26734 research outputs found
Sort by
AI Literacy for Student Success Series Workshop #2
No full textThis workshop examines how artificial intelligence (AI) tools can support the literature review process in academic research. Participants will explore how tools such as ChatGPT, Scite, and ScopusAI can assist with topic exploration, research question refinement, identifying key themes, summarizing sources, and generating synthesis. The session will emphasize how AI can help with organizing citations and structuring arguments, while underscoring that critical thinking and scholarly judgment remain essential. Attention will also be given to the limitations of AI, including hallucination risks, lack of access to paywalled content, and outdated training data. By comparing AI-generated outputs with results from academic databases, participants will develop strategies for evaluating and responsibly integrating AI into the research process.
Learning Objectives
By the end of this workshop, participants will be able to:
Apply AI tools to support topic exploration, research question refinement, and identification of key themes.
Compare and contrast AI-generated outputs with results from academic databases to assess credibility, accuracy, and comprehensiveness.
Evaluate the affordances and limitations of AI tools in relation to the literature review process, including issues of hallucination, access restrictions, and outdated data.
Develop strategies for integrating AI responsibly into citation management, synthesis, and argumentation while maintaining critical engagement with scholarly sources
Real-World Gait Analysis for Older Adults: Identifying Optimal Gait Bouts and Reconstructing Lower-Back Signals from Wrist-Worn Devices
No full textEarly detection and monitoring of cognitive impairment remain major challenges in the management of Alzheimer’s disease and related conditions. While traditional biomarkers provide valuable insights into underlying pathology, they are often costly, invasive, and limited to discrete clinical visits. Gait has emerged as a promising digital biomarker of cognitive function, reflecting the interaction between motor control and higher-order cognitive processes. Wearable sensors enable continuous, real-world gait monitoring; however, methodological challenges related to data heterogeneity, noise, and sensor placement have limited their clinical utility. This thesis investigates real-world gait assessment in clinical populations using wearable inertial sensors, with a focus on improving interpretability and scalability. First, an unsupervised framework is proposed to identify representative, lab-quality walking bouts from continuous real-world accelerometer data. This approach isolates steady-state gait cycles without reliance on predefined templates or labelled data, enabling robust gait characterization under naturalistic conditions. Second, frequency-domain methods are used to reconstruct lower-back kinematic gait information from wrist-worn sensor data, addressing the trade-off between biomechanical fidelity and user compliance. The proposed methods were evaluated using synchronized wrist and lower-back inertial measurement units collected during both controlled laboratory assessments and long-term real-world monitoring in older adults across clinical groups. Results demonstrate that representative gait patterns can be reliably extracted from real-world data and that wrist-derived signals capture key frequency-domain characteristics of trunk motion. Together, these findings support the feasibility of low-burden, wearable-based gait monitoring as a scalable approach for assessing mobility and cognitive health. This work contributes methodological advances toward the development of real-world gait-based digital biomarkers for cognitive impairment
Tailored g-C3N4 Architectures for High-Performance Photodetection, Ultra-Trace Heavy Metal Sensing, and CO2-to-Methanol Photocatalysis
No full textGraphitic carbon nitride (g-C3N4) is a metal-free 2D semiconductor widely recognized for its low cost, chemical stability, and visible-light activity, offering a sustainable platform for next-generation energy and sensing technologies. This dissertation establishes a materials engineering framework to tailor g-C3N4 at structural, electronic, and interfacial levels for high-performance multifunctional applications, including optoelectronics, environmental sensing, and solar fuel conversion. In the first part of this work, a nitrogen (N) self-doped g-C3N4 (g-C3N4+) was synthesized and coupled with carbon quantum dots (CQDs) to create a 2D/2D p–n heterostructure for visible (Vis) -light organic photodetector (OPD). The N self-doping enriches the conduction band density of states and narrows the bandgap (to ~1.2 eV), while CQDs provide size-tunable band alignment and fast carrier transport. Under zero bias, this OPD delivered ultrahigh alternating current performance across the Vis spectrum—achieving a specific detectivity of 4.6 × 1018 Jones, responsivity of 1.43 × 107 A W⁻¹, and external quantum efficiency of 43 × 107 % at an optical intensity of 3.56×10-4 mW/cm2 and a wavelength of 405 nm while delivering competitive performance at 532 nm and 635 nm as well. Frequency-domain analysis revealed symmetric rise and decay times (~2.9 ms) at high modulation frequencies and stable signal generation even at subzero temperatures, underscoring the potential of this platform for low-power, high-frequency photodetection in harsh environments. Building on this