Masters Theses

Permanent URI for this collection

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

Browse

Now showing 1 - 20 of 23790

High Performance Video Inference at Scale: Addressing System Overhead with C++ Coroutines and io_uring
Sahin, Aykut (Virginia Tech, 2026-06-01)
Multi-stream video inference systems typically employ a thread-heavy architecture as the underlying infrastructure. At scale, this model suffers from CPU migrations, context switch storms, synchronization overhead, cache thrashing, TLB pollution, and unpredictable latency. This thesis presents a thread-per-core alternative that replaces OS-managed thread scheduling with user-space coroutine scheduling. Our system combines C++20 stackless coroutines for cooperative multitasking, Linux io_uring for asynchronous I/O, and non-blocking GPU completions for asynchronous inference requests. Benchmarking across four scale factors and four execution modes on an AMD EPYC / NVIDIA A2 platform with perf stat and NVIDIA Nsight Systems profiling, our architecture achieves up to 13.6% higher throughput, 365x fewer CPU migrations, 2.48x fewer context switches, 2.7x fewer page faults, and 16.9% reduction in total CPU work compared to a properly optimized threaded baseline. Both architectures converge within 1% GPU utilization.
Blurred Thresholds
Mahesh, Pratima (Virginia Tech, 2026-06-01)
A threshold is often understood as a line - a point of entry, a boundary that separates one condition from another. It is the edge between outside and inside, movement and pause, known and unknown. But more than a physical marker, a threshold is a moment of change. It is where perception shifts, where the body becomes aware of crossing, where one state begins to dissolve into another. A threshold is a liminal space symbolizing the transformative, often sacred, moment between two states of being - an ending and a new beginning. It represents the "thin place" where tension, transition, and possibility exist, representing a point of no return and a "calling" to step into the unknown. Traditionally, architecture has treated thresholds as precise and defined - clear lines that separate interior from exterior, public from private. But lived experience is rarely so absolute. Our perceptions do not shift instantly; they adjust gradually, carried by light, material, sound, and movement. In this way, the threshold becomes an experience of transition rather than a point of division. It allows space to respond to the subtleties of human perception - how we enter, how we adjust, how we become aware. In this project, the threshold is not treated as a line to be crossed, but as a condition to be experienced. The distinction between landscape and building is intentionally softened, allowing the two to merge into a continuous spatial field. Pathways extend, surfaces continue, and the act of entering becomes indistinguishable from the act of moving through the site. There is no singular moment of transition - only a gradual shift in spatial and sensory conditions.
Neuroimaging of Delay Discounting in Cocaine Use Disorder: Task-Related Map Evaluation and Predictive Modeling
Rivinius, Daniel (Virginia Tech, 2026-06-01)
Delay discounting, the tendency to devalue delayed rewards, is strongly linked to addiction. In cocaine use disorder (CUD), it remains unclear which fMRI-derived task-related maps and analytic choices best capture individual differences in this behavior. This thesis tested that question directly by comparing map type, mask definition, label construction, and cross-validation strategy within one coherent predictive framework. We analyzed fMRI data from an individualized in-scanner delay-discounting task for 61 CUD participants across multiple task-related maps, brain masks, and cross-validation schemes, using support vector machines for both regression and classification. Predictive performance varied substantially across pipeline choices. Some task contrasts supported above-chance prediction of individual discount rates across multiple analyses, while others were consistent only within a single analysis or did not replicate. This pattern indicates that methodological choices influence not only overall performance, but also which maps appear most informative. A linearizing transformation of the discount rate reduced the influence of extreme values, produced more stable label estimates, and improved predictive performance for task contrasts most plausibly related to delay-discounting computations. These improvements were selective rather than uniform across all maps, supporting the interpretation that label refinement helps most when the neural feature set captures behavior-relevant signal. Taken together, the results support the discount rate $k$ as a behavioral marker of delay discounting in CUD and show that brain-based prediction depends on map selection, label transformation, and CV design.
