, using rheological properties as descriptors for optimization. In conclusion, this dissertation underscores the importance of advanced methods for characterizing viscoelastic materials, which are crucial to a diverse range of applications. The integration of fractional calculus and Bayesian Optimization not only enhances our understanding of viscoelastic behavior but also addresses the challenges posed by traditional modeling techniques. The innovative approaches presented across the five publications demonstrate their practical applicability and effectiveness in bridging the gap between theoretical frameworks and real-world material design.
  • Resolving functional connections between visually guided behavior and the retinal signals in dim light
    (2026) Martyniuk, Nataliia
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-03-27
    All visual information sent from the eye to the brain is encoded in spike trains by retinal ganglion cells (RGCs), but how signals from different RGC types are used to drive visually guided behavior remains largely unclear. In the first two studies of this thesis, I investigate the relationship between responses of specific RGC types and behavior at the sensitivity limit of vision, taking advantage of this maximally simplified situation in which detection is mediated solely by a small proportion of extremely sensitive RGCs. We show that ON and OFF alpha RGCs are the most sensitive RGC types, the former increasing, the latter decreasing firing in response to light increments. Using a strain of transgenic mice in which rod single-photon responses are reduced in amplitude, it was possible to amplify the difference between the less sensitive but more reliable ON alpha RGCs and the more sensitive but noisier OFF alpha RGCs. For a putative behavioral correlate, we measured the detection performance of swimming mice guided to an invisible escape ramp in one of six arms of a water maze by either a light spot in darkness (Paper I) or a dark spot, a “shadow”, on a weak background (Paper II). We found that behavioral sensitivity for detecting shadows was consistent with the high sensitivity of OFF alpha RGCs, whereas for detecting light increments, behavior relied solely on ON alpha RGCs. Alternatively, it may be hypothesized that behavior uses a decoding strategy based solely on increases in spiking. Paper II delves further into the neural limits of shadow detection in dim light. Reconstruction of the stimulus trajectory on the retinal mosaic of OFF alpha RGCs during swimming (based on markerless head and eye tracking) and modelling that accounted for known sources of noise and loss of single-photon signals showed that detection performance approached the theoretical limits. In Paper III, the focus is shifted to population coding and its structural basis. To overcome limitations of existing recording methods, we developed a four-electrode patch clamp system integrated with two-photon imaging, enabling simultaneous measurement of excitatory input currents, spike outputs, and dendritic morphology from up to four ON alpha RGCs. We demonstrated this technology in two key ways. First, we found that while physiologically measured receptive field overlaps could predict the strength of pairwise noise correlations in spike outputs, incorporating detailed morphological dendritic structure and input current recordings provided a much more accurate estimate of pairwise noise correlation strength. Further, dendritic reconstructions revealed that cells sample visual space more uniformly than expected by chance. This demonstrates that understanding fine-scale morphological organization is critical for explaining population coding. Second, we applied a Gaussian copula model to the simultaneously recorded excitatory input currents, confirming that pairwise noise correlations can explain the observed higher-order noise correlations within small RGC populations. This provides the first experimental evidence that copula modeling is valid for describing higher-order noise correlations in excitatory input currents. Together, these results establish a mechanistic link between neuronal structure and correlated variability in neural code.
  • High resolution morphological characterization of solid polymer electrolytes
    (2026) Król, Monika
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-03-27
    The global need for transition towards the sustainable energy sources accelerated the search for next-generation energy storage technologies that can address the limitations of the conventional, liquid electrolyte-based lithium-ion batteries. Block copolymer electrolytes (BCPEs) emerged as possible successor due to their unique properties: namely, BCPEs are able to form various self-assembled morphologies with features at the nanoscopic length scales. Moreover, the recent developments in the available synthetic routes allow to i) precisely control the molecular weight of constituent blocks; ii) design different topological architectures; iii) access a broad range of blocks’ chemistries tailoring the materials’ properties. In case of BCPEs, the soft block with ability to conduct ions tends to be jointed with the hard block, ensuring the mechanical stability. Nano-segregated structure allows to decouple those properties. Gaining insights into the relationship between the morphology and the functional performance is essential to unlock the full potential of BCPEs as advanced materials for the energy storage applications. This thesis investigates the structure-property relationship in emerging BCPE systems and focuses on different aspects affecting the self-assembly. Advanced microscopic and scattering characterization techniques, including transmission electron microscopy (TEM) and small-angle X-ray scattering (SAXS), are employed to map the microstructural features. In Publication I, the effect of BCPE topology (star versus