Graduate School of Engineering and Science
Permanent URI for this collectionhttps://hdl.handle.net/11693/115678
Browse
Recent Submissions
Item Open Access Adaptive modularity: deep reinforcement learning for optimized modular housing massing(2025-07) Ekici, Betül KeleşThe advancement of prefabrication technology has positioned modular housing as a precise, sustainable construction method that reduces greenhouse gas emissions and resource intensiveness while enabling a circular economy and fostering energy efficiency in meeting increasing housing demands. The design of massing configurations from modular units presents challenges due to the complexity of optimizing spatial arrangements to simultaneously balance architectural, structural, and environmental performance criteria and constraints. This thesis addresses the challenge of modular housing design, tackling the considerable complexity of optimizing spatial arrangements. The aim is to improve the design process through Deep Reinforcement Learning (DRL) to balance architectural criteria and constraints. The thesis investigates how DRL can enhance modular housing design by optimizing spatial configurations, overcoming the limitations of human-driven iterations, and leveraging the benefits of a feedback-driven system. It proposes a simulation environment based on a Cellular Automata (CA) voxel space, where an AI agent, trained using the Proximal Policy Optimization (PPO) algorithm, autonomously places massing cells to generate optimized housing configurations. These configurations are evaluated based on multiple performance objectives, such as ensuring efficient space utilization by maximizing the number of housing units within the buildable area, minimizing circulation space while maintaining accessibility to all units, and promoting the creation of housing units that are more compact in shape. This thesis focuses on the effort to optimize modular mass configuration for housing design. It outlines the steps taken in the process, the challenges faced along the way, and evaluates the success of the final configurations that were developed.Item Open Access Analysis of erbin gene expression regulation by micrornas in breast cancer(2025-08) Yüksel, ÖzgeBreast cancer is a heterogeneous disease characterized by distinct molecular subtypes, defined by specific gene markers and mutations associated with different prognoses and treatment responses. The ERBIN (ERBB2IP) gene encodes a basolateral scaffold/adaptor protein that maintains epithelial cell polarity and adhesion, while also regulating key signaling pathways, including MAPK and TGF-β signaling. Its expression changes across diseases, including HER2-positive breast cancer. However, the mechanisms underlying aberrant ERBIN expression levels remain poorly investigated. This study aimed to identify specific miRNAs regulating ERBIN expression in distinct breast cancer subtypes and to understand further regulatory interactions within the miRNA-ERBIN axis. ERBIN expression patterns in publicly available breast cancer patient datasets were analyzed across molecular subtypes, TP53 mutation status, and hormone receptor status to address this aim. Expression correlation analysis between ERBIN and miRNA expression, and the results from miRNA target prediction tools, identified potential miRNAs that may regulate ERBIN expression in breast cancer. These candidates were further analyzed for their expression in aggressive subgroups to evaluate their inverse association with ERBIN, considering clinical characteristics. In addition, miRNA-seq analysis following ERBIN knockdown in MDA-MB-231 cells revealed potential regulatory feedback loops between ERBIN and selected miRNAs. Overall, the results provide comprehensive insight into the post-transcriptional regulation of ERBIN in breast cancer and suggest that the miRNA-ERBIN relationships provide a potential biomarker or therapeutic target.Item Open Access Fast and efficient optical field control via binary amplitude modulation: applications to optical holography and speckle customization(2025-08) Alıcı, HakanSpeckle patterns, which arise from the scattering of coherent light by disordered media, play a central role in applications such as imaging, sensing, and optical communication. However, controlling speckle patterns with both high speed and high efficiency remains a major challenge. Liquid crystal spatial light modulators provide precise phase control but operate at frame rates of only 50–100 Hz, while Digital Micromirror Devices (DMDs) can reach kHz speeds but typically suffer from low diffraction efficiency. This thesis introduces Diffuser-integrated Binary Amplitude Modulation (DiBAM), a framework that combines the kHz switching speed of DMDs with the strong mode mixing of a diffuser to enable rapid, efficient, and accurate speckle pattern engineering. Through purely binary amplitude modulation, DiBAM supports a wide range of wavefront-shaping applications, including point focusing, high-efficiency holography, and non-Rayleigh speckle customization. The diffuser ensures that each micromirror contributes to all output channels, enabling complex field synthesis with high modulation fidelity. Numerical and experimental demonstrations show that DiBAM achieves point focusing without phase modulators, surpasses state-of-the-art holographic methods in balancing efficiency and fidelity, and enables fast speckle customization with tailored intensity probability density functions. These findings establish DiBAM