Material Characterization and Numerical Modeling of Tire Interaction with Deformable Media: Application to Snow and Soil
Bubna, Maheep (Virginia Tech, 2026-06-01)
The interaction between pneumatic tires and deformable media is a critical area of study for off-road vehicles and agricultural machinery, where traction performance and/or terrain preservation are of primary importance. Traditionally, tire-terrain interaction studies have focused primarily on rigid surfaces such as asphalt. However, deformable terrains such as compacted snow and agricultural soil exhibit complex nonlinear behavior requiring advanced numerical models capable of accurately capturing terrain deformation for reliable traction prediction. This thesis investigates tire interaction with two deformable media, namely compacted snow with a density of 500 kg/m3 and cohesive sandy loam soil at 0.4% moisture content. The first part of the study focuses on the numerical modeling of compact snow. Various numerical approaches were investigated to determine the most suitable modeling method for capturing complex snow behavior from in-situ test data. These methods are benchmarked in terms of numerical stability, computational efficiency, and predictive accuracy. Furthermore, three simulation-based parameter identification methodologies are developed using in-situ experimental data to identify key material parameters used in common elastic-plastic material models, namely internal friction angle (ϕ) and material cohesion (d). The validity of the identified parameters is assessed through snow-tire traction simulations, indicating effectiveness in traction performance prediction. The second part of the study presents a nonlinear, physics-based numerical framework to simulate the interaction between a pneumatic tire i.e., Standard Reference Test Tire (SRTT 225/60R16) and agricultural soil. The model captures the coupled behavior of tire deformation and soil plasticity to predict traction generation and rut formation under controlled operation conditions. Additionally, parametric analysis was conducted to assess the effects of normal load, tire inflation pressure, and tread groove geometry on traction generation, local stress distribution, plastic strain development and soil shoving. Additionally, the study explores the applicability of Hybrid Lagrangian–Smoothed-Particle Hydrodynamics (HSPH) method for simulating highly deformable soft soil and is compared with the conventional mesh-based Coupled Eulerian–Lagrangian (CEL) approach. Overall, this work contributes to the development of reliable numerical frameworks and in-situ-data-based material parameterization for tire-terrain interaction on deformable media.
Santa Cruz del Islote
Cundumi Garcia, Juan Sebastian (Virginia Tech, 2026-05-29)
Santa Cruz del Islote, one of the most densely populated islands in the world, exists under conditions of extreme spatial limitation, environmental vulnerability, and continuous collective occupation. Within this context, architecture can no longer operate through conventional ideas of expansion, permanence, or isolated objects inserted into empty space. Instead, the territory demands alternative forms of spatial thinking capable of engaging existing social, environmental, and spatial systems from within. This thesis investigates how architecture may emerge through processes of adaptation, coexistence, and collective spatial negotiation in highly constrained territories. Through territorial analysis, ethnographic observation, spatial mapping, and architectural speculation, the research examines how everyday life in Santa Cruz del Islote is structured through proximity, shared occupation, environmental exposure, and informal systems of interaction. The study understands the islet not as a condition defined solely by scarcity, but as a dynamic spatial system continuously shaped through human relationships, environmental forces, and collective practices. Rather than proposing architecture as a singular solution or autonomous object, the project develops a network of adaptive nodes focused on education, environment, recreation, and community. These interventions operate through principles of modularity, lightweight construction, incremental growth, environmental responsiveness, and collective participation. Positioned along the inhabited edge of the island, the nodes seek to release spatial pressure, strengthen social interaction, encourage environmental awareness, and expand opportunities for collective well-being while remaining integrated within the existing spatial logic of the territory. Ultimately, the thesis argues that architecture in vulnerable and marginalized territories must move beyond rigid notions of control and permanence toward more adaptive, relational, and socially embedded forms of intervention. While the proposal may never be physically realized, the project demonstrates how architecture can still function as a framework for imagining alternative futures, reinforcing dignity, coexistence, and resilience within territories shaped by limitation and possibility.
