Masters Theses

Permanent URI for this collection

https://hdl.handle.net/10919/9291

Browse

Now showing 1 - 20 of 23265

QuOTE: Question-Oriented Text Embeddings
Neeser, Andrew Kyle (Virginia Tech, 2025-06-13)
We present QuOTE (Question-Oriented Text Embeddings), a novel enhancement to retrieval- augmented generation (RAG) systems, aimed at improving document representation for accurate and nuanced retrieval. Unlike traditional RAG pipelines, which rely on embed- ding raw text chunks, QuOTE augments chunks with hypothetical questions that the chunk can potentially answer, enriching the representation space. This better aligns document embeddings with user query semantics, and helps address issues such as ambiguity and context-dependent relevance. Through extensive experiments across diverse benchmarks, we demonstrate that QuOTE significantly enhances retrieval accuracy, including in multi-hop question-answering tasks. Our findings highlight the versatility of question generation as a fundamental indexing strategy, opening new avenues for integrating question generation into retrieval-based AI pipelines.
Development of a Workflow to study Neuronal Injury In Vivo in the Retina and In Vitro in Collagen Hydrogels
Kaplan, Amber Leigh (Virginia Tech, 2025-06-13)
Ocular trauma affects 1.5-2 million individuals and is the fourth leading cause of blindness in the United States (Bourne et al., 2021a; Hashemi et al., 2023). In the military, the frequency of blasts has increased during recent military conflicts where 186,555 active-duty military personnel are now diagnosed with ocular injuries (Hilber, 2011). This study used in vitro and in vivo models to determine the cellular responses to mechanical trauma. A whole-body blast injury model in rats was used to observe changes in retinal structure and function in Aim 1. To study the mechanical mechanisms of injury in isolation, an in vitro model of compression to neuronal cells was developed in Aim 2. The first aim of the study employed an established preclinical blast model to expose the entire body of rats to three blast waves, each one hour apart. Immunohistochemistry was performed on retinas isolated twenty-nine days after blast exposure. Major differences in immunolabeling were found between retinas from the blast-exposed group and the sham group. Immunolabeling against RNA Binding Protein with Multiple Splicing (RBPMS), revealed significantly fewer retinal ganglion cell (RGC) somas in the blast-exposed group than the sham group (P < 0.001). Nitrotyrosine, an indicator of oxidative stress, was elevated in the ganglion cell layer of the blast-exposed group. Müller cells of the retina express glial fibrillary acidic protein (GFAP). GFAP expression was similar between the two groups. The whole-body blast model resulted in degeneration of RGCs and heightened oxidative stress in the ganglion cell layer, but no activation of Müller cells 29 days after exposure to blast. Therefore, this rat model of traumatic injury results in pathology of RGCs in the retina and needs to be further studied to determine the mechanisms underlying blast-induced retinal injury. An in vitro compression model was developed to study the effects of mechanical stress on neuronal cells using a 3D platform. Additionally, a protocol was established to differentiate the SH-SY5Y neuroblastoma cell line into neurons in 3D collagen hydrogels. The protocol determined the optimal collagen concentration and seeding density. SH-SY5Y differentiation was effective in a 0.5 mg/mL and 1.0 mg/mL collagen hydrogels seeded at a density of 6x105 cells/mL. Next, a method to statically compress collagen hydrogels between 0-18% was developed. This compression model can be used to study the mechanical response of neurons, such as differentiated SH-SY5Y cells, or retinal ganglion cells, in a 3D environment. Together, the two aims present the opportunity to better understand the mechanisms underlying neuronal injury caused by mechanical stress both in vivo in the retina and in vitro in a 3D environment.
Multimaterial, multifunctional fiber-based designs for integration into minimally invasive and wearable, translatable medical devices
Charlton, Alyssa Mary (Virginia Tech, 2025-06-13)
Fiber-based medical devices have the potential to contribute to both research and clinical applications through their multifunctionalities, material properties, compact sizes, and customizable geometries. This thesis discusses current developments and testing of such devices. A minimally invasive interstitial fluid glucose biosensor as well as a wearable sweat sensing glucose biosensor are presented. Moreover, a conductive fiber design is evaluated for pH sensing and integrative capabilities. Additional work and future areas of interest are discussed, exploring the versatility and reach of fiber-based devices.