linear) on the morphology and the ionic conductivity was explored. The study revealed that star architecture with the minor conductive phase promoted the long-range order of nanodomains which correlated with the enhanced ion-conduction performance. As continuation of this study, in Publication II, those BCPEs were used as a doping agent in nanostructured photopolymerized films, where the conductive phase is the major component. As a result, high-performance and robust solid polymer electrolytes were obtained. The morphological features were highly dependent on the salt concentration, where elongated nano-objects were formed at the higher loadings. Among explored architectures, BCP dopant with linear topology turned out to provide the best compromise between the functional properties: mechanical strength and conductivity. Publication III explored how given thermal treatment affects crystallization kinetics, phase separation and crystalline phase composition of semi-crystalline ABA triblock. Here, polymorphic poly-l-lactic acid crystalline domains contributed to mechanical integrity while the carbonate based conductive midblock facilitated ion transport. Used thermal protocol led to the formation of hierarchical microstructures at micro and nanoscale. The ratio of crystalline polymorphs and the overall crystallinity were found to be depended on both the annealing time and the salt concentration. Finally, Publication IV introduced a strategy to induce long-range anisotropy in liquid-crystalline BCPEs by applying the electric field during the solvent casting. Ten-fold increase of ionic conductivity was observed in comparison to the film casted without the external stimuli, which was ascribed to enhanced phase separation and favourable orientation of liquid crystalline phase in regard to the ion transport. Overall, this work emphasizes the interplay between molecular design, morphology and resulting functional (electrochemical and mechanical) performance. Acquired insights into the microstructural features promote the rational, holistic design of BCPEs for high-performance solid-state batteries.
  • Machine learning applications in enhancing sustainable supply chains—a foresighted empirical study
    (2026) Farshadfar, Zeinab
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-03-20
    This dissertation explores the role of machine learning (ML) in advancing sustainable and circular supply chains (CSCs). By drawing on the sustainable supply chain management (SSCM) field and applying the triple bottom line (TBL) as a classification lens, the doctoral dissertation identifies gaps in current research. It is worth noting that this doctoral dissertation does not aim to develop a new theory in SSCM; rather, it identifies and addresses specific gaps within the existing SSCM literature through empirical and methodological contributions. Although literature on ML in supply chain management (SCM) has expanded rapidly, most research remains conceptual or simulation-based, providing limited empirical evidence of real-world impact. This study addresses that gap by combining a systematic literature review (SLR) with both real-world and simulation-based pioneer case studies to assess how ML contributes to the sustainability aspects of supply chains (SCs). Two main shortcomings emerge from previous research. First, ML is often used at the decision-support level—predicting demand, ranking suppliers, or creating optimisation inputs—where its sustainability impact relies on managerial action rather than operational integration. Second, even when operational ML is examined, its effects on sustainability outcomes are rarely measured directly. Instead, improvements are credited to broader optimisation frameworks, leaving ML’s unique role unclear. These limitations matter because, without real-world deployment and cost–benefit analysis, ML’s transformative potential for circularity remains largely aspirational. This dissertation adopts a foresighted empirical approach, concentrating on three SC functions with high sustainability potential: construction waste management, aviation spare parts (SPs) provisioning, and CSC optimisation. The pioneering case studies demonstrate that ML-enabled waste sorting surpasses traditional methods by increasing recycling rates and lowering long-term costs in high-wage environments. Furthermore, the sensitivity analyses on the same ML-enabled waste sorting case emphasise that labour costs, discount rates, and equipment investments are critical factors for the competitive advantage of ML-enabled waste sorting. In aviation, SP inventory pooling, ML-enabled generative design (GD), and additive manufacturing reduce CO₂ emissions through lightweighting and determine the cost and durability thresholds necessary for the economic viability of ML-enabled GD. Beyond the mentioned insights, the dissertation presents a structured synthesis of ML applications across SC, highlighting the predominance of ML supervised learning methods and the limited empirical evidence regarding their operational integration. By evaluating the impacts of ML using cost models and sensitivity analyses, this doctoral research provides a more accurate quantification of ML’s role in advancing SC sustainability. Overall, this doctoral research bridges the gap between conceptual enthusiasm and empirical evidence, demonstrating that ML-enabled solutions can serve not only as a decision-support tool but also as an operational technology that reduces costs, minimises waste (based on article 2 findings), extends product lifecycles (based on article 4 findings), and supports circular flows (based on article 1 findings). Simultaneously, this doctoral research identifies the enabling conditions—economic, technological, and policy-related—that influence adoption, providing a foresighted perspective for both scholars and practitioners on the future of sustainable and CSCs.