as a versatile, high-performance platform for next-generation adaptive optics and computational imaging.Item Open Access Physics-driven deep learning for medical image reconstruction(2025-08) Kabaş, BilalMedical image reconstruction from undersampled acquisitions is an ill-posed problem involving inversion of the imaging operator linking measurement and image domains. Physics-driven (PD) models have gained prominence in reconstruction tasks due to their desirable performance and generalization. These models jointly promote data fidelity and artifact suppression, typically by combining data-consistency mechanisms with learned network modules. Artifact suppression depends on the network’s ability to disentangle artifacts from true tissue signals, both of which can exhibit contextual structure across diverse spatial scales. Convolutional neural networks (CNNs) are strong in capturing local correlations, albeit relatively insensitive to non-local context. While transformers promise to alleviate this limitation, practical implementations frequently involve design compromises to reduce computational cost by balancing local and non-local sensitivity, occasionally resulting in performance comparable to or trailing that of CNNs. To enhance contextual sensitivity without incurring high complexity, we introduce a novel physics-driven autoregressive state-space model (MambaRoll) for medical image reconstruction. In each cascade of its unrolled architecture, MambaRoll employs a physics-driven state-space module (PD-SSM) to aggregate contextual features efficiently at a given spatial scale, and autoregressively predicts finer scale feature maps conditioned on coarser-scale features to capture multi-scale context. Learning across scales is further enhanced via a deep multi-scale decoding (DMSD) loss tailored to the autoregressive prediction task. Demonstrations on accelerated MRI and sparse-view CT reconstructions show that MambaRoll consistently outperforms state-of-the-art data-driven and physics-driven methods based on CNN, transformer, and SSM backbones.Item Embargo Generation of liver assembloids and mechanoregulation of liver regeneration(2025-08) Uzun, BelizLiver has been known to be capable of regenerating its cell population upon damage. Upon persistent liver damage such as fibrosis, coupled with the impairment of hepatocyte proliferation, cholangiocyte population shows a leniency to acquire an expression profile resembling hepatocytes. However, this leniency is mainly shown in animal models and has not yet been demonstrated in an organoid system without the external regulation of the pathways to push cells to hepatic fate. Liver cells also have been shown to respond to mechanical stimuli upon changes in the micro-environment including fibrosis but the response of cholangiocytes to mechanical cues is a topic that requires further investigation. In this study, we take advantage of the bi-phenotypic nature of cholangiocytes to generate hepatocyte-like organoids from ductal organoids isolated from both TAA and control livers. By assembling these organoids in a kinetic culture system, a novel attempt to generate a liver assembloid model that recapitulates native liver architecture is performed. Here,we propose a novel mechanotransduction pathway in cholangiocytes, mediated by PIEZO1, that orchestrates liver tissue organization. We demonstrate that PIEZO1 which is localized in the primary cilium regulates YAP1 signaling, influencing cellular responses to fibrosis and guiding cholangiocyte-driven tissue remodeling. Employing both in vitro fibrosis organoid cultures and in vivo fibrosis models, we reveal distinct expression profiles in fibrotic versus control conditions. This study advances our understanding of hepatic cell plasticity and offers a powerful platform for therapeutic exploration in liver disease.Item Embargo Accurate visible light positioning through power adaptation of led arrays(2025-08) Yaz, RamazanIn this thesis, we develop power allocation strategies for asynchronous visible light positioning (VLP) systems employing light-emitting diode (LED) arrays with the aim of improving localization accuracy. We first review the literature on visible light positioning in indoor scenarios. Then, we formulate an optimization problem for minimizing the Cramér-Rao lower bound (CRLB) related to the estimation of the receiver location subject to practical constraints on total and individual transmission powers and required illumination levels. Due to the nonconvexity of the proposed formulation, we devise a two-step methodology. First, a convex optimization problem is employed to determine the power allocation that maximizes the total received power. Subsequently, a projected gradient descent algorithm is utilized to refine the power allocation according to the CRLB metric under the power and illumination constraints. In addition, a framework for iterative power allocation and localization is developed for practical deployment, in which each step of the optimization process uses the most recent position estimate. Simulation results indicate that the proposed method can outperform both the uniform and power maximization based allocation schemes, resulting in improved localization accuracy over a wide range of operating conditions. Hence, it can provide robust and high-precision indoor positioning, even in circumstances involving asynchronous transmission and stringent illumination requirements.Item Open Access Geometry, topology, and emergent quantum