Tenko: A Zero Trust Inspired Framework for Real-Time Network Defense via Intelligent Thresholding and Node-Level Anomaly Scoring
Bhola, Sahil (Virginia Tech, 2025-06-13)
Evaluating Model-Estimated Shoulder Muscle Activity During Overhead Work with Varied Task Demands and Exoskeleton Use
Li, Lingyu (Virginia Tech, 2025-05-09)
Passive arm support exoskeletons (ASEs) have emerged as an intervention that can reduce shoulder stress during overhead tasks. However, the effects of these devices are task- and device-specific, and current evaluation protocols remain time-consuming and resource-intensive. Musculoskeletal modeling could simplify the process of ASE evaluation, by replacing electromyography (EMG) sensors with estimates of muscle activation. However, there is no existing evidence to determine whether model performance with ASE is sufficient or consistent under varied task demands. In this study, I evaluated estimates of shoulder muscle activity generated by one commercial biomechanical model, during dynamic overhead push tasks at different heights and directions, both with and without an ASE. Kinematics and external load data were input into the AnyBody Modeling System to simulate muscle activation. Model estimates were then compared to normalized EMG using pattern similarity and magnitude difference metrics. Overall, the results obtained demonstrated good model performance with relatively smaller arm elevation, but that model performance decreased as arm elevation increased and that ASE use further impaired model performance. These findings indicate that model-estimated shoulder muscle activity is reasonably accurate under specific task conditions. However, improvements to musculoskeletal models are necessary to make these models suitable for a broader range of tasks.
Reinforcement Learning with a Lost Person Model for Search and Rescue Path Planning
Howell, Bryson L. (Virginia Tech, 2025-05-12)
In this thesis, we train a reinforcement learning agent to plan paths for search and rescue applications using a model of lost person behavior trained on past search incidents. We propose an improved method for producing occupancy maps from the trajectories of an agent-based lost person model. We demonstrate that through an end-to-end learning approach our agent can generalize to novel search incidents without directly observing the probability distribution describing search risk.
On Privacy in Group Testing
Liu, Shuqi (Virginia Tech, 2025-05-06)
Group testing is a method of designing collections of samples of individual items and assessing them as collections (rather than individuals) with the goal of revealing individuals who are positive for some attribute. It has been used to test for highly contagious diseases, such as coronavirus in recent years, often with the goal of minimizing processing time or the number of tests. Privacy of personal information is important, particularly when it comes to medical history or test results for diseases. Our research studies group testing designed with parity-check matrices of Hamming codes and constant row weight d-disjunct matrices. We consider partial knowledge that an eavesdropper needs to know from the group testing matrix to obtain personal medical data. We also evaluate the leakage risk of the information under certain assumptions of the eavesdropper's abilities. The thesis concludes by proposing future directions such as handling noise, correlated individuals, and decentralized testing designs.
A Comparison of Autonomous Navigation Methods for Earth-Moon Halo Orbits
O'Leary, Colin M. (Virginia Tech, 2025-05-09)
With a renewed interest in lunar travel on the horizon, it is becoming clear that our current methods of satellite tracking for spacecraft around the Moon are unequipped to handle this increase in demand. Therefore, a reliable method of performing autonomous navigation in lunar space is required to facilitate this new interest. Many methods of performing this autonomous navigation have been devised, but little work has been done to compare these methods to one another using standardized, relevant test conditions. In this study, we compared three different methods of performing autonomous navigation in six different test cases. The methods being tested were the star occultation method, the optical navigation method, and the lunar mirror method. These six test cases were a selection of halo orbits in the Earth-Moon system (a set of sensitive orbits dependent on Earth-Moon dynamics). We simulated the true state information for each of these test cases for a period of 30 days. We also simulated the true values of the measurements that would have been obtained by each of our three methods using this same software. Then to test the methods, we propagated the initial state of each test case using an estimator equipped with one of our measurement methods under test. This was repeated for each method, for a total of 18 test results. It was found that the optical navigation method was the best performing for all six of our test cases.