  • Measurement techniques for quantum microwaves
    (2026) Keränen, Aarne
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-03-13
    Superconducting quantum circuits provide a versatile and powerful platform for quantum computing as well as for the study of fundamental physics and quantum technologies in general, such as radiation sensors. The development of novel detectors and measurement techniques is important for advancing quantum technologies by enabling detection schemes that are faster, more accurate, or less complex than the current state of the art. In this thesis, we improve upon the existing measurement techniques in superconducting quantum circuits by employing novel methods on a radiation sensor based on superconductor–normal-metal–superconductor (SNS) junctions, in addition to superconducting quantum bits (qubits). The presented methods are experimentally demonstrated in the five publications of this thesis. First, we demonstrate that an SNS sensor, essentially a power or a photon detector, has the capability to directly measure the variance of itinerant microwave photons. Experimentally, we show the measurement of the zero-delay second-order correlation function of interfering thermal and coherent photons displaying bunching with increasing proportion of thermal photons. We combine the signal field with a homodyne field and carry out Wigner function tomography by utilizing techniques from quantum optics. Second, we show a new readout technique for superconducting qubits by multichannel driving which utilizes effectively longitudinal coupling of the qubit and a readout cavity. We find that applying this technique reduces the readout error by 50%, while being roughly 20% faster than the conventional dispersive readout. We finish the thesis by displaying methods for making quantum circuits more compact. We demonstrate a rigorous process for the design of lumped-element resonators without utilizing any computer simulations. The fabricated resonators have a small physical footprint compared to that of coplanar waveguide resonators ubiquitous in superconducting circuits. We achieve an accuracy of 99% in the designed resonance frequency with a standard deviation of 1%, demonstrating the applicability of the model. The final publication demonstrates multiplexed readout of the SNS sensors, reducing the number of required control and measurement lines.
  • Decision models for preventive maintenance of technical systems
    (2026) Leppinen, Jussi
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-02-27
    The reliable operation of technical systems and their components requires maintenance, the costs of which can constitute a substantial share of a system’s total operating expenses. Rather than relying on corrective maintenance, systems are often best serviced preventively with the aim of eliminating failures, since downtime can be expensive. However, components should not be replaced more frequently than necessary. Consequently, optimizing component maintenance schedules can yield significant cost savings. Yet, optimization is complicated by dependencies between components and by incomplete information about their remaining useful life. This dissertation presents novel analytical approaches to improve the maintenance management of technical systems. It has two main contributions. The first contribution is a Markov decision process model to determine cost-effective maintenance schedules for systems when there are economic and structural dependencies between components. The second contribution is a structured decisionmaking process that supports the development of AI-based solutions for predicting the remaining useful life of the system. The proposed Markov model is the first to account for the threshold for system reliability and the structural dependencies arising from the disassembly and replacement of components. In addition, it considers reliability structures derived from various component configurations and the expected costs of system downtime when maintenance can be performed periodically. The optimal maintenance policy is solved using a modified policy iteration algorithm enhanced with Anderson acceleration. The resulting maintenance schedule leverages the simultaneous servicing of components, achieving improved cost efficiency in comparison with many heuristic policies. The model supports operational, tactical, and strategic maintenance decision-making by providing insights into the use of individual maintenance actions, long-term total costs of the optimal policy, and, for example, the impact of the maintenance period length on overall costs. Furthermore, the model offers guidance on how to design the system structure. No process has previously been established for developing AI-based solutions aimed at predicting a system’s remaining useful life while taking into account both technical and organizational criteria in the early stages of development. In the proposed stage-gate process, the development of AIbased solutions progresses from use case definition through exploratory testing and implementation planning to final deployment. At each decision gate, the candidate solutions to advance are selected. Candidates are evaluated against both technical and organization-specific criteria, which are derived from six development objectives. Since uncertainties may exist in the evaluation of candidates with respect to criteria and prioritization of objectives during early development stages, robust decisionmaking is employed to support choices at the gates. This reduces the risk that selected candidates later become infeasible. In addition, the benefits of the development effort are assessed in terms of the organization’s capability advancement.