phases: quasicrystals, Moiré magnets, and spin liquids(2025-08) Keskiner, Mehmet AkifThis thesis explores how quasiperiodic geometry, moiré superlattice, and fractionalized spin excitations generate novel electronic and magnetic behavior. Through four theoretical studies, it examines: (i) strictly localized states in the Socolar dodecagonal quasicrystal; (ii) magnetic textures in a twisted moiré superlattice; (iii) Kitaev-type spin liquids on the dual Ammann–Beenker quasicrystal; and (iv) magnetic order induced by Kondo coupling to quantum spin liquids. Localized states: In the Socolar dodecagonal lattice (SDL), a quasicrystal with twelvefold symmetry, we identify 18 distinct types of strictly localized states (LS), accounting for approximately 7.58% of the Hilbert space, closely matching numerical estimates of 7.61%. Through perpendicular space analysis, we demonstrate that at least 3.9% of sites are forbidden from hosting LS due to local connectivity constraints—revealing behavior intermediate between Penrose and Ammann–Beenker quasicrystals. Moiré magnetism: In twisted heterostructures composed of a Mott insulator and a semimetal, we study the emergence of spatially modulated magnetic order arising from nonuniform RKKY interactions. Our Monte Carlo simulations reveal a rich phase structure: AA-stacked regions exhibit antiferromagnetic (AFM) order, AB-stacked regions favor ferromagnetic (FM) alignment, and the intervening regions host ferromagnetic chains coupled antiferromagnetically (FMC). The spatial extent and coexistence of these domains are governed by the inverse decay length, α, of the Kondo interaction—where small α favors extended FMC regions, while larger α leads to the coexistence of FM, AFM, and FMC textures across the moiré unit cell. Quasicrystalline spin liquid: We formulate an exactly solvable Kitaev-type model on the dual Ammann–Beenker lattice (dABL), exploiting its fourfold coordination and partite bond structure. Our comprehensive study uncovers a rich variety of phases, including both gapless and gapped quantum spin liquids with chiral and abelian characteristics, analyzed via Monte Carlo methods and variational techniques. Additionally, incorporating an onsite perturbation refines the ground state selection to 21 unique vison configurations while preserving integrability. This work highlights the complex relationship between quasiperiodicity and emergent quantum magnetic phases. Spin-liquid-mediated magnetism: We investigate how magnetic order emerges among localized spins that interact exclusively via their coupling to a Kitaev-type spin liquid. Studying Kitaev, Yao-Lee, and square-lattice generalization models, we derive effective spin interactions mediated by fractionalized Majorana fermions. Short-range couplings stabilize the spin liquid in the Kitaev model, while the Yao-Lee model exhibits long-range RKKY-like antiferromagnetic order and partial Majorana gap formation. The square-lattice model shows competing anisotropic interactions, leading to dimerized quantum paramagnetism or Ising antiferromagnetism depending on parameters. These results reveal the rich magnetic phases enabled by Kitaev-type spin liquids.Item Embargo Broadband light-matter interaction using subwavelength metasurface designs for multi-spectral camouflage and polarization conversion application(2025-08) Elshurafa, Reham IMPhotonic metamaterials are one of the most promising solutions to achieve broadband optical response. Broadband response is an essential part of many technological advancements, from microscopy, lasers, communication, to energy harvesting, thermal management, camouflage, and many other applications. In this thesis, I have presented a noble metamaterial design for the multispectral camouflage application, which is fabricated, and the performance of the device is checked for the desired application. In addition, I have showcased the potential of the multistaged particle swarm optimization (MPSO) algorithm by applying it to design an ultra-broadband polarization converter. Due to the diverse detection mechanisms, it is important to make the optical response of a camouflage surface multifunctional, e.g., different reflection and emission properties in different frequency bands, to suppress its detectable signatures. A multispectral camouflage device requires low visible light absorption (to avoid solar induced heating), low reflection in short-wave infrared (SWIR) (0.9-1.7µm), low reflection at 1.06 µm and 1.55 µm against laser guiding, low emission in the mid infrared (MIR) (3.5-5, 8-12 µm), and high emission in the nontransmissive atmospheric window for radiative cooling. Various attempts with different material combinations have been made to achieve multispectral camouflage functionality, all of which could partially address a few spectral bands among all the bands described above. In this dissertation, I designed a noble metasurface comprising a layer of thick bottom Al followed by an ITO layer and a Si-Al grating on top that achieves the desired optical response in most of the multispectral camouflage bands. A highly stable device is fabricated using the nanofabrication techniques, such as electron beam lithography and physical vapor deposition, according to the design materials and geometric parameters. The spectral measurement of the fabricated device with optimum geometry shows a very low SWIR reflection (0.18 a.u.), low laser reflection (0.02 a.u. for 1.06 µm and 0.12 a.u. for 1.55 µm), low MIR average emission