33760 research outputs found
Sort by
Masters Thesis
No full textThis study explores the spatiotemporal variability of stable hydrogen (δ²H) and oxygen (δ¹⁸O) isotope values in Maryland’s rivers and streams during a two-year period (2022-2024), emphasizing the influence of precipitation sources, physiographic features, and hydrological processes. Rivers in western Maryland exhibited lower δ²H and δ¹⁸O values, likely due to long-distance moisture transport and altitude effects. In contrast, eastern rivers and streams displayed higher isotopic compositions, likely influenced by local moisture recycling, higher temperatures, and greater warm-season precipitation inputs. A Local Meteoric Water Line (LMWL) was derived for Maryland as δ²H = 7.84·δ¹⁸O + 12.86 (R² = 0.99), deviating slightly from the Global Meteoric Water Line (GMWL) because of regional climatic influences such as atmospheric vapor recycling, and sub cloud evaporation. Elevation demonstrated a clear isotopic control on river-water isotope values, with a 0.6‰ decrease per 100 m of increase in elevation for δ²H and 0.1‰ per 100 m for δ¹⁸O. Seasonal patterns were also evident, with lower isotopic values during winter due to cold-temperature isotopic fractionation and remote moisture sources and more positive values in summer as a result of convective storms and evaporation. During the drier year (2024) with a reduced moisture surplus, river systems relied more on stored winter precipitation, emphasizing the buffering role of groundwater. Deuterium excess (d-excess) values further showed regional differences in moisture sources. Higher d-excess in western Maryland pointed to potential lake-effect precipitation derived from the Laurentian Great Lakes and long-distance transport, whereas lower values in the east reflected enhanced local evaporation. These findings establish a regional baseline that enhances our understanding of hydrological and climatic controls of river-water isotope values across Maryland. They also imply that a reduction in winter precipitation could diminish groundwater recharge and baseflow, affecting dry-season water availability. Meanwhile, more intense summer rainfall may increase surface runoff and nutrient loading, heightening flood risks and degrading water quality.Funding support by the Maryland Water Resources Research Center (UMD) through the Water Resources Research Act (USGS)
The Association Between Ghost Gun Usage and Neighborhood Disadvantage
No full textThis study examines the relationship between Privately Made Firearm (PMF) usage and multiple characteristics of neighborhood disadvantage in a large metropolitan city. PMFs, also known as ghost guns, are unserialized firearms typically ordered as parts and constructed by hand at home. Since they are unserialized, they are untraceable by the Bureau of Alcohol, Tobacco, Firearms and Explosives (ATF). Thus, PMFs are frequently owned by individuals who are unable to purchase a firearm through legitimate means. Police data containing PMF recovery incidents, categorized as criminal or non-criminal incidents, from 2020 to 2023 were obtained and mapped onto census data to find the PMF rate per 1,000 people in each census tract. A linear regression analysis was conducted to determine neighborhood disadvantage from 2015 to 2019 was predictive of PMF use. Neighborhood disadvantage was measured through unemployment rate, poverty rate, and rate of residents over 25 without a high school diploma. Results show that higher neighborhood disadvantage is associated with a higher PMF rate per census tract for all PMF recovery incidents, criminal incidents alone, and non-criminal incidents. This research provides important contributions to firearm research and demonstrates that PMF use is associated with neighborhood disadvantage in the same way as violent crime.University of Maryland Honors College Research Gran
Evaluation of Pv11 Cell Immobilization Techniques Across Substrates
No full textAvidin-biotin binding, Cell-Tak, and crosslinked alginate hydrogel were compared to collagen-BAM, our standard method, to assess how well they immobilized the cells and how the cells subsequently responded to odorants. These alternate methods exhibited high, although inconsistent, immobilization of cells, and the cells remained responsive to pentyl acetate, potentially offering new alternatives for Pv11 cell immobilization.National Science Foundation CBET BIOSENS 2316199
EFRI ELiS 231802
Bridging the Gap: Connecting the University of Maryland Extension Community with Library Resources
No full textIn conjunction with Isabella Baxter, UMD’s Agriculture and Natural Resources (AGNR) Librarian, I have enhanced the Library’s AGNR Research Guide to include robust resources specifically curated for University of Maryland Extension (UME) staff and community members. UME is a non-formal education system within the College of Agriculture and Natural Resources and Maryland Eastern Shore campus. Through an analysis of Ref Analytics on LibInsights, I identified three needs of the UME community that UMD Libraries can support: accessing library resources off-campus, professional development, and research tools. UMD’s AGNR program is headquartered at the College Park campus, but UME operates at UMES, since the UME community is connecting to the libraries remotely, the LibGuide serves as an asynchronous learning tool that will answer many of their frequently asked questions and bridge the gap between their programming and UMD Libraries. My poster presentation will include information about how the libraries can continue to identify and support the needs of UME and off-campus user groups more broadly through asynchronous instruction
The Impact of COVID-19 School Learning Modality on Child Mental Health