  • Towards automated programming feedback with open-weight language models
    (2026) Koutcheme, Charles
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-02-20
    The increasing demand for computer science experts highlights the importance of supporting novice learners effectively, particularly in introductory programming courses. A key factor in maintaining learner engagement and progress is the provision of timely, constructive feedback on student code. Yet, delivering such feedback at scale remains a significant challenge: human-centered support does not scale easily, and existing automated assessment systems often lack the ability to provide meaningful and nuanced guidance. Recent advances in language model (LM) research offer promising new avenues to address this gap. Large language models (LLMs) in particular have demonstrated strong capabilities for generating educational feedback. Much of the work in this space has relied on proprietary models such as those behind ChatGPT. However, reliance on these models raises concerns around cost,control, and long-term accessibility. These challenges are motivating a shift toward open-weight alternatives. This dissertation addresses several challenges in integrating open-weight language models into educational feedback systems, presenting contributions across three complementary dimensions. First, it explores methods for enabling pre-trained LMs to repair student programs. These methods combine infilling models with search algorithms to repair programs in place and leverage automated repair tools with distillation pipelines to bootstrap training. Second, it proposes two automated evaluation approaches to assess feedback capabilities without requiring human annotation: (i) an LLM-as-a-Judge approach that leverages language models'reasoning abilities to score feedback, and (ii) a repair-as-proxy approach that uses program repair performance to measure feedback proficiency. Third, it contributes reinforcement learning techniques to align small language models (SLMs) with pedagogical objectives. These approaches allow practitioners to choose between two forms of automated supervision: alignment via model preferences (i.e., AI-generated rankings) and self-supervised alignment using verifiable program repairs. Empirical evaluations on student code datasets and public programming benchmarks demonstrate that small, open-weight models can be selected, tuned, and evaluated almost automatically to generate useful explanations, hints, and corrections for novice programmers.
  • Advancing research methodologies in digital phenotyping for mental health
    (2026) Ikäheimonen, Arsi
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-02-13
    Digital phenotyping is an evolving field that fuses the disciplines of data science, behavioral science, and medicine. Digital phenotyping research utilizes data from personal digital devices and online platforms to investigate various behavioral, social, and health-related aspects of an individual's life. By analyzing these data, research aims to find novel insights into behavior, health, and well-being. These insights may lead to new complementary methods and applications for more effective healthcare solutions. As a new field, several factors hinder digital phenotyping from reaching its full potential. This thesis aims to advance the field of digital phenotyping through two perspectives: methodological and practical. From a methodological perspective, the thesis focuses on the barriers due to technical and research methodology-related challenges. In turn, the practical perspective evaluates the potential of using digital phenotyping to monitor and predict depressive symptom severity using data collected from outpatients diagnosed with ongoing depressive episodes. This thesis comprises five research articles, two of which focus on advancing the methodology, and the remaining three focus on practical aspects. The first article introduces Niimpy, an open-source Python toolbox for analyzing behavioral data. The second article outlines a datadriven workflow that facilitates research and improves its generalizability. The third article examines how accurately digital phenotyping, using smartphone-sensed data, can assess depression severity. The fourth article inspects the feasibility of digital phenotyping by analyzing the behavioral differences between the healthy control and patient cohorts. The fifth article explores the inherent variability in depressive symptomology and associated mobile-sensed behaviors. Taken together, these articles provide standardized tools and data analysis pipelines that facilitate behavioral data analysis, thereby lowering the barrier to entry into digital phenotyping research. This thesis proposes concrete, actionable guidelines for improving research replicability and reproducibility through the use of analysis workflows. The work demonstrates that machine learning models, using digital phenotyping data, can predict future depression severity. In addition, the feasibility of using smartphone data is supported by both study groups, healthy controls and patients diagnosed with depression, providing behavioral data with no differences in participation adherence or data quantity. Finally, the thesis demonstrates how behavioral differences, both between and within the participants diagnosed with depression, are connected with their self-reported symptoms. To conclude, this thesis advances digital phenotyping research by introducing tools and workflows to facilitate analysis. Furthermore, by building on previous work in the field, the thesis contributes to the research by presenting research outcomes obtained through the analysis of depressive outpatient behavioral data.
  • Hydrogel-based 3D culture systems for modeling breast cancer
    (2026) Heilala, Maria
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-30
    Breast cancer is a significant cause of cancer-related death in women. To develop more targeted therapies for advanced breast cancer, a detailed knowledge of mechanisms involved in breast cancer progression is required. Hydrogel-based 3D culture systems are particularly promising models for studying microenvironment-regulated changes in cellular phenotypes during metastasis. This thesis compiles three publications that investigated how cells can be guided to model prometastatic phenotypes by tailoring the properties of biopolymer-based hydrogel scaffolds. In Publication I, fibrin gels were used to study the interaction between fibrin and circulating tumor cell (CTC)-like cells representing triple negative breast cancer (TNBC). The phenotypes observed in fibrin gels were clearly distinct from those in Matrigel and suspension culture, evidenced by differences in gene and protein expression and cell morphology. By varying fibrin concentration and degradation rate, the association between fibrin stiffness and the regulation of cellular protrusions was revealed. In addition to cancer cells, immunosuppressive immune cells are associated with poor prognosis in breast cancer. In Publications II and III, plant-derived nanofibrillar cellulose (NFC) gels were used as the hydrogel scaffolds. In Publication II, unmodified NFC gels with varying stiffnesses were used for breast cancer patient-derived explant culture (PDEC). It was discovered that soft matrix influenced the balance of T cells and promoted macrophage polarization to M2-like phenotype through the fibroblast growth factor 2 (FGF2) secreted by the cancer cells. Publication III investigated the immunomodulatory effects of NFC chemical modifications on primary immune cells. NFC surface chemistry was found to influence the gene expression profiles of peripheral blood mononuclear cells (PBMCs) and the differentiation of monocytes to M2-like macrophages. Notably, macrophages generated in phosphorylated NFC gel expressed M2 marker CD206 at similar levels as M2a macrophages differentiated with interleukin-4 (IL-4) treatment in conventional 2D culture. Moreover, the cytokine profile was differentially modulated in cells in phosphorylated NFC gel compared to cells in 2D culture. In conclusion, diverse prometastatic phenomena were recapitulated in the hydrogel-based 3D culture models developed in this thesis. Importantly, the models expanded the phenotypes captured by the 2D culture setup and the commercial Matrigel matrix. These findings contribute to the development of more versatile breast cancer models, which may lead to the discovery of novel treatment modalities to prevent and treat metastatic disease.