Uncertainty Quantification in Security Aware Data Pipelines
Dadeboe, Alberta O. (Virginia Tech, 2025-05-09)
With the recent rise in connected devices through the Internet of Things and interconnected cyberphysical systems, the diversity and volume of data have expanded. Proper management of sensitive information collected and processed through data pipelines is crucial. Traditional data pipelines usually perform error analysis of the final pipeline output after a detection model. As a result, they miss malicious attacks or data corruption that occur earlier in the pipeline. Providing assurance of security throughout all stages of pipeline processing can improve credibility at a more fine-grained level. This thesis introduces a combination of data pipeline augmentation capabilities aimed at estimating the uncertainty of computations with constant monitoring of trends in shifts in data at every pipeline stage. The proposed framework integrates uncertainty quantification (UQ), data provenance tracking, sensitivity analysis, and tunable alerts to understand parameter influence on function outputs, methodically detect potential corruptions, maintain a meticulous audit trail, and prompt observers during suspicious activity. This contribution advances conventional data pipeline anomaly detection by providing combined fault-sensitive execution and full-fault traceability with continuous estimation of uncertainty for each pipeline stage.
Investigating the Roles of CwlD and GerS in Clostridioides difficile Spore Cortex Modification
Nawar, Shaeri (Virginia Tech, 2025-05-12)
Clostridioides difficile is a notable nosocomial pathogen distinguished by its capacity to produce spores, essential for its survival and dissemination. The spore cortex, consisting of a modified peptidoglycan layer, is essential for dormancy and germination. This work examines the enzyme CwlD, which facilitates the transformation of N-acetylmuramic acid (NAM) into muramic-δlactam (MAL)—a change crucial for adequate cortex production and germination. The lipoprotein GerS is essential for CwlD function in C. difficile; however, no GerS homolog is present in Bacillus subtilis, a well-established model for spore biology. To assess the species requirements for CwlD and GerS, we expressed FLAG-tagged versions of C. difficile CwlD, both in the presence and absence of GerS, in a B. subtilis ΔcwlD background. Despite the effective assembly and integration of all genetic constructs, no expression of FLAG-tagged protein was seen. The data indicates that factors such as translational inefficiency or protein degradation may impede the expression of these heterologous proteins in B. subtilis. This contrasts with prior research demonstrating observable production of variants even though they remained non-functional in B. subtilis. These findings highlight the significance of the native cellular environment for the functional expression of CwlD–GerS and enhance the comprehension of species-specific regulatory mechanisms in spore life. These results may guide future therapy strategies aimed at targeting C. difficile germination by inhibiting CwlD-GerS-mediated MAL production. This would impede spore germination, presenting a potential approach to avoid infection initiation and diminish recurrence by maintaining remaining spores in a dormant state.
Transport of Per- and Polyfluoroalkyl Substances (PFAS) at a Managed Aquifer Recharge Site
Nice, Shannah Marie (Virginia Tech, 2025-06-12)
In response to aquifer overuse and its related issues in eastern Virginia, the Hampton Roads Sanitation District (HRSD) is conducting a managed aquifer recharge (MAR) project called the Sustainable Water Initiative For Tomorrow (SWIFT). The objective of the SWIFT program is to treat wastewater to drinking water standards before injecting it into the underlying Potomac Aquifer System (PAS). SWIFT is an aquifer long-term replenishment (ALTR) designed to increase the storage and potentiometric surface of the PAS. Previous work used simple analytical models to simulate the transport of conservative constituents present in SWIFT Water and delivered to the PAS at the HRSD SWIFT Research Center (SRC). Using the previously calibrated model, this study describes adaptations to the model for simulating the transport and attenuation of selected per- and polyfluoroalkyl substances (PFAS) in groundwater at the SRC. The transport of seven PFAS detected at the SRC is examined in this study: perfluorobutanoic acid (PFBA), perfluorobutane sulfonic acid (PFBS), perfluoropentanoic acid (PFPeA), perfluorohexanoic acid (PFHxA), perfluorohexane sulfonic acid (PFHxS), perfluoroheptanoic acid (PFHpA), and perfluorooctanoic acid (PFOA). Among these PFAS, five are short-chain compounds (PFBA, PFBS, PFPeA, PFHxA, and PFHpA) and two are long-chain compounds (PFHxS and PFOA). The fluorocarbon chain length among these compounds ranges from four (PFBA) to seven (PFOA). Two compounds contain a sulfonic acid functional group (PFBS and PFHxS) and five compounds contain a carboxylic acid functional group (PFBA, PFPeA, PFHxA, PFHpA, and PFOA). Sorption is modeled at the SRC using retardation factors, which were calibrated for each PFAS to best fit the observed concentrations at a PAS monitoring well. Calibrated retardation factors are then used to calculate PFAS-specific partition coefficients (Kd) and organic carbon partition coefficients (Koc) and compared to literature values. In general, best-fit retardation factors increase with PFAS chain length, with the exception of PFHxS, whose calibration is challenged by lack of data and low levels of detection. Calculated Kd values fall largely within the range of representative values reported in the literature, with the exception of PFBA. Some calculated Koc values approach or exceed the upper ranges of values observed in the literature, suggesting that sorption to natural organic carbon in aquifer sediment does not fully account for PFAS sorption in the PAS. Overall, this study presents a unique data set and novel analysis of PFAS transport and attenuation in an oligotrophic aquifer under complex field conditions.