optoelectronic characteristic in the second part of this work, a dual-functionalized CQD-(NH2-COOH)/g-C3N4 heterostructure was developed for ultrasensitive and selective detection of hexavalent chromium (Cr6+) in water. CQD-(NH2-COOH), produced from spent coffee grounds via a one-step ultrasonic process, were non-covalently assembled onto g-C3N4 nanosheets to construct a 2D/2D hybrid interface with covalent, hydrogen-bonding, and π–π interactions. This interfacial engineering introduced mid-gap bands that broadened emission bandwidth and enhanced fluorescence quenching efficiency, enabling trace-level Cr6+ detection down to 70 pM, well below the WHO guideline of 96 pM, across a wide dynamic range (0.1 nM–100 µM) in the presence of 12 competing ions. This sustainable sensing approach demonstrates the ability of engineered CQD-(NH2-COOH)/g-C3N4 to achieve a portable, low-cost, field-deployable environmental sensor. To advance the material platform toward energy conversion, a 2D cobalt (Co) and N doped heptazine-based g-C3N4 (Co-g-C3N4+) photocatalyst was designed for visible-light-driven artificial photosynthesis of methanol from carbon dioxide (CO2). In this work, both heptazine- and triazine- based g-C3N4 framework, as well as effect of single and dual doping, were systematically investigated. The heptazine-based-g-C3N4 framework was selected over triazine structures due to their ~30 kJ mol⁻¹ higher thermodynamic stability, which enhances long-term durability, and offers a more extended π-conjugation network for charge transport. The heptazine units’ larger pore sizes and altered electronic configuration facilitate stronger CO2 adsorption and activation, while cobalt dopants enrich active sites offering variable oxidation states that also promote the adsorption of CO2. Nitrogen doping created localized midgap-band and enhances the visible light absorption and improves electron mobility. This catalyst achieved a methanol production rate of 2-3 mmol g-1h-1 and methanol yield of 27.72 %, highlighting its promise for solar-driven carbon recycling and sustainable fuel production. Together, these three parts demonstrate a cohesive design strategy for multifunctional g-C3N4-based nanostructures, linking electronic band structure engineering, heterointerface design, and surface catalytic modification
Koszul Duality for Generalized Steinberg Representations of p-adic Groups
No full textLet G be a semisimple group, split over a p-adic field F. We prove that the category of modules over the extension algebra of generalized Steinberg representations of G(F) is equivalent to a full subcategory of equivariant perverse sheaves on the variety of Langlands parameters for these representations. Specifically, we establish an equivalence Mod(Ext•G(Σλ, Σλ)) ≃ Per◦bG(Xλ), where Σλ is the direct sum of generalized Steinberg representations and Per◦ bG(Xλ) is the subcategory of perverse sheaves on the variety of Langlands parameters Xλ corresponding to these representations under Vogan’s geometrization of the Langlands correspondence. Furthermore, we demonstrate that this equivalence is a true Koszul duality by showing that the extension algebra of generalized Steinberg representations is Koszul dual to the endomorphism algebra of the direct sum of corresponding equivariant perverse sheaves, taken in the equivariant derived category DbbG(Xλ)
Fairness Engineering for Machine Learning Systems
No full textIn machine learning (ML) research and AI regulatory documents, the fairness of AI and ML models and their safe adoption into real-world systems are strongly emphasized. However, little work has focused on the engineering of ML systems that aim to integrate such fairness-aware models into production workflows. In practice, fairness and performance are often orthogonal: attempts to optimize one can destabilize the other. Moreover, existing tools provide limited guidance for incorporating fairness constraints into automated model development pipelines or for evaluating and mitigating bias in free-form text generated by large language models (LLMs). As a result, practitioners lack comprehensive, end-to-end methods for building AI systems that are both accurate and aligned with emerging fairness requirements. This thesis addresses these gaps through two complementary lines of contribution spanning structured ML pipelines and generative LLM systems. The first contribution improves the fairness–aware automation of structured ML models. FairSpace enhances upstream AutoML search by introducing LLM-assisted fea-ture engineering, a joint fairness–accuracy optimization objective, and selective pruning to reduce search ine!ciency. It achieves win–win (fair and accurate) outcomes in 63% of cases, reduces computation time by approximately 25%, and places 94% of evaluations in win–win and good trade-o” regions. Complementing this, FairGA refines downstream model selection using an evolutionary search guided by fairness signals extracted from the AutoML run history. It delivers win–win outcomes in 64% of cases and positions 100% of results within the win–win and good trade-o” regions across benchmark datasets, demonstrating stable and reliable fairness improvements beyond the limits of standard AutoML. The second contribution focuses on fairness evaluation and mitigation in LLM-generated text. The thesis introduces Biascore, a composite metric that measures polarity, stereotype cues, counterfactual asymmetry, and semantic framing to capture subtle forms of linguistic and contextual bias. Building on this metric, the thesis presents AutoFairer, an automated pipeline for detecting and sanitizing biased text using binary classification, token-level extraction, controlled rewriting, and recalibration through Biascore. Across one hundred biased text samples, AutoFairer achieves 96.2% agreement under human judgment and 98.1% agreement under LLM evaluation, outperforming all existing debiasing frameworks. Together, FairSpace, FairGA, Biascore, and AutoFairer provide an integrated approach to engineering fairness-aware ML and LLM systems. These contributions improve the consistency, scalability, and respon-sibility of automated AI workflows, o”ering concrete and practical methods for aligning ML development with emerging standards for fair and accountable AI