  • Beyond critical points: Critical manifolds in self-organizing systems
    (2026) Sormunen, Silja
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-29
    Self-organized criticality describes the spontaneous evolution of a system to the verge of a phase transition, where the system balances between two macroscopically distinct phases. At this critical threshold, microscopic dynamics give rise to universal behavior characterized by scale-invariant spatial and temporal correlations spanning the whole system. Traditionally, self-organized criticality has been depicted as a one-dimensional phenomenon in which a single parameter is driven to its critical value. However, as the number of parameters in most real-world systems is vast, the states of the system corresponding to criticality can be expected to form a highdimensional manifold. This thesis explores the consequences of this multidimensionality, illustrating the topological and dynamical flexibility inherent in the self-organized critical state. Positioned at the intersection of network science and dynamical systems theory, this doctoral thesis makes both theoretical and methodological contributions to the study of self-organized criticality and its signatures. Using simulations and theoretical analysis, the first study demonstrates the phenomenon of a self-organized critical drift, where the self-organization process drives an adaptive network to drift on the critical manifold. The second study shows that this topological flexibility corresponds to dynamical richness, providing the first demonstration of an adaptive system self-organizing to multicriticality, where the system resides simultaneously on the verge of two different phase transitions. Finally, the third study addresses the methodological challenge of identifying the power-law distribution, a core hallmark of criticality. Power laws are notoriously difficult to identify, and this task is further complicated when data is subsampled. Our results show that two state-of-the-art methods fail to reliably distinguish between subsamples from power-law and some other heavy-tailed distributions, calling for caution in their use. Together, the results of this thesis provide a novel framework for understanding self-organized criticality that both opens new perspectives and unifies research directions previously regarded as competing. In previous research, it has been suggested that the brain would tune itself to criticality, but a lack of consensus on the phase transition in question has slowed down progress in the field. By demonstrating the inherent flexibility of the criticality hypothesis, this thesis shows how a system can seamlessly move to a new transition without leaving another, providing new insight into systems that may rely on multiple phase transitions to support complex adaptive behavior. As this work concerns the theoretical foundations of self-organized criticality, the findings have broad relevance across disciplines, generally applying to all self-organizing systems from biological brains and artificial neural networks to ecosystems and human contact networks.
  • Bayesian ordinary differential equation and Gaussian process modeling of biomedical data
    (2026) Timonen, Juho
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-23
    Probabilistic models of biomedical data account for uncertainty and variability in biomedical systems by describing the data and unknown quantities through probability distributions. Probability-based models are particularly valuable in biomedical contexts, where data are often noisy, incomplete, or heterogeneous, and where capturing uncertainty is critical for decisionmaking. This thesis focuses on complex probabilistic models, particularly those involving Gaussian processes (GPs) and ordinary differential equations (ODEs). The aim is to develop interpretable and practical models for biomedical data and address computational challenges that arise from both the complexity of the models and the large scale of the data involved. The first line of research advances inference techniques for Bayesian ODE models. Such models are useful for example for modeling intracellular regulation mechanisms, distribution of a drug in the body, or spreading of infectious diseases. However, simulating the time evolution of such models typically requires computationally intensive numerical solvers, which provide only approximate solutions and require that the functional form defining the ODE system is known. In practice, however, the kinetic parameters—and often the structure—of the system are unknown and must be inferred from data, introducing significant computational complexity. While the numerical approximation errors of ODE solvers have been extensively studied in classical numerical analysis, their