Evaluating the Effects of Financial Deregulation on Bank Risk using Double Machine Learning
Shah, Gaurav Kandarp (Virginia Tech, 2025-06-12)
This work examines the causal impact of deregulation within the U.S. banking sector, fo- cusing on the rollback of a specific provision of the Dodd-Frank Act through the Economic Growth, Regulatory Relief, and Consumer Protection Act of 2018 (EGRRCPA). Originally enacted in response to the 2008 financial crisis, the Dodd-Frank Act introduced extensive regulatory reforms aimed at mitigating systemic risk. However, the partial repeal of its provisions has prompted renewed interest in assessing the implications for bank risk and financial stability. Our research contributes to this growing body of work by employing recent developments in causal inference, particularly Double Machine Learning (DML), to more accurately estimate treatment effects. DML leverages machine learning algorithms to flexibly model both treatment and outcome processes, controlling for bias via orthogonaliza- tion techniques and sample-splitting strategies. By applying DML to panel data, we address the complexities of policy evaluation with panel data and aim to improve the robustness of causal estimates. We conduct a reanalysis of Chronopoulos et al. [12], comparing estimates produced using traditional fixed effects with linear regression models and those generated by modern machine learning based estimators. Furthermore, we investigate the implications of key implementation choices such as panel data transformation techniques, cross-fitting pro- cedures, and hyperparameter optimization on the performance and interpretability of DML in applied policy settings. Our work underscores the value of integrating modern computa- tional tools into empirical regulatory analysis, offering insights for policymakers and causal researchers.
The Urban Sublime: A Park Lodge for the National Mall
Sheffield, Aaron Cristopher (Virginia Tech, 2025-06-12)
The National Parks are a beloved element of the United States geography. They are characterized by monumental natural elements such as the Grand Canyon, Niagara Falls, and Old Faithful geyser at Yellowstone. The National Park System was founded to provide public access to these natural wonders. The original parks all centered around Great Lodges that allowed visitors to not merely to see the scenery, but to live within it – creating profound, sublime experiences of nature. The National Mall in Washington, D.C. is a National Park. Not only in name, but also in spirit. The Mall is a landscape consisting of monumental structures connected by parkland and accessible mainly by foot. However, the National Mall lacks a Great Lodge of its own. What would it mean to create a Lodge for the Mall? What should a Lodge for the Mall look like? What relationships would be important? Where on the Mall should it go? This thesis proposes a new Lodge for the National Mall to be built on the hill between the Vietnam Memorial and the Constitution Gardens pond. Guided by the sublime character of the Mall – a grandeur not of wilderness, but of human-made landscape – the project seeks to reveal the Mall's unique place within the National Park System and offer visitors a new way to inhabit this symbolic terrain.
What if Structures Could Speak?
Fleming, Alec Paul (Virginia Tech, 2025-06-12)
Architecture constantly plays with the tectonics of form through competing forces from the structure and the façade. In the English language we define façade as a false front, whereas façade in architectural language defines the barrier between the outside and the inside. Often the façade blurs the line of truth behind the structure, creating a mask for it. This project explores the direct use of the structure to define the form and the facade through the wrapping of structure, allowing the structure to speak for the form of the building with no interruption. With this continuity between structure and façade, the occupancy of the building becomes highlighted through shadow, allowing expression of internal use to the outside.