in trnasmissive window (0.24 a.u. average absorption between 3.5-5 µm, 0.05 a.u. average absorption between 8-12 µm), and high MIR average emission (0.7 a.u. average absorption) in the non-transparent atmospheric window (2.4-3.5 µm). Furthermore, exceptionally high SWIR absorption has been demonstrated to result from so-called Fano resonance. In this range, the structure yields double Fano resonance in absorption mode coupled to a Lorentzian that enables broadband absorption. On the other hand, polarization converters play a significant role by manipulating one of the fundamental characteristics of light and can enhance imaging contrast, especially for deep structures in highly scattering materials. Efforts have been made to achieve ultra-broadband polarization converters based on metasurfaces. However, the Conventional simulation-driven forward design approach demands a time-consuming optimization pathway, and the process can lead to suboptimal device performance. We implemented multi-stage particle swarm optimization to achieve optimum metasurface for ultra-broadband THz polarization conversion in transmission mode, as many real-world applications, i.e., THz communication systems, require polarization control in transmission mode. We f irst proposed a design that comprises a tri-layer metasurface structure, where a split-bridged disk-like resonator is sandwiched between two perpendicularly oriented grating layers separated by dielectric spacers. This configuration enables an efficient conversion of linearly polarized waves into their cross-polarized counterparts under normal incidence, achieving a high polarization conversion ratio exceeding 90% across the frequency range of 1.19–9.0 THz, corresponding to a relative bandwidth of 153%. Through the surface current distribution analysis at different operating frequencies, I found that the induced magnetic and electric dipole moments contribute to ultra-broadband polarization conversion.Item Embargo Characterization of secretion signals within the novel carboxy-terminal domain of mutant calreticulin in myeloproliferative neoplasms(2025-08) Puka, BeanaMyeloproliferative neoplasms (MPNs) are clonal hematopoietic disorders marked by excessive production of mature myeloid cells. In a subset of patients, mutations in calreticulin (CALR) drive the disease by constitutive activation of the JAK/STAT pathway. All CALR mutations result in a +1 frameshift that generates a novel carboxy- terminal sequence enriched in basic residues and lacking the KDEL ER-retention signal. While CALR normally resides in the endoplasmic reticulum as a multifunctional chaperone, in MPNs mutant CALR is aberrantly secreted, however the precise mechanism underlying this process remains not fully understood. In this study, we aimed to investigate the secretion mechanism of CALR mutant to better understand its role in MPN development. To this end, we employed both in silico and wet lab analysis such as ELM and CELLO prediction tools, site directed mutagenesis and Western blotting techniques. Contrary to the current understanding suggesting that CALR mutant secretion is mainly caused by the loss of the KDEL retention signal, our results indicate that the secretion is largely driven by the newly generated motifs rich in arginine residues. Furthermore, prediction scores suggest that these motifs may also direct the mutant protein to other subcellular compartments, particularly the nucleus. Understanding this mechanism is crucial, as it highlights specific motifs as potential therapeutic targets that could be exploited to develop treatments for CALR-mutant MPN patients.Item Open Access Human activity classification via trainable fractional fourier transform(2025-08) Uğur, ÖzençThe human activity classification regarding micro-Doppler characteristics has been a captivating research area since the introduction of the micro-Doppler concept in various fields such as security surveillance, healthcare monitoring/diagnostics, and gait analysis. The extraction of micro-Doppler signatures through radar sensing and subsequent time-frequency analysis of humans or other targets enables precise and reliable characterization of their motion dynamics, allowing for robust classification even in cluttered or low-resolution environments. Compared to conventional vision-based or wearable sensing methods, radar systems are favored due to their inherent ability to preserve target privacy without capturing identifiable visual features. To extract the meaningful micro-Doppler signatures embedded in the radar data, assorted time-frequency analysis approaches have been employed in the literature, especially the Short-Time Fourier Transform (STFT), due to simplicity and computational efficiency. The Fourier transform-based approach might be broadened by utilizing the Fractional Fourier Transform (FrFT), which generalizes the classical Fourier transform by considering the continuum of infinitely many representations between the time and frequency domains. Thanks to the opportunity to work in intermediate domains provided by the Fractional Fourier transform, more distinguishable representations of the micro-Doppler signatures might be obtained. Motivated by the limitations of conventional approaches that depend on fixed time-frequency representations, this thesis proposes a deep learning-based classification methodology for the micro-Doppler signature classification of human activities based on the Fractional Fourier transform. Instead of using the singular