impact on probabilistic inference has received less attention. This thesis proposes a computationally efficient workflow for obtaining reliable Markov chain Monte Carlo (MCMC)–based Bayesian inference results for nonlinear ODE models. It highlights and analyses the limitations of common adaptive ODE solvers in probabilistic contexts and introduces a scalable method for probabilistic network inference under structural uncertainty of the system. The second line of research focuses on interpretable, nonparametric modeling of longitudinal data using additive GPs. Such data typically involve a mix of relevant and irrelevant continuous and categorical predictor variables. In this thesis, flexible yet explainable models of longitudinal data are constructed using additive Gaussian processes. The thesis proposes novel modifications to the GP covariance functions to enhance the interpretability of model components. To accommodate large datasets, it is shown how to apply a reduced rank approximation scheme for GPs whose covariance depends on both categorical and continuous predictor variables. To increase model parsimony and interpretability, model reduction techniques are developed, adapted and evaluated for their performance in identifying relevant predictors and retaining predictive power while simplifying the model.While the core contributions are methodological, the developed techniques are applied to realworld biomedical problems, including the inference of gene regulatory networks involved in cell differentiation and the modeling of biomarkers associated with the development of type 1 diabetes
  • Hybrid Josephson junctions and electrothermal effects in graphene devices
    (2026) Haque, Mohammad Tasnimul
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-23
    In this thesis, I explore hybrid graphene-based Josephson junction (SGS) devices integrated with superconducting microwave cavities. The field of quantum computing has advanced rapidly in recent times, with realisations of various hardware platforms such as ion-traps, ultracold atoms, spin qubits, topological systems, and Josephson junction-based superconducting qubits, each aiming to reach the fault-tolerant quantum computing regime. Alongside these developments, two-dimensional materials such as graphene exhibit unique mechanical and electronic properties, such as atomically thin, high carrier mobility, and gate-tunable proximity-induced supercurrent, which makes SGS junctions an attractive architecture for long-lived, magnetic-field-resilient quantum circuits. Critical current fluctuations are a key dominant factor for decoherence in a Josephson junction. In the first experiment, I investigate low-frequency 1/f noise in the critical current of a hexagonal boron nitride encapsulated edge-contacted SGS junction. Fluctuations in the inductance of the SGS junction were tracked using microwave reflection measurements, and then the critical current noise was extracted from these measured fluctuations. We find significant fluctuations in the critical current, on the order of per unit band of 1 Hz. The critical current noise grows away from the charge neutrality point, which we attribute to the variation of the proximity-induced gap in the SGS junction. In a second set of experiments, I study various electrothermal effects that emerged in graphene devices. A superconductor–insulator–normal-metal–insulator–superconductor (SINIS) junction is formed from a monolayer graphene flake and coupled to a superconducting cavity. The device exhibits thermal self-oscillations arising from a highly nonlinear temperature-dependent resistance. We show that the modelling of these thermal oscillations provides a method to evaluate electronphonon coupling in graphene. In addition, these oscillations can be harnessed for the parametric amplification of microwave signals. In another device, we demonstrate the generation of a thermoelectric current in a graphene-based Cooper pair splitter, which is formed from two graphene quantum dots connected to an aluminium superconductor. Finally, I investigate microwave quantum optomechanics utilizing the Josephson capacitance of a Cooper pair box and find several orders of magnitude enhancement in both the optomechanical and the cross-Kerr couplings. Mediated by the Josephson capacitance, this three-partite system reaches the single photon ultrastrong coupling regime, enabling the generation of non-classical states of light and mechanical motion, as well as providing a platform for a single phonon counter. Together, these results demonstrate the potential of graphene–superconductor hybrids as a versatile building block for future quantum circuits, combining the design flexibility of superconducting platforms with the exceptional properties of two-dimensional materials.
