Dissertations

The past decade of computing has been marked by