Multi-modal Multi-Level Neuroimaging Fusion with Modality-Aware Mask-Guided Attention and Deep Canonical Correlation Analysis to Improve Dementia Risk Prediction
Singh, Swapnil Satyendra (Virginia Tech, 2025-06-11)
Alzheimer's Disease (AD) is a progressive neurodegenerative disorder characterized by structural and molecular changes in the brain. Early diagnosis and accurate subtyping are essential for timely intervention and therapeutic planning. This thesis presents a novel multimodal deep learning framework that integrates T1-weighted MRI and Amyloid PET imaging to improve the diagnosis and stratification of AD. The proposed architecture leverages a two-stage pipeline involving modality-specific feature extraction using ResNet50 backbones, followed by middle fusion enhanced with a Modality-Aware Mask-Guided Attention (MAMGA) mechanism. To address missing modalities and inter-modal misalignment, the model incorporates Random Modality Masking and Deep Canonical Correlation Analysis (DCCA) for cross-modal feature alignment. Experiments on the ADNI dataset demonstrate that the proposed MRI+PET (MAMGA+DCCA) model achieves a balanced accuracy of 0.998 and an AUC-ROC of 0.999 in distinguishing stable normal cognition (sNC) from stable Alzheimer's Disease (sDAT). For the more challenging task of separating stable and progressive MCI (sMCI vs. pMCI), the best-performing fusion model achieved a balanced accuracy of 0.732 and an AUC of 0.789. Extensive ablation studies confirm the contributions of MAMGA, DCCA, and dual-optimizer strategies in enhancing diagnostic robustness. This work highlights the clinical potential of multimodal deep learning frameworks in improving early Alzheimer's detection and stratification.
Los Angeles 2028: Inglewood Station and Ballfield
Utter, Larry Ryan (Virginia Tech, 2025-06-11)
Deviating from traditional zoning segregation, many municipalities around the country increasingly seek to accommodate large congregations inside their town limits based on hybrid land use. Especially sports stadiums are locations with large infrastructural investments but often active only during specific events. Over the course of a year, the sport or concert activities denote only very short portion in the life of a stadium complex, whereas most of the time, the restricted typology contributes little to urban qualities. This thesis proposes a stadium typology based on hybrid land use, to include rail-based public transportation, a year-round retail zone, and space for social programs in its perimeter. The proposal aims to provide both the event space and offer an enduring lively environment and respectively a better service to its community. Los Angeles, poised to host the 2028 Summer Olympics, offers itself the territory for a site that promises to continue its urban contributions beyond the games.
The Four Thresholds
Massey, Airii (Virginia Tech, 2025-06-11)
This thesis investigates the architecture of thresholds as a poetic medium for shaping the transitional experience within the built environment. Through the careful orchestration of light, materiality, and spatial progression, the project seeks to cultivate a deliberate slowing of movement, inviting users to engage more deeply with each moment of passage. Drawing inspiration from the contemplative rhythms of Japanese living, where the act of transition is as meaningful as the destination itself, and contrasting them with the often hurried patterns of American habitation and relaxation, this study proposes a bath house nestled into the terraced landscape of Great Falls, Virginia. The architecture unfolds alongside the descending waterfront hillside, where spaces are carved, layered, and revealed gradually, blurring the boundaries between interior and exterior. Here, the architecture acts not merely as a vessel for activity but as an active participant in a ritual of slowing down, reconnecting body, mind, and environment through the measured unfolding of space.