  • Enabling cryogenic technologies for superconducting quantum devices
    (2026) Hätinen, Joel
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-23
    Low-temperature refrigerators cool systems down to cryogenic temperatures near absolute zero, where thermal noise and decoherence are suppressed. This allows quantum phases, such as superconductivity, to emerge in certain materials and enables the harnessing of individual quantum states for scientific and high-performance applications. However, the refrigerators used for these purposes are large and rely on cryoliquids, such as scarce and expensive 3He, which can be a limiting factor depending on the technological application. To enable more scalable, costeffective cryogenic platforms, new refrigeration technologies must be developed. To this end, chip-scale coolers based on superconducting tunnel junctions have been envisioned to provide a fully solid-state alternative. Proof-of-principle operation of these coolers has been demonstrated at temperatures below 1.5 K, but to link them with commercially available 4He pulse tube cryocoolers, stage operating above 2.0 K is required. Additionally, thermally isolating and electrically conducting methods are needed to cascade coolers operating at different temperature ranges. In this thesis, the fundamental components of a multi-stage chip-scale cooler operating at temperatures compatible with 4He pulse tube cryocoolers are developed. A superconducting flipchip assembly fabricated with In-bumps was characterized in the sub-kelvin temperature range, and the inter-chip thermal resistance was found to be suitable for chip-scale cooling applications. A through-chip signal routing method utilizing ALD TiN-based TSVs was developed, and the demonstrated critical temperature of 2.0 K enables dissipationless DC transport for multi-chip assemblies, such as cascaded coolers. Additionally, ALD MoCx was shown to exhibit a superconducting transition temperature up to 4.4 K and high conformality, showing promise as a TSV-compatible material. The key achievement of electronic cooling of Al thin film from a bath temperature of 2.4 K down to 1.6 K was demonstrated using Nb-based superconducting tunnel junctions, probed by an onchip junction thermometer. Thermal model calculations highlighted the emergence of superconductivity in the Al beneath the cooler junctions, persisting up to a bath temperature of 2.4 K: one kelvin higher than the nominal critical temperature of the Al thin film. The single-stage cooler operating above 2.0 K enables solid-state on-chip cooling from 4He pulse-tube compatible temperature without the use of magnetic fields. Additionally, Al- and V-based tunnel junctions were fabricated at the wafer scale using degenerately doped Si as the normal electrode. The junctions exhibited suitable low-temperature electrical characteristics for cooling applications. From superconducting interconnects to tunnel-junction components supporting high cooling power density above 1 K, the achievements presented in this thesis enable modular design of chip-scale cascade coolers. This technology is envisioned to support the scaling of several superconducting quantum devices from proof-of-principle to multi-component systems beyond experimental lab environments.
  • Thermometry based on a superconducting qubit
    (2026) Lvov, Dmitrii
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-20
    In this thesis, we demonstrate the operation of a thermometer based on a superconducting transmon qubit. We experimentally show that by measuring the qubit’s population distribution and its respective effective temperature, it is possible to measure temperature of an object with which the qubit is thermalized. However, due to the qubit’s quantum nature, it is very sensitive to any other source of excitation present in the system, which can be treated as a separate noise source or a heat bath. Consideration of a qubit coupled to several uncorrelated noise sources constitutes an important part of the thesis. We investigate the applicability range of the qubit thermometer in terms of dynamic range, operation speed and precision. While the upper bound on the dynamic temperature range is mainly defined by the material properties of the superconductor used in the qubit fabrication, namely, by the superconducting gap, the lower temperature is set by coupling to parasitic excitation sources. The population measurements of a transmon qubit are based on an algorithm which uses π-pulses for swapping the populations of the three lowest energy levels of the transmon. The largest impact on the precision of the measurement is defined by the signal-to-noise ratio and the quality of the qubit control pulses. The precision limit is ultimately defined by the quantum Cramér-Rao bound, highlighting the statistical origin of this thermometry method. Finally, we utilize the transmon qubit for thermometry of a mesoscopic heat bath located on the same chip. The heat bath is a normal metal resistor, whose temperature was controlled with normal metal/insulator/superconductor (NIS) junctions. By coupling the transmon to the resistor capacitively and performing the population measurements, we could observe linear dependence between the qubit effective temperature and the temperature of the resistor. Moreover, while the designed mechanism of the qubit-resistor coupling was photonic, we managed to observe another channel of interaction. At large bias voltages applied to the NIS-junction, which sets the temperature of the resistor, the junction starts to emit nonequilibrium phonons that can break Cooper pairs in the superconducting qubit. This is a nonlocal effect, in which nonequilibrium quasiparticles lead both to change of the qubit population distribution and significant suppression of the relaxation time.