usage of the fractional domain representations by searching for optimal transform order empirically, trainable Fractional Fourier transform blocks are engaged in different deep learning architectures. Spectrograms representing the time-frequency characteristics are constructed within the deep-learning architectures, aiming for optimum representations via trainable fractional Fourier transform blocks dynamically. The proposed approach is evaluated with multiple radar-based datasets, including both simulated and real-world measurements with different radars and configurations. A comprehensive set of experiments is conducted with various deep learning architectures, including single-branch CNN, LSTM, and GRU based architectures, with the multi-branch configurations including the simultaneous use of different time-frequency representations obtained via fractional Fourier transform and multi-input Siamase-based models. The proposed approach is compared in detail with the conventional Fourier transform approach in distinct experiments, and the behavior of the fractional representations across the models and datasets is examined. As far as we are aware, this study is the first to combine a trainable Fractional Fourier Transform with micro-Doppler signature-based human activity classification. The experimental results demonstrate that the proposed approach consistently outperforms traditional Fourier Transform-based methods. Moreover, the simultaneous utilization of multiple time-frequency representations with distinct FrFT orders is, to the best of our knowledge, a novel contribution to the literature, yielding considerable performance improvements. It is believed that the findings of this study pave the way for future research on adaptive time-frequency analysis in radar-based sensing applications.Item Embargo A novel acoustofluidic platform for enhanced microstreaming and chaotic flow in single-phase and droplet microfluidics(2025-08) Vardin, Ali PourabdollahMicrofluidics has transformed the landscape of biomedical and chemical re search by enabling precise control over small fluid volumes, facilitating rapid mix ing, particle manipulation, and reagent economy in lab-on-a-chip systems. How ever, achieving efficient mixing and tunable reaction conditions within microchan nels remains a persistent challenge due to the laminar flow regime that dominates at these scales. To overcome these limitations, acoustofluidics—an emerging field that harnesses acoustic forces to manipulate fluids and particles—offers a power ful, contactless strategy for enhancing microscale operations. This thesis presents the development of novel acoustofluidic platforms tai lored for diverse biological and chemical applications, with an emphasis on li posome synthesis, flow control, and particle-based assays. In the first study, a high-efficiency liposome synthesis method is demonstrated using a sharp-edged acoustofluidic micromixer. By introducing glycerol into the aqueous phase, the size of liposomes can be precisely controlled, and by adjusting the glycerol per centage, size-tunable vesicles with improved dispersity are obtained. In the second study, a novel acoustofluidic platform is developed that combines an oscillating thin elastic membrane with vibrating trapped air bubbles to gener ate enhanced acoustic streaming. The working principle and mixing mechanism are examined both numerically and experimentally, with simulations guiding the optimization of structural parameters and predicting internal flow patterns. The device achieves exceptional mixing performance for both aqueous and viscous so lutions at flow rates up to 8000 µL/h, enabling high-throughput production of monodisperse lipid nanoparticles using both solvent-based and solvent-free meth ods. This high mixing efficiency also prevents nanoparticle aggregation, making the platform uniquely suited for synthesizing monodisperse liposomes. As a proof of concept, the effects of phospholipid type and concentration, flow rate, and glycerol content are systematically investigated, revealing a dramatic reduction in liposome size—from approximately 900 nm to 40 nm—by simply introducing 75% glycerol into the reagents. The simplicity of fabrication and operation, com bined with rapid mixing and intense agitation, positions this device as a versatile and indispensable acoustofluidic tool for nanoparticle and lipid nanoparticle pro duction, offering fine control over synthesis outcomes. In the final study, a pulsatile acoustofluidic platform is introduced to induce chaotic advection in both single-phase and droplet-based (multiphase) microflu idic systems via controlled acoustic excitation cycles. By tuning the frequency and duration of acoustic pulses, dynamic flow modulation is achieved, transition ing between laminar and chaotic regimes. This enhances mixing efficiency, droplet homogenization, and particle manipulation within confined microscale volumes. Beyond flow control, the platform is applied to a fluorescence-based protein–drug binding assay (RB–BSA interaction), where pulsatile actuation markedly im proves binding kinetics. Cell viability analysis using MCF-7 cells further confirms the system’s compatibility with biological samples. Together, these platforms demonstrate the versatility and adaptability of acoustofluidic systems for advanced microscale applications, offering scalable, tunable, and biocompatible solutions for nanoparticle synthesis, biochemical pro cessing, and droplet-based manipulation.Item