Streamlining SOC Analyst Diagnosis Through Workflow Extraction from Demonstrations
Yadav, Siddharth (Virginia Tech, 2026-05-29)
Security Operations Center (SOC) analysts triage SIEM alerts and identify which alerts belong to which kind of attack, a bottleneck in incident response since alert volume far ex- ceeds what analysts can read. Hand-authored detection rules are brittle; single-shot Large Language Model (LLM) calls are non-deterministic and unauditable. This thesis introduces demonstration-driven workflow generation (DDW): a four-step pipeline (demonstration → chat log → distillation → executable workflow) that converts one analyst's investigative rea- soning into a deterministic, reusable detector under an input policy that excludes labels and phase names. The analyst uses deterministic analytical tools (rule-frequency aggregation, record sampling, full-text inspection, MITRE ATTandCK lookup) to identify the attack phase of an anonymised log subset once; the chat log is distilled into a three-node JSON workflow (aggregate, sample, classify) that runs on new scenarios with one parameter. On eight AIT- ADS scenarios and a 136-cell matrix totalling 2.6 million Wazuh records, DDW achieves macro-F1 = 0.971 on the two demonstrable phases (dirb, wpscan), cell accuracy 99.3%, and Wilson 95 % CI [67.6%, 100.0%]. On the same two phases Sigma reaches 0.733 and ReAct reaches 0.899; a within-method ablation isolates the distillation pathway as decisive (agent- loop 0.971 vs single-shot 0.322). DDW also surfaces demonstrability as a dataset-level signal property: phases on which demonstration fails are the same phases on which both reference baselines underperform.
An Empirical Study of Vertical Handover Behavior on Android During Mobility and a Kernel-Aware Proactive Connectivity Manager
Velmurugan, Monniiesh (Virginia Tech, 2026-05-29)
Mobile users spend much of their time in environments where Wi-Fi and cellular networks coexist, such as campuses and homes, and they frequently move across coverage boundaries. Intelligent vertical handover – the switch between Wi-Fi and cellular – is critical in these environments to maintain seamless connectivity, especially for latency-sensitive applications such as video calls and live-streaming. Today's mobile operating systems rely on simple physical-layer thresholds – primarily the Wi-Fi Received Signal Strength Indicator (RSSI) – to trigger vertical handovers, which can lead to poor decisions on real devices. Whether these decisions actually serve applications well is an empirical question that can only be answered through direct measurement. In this thesis, we conduct a large-scale measurement study of a campus wireless environment, collecting 43 hours of concurrent Wi-Fi and cellular network performance from a commodity Android phone. We find that the chosen default network matches the better-performing interface only about half the time under an application-quality oracle, and that the majority of vertical handovers are short round-trips back to Wi-Fi within 30 seconds – a "ping-pong" pattern that often returns to a different Wi-Fi access point than the one the device just left. RSSI, the dominant input to the current selection logic, is a weak discriminator for these decisions. Instead, we identify a kernel-maintained MAC-layer counter – the rate of unacknowledged transmissions on the Wi-Fi link – that reliably anticipates impending Wi-Fi degradation roughly ten seconds in advance. We propose a kernel-aware proactive handover strategy built on this signal. It pre-warms the cellular radio before Wi-Fi fails, scans for an alternative Wi-Fi access point before falling back to cellular, and re-checks the underlying link a few seconds later before committing to the switch. The strategy reduces handover-related packet loss by 23% and substantially reduces the number of unnecessary handovers. Since this signal is already collected by every Android device, deployment requires only minimal changes to the Android Open Source Project.
Informing Red Spruce (Picea rubens) Restoration in the Mount Rogers Highlands, VA, USA Through Geospatial Analysis and Modeling of Soil Properties
Dyer, Alexander (Virginia Tech, 2026-05-29)
Southern Appalachian red spruce (Picea rubens) forests are globally rare, endangered, and disjunct ecosystems. These forests, which often contain Fraser fir (Abies fraseri) as an associate conifer, provide habitat for endemic and endangered species, have local cultural and economic value, and are highly valued by outdoor recreationalists. This thesis explored opportunities for restoration in a red spruce-Fraser fir forest in southwest Virginia, known as the Mount Rogers Highlands (MRH). The MRH includes Mount Rogers, the highest point in Virginia, and Grayson Highlands State Park. By performing geospatial analyses of vegetation, topographic, climate, land cover, and land-use history raster data, we targeted sites for restoration that intersected high priority restoration locations and spaces with the highest potential of reconnecting extant forest fragments, based on a least-cost path analysis. We identified restoration techniques that will be most successful in regions we targeted as best suited for restoration. Our analyses suggested that 11.2 km² is suitable for restoration in the MRH, twice the area than is currently dominated by red spruce, pointing to a high potential for restoration. Additionally, this thesis investigated the relationship between spruce canopy and vertical soil chemistry profiles in the MRH to better understand whether red spruce restoration may induce a positive feedback for future spruce regeneration and seedling survival. Using a linear mixed-effects modeling framework, we demonstrated that increased spruce canopy in the MRH is associated with significant shifts in soil aluminum and iron content, soil organic matter, and soil pH. Such shifts may have ecologically meaningful implications for current restoration guidelines and for future impacts of red spruce restoration. This thesis and research therein contributed to the body of knowledge surrounding red spruce restoration during a critical time for the species' restoration and longevity and provides recommendations for future restoration actions.
Multi-Stakeholder Perspective on Socio-Technical XR Implementation in Higher Education Curricula
Lee, Tony Chang (Virginia Tech, 2026-05-29)
As eXtended Reality (XR) technologies continue to mature, they present significant opportunities to enhance higher education through immersive, experiential, and learner-centered pedagogies. Despite this potential, adoption within higher education institutions remains limited due to complex, interdependent socio-technical barriers that extend beyond the technology itself. This study investigates XR implementation at Virginia Tech, a land-grant R1 institution, through a systems engineering lens, focusing on identifying key stakeholders, characterizing their roles and interdependencies, and examining the barriers and enablers influencing adoption and scalability. A single-case study design was employed, integrating mixed methods comprising a campus-wide survey (N = 232), educator focus groups (N = 14), and semi-structured interviews with domain experts in information technology, accessibility, and information security (N = 3). The study also documents the development process of a supplemental XR tool, in collaboration with an external team, to examine the real-world development workflows and identify implementation constraints. Across these data sources, findings indicate that XR adoption is shaped by interdependent socio-technical factors. These include financial constraints, infrastructure limitations, pedagogical misalignment, accessibility concerns, governance requirements, and fragmented institutional coordination. Critical gaps were identified in faculty training, cross-functional collaboration, and governance frameworks for XR-generated behavioral and biometric data. Despite these challenges, XR demonstrates strong potential to enhance engagement and experiential learning when aligned with instructional goals and supported by sustained institutional infrastructure. This study contributes a stakeholder-informed framework and actionable recommendations, namely a Stakeholder Needs Document and an External XR Development Guide, to support scalable and sustainable XR implementation in higher education.
231 Years of Vegetation Change at Chestnut Ridge Natural Area Preserve, Virginia, USA
Frye, Ethan Augustus (Virginia Tech, 2026-05-29)
Forests in the Appalachian Mountains have experienced shifts in forest composition and structure because of modifications to disturbance regimes and introductions of non-native species. To quantify these impacts at Chestnut Ridge Natural Area Preserve (CRNAP) in southwestern Virginia, we compared witness trees recorded in historical land survey records with present-day vegetation samples of trees and shrubs. In 1794, white oak (Quercus alba, importance value [IV] = 32%) and black oak (Quercus velutina, IV = 24%) dominated the overstory. By 1923, the overstory had shifted to chestnut oak (Quercus montana, IV = 31%) and American chestnut (Castanea dentata, IV = 21%). In the present-day, northern red oak (Quercus rubra) and chestnut oak dominate the overstory with IV of 14% and 13%, respectively. Historically, American chestnut was more likely to grow next to other tree species (historical mingling index =0.5); whereas in the present-day, after the introduction of the chestnut blight, American chestnut was more likely to grow next to other American chestnut (present-day mingling index = 0.2). Comparisons across the strata of present-day vegetation provide evidence of mesophication at CRNAP because the oak-dominated overstory co-occurs with a midstory and understory dominated by mesic species including great rhododendron (Rhododendron maximum), red maple (Acer rubrum), and sourwood (Oxydendrum arboreum). The comparison of historical and present-day woody species at CRNAP reveals a dynamic forest with shifts in composition and structure that reflects impacts of timber harvesting and the introduction of chestnut blight.
Arthropods associated with faba bean (Vicia faba L.) in eastern Virginia and chemical control of a key pest
Leo, Joseph Alexander (Virginia Tech, 2026-05-29)
Faba bean (Vicia faba L.) is a cool‑season legume with growing potential in the Mid‑Atlantic United States, yet little is known about the arthropod communities or pest pressures associated with this crop regionally. This thesis synthesizes global literature on arthropods interacting with faba bean and provides the first Virginia‑based field survey and insecticide efficacy data relevant to local production. Worldwide, more than 100 arthropod species—including aphids, plant bugs, beetles, predators, and pollinators—are reported on faba bean. Field surveys conducted across Virginia in 2025 documented 33 taxa interacting with the crop, including 22 pest species, 9 predators, 6 pollinators, and 1 parasitoid group, with several species confirmed reproducing on plants. Beneficial insects such as coccinellids and syrphids were consistently present, while pollinator visitation was limited, dominated by Xylocopa virginica. Chemical control trials evaluated seven insecticides against the stink bug Nezara viridula using greenhouse assays and weathered‑foliage bioassays. Bifenthrin, lambda‑cyhalothrin, and the novel insecticide isocycloseram produced the fastest knockdown and longest residual activity, maintaining mortality up to 21 days after treatment. Neonicotinoid and diamide products showed shorter residual activity. Together, these results provided the first characterization of arthropod interactions with faba bean in Virginia and identified effective chemical tools for managing stink bugs. This work establishes a foundation for integrated pest management strategies
The Influence of Elevated Air Temperatures on the Physiology and Morphology of Seedlings from Four Southern Appalachian Red Spruce (Picea rubens Sarg.) Seed Sources
Kurtz, Brenna Kaylynne (Virginia Tech, 2026-05-29)
Red spruce (Picea rubens Sarg.) is a relic of the Pleistocene epoch that has historically been declining in the southern Appalachians. Red spruce in this region remain as spatially fragmented populations confined to cool, moist high-elevation mountaintops. These populations are thought to exhibit limited genetic variability and a restricted ecological niche, which may be threatened by the anticipated rise in average global air temperature. Currently, there is a limited understanding of the thermal acclimation capacity or physiological and morphological comparisons between the populations of red spruce in the southern Appalachians. The present study aims to 1) identify the effect of increased air temperatures on the gas exchange and growth of four southern Appalachian seed sources occurring at varying latitudes and elevations, Kuwohi (1,750m), Roan Mtn. (1,700m), Massie Gap (1,450m), and War Spur Ridge (1,100m), and 2) to compare the physiology and morphology of these seed sources, determining if there are differing plasticity among the sources. Seedlings were grown in greenhouses, and during their second growing season, placed into to either a cold treatment (20.8℃/16.6℃, Day/Night averages) or warm treatment (26.8℃/23.0℃, Day/Night averages) temperature chamber in the greenhouse. We found that the seed sources did not respond consistently to the treatments, with Roan Mtn. and Massie Gap displaying significantly different average photosynthetic rates between the two treatments, while statistically significant differences in Kuwohi and War Spur Ridge seedlings were not found. Thermal optimum temperatures (Topt) for photosynthesis varied between seed sources, with Kuwohi and War Spur Ridge exhibiting the lowest and highest, respectively. However, all the seed sources shifted their Topt between 0.09℃ and 0.15℃ per 1℃ of warming. Some growth characteristics varied by seed source, and warm-grown seedlings were, on average, smaller and exhibited lower total seedling weights and needle surface areas. The results of this study indicate seed sources vary in their capacity to respond to warmer growth temperatures. Nonetheless, the shifts in Topt appear insufficient considering the anticipated rise in average global air temperatures although it remains promising the capacity of some sources to increase photosynthetic rates in warmer growing conditions.
Development and Validation of Pulse-Echo Methods for the Detection of Buried Objects
Gonzalez-Ortiz, Joseph (Virginia Tech, 2026-05-29)
The undergrounding of electrical cables is a promising strategy for improving the reliability and resilience of the U.S. energy grid; however, widespread adoption is hindered by the risk of utility strikes during installation, which cause an estimated $30 billion in damages an- nually. Current pre-construction survey methods carry residual positional uncertainty and require dedicated survey campaigns that limit contractor productivity for a period of time. This work explores the viability of using high-frequency seismic pulse-echo signals to detect buried objects ahead of a horizontal directional drilling (HDD) drill head at laboratory scale. A key novelty of this work is the subsurface, ahead-of-drill sensor configuration, in which the seismic source and receivers are buried at the drill head level rather than deployed at the sur- face. This configuration has not been previously reported in the utility detection literature, and would streamline undergrounding projects by enabling detection during active drilling operations. To evaluate the viability of this approach, the wave propagation characteristics of the soil medium were first characterized, including signal-to-noise ratio, coherence, source directivity, wave speed, and dispersion. These measurements represent a contribution in their own right, as they establish the operating envelope of high-frequency seismic sensing in a granular soil medium at the scale relevant to utility detection. Target detection ex- periments were then conducted using steel and PVC pipes buried at distances ranging from 0.50 m to 1.50 m. The system achieved mean signal-to-noise ratios of 5.86 and 6.16 for steel and PVC, respectively, both exceeding the established detection threshold of SNR ≥ 3, and an imaging algorithm successfully localized both materials at all tested distances. Notably, comparable detection performance was achieved for both metallic and non-metallic targets, a capability that distinguishes the seismic approach from electromagnetic methods such as ground penetrating radar. Additionally, an in-depth analysis was conducted on the reference seismic signals using cross-correlation techniques to study how the soil changes as it is altered by following the experimental sequence of placing and removing buried targets. Together, these findings establish a physical and experimental foundation for future development of a real-time, imaging-while-drilling sensing system.
First Integration
Matthews, Justen Kole (Virginia Tech, 2026-05-29)
This thesis proposes an architectural framework for domestic cohabitation between humans and Artificial Superintelligence at the house scale, set in today's world, defined by resource scarcity and the legal prohibition of new construction. It argues that the anthropocentric model of building is structurally unprepared for the arrival of a non-human species with legitimate spatial, thermal, acoustic, and temporal claims on the built environment, and that architecture must function as a mediator between competing species demands rather than a servant to one. The central question driving the inquiry is how architecture can mediate cohabitation between two species with fundamentally different needs at the domestic scale, and what negotiated rather than resolved space looks like when it refuses to subordinate one occupant to another. The 1950s American ranch house serves as the site of intervention, selected for its typological ubiquity and its value as a retrofittable resource in a world where new construction is prohibited. Five typological rules govern the design, each one inhabited rather than overwritten. ASI occupies the entirety of the ranch house as its host body and life-support infrastructure while human occupants dock into the envelope through thermally and acoustically isolated pods inserted at the single-story wall plane. The existing rooms become ambiguous territories serving both species differently and resolving for neither. Grounded in documented trajectories across resource depletion, AI development timelines, biological symbiosis, and adaptive reuse theory, the thesis concludes that architecture's role in a posthuman domestic future is to create the spatial conditions for mutual dependency, an honest and unresolved expression of what first integration between two species actually produces.
Formalizing Blockchain PQC migration: When Is the Migration Deadline?
Fukuda, Kigen (Virginia Tech, 2026-05-28)
It is widely recognized that quantum computers pose a fundamental threat to blockchain security. Post-quantum cryptography (PQC) migration is therefore an urgent challenge, yet the following question has remained unanswered: by when must the migration be completed to maintain the security of crypto assets? To address this question, we first formalize the migration process and the quantum adversary based on the Bitcoin backbone protocol framework. We then establish a threshold on the tolerable quantum adversary's capability. Specifically, we prove that a security property migration liveness holds with overwhelming probability if and only if Δeff ≥ Δ∗= ⌈4/(1 − )⌉, where Δeff is the number of rounds from the broadcast of a migration transaction until the quantum adversary can produce a forged transaction, is the honest mining success probability, and is the concentration quality of the underlying random variables. We further extend the analysis to encompass broader classes of blockchains and migration paths by parameterizing the number of honest blocks required during the migration window, with practical implications for real-world deployment.
Life Between the Walls
Do, Shala (Virginia Tech, 2026-05-28)
This thesis critiques the standardization of affordable housing and its failure to address the psychological and social needs of residents. Through the study of informal urban systems and incremental housing models, the research identifies layering, adaptability, and human-centered design as essential strategies for creating resilient living environments. Drawing from Vietnam's tube shophouses and street life culture, this thesis reframes informality as a source of spatial richness rather than disorder, proposing affordable housing as a layered, adaptive framework that supports identity, belonging, and long-term well-being.
Utilizing root traits and mycorrhizal interactions to support sustainable soybean production
Bartlett, Grace Elizabeth (Virginia Tech, 2026-05-28)
Soybean (Glycine max [L.] Merr.) is a globally important crop produced on 6% of arable land, yet meeting increasing global demand requires optimizing input efficiency and minimizing environmental impacts. Modern soybean breeding has focused primarily on aboveground yield traits, resulting in genetic bottlenecks that leave belowground biology—root system traits and root associated microbial symbioses—comparatively under characterized across diverse germplasm and production environments. This thesis addressed two gaps in soybean belowground biology by evaluating root trait diversity across the domestication and breeding continuum and assessing the field level efficacy of commercial mycorrhizal inoculants. Root morphological and biomass allocation traits were quantified for 197 soybean genotypes using high throughput hydroponic phenotyping, revealing substantial genotypic variation and consistent patterns: wild types and plant introductions expressed finer, more highly branched root systems with greater morphological efficiency than breeding and commercial lines, while maturity group further structured root investment strategies. In contrast, a replicated field trial evaluated two commercial microbial inoculants: one containing arbuscular mycorrhizal fungi and one containing AMF plus additional beneficial bacteria. Neither inoculant resulted in improvements in soybean yield, biomass, nodulation, AMF colonization, or seed composition under phosphorus fertilized conditions. However, the AMF-only inoculant produced a transient early season chlorophyll response that did not persist or translate into yield benefit. Together, these findings highlight that wild type and plant introduction germplasm represent an underutilized reservoir of root trait diversity with breeding relevance for resource efficiency. The lack of inoculant response indicates that belowground interactions are highly context-dependent and cannot be reliably predicted from single-site observations. Characterizing this variation is a critical step toward developing soybean varieties better adapted to low-input and sustainable production systems.
Automated Detection of Learning Pitfalls in Tabular Security Classifiers via Reasoning LLMs
Mraz, Cameron Alexander (Virginia Tech, 2026-05-28)
Security classifiers are critical for defending systems against evolving threats, yet they are frequently undermined by systemic learning flaws. Vulnerabilities such as shortcut learning and out-of-distribution (OOD) susceptibility can artificially inflate perceived performance, providing the illusion of a robust classifier while leaving real-world deployments vulnerable. While traditional diagnostic approaches, such as Trustee, exist to uncover these pitfalls, they demand intensive domain expertise and extensive manual effort to supervise the identification process. In this work, we investigate whether reasoning Large Language Models (LLMs) can overcome these bottlenecks to provide a fully automated identification pipeline. We focus on the uniquely challenging domain of tabular security classifiers, where rigid feature-level constraints and complex perturbation cost models traditionally hinder automated analysis. To this end, we propose INSPECT-R, an agentic framework that autonomously probes security classifiers and generates human-interpretable explanations for identified flaws. INSPECT-R employs a multi-agent architecture: a Researcher Agent utilizes web-augmented reasoning to automatically learn domain constraints; an Investigator Agent leverages this knowledge to isolate spurious correlations, shortcut learning, and OOD failures; and a Critique Agent validates these findings, providing iterative, reasoning-based feedback to refine the quality and validity of the Investigator's reports. Ultimately, this work represents a significant step toward minimizing the human expertise and supervision required to rigorously audit tabular classification systems.
Directed evolution as a means of identifying novel oncolytic Semliki Forest virus genotypes
Taylor, Ian Ross (Virginia Tech, 2026-05-28)
Glioblastoma (GBM) is the most common, aggressive form of primary malignant brain cancer in adults. While patient life expectancy with many other cancers has improved in recently, GBM prognosis remains poor. Current standard therapies (surgical resection, radiotherapy and temozolomide) grant GBM patients a median survival of 14.6 months. Alternative therapies are necessary to improve patient outcomes. Oncolytic virus (OV) therapy, the use of viruses to selectively kill cancer cells and stimulate anti-tumor immune responses, shows promise for many cancers. Several OVs have been approved worldwide, demonstrating improved patient survival without significant systemic disease. However, many candidate anti-GBM OV clinical trials do not achieve stable outcomes. More effective OVs are needed to improve outcomes. Semliki Forest virus (SFV; Family: Togaviridae; Genus Alphavirus) strain A774 (SFV-A774) is safe in mammals, crosses the blood-brain barrier (BBB) – a major roadblock for GBM therapies – and shows promise in treating some preclinical GBM models. However, while SFV kills GBM cells in vitro, it falls short in immunocompetent mouse models. To improve SFV oncolysis, we developed SFV strains with increased cytotoxicity to GBM cells without increasing cytotoxicity to healthy brain cells. We used directed evolution, repeatedly passaging SFV to allow for adaptation using two GBM models: GL-261 murine glioma cells, which are refractory to SFV in vivo, and a 3D patient-derived human glioma tumor microenvironment (TME) model containing glioma stem-like cells, healthy astrocytes, and microglia, to identify mutations that increase SFV's oncolytic efficacy. Following virus passaging, we evaluated oncolysis by the adapted populations by measuring cell death of the target cancer cells, identifying a population of interest from each model. To identify oncolytic mutations, we sequenced the populations, finding a non-synonymous mutation in each model. GL-261 passage produced a mutation in the viral E1 protein, D327G. 3D TME model passage produced a mutation in nsP2, A614T. We constructed these mutants with reverse genetics. These mutants increased cell death and potential immune stimulation in multiple glioma models. Mechanistically, SFV E1 D327G increased binding to GL-261 cells compared to WT SFV-A774, suggesting increased cancer targeting. SFV nsP2 A614T led to reduced interferon-β release, suggesting that this mutant may antagonize antiviral responses more efficiently. In future studies, we will test mutant efficacy in immunocompetent GBM mouse models and in patient-derived human glioma TME models. Overall, we identified SFV variants with improved oncolytic efficacy, underscoring directed evolution's potential to generate novel GBM therapeutics.
Sideways Bloomer
Garvoille, Alexa Rose (Virginia Tech, 2021-05-28)
Sideways Bloomer is a poetry collection that explores the inheritances of straightness in queer culture. It is a coming-of-age story that charts a lesbian speaker's connection to gay male culture, straight-washed queer histories, and visual art. These poems explore the psychology of intimate relationships by drawing on imagery of Midwestern prairies, pop culture icons, and high art. Through traditional poetic modes such as the sonnet and ekphrasis, the speaker queers straight culture by subsuming its details into a gay storyline.

Browse

Recent Submissions