Modeling and Forecasting Extreme Electricity Prices in Competitive Markets
No full textThis thesis examines the problem of modeling and forecasting electricity price spikes in competitive electricity markets. These extreme price events create both risks and opportunities for the market participants. Spike modeling supports risk-management decisions. The challenging nature of price spikes highlights the need for continuous exploration of diverse methodologies. The modeling frameworks presented across the chapters in this thesis combine machine learning, including deep learning, and statistical approaches. Before introducing these frameworks, the thesis provides a dedicated foundations chapter that reviews key concepts in electricity markets, price behavior, and spike analytics. The first framework, focused on spike occurrence forecasting, introduces a statistical–economic analysis that evaluates multiple models under different spike threshold definitions. The second framework applies survival analysis techniques, specifically the Kaplan–Meier estimator, to quantify spike duration. Both chapters build on the fundamental definition of a price spike as a two-dimensional phenomenon characterized by magnitude and duration. A third framework integrates deep learning with dynamic sparse training to forecast normal electricity prices. Extending this approach, the fourth framework combines deep learning, dynamic sparse training, and extreme value theory to predict extreme prices using the block maxima method. The proposed methodologies are tested on several electricity markets featuring both day-ahead and real-time structures. In summary, no single approach can fully address the complexity of modeling and forecasting electricity price spikes, making it essential to explore diverse methodologies. Furthermore, as operational uncertainty in electricity markets continues to grow, advancing this research requires an ongoing development of techniques that can adapt to the evolving market conditions and support decision-making
Targeted deletion of EMMPRIN in microglia/macrophages mitigates neuronal death in intracerebral hemorrhage
No full textAbstract Background Intracerebral hemorrhage (ICH) is a devastating subtype of stroke with high mortality and limited therapeutic options. Microglia and macrophages are rapidly recruited to the lesion site and contribute substantially to secondary brain injury. However, the key molecular mediators that drive their neurotoxic effects remain incompletely understood. Methods We investigated the role of extracellular matrix metalloproteinase inducer (EMMPRIN, also known as CD147) in promoting microglia/macrophage-mediated neurotoxicity after ICH. EMMPRIN was selectively deleted in myeloid cells using both AAV-mediated knockdown and CX3CR1Cre:EMMPRINfl/fl mice. Neuronal survival and functional outcomes were assessed using histological, molecular, and behavioral analyses. Results Targeted deletion of EMMPRIN in microglia/macrophages significantly reduced neuronal death and improved neurological recovery following ICH. Mechanistically, EMMPRIN-mediated neurotoxicity was associated with elevated expression of matrix metalloproteinases and enhanced activation of the p38 mitogen-activated protein kinase (MAPK) pathway, and with downstream engagement of myocyte enhancer factor 2 C (MEF2C) and B-cell lymphoma 2 (Bcl2). Notably, EMMPRIN deletion also enhanced neurogenesis and oligodendrogenesis in the perihematomal region, suggesting a potential role in promoting endogenous brain repair. Conclusions These findings establish EMMPRIN elevation in myeloid cells as a prominent regulator of ICH pathophysiology and a promising therapeutic target to limit secondary injury and promote brain repair. Graphical Abstrac
Exploring Mitochondrial Dysfunction in Autistic Individuals
No full textMitochondrial dysfunction has been widely implicated in neurodevelopmental conditions such as autism and attention-deficit/hyperactivity disorder (ADHD), yet findings across studies remain inconsistent. This thesis examines whether mitochondrial involvement reflects categorical pathology or biologically meaningful inter-individual variation by integrating non-invasive mitochondrial genomic and functional analyses in a pediatric cohort. Using a novel buccal and urine-derived cell collection and culture protocol, mitochondrial DNA (mtDNA) sequencing and high-resolution respirometry were performed in autistic children, children with ADHD, and neurotypical controls. Total mtDNA variant burden was largely conserved across diagnostic groups, indicating that global mitochondrial disruption is not a defining feature of diagnosis. In contrast, locus-specific variation, particularly within Complex I–encoding genes and