Embargo Mapping urban change using streetview imagery and computer vision : case of Karaköy(2025-08) Kancan, ElifnazTraditional urban studies have relied on fieldwork and archival research. Yet, recent advances in AI and remote sensing provide scalable methods to monitor urban transformation. By com paring historical and recent Street View imagery, computer vision enables the systematic detec tion of spatial modifications. However, if not critically linked to socio-historical frameworks, such outputs risk becoming decontextualized. This study introduces the concepts of pings and signals within a mapping-based model that combines computational detection with contextual interpretation. Micro-scale interventions, such as demolitions, sign changes, or function shifts, are treated as pings. When accumulated over time, these pings evolve into signals that indicate broader shifts in land use and socio-economic dynamics. Karak¨ oy, Istanbul, is examined as a site where redevelopment and the growth of creative industries have reshaped the urban fabric. These processes have altered the area’s identity, accessibility, and spatial practices. The findings suggest that Google Street View, when used in conjunction with diachronic models, can effec tively support spatial analysis by capturing patterns of change. Nevertheless, this approach has several limitations. These include data biases, uneven temporal coverage, and limited qualita tive integration, all of which require careful and critical interpretation. Despite these challenges, street-level imagery remains a valuable source. This article demonstrates that the integration of computational and experiential data enables richer interpretations of urban change. In this con text, mapping emerges as a relational practice that links physical traces to broader contextual narratives.Item Open Access Mott transition and electron correlation effects in an octagonal quasicrystal(2025-07) Yelesti, EfeFlat-band systems, where electronic kinetic energy is quenched, provide fertile ground for studying strong correlation effects. In such systems, interactions dom inate the physics, giving rise to rich phases. While most existing studies have focused on periodic or translationally invariant systems, quasicrystals present a unique, intermediate class of materials that combine long-range order with ape riodicity. Specifically Ammann–Beenker, an octagonal aperiodic tiling, support strictly localized states due to interference effects and offer a rich platform for exploring interaction-driven physics. In this thesis, we investigate the effects of electron–electron interactions on the electronic properties of the periodic approximant of the Ammann–Beenker tiling, obtained with cut and project method, using the slave-rotor mean-field theory. We focus particularly on the metal–Mott insulator transition and the behavior of strictly localized states under correlation effects. Our analysis reveals a first order Mott transition at half-filling, with the spatial distribution of quasiparticle weights and Lagrange multipliers reflecting the underlying local geometry of the tiling. Furthermore, we demonstrate that interactions induce energy splitting between different types of localized states and can partially delocalize them via hybridization with extended states. To establish confidence in the method, we benchmark the slave-rotor approach on simple models, starting from a two-site Hubbard system and extending to a one-dimensional chain before applying it to the Ammann–Beenker tiling. By solving the resulting self-consistent equations numerically, we map out the phase diagram and explore the evolution of localized states under increasing interaction strength. These findings contribute to the broader understanding of correlated phases in aperiodic systems and highlight the role of local connectivity in shaping their electronic behavior.Item Embargo Novel joint optimization of gradient boosting decision trees and SARIMAX models for nonlinear time series regression using a state space approach(2025-08) Koç, Ahmet BerkerWe investigate nonlinear regression or forecasting in an online setting and propose an end-to-end ensemble model that addresses two key challenges: the trade-off between linear and nonlinear modeling, and the disjoint nature of the optimiza tion process. The proposed architecture seamlessly integrates a linear time series model and a nonlinear soft decision tree-based model within a unified struc ture. Specifically, the architecture employs a Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX) model to capture key linear patterns in time series data, such as seasonality and trend, alongside a soft gradient boosting decision tree (SGBDT), which is adept at modeling nonlinear dependencies and extracting features directly from raw inputs. Different from existing state-of-the-art hybrid approaches that typically rely on sequential and disjoint training strategies, for the first time in the literature, we introduce a jointly optimized hybrid model in which both the linear and nonlinear compo nents are trained simultaneously. This is achieved through a novel formulation based on state-space representations, which allows the integration of the ARMA based SARIMAX model and the SGBDT into a single state-space framework. We employ the Extended Kalman Filter (EKF), a nonlinear optimization tech nique, to efficiently perform the joint training process by leveraging derived state transition and measurement equations. Furthermore, the