  • Electrical properties and transport characteristics of single-walled carbon nanotube bundles
    (2026) Khan, Md Abu Taher
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-16
    As silicon approaches its fundamental performance limits, single-walled carbon nanotubes (SWCNTs) are emerging as a promising alternative channel material for transistors. Floating catalyst chemical vapor deposition (FC-CVD) is a highly effective technique for growing highly crystalline, defect-free SWCNTs while enabling their direct deposition from the reactor onto the wafer. During growth in an FC-CVD reactor, carbon nanotubes agglomerate due to Brownian diffusion, forming bundles, as predicted by simple aerosol models. In this dissertation, a fast and clean dry deposition method was developed to fabricate hundreds of functional SWCNT field-effect transistors (FETs), aiming to obtain a large number of single-bundle FETs. This method enabled the characterization of the pristine electrical properties of assynthesized FC-CVD SWCNTs and was also used to investigate the transport characteristic and electrical performance of SWCNT bundles. FETs were fabricated using two sets of FC-CVD-grown SWCNT bundles: Small Bundle Small Diameter (SBSD) SWCNTs and Large Bundle Large Diameter (LBLD) SWCNTs. The mean diameters of SBSD and LBLD SWCNTs were 4.1 ± 2.1 nm and 7.1 ± 3.7 nm, respectively, with nanotubes having mean diameters of 1.4 nm and 1.9 nm, respectively. Electron diffraction (ED) analysis determined that as-synthesized SBSD and LBLD SWCNTs contained metallic fractions of 38% and 46.3%, respectively. Due to the presence of a large fraction of metallic SWCNTs, bundles consist of a mixture of metallic and semiconducting nanotubes, leading to the expectation that FETs would lose their semiconducting switching efficiency. Interestingly, our experimental data show the opposite. The fraction of semiconducting FETs (s-FETs) was higher than the fraction of assynthesized semiconducting SWCNTs. The SBSD SWCNT FETs exhibited a semiconducting fraction of 71.5% out of 1,887 functioning FETs, while the LBLD SWCNT FETs showed a semiconducting fraction of 62% out of a total of 1,839 functioning FETs. The charge carrier mobility of SWCNT s-FETs was extracted using two rigorous methods: the peak transconductance (𝜇𝜇𝑃𝑃𝑃𝑃𝑃𝑃) and the Y-function (𝜇𝜇𝑌𝑌𝑌𝑌) methods. The ohmic-contact SBSD SWCNT s-FET exhibited mean 𝜇𝜇𝑃𝑃𝑃𝑃𝑃𝑃 and 𝜇𝜇𝑌𝑌𝑌𝑌 mobility values of 1,061 cm2V-1S-1 and 2,817 cm2V-1S-1, respectively, while the ohmic-contact LBLD SWCNT s-FET exhibited mean 𝜇𝜇𝑃𝑃𝑃𝑃𝑃𝑃 and 𝜇𝜇𝑌𝑌𝑌𝑌 mobility values of 1,854 cm2V-1S-1 and 5,378 cm2V-1S-1, respectively. These values are several orders of magnitude higher than the performance of SWCNTs grown using other methods. Similarly, the mobility of ohmic-contact single-junction SWCNT s-FETs was extracted for both SBSD and LBLD SWCNTs. Compared to single-bundle FETs, the mean 𝜇𝜇𝑃𝑃𝑃𝑃𝑃𝑃 and 𝜇𝜇𝑌𝑌𝑌𝑌 of single-junction SBSD SWCNT s-FETs decreased about fourfold to 255 cm2V-1S-1 and 737 cm2V-1S-1, respectively, while for single-junction LBLD SWCNT s-FETs, these values decreased about threefold to 856 cm2V-1S-1 and 1,732 cm2V-1S-1, respectively. Furthermore, both SBSD and LBLD SWCNT s-FETs exhibited an on-off ratio of up to 108.
  • Exploring movement and dance in virtual reality
    (2026) Laattala, Markus
    School of Science | Doctoral thesis (article-based) | Defence date: 2026-01-16
    Virtual Reality (VR) allows the creation of new types of unique, interactive, and immersive experiences. However, designing for the medium brings its own strengths and limitations. While a player in a virtual world can see and hear anything, their material body's movement is still limited by both their physical capabilities and the space they are in. Moreover, today's mainstream consumer VR devices still struggle to accurately capture subtle human movements or provide realistic force feedback when interacting with virtual objects. These limitations highlight the need for more advanced interaction methods tailored to immersive environments. This dissertation investigates these types of novel techniques for user movement in virtual environments and explores how motion can be guided in ways that are unique to Virtual Reality. From the five publications included in this dissertation, Publications I, II, and III focus on dance experiences. Publication I introduces and evaluates a new way of instructing choreographies, allowing the user to anticipate incoming moves in games and other experiences. Publication II explores solutions for creating multiplayer dance games in Virtual Reality with limited dance space and movement tracking. Publication III examines the dance experience in VR compared to Mixed Reality, where a virtual partner is embedded in the user's real surroundings and the user sees their own body similar to real-life. The remaining two publications IV and V examine how familiar ovement actions from traditional video games can be reimagined in more embodied forms that suit the physical and technical realities of VR. To achieve these goals, this dissertation employs a mixed-methods approach to exploring and developing novel movement- and dance-based game mechanics and interactions, describing the design processes and rationale, and evaluating the mechanics and interactions through user studies that combine quantitative data with qualitative interviews. This dissertation explores different novel VR experiences and highlights unexplored design space for movement-based experiences in VR. It also demonstrates how the medium's technological limitations and non-physical nature can serve as strengths, inspiring more creative approaches to movement design and a deeper focus on user experience in VR. The main contributions of this dissertation are the design insights gathered from the projects as well as the projects themselves that serve as groundwork for future dance-centered and movement-centered games, applications, and research in VR.