mitochondrial RNA regions, differentiated individuals independent of diagnostic category. Functional bioenergetic assessments revealed substantial inter-individual variability in ATP-linked respiration, coupling efficiency, proton leak, and reserve capacity, often exceeding between-group differences. Subtle Complex I–associated variation was linked to specific respiratory parameters, providing a mechanistic connection between genomic variation and functional bioenergetic regulation. Collectively, these findings support a model in which mitochondrial biology acts as a modulatory system, shaping bioenergetic flexibility rather than producing uniform dysfunction. Overlapping mitochondrial profiles between autism and ADHD further support a transdiagnostic framework in which shared bioenergetic mechanisms contribute to neurodevelopmental variability
Computational Thermodynamics of Systems Containing Associating Molecules
No full textThermodynamic modeling is a powerful tool in studying the behavior of physical systems, guiding experimental design and operational conditions, and predicting equilibrium properties. A powerful thermodynamic model saves time and resources in experimental studies and in the design of industrial processes, enabling more efficient and safer operation. In this thesis, the phase equilibria of systems containing bitumen, water, light and intermediate alkanes, toluene, hydrogen, dimethyl ether, and monoethynene glycol are studied by using two- and three-phase flash calculations and stability analysis. The utilized EoSs are based on Wertheim’s thermodynamic perturbation theory (TPT). The cubic plus association (CPA) EoS, based on the first-order thermodynamic perturbation theory (TPT1), is applied to normal alkane/bitumen, water/bitumen, water/normal alkane, and water/bitumen/normal alkane. Initially, the CPA EoS is adjusted by using the liquid-liquid equilibria (LLE) data, which allows the LLE-tuned model to predict vapor–liquid equilibrium (VLE) behavior successfully. Subsequently, the CPA EoS is further parametrized using experimental data of normal alkane/bitumen, water/bitumen, and water/normal alkane systems. The model's predictive capability is then investigated using experimental three-phase data for the water/bitumen/normal alkane system. The model's predictive capability motivated and enabled the development of a ternary phase diagram of water/bitumen/normal alkane. The phase equilibria regions are constructed as a function of feed composition. This computational framework is then validated by experimental data. In addition to the CPA EoS, the perturbed-chain statistical associating fluid theory (PC-SAFT) is employed to evaluate its predictive capability for the phase behavior of the water/synthetic condensate system, using binary-tuned models to predict the multicomponent mixture behavior. The Tz diagram of the system is constructed using both models, and their predictions are validated against experimental data. The models successfully detected the phases in all experiments and predicted the trend in molar composition. Subsequently, higher-order terms of TPT are used. The theory is developed for a spherical particle with four patchy sites and validated by Monte Carlo (MC) simulation. The model is applied to pure monoethylene glycol and its binary mixtures with light gases. The theory outperformed the TPT1-based model in predicting liquid density and vapor pressure
Modelling Temporal Lobe Epilepsy with Sclerosis in the SSP-Saporin ‘Trojan Horse’ Model
No full textTemporal lobe epilepsy with hippocampal sclerosis (TLE-HS+) is a common and often refractory form of human epilepsy. Stable Substance P-Saporin (SSP-SAP) is a neurotoxin that selectively targets inhibitory interneurons and has been used to determine whether a focal GABAergic defect in an otherwise normal brain is sufficient to initiate the epileptogenic process that results in TLE-HS+. When injected unilaterally into multiple sites along the longitudinal axis of the rat dentate gyrus, SSP-SAP caused selective inhibitory neuron death, followed by reactive seizures, and eventually self-generated hippocampal-onset epileptic seizures. In this study, we first examined how long SSP-SAP is detectable in inhibitory interneurons following unilateral injection. Next, we analyzed the EEG characteristics and behavioral expression of the “reactive” seizures that developed several days after SSP-SAP injection. We quantified the progressive increase in frequency of self-generated 7-Hz epileptiform activity over 3 months (endogenous kindling). Finally, we counted the numbers of principal cells and astrocytes up to 2 months post-injection. We observed that SSP-SAP was internalized within 2 hours following delivery and was undetectable within 5 days. Rats exhibited reactive behavioural seizures between days 4-6 following SSP-SAP administration, with most seizures having a Racine score of either 1 or 2. We observed that 7-Hz epileptiform activity, associated with behavioural arrest, continued after the reactive seizures had abated and became more frequent over the 3-month observation period. We also observed the continuous loss of principal cells and continuing increase in astrocytes over time, two defining features of hippocampal sclerosis, a common pathological observation in TLE patients. This study confirms that the SSP-SAP model reproduces the defining features of human TLE and that a primary and selective GABAergic defect is sufficient to trigger epileptogenesis that results in TLE-HS+