Identifying Biophysical Drivers of Evapotranspiration for Forest Cover in a Mountainous Region
Nicolai, Lydia Rose (Virginia Tech, 2025-06-11)
Evapotranspiration (ET) is critical for understanding the impacts of climate change and land-use/land-cover change on water availability, ecosystem health, and agricultural productivity. However, point-based, field-measured ET data often lacks sufficient spatial and temporal coverage, especially in complex mountainous terrains such as the Appalachian Mountains. Consequently, characterizing ET rates across diverse land cover types and changing climate conditions remains challenging. This study uses remote sensing-derived ET data from the METRIC model for four selected watersheds in Virginia's Appalachian Mountains. Landsat-derived ET data with a 30-meter resolution spanning from 2015 to 2020 were obtained through the Earth Engine Evapotranspiration Flux (EEFLUX) platform on Google Earth Engine. Using supplementary GridMET reference evapotranspiration (ETr) data, temporal interpolation methods were applied to generate pixel-level daily ET profiles for the entire study area. The main objectives included comparing ET rates across land cover types from the National Land Cover Database (NLCD) and quantifying relative differences among land covers. Within forested land covers specifically, I further examined how topographic, soil, and vegetative factors influence ET variability. Generalized Least Squares and Random Forest models were employed to assess the relationships between selected biophysical variables and ET, highlighting both linear modeling with correlated error structures and the identification of non-linear patterns. Results from both models highlighted the significant roles of aspect, slope, and tree canopy cover in influencing ET variability, providing valuable insights into landscape-scale hydrological processes. Additionally, these models can be used to potentially fill gaps in ET estimates when satellite-derived data are limited due to cloud cover or other data availability constraints.
A Cultural and Dynamic Landscape Design
Zhao, Yixuan (Virginia Tech, 2025-06-11)
Washington, D.C.'s state-named streets honor the nation's history, yet they lack meaningful cultural representation. This project proposes a new urban design framework to transform these streets into immersive cultural experiences. The ultimate goal is to design all 51 state-named streets, ensuring that each one tells the unique story of its respective state. Due to time constraints, Louisiana Avenue serves as the first prototype, demonstrating how landscape design can blend reality and abstraction to express a state's geography and culture. Inspired by Mud Island Park's realistic geographic modeling and California Scenario's abstract artistic approach, this design integrates Louisiana's topography, Mississippi River dynamics, and Creole cultural symbols into the urban environment. The project also incorporates flood-resilient drainage systems, including Louisiana-style water wells, addressing climate challenges. This scalable strategy can be applied to future projects, ensuring that all 51 state-named streets in Washington, D.C. evolve from simple names into meaningful, story-driven public spaces.
Impact of Sex and Stretch-Shortening Cycle Induced-Fatigue on Limb Stiffness and Limb Stiffness Asymmetry During the Stop-Jump
Rebholz, Victoria Anne (Virginia Tech, 2025-06-11)
Sports-related injuries have a 67% incidence of occurring during competitive play, particularly in sports where fatigue is thought to negatively affect the way that the lower extremities attenuate forces being placed on the body.72 With the overall incidence of anterior cruciate ligament (ACL) injuries increasing, there is a growing incidence of these injuries occurring in the female population.117,12 Landing mechanics that are more erect, and those that are more asymmetrical, are known to increase the risk of injury, and may be heightened by the onset of fatigue.32 This study aimed to assess the effect of sex and fatigue induced by the stretch-shortening cycle on lower limb mechanics and asymmetry values in the stop-jump task. The first purpose of the study was to assess the effect of sex and stretch-shortening cycle fatigue on limb stiffness and limb stiffness asymmetry. The second purpose was to assess landing mechanics that are known to be indicators of how forces are being attenuated by the limbs. The components of limb stiffness were also assessed, including the resultant ground reaction force (rGRF) and change in limb length. A significant interaction was found for nondominant limb change in limb length (p=0.005), where males showed an increase in the change in limb length, while females showed a decrease following the fatigue protocol. The rGRF of both limbs was also different between pre- and post-fatigue conditions, decreasing in both sexes with the onset of fatigue. Asymmetry values for peak knee flexion angle, absolute value of knee flexion angle at initial contact (IC), and loading rate were also assessed before and after fatigue. Significant interactions for asymmetry values of peak knee flexion angle and absolute value of knee flexion angle at IC indicated that only female participants had an increase in asymmetry of knee flexion at IC and peak knee flexion values after the fatigue protocol. These results suggest that females adopt a more asymmetrical landing strategy than males after fatigue. A significant increase in peak knee flexion was also found for both sexes after fatigue. Thus, the decrease in rGRF may be due to the increase in peak knee flexion, which aids in the attenuation of the forces placed on the body. The results of this study indicate that, with fatigue, female participants may adopt landing strategies that put them at greater risk of sustaining lower extremity knee injuries during sport.

Browse

Recent Submissions