architecture is modular and extensible, allowing for the substitution or addition of alternative nonlinear models and other linear time series techniques, such as Exponential Smoothing (ETS), as long as state representations are obtained. The optimization compo nent is also flexible, enabling the replacement of the EKF with other techniques such as the Unscented Kalman Filter (UKF) or Particle Filters (PF). Through extensive experiments on well-known real-world datasets, the proposed approach demonstrates superior performance. We provide the source code as a publicly available repository to support further research and ensure reproducibility.Item Embargo Wearables-based user identity recognition through image representation of motion sensor data sequences and pretrained vision models(2025-07) Ünlü, Rabia ElaThe common methods employed in User Identity Recognition (UIR) and verifi cation are often vulnerable to cyber attacks, requiring more robust solutions. Mo tion sensor data and biometric data are used in tackling both the UIR and Human Activity Recognition (HAR) tasks. These tasks are mostly accomplished by using Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), and CNN-LSTM hybrid models. We propose a method that employs pretrained CNN and vision transformer-based models to achieve the UIR task by classifying image representations of sensor data. We conduct a comparative study by evaluating the performance of various pretrained networks in the image classification task by pro cessing four activity datasets comprising raw data sequences. We construct a new hybrid architecture which combines DeiT-B and DenseNet201 models in a parallel configuration. This study also compares two kinds of preprocessing methods which are spectrogram and wavelet spectrogram and introduces a novel approach that is fundamentally distinct from these methods. This technique fuses raw data, spectro gram, and wavelet spectrogram information. The DeiT-B model obtains the highest accuracy as 99.76% on the DSA Dataset; however, our new hybrid architecture that combines DeiT-B and DenseNet201 performs superior.Item Open Access Localized signal suppression in magnetic particle imaging(2025-08) Kor, EgeMagnetic Particle Imaging (MPI) is a tracer-based imaging modality offering high temporal and spatial resolution without any signal from the background tissue. MPIimages the spatial distribution of magnetic nanoparticles (MNPs). However, these MNPs can accumulate in off-target organs, such as the liver and spleen, and the resulting strong signal can obscure nearby target signals. This problem is fur ther exacerbated by the broad point spread function of the imaging system, which causes even distant off-target regions to interfere with the desired signal. This thesis proposes localized signal suppression through the use of additional DC saturation fields that locally saturate MNP magnetization. After deriving the relevant imaging equations, the requirements for the saturation field strength are determined via simulations. Next, several coil designs are evaluated for localized MNP saturation. Quantitative results show that there is a balance between sup pression efficiency and spatial spillover, guiding optimal coil design. Electromag netic interference caused by interactions between the saturation coil and the rest of the MPI hardware is investigated by analyzing the sources of secondary field inductions. Among several potential solutions to address this challenge, an L choke component is incorporated into the final circuitry to minimize interference. Imaging experiments on a custom MPI scanner demonstrate that the proposed approach effectively reduces off-target signals and enhances target-to-background contrast in MPI images. The results of this thesis demonstrate localized signal suppression as a viable and successful approach for improving contrast in MPI in the case of MNP accumulation in off-target tissues.Item Open Access Privacy-preserving data normalization techniques with multiparty homomorphic encryption for federated learning(2025-08) Coşğun, MelihData normalization is a crucial preprocessing step for enhancing model per formance and training stability. In federated learning (FL), where data remains distributed across multiple parties during collaborative model training, normal ization presents unique challenges due to the decentralized and often heteroge neous nature of the data. Traditional methods rely on either independent client side processing, i.e., local normalization, or normalizing the entire dataset before distributing it to parties, i.e., pooled normalization. Local normalization can be problematic when data distributions across parties are non-IID, while the pooled normalization approach conflicts with the decentralized nature of FL. In this the sis, we explore the adaptation of widely used normalization techniques to FL and define the term federated normalization. Federated normalization simulates pooled normalization by enabling the collaborative exchange of normalization parameters among parties. Thus, it achieves performance on par with pooled normalization without compromising data locality. However, sharing normaliza tion parameters such as the median introduces potential privacy risks, which we further mitigate through a robust privacy-preserving solution. Our contributions include: (i) We systematically evaluate the impact of various federated and local normalization techniques in non-IID FL scenarios, (ii) We propose a