No full textThe COVID-19 pandemic has affected the health, economic life and social norms of millions worldwide. The success and effectiveness of non-pharmaceutical intervention (NPIs) are important to our understanding of future pandemics. In the US, a prominent yet controversial NPI, remote-hybrid learning, was widely adopted to reduce transmission risks while vaccines were still being developed and distributed. The adoption of remote-hybrid learning in the US varied by geographic region and time. Considerable evidence has shown that it decreased student learning and widened achievement gaps. However, current evidence on its impact on children’s mental health is mixed. Remote-hybrid learning might improve mental health outcomes as it reduces negative exposures often experienced at schools. Or it may negatively impact outcomes as remote-hybrid learning reduced social interactions and support. Using quasi-experimental methods and data from the Healthcare Cost and Utilization Project and Monitoring the Future, my results suggest that marginal week of cumulative remote-hybrid learning was associated with a 0.4% decline in psychiatric ED visits relative to the mean. Males, younger children, and children from moderate or higher poverty districts experienced larger reductions and no subgroups experienced an increase in psychiatric ED visits. I do not find strong evidence that remote-hybrid learning changes student reported internalizing and externalizing symptoms. Findings from this dissertation contribute to our understanding of the costs and benefits of remote-hybrid learning during public health emergencies. My results suggest that while remote-hybrid learning, as implemented during COVID-19, may not pass a cost-benefit test due to learning loss and other deleterious outcomes (e.g. parental wellbeing), the intervention does not harm child mental health. Policymakers and school administrators can also use these findings to advocate for reforming learning context to support child mental health
Science of Deep Learning: From Initialization to Emergent Structures
No full textAs artificial intelligence (AI) systems grow increasingly powerful and permeate every aspect of our lives, their impact on both individuals and society is an urgent concern. Questions of safety and robustness in AI stem largely from our limited understanding of deep learning. Research in this domain has traditionally followed two parallel paths: an empirical approach that prioritizes practical advancements and a theoretical approach that seeks a mathematical understanding from first principles. Despite notable progress, a significant gap remains between deep learning practice and its theoretical underpinnings. This dissertation advocates for a phenomenological approach to understanding AI systems -- one that integrates empirical observations with theoretical model-building. This methodology has been instrumental in the physical sciences, and it holds similar promise for advancing the science of deep learning. Over two broad parts, this work demonstrates the effectiveness of this approach in characterizing model architectures and their emergent capabilities.
In the first part, we explore how signal propagation analysis in large-N limits can inform the design and initialization of model architectures. We develop a diagnostic observable that distinguishes between ordered and chaotic behaviors in neural networks, guiding optimal parameter initialization for training. Our analysis establishes the theoretical soundness of this observable in simple networks and confirms its empirical utility in state-of-the-art architectures. The findings reveal an architecture design paradigm that eliminates the need for careful initialization, shedding light on widely used heuristic practices. Additionally, we introduce an algorithm that automates initialization across diverse model architectures, enhancing their trainability.
In the second part, we highlight the importance of the systems identification approach for characterizing AI systems. We explore several stylized setups where model capabilities emerge as a function of compute, data quantity, and data diversity. Using arithmetic and cryptographic tasks as examples, we demonstrate that emergent abilities such as grokking and in-context learning arise alongside the formation of interpretable structures within the model’s parameters, hidden representations, and outputs. Through targeted experiments, we identify these structures using (i) black-box probing, which examines model responses to characteristic inputs, and (ii) open-box analysis, which leverages curated task-specific observables and metrics to study internal model states.
This dissertation promotes a paradigm for understanding deep learning that complements both heuristic-driven and hypothesis-driven approaches. By integrating experimental methodologies and analytical tools from established scientific disciplines, this framework has the potential to steer the field toward safer, more robust, and more efficient AI systems
DEVELOPMENT OF COMPUTATIONALLY AIDED METHODS FOR INTERPRETING MOLECULAR INFORMATION
No full textModern electronics have enabled a highly connected world due to the immense amount of information and communication that is readily available. Comparatively, biological information lies within biomolecular states and their transitions rather than in electrons and photons. Thus, rendering the process of understanding and extracting biological information difficult and time consuming; often requiring labor intensive and complex analytical methodologies. The development of rapid approaches for designing, controlling, and monitoring biologics are critical to both clinical and industrial applications of biotechnology. Redox reactions are ubiquitous across biology and offer a bridge to electronically connect into biologics through their electron exchanges. By leveraging redox molecules to interface between biomolecular and electronic communication, we can implement electronic signaling in biological systems to rapidly monitor and induce functional changes. Further, we can incorporate computational methods to model biological dynamics and parse electronic outputs from these systems, providing approaches for improved biological design and control.This dissertation focused on developing synthetic means for information propagation within biological bacterial co-cultures that made use of multi-modal molecular signals. We showed how molecular information could be initiated and propagated through these co-cultures, and controlled through modulating species composition. A graph network model was then developed for simulating multi-modal signaling in synthetic bacterial consortia to study the effects of spatial conformation within co-cultures on information transmission. Finally, we developed electrochemical methods for rapidly extracting biological information using redox-mediated “probing”. A novel machine learning methodology was implemented to extract biological information from this electronic data. These methodologies were developed for specific industrial and clinically relevant use cases such as: 1) Biomanufacturing of antibody therapeutics, to quantify a critical nutrient, L-cysteine, and antibody fragmentation as a marker for product quality; and 2) Female reproductive health monitoring, to classify vaginal microbiome states, where dysbiosis is linked to higher risk of health complications such as pelvic inflammatory disease and preterm labor. In sum, this work contributes molecular and electronic tools for designing and monitoring complex biological systems to predict responses and enhance clinical and industrial manufacturing applications
PALS 2025 : Deliverable 6
No full textThis report documents the evolution and impact of a creative placemaking partnership between Dance Exchange, the University of Maryland (UMD), the Purple Line Corridor Coalition (PLCC), and the Takoma Langley Crossroads Development Authority (TLTCDA). Together, these partners engaged artists, students, faculty, residents, and local leaders in reimagining public space and community belonging in the context of rapid transit-oriented development and
neighborhood change.Acknowledgements: Dance Exchange, School of Architecture and Urban Planning, Creative Placemaking Minor, PALS (Partnership for Action Learning in Sustainability), Purple Line Corridor Coalition, Takoma Langley Crossroads Development Authorityhttps://drive.google.com/file/d/1PAjHILtFldcN2huFeITH5UKemPhnl_s4/view?usp=sharin
Improving Round and Communication Metrics in Consensus
No full textByzantine Consensus protocols, aka Byzantine Agreement (BA), Byzantine Broadcast (BB), allow a set of n mutually distrusting parties to share input values and agree on the same output value. Notable practical applications of consensus in blockchains and MPC require efficient, practical implementations of consensus protocols. The rise of blockchain systems and the general trend towards decentralized services, which inherently utilize consensus, as well as the increased demand of consensus as a primitive for implementing other cryptographic protocols (e.g. in MPC), brought once again this topic in the forefront of research.
A recurring trilemma of consensus when utilized in practice is striking the balance in order to construct protocols that are: i) efficient, ii) resilient under harsh adversarial conditions, and iii) with minimal assumptions for the corresponding setting. These properties are usually negatively associated; efficient protocols often either require stronger assumptions, or are less resilient to strong adversaries.
In this dissertation we aim to improve the efficiency of consensus protocols under harsh adversarial conditions and weak assumptions. We explore directions towards improving the two major metrics of efficiency of such protocols; i.e. their communication and round complexities. We focus on protocols operating in a synchronous communication network, and show how to achieve efficiency of either rounds or communication under different assumptions, against weakly adaptive adversaries with high corruption threshold. We will construct i) (Parallel) Broadcast protocols with improved state of the art communication, ii) the first Broadcast protocol with sublinear rounds without trusted setup, and iii) the first Deterministic Byzantine Agreement protocol with adaptive O(n·f) communication
Modeling and analysis of canonical turbulent shear flows
No full textTurbulent flows are ubiquitous in engineering systems, however analyzing such flows around engineering-relevant geometries using turbulence-resolving simulations is difficult due to computational cost constraints. Therefore, to advance our ability to predict these flows, we need physics-based models that alleviate the computational cost while retaining high accuracy. The thesis contributes to this broad challenge by introducing two models for high-speed wall-bounded turbulent flows and through a detailed study of 3D effects in turbulent shear flows.
In Part I of this thesis, the modeling of high-speed wall-bounded turbulent flows is explored using different solution fidelities. First, a low-cost modular method is developed that estimates the mean velocity and temperature profiles, and hence the friction and heat transfer coefficients for a given Mach number, wall temperature and Reynolds number. The predictionsmade by the proposed method produce up to 8 and 11% error in wall shear stress and heat flux with respect to DNS data, hence making it a useful tool for the preliminary engineering design
calculations of high-speed vehicles. Second, a new wall-model is developed with the goal of improving the prediction accuracy of wall heat flux over the current state-of-the-art for wall-modeled large eddy simulations of high-speed wall-bounded flows. The proposed model introduces two new modeling components: a simple model for the near-wall diffusion
of the turbulence kinetic energy, and an altered near-wall damping of the thermal eddy diffusivity. Both a priori and a posteriori tests are performed using reference DNS data of boundary layers up to Mach 10. The a priori errors for the proposed model are confined within 5%, although the a posteriori errors are larger, partly due to about 2% and 5% commutation error in the wall friction and heat transfer coefficients emerging from applying the wall-model on instantaneous data instead of averaged data.
In Part II of this thesis, a detailed study on three-dimensional effects in turbulent shear layersis performed with the objective of increasing our basic understanding. The skewed shear layers are generated by shearing two misaligned boundary layers at their interface. In the long-time, a skewed shear layer approaches the planar shear layer state, provided that it is analyzed in the mean shear frame. The most prominent three-dimensional effect is the presence of a planar jet-like flow orthogonal to the mean shear direction, that decays slowly in time. Nonetheless, the decaying jet has limited influence on the overall shear layer once its relative strength diminishes. Consequently, the mean statistics of a skewed shear layer eventually evolve similar to a planar shear layer despite the continued presence of a spanwise jet