novel ho momorphically encrypted k-th ranked element (and median) calculation tailored for the federated setting, enabling secure and efficient federated normalization, (iii) We propose privacy-preserving implementations of widely used normalization techniques for FL, leveraging multiparty fully homomorphic encryption (MHE).Item Embargo Physics-constrained unsupervised deep learning for accelerated diffusion MRI(2025-08) Topcu, AtakanDiffusion Magnetic Resonance Imaging (dMRI) is a noninvasive technique that probes the microscopic Brownian movement of water molecules within neural tissues, providing insights into the underlying microstructural architecture. In dMRI, the displacement of spins is encoded in a domain called q-space through the use of diffusion-sensitizing gradients. Classical dMRI models, such as diffu sion tensor imaging (DTI), require only a few samples in q-space, but fall short in resolving crossing or diverging fiber bundles. To address these limitations, High Angular Resolution Diffusion Imaging (HARDI) was introduced to enhance f iber characterization by densely sampling the q-space across multiple spherical shells defined by different b-values, thereby detecting several fiber orientations within a single voxel. Building on this framework, advanced multi-shell tech niques such as Multi-Shell Spherical Deconvolution (MSMT-CSD) and Neurite Orientation Dispersion and Density Imaging (NODDI) have been developed, of fering refined insights into complex microstructural features. Nevertheless, the requirement for densely sampling q-space renders advanced dMRI techniques ex tremely time-consuming and impractical for clinical use. This thesis proposes a deep unsupervised Q-space Upsampling via physics-Constrained Coordinate based Implicit network (QUCCI) to accelerate multi-shell dMRI. QUCCI models the underlying volume as a continuous function in both spatial coordinates and q-space, enabling the sampling of q-space along arbitrary directions without the constraints of fixed sampling schemes. An encoder maps coordinates to a la tent code, and an MLP predicts the signal, allowing arbitrary q-space sampling without large training datasets or vendor harmonization. Physics-based regu larization stabilizes learning. Tested on 10 subjects at R = 10, 15, 22.5, and extended to joint q-space interpolation plus in-plane super-resolution for submil limeter whole-brain dMRI, QUCCI surpasses a recent deep-learning competitor, a least-squares baseline, and raw undersampled data. Slice-, subject-, and metric level evaluations, and downstream DTI, MSMT-CSD, and NODDI maps confirm its superior fidelity. QUCCI enables accelerated dMRI with minimal information loss, advancing the clinical feasibility of advanced multi-shell methods.Item Embargo Optimization over trained graph neural networks with an application in brain tumor resection(2025-08) Çakıroğlu, KaanLow-grade gliomas (LGGs) present a critical challenge in neurosurgical oncology due to their slow progression and the brain’s adaptive neuroplastic reorganization. While complete resection reduces recurrence risk, it may disrupt reorganized functional networks. Emerging evidence shows partial resection can preserve neurological outcomes without compromising tumor control. However, surgeons lack quantitative tools to preoperatively evaluate these tradeoffs. The absence of a computational framework to model resection trade offs leaves surgical planning reliant on subjective intraoperative assessments rather than predictive network analysis. Our research addresses this gap by formulating the brain tumor resection problem as a mathematical optimization model using graph neural networks (GNNs), where the brain is represented as a multiplex network comprising regions of interest (ROIs) with structural and spatial connectivity. In our formulation, GNNs approximate brain functionality metrics such as global efficiency and modularity, serving as surrogate objective functions within a mixed-integer programming framework. We extend this methodology to bi-objective optimization, systematically analyzing trade-offs between several brain functionality metrics. Our framework provides surgeons with data-driven strategies that balance maximal tumor control with minimal network disruption. It achieves solutions whithin 0.34% of the true optimium and efficiently approximates the Pareto frontier, enabling precision neurosurgery tailored to individual brain networks.Item Open Access Predictive modeling of vehicle failures with hierarchical Bayesian methods for workforce planning(2025-08) Koçak, Doğuş BerkVehicles that operate under demanding conditions need an understanding of failures to ensure reliability and take appropriate actions. To address this, a statistical framework is developed for modeling failure times using real-world operational data. The approach employs Bayesian Generalized Linear Mixed Models to capture unit and vehicle-level effects, and intervention effects. A sequential simulation framework models temporal dependencies and generates multi-step failure predictions with full uncertainty quantification. The proposed model and simulation approach are evaluated to demonstrate both calibration and predictive performance. Additionally, the work shows how predictive outputs can inform decision-making by deriving new system-level metrics and assessing their reliability. Finally, the results are applied in a representative sequential decision-making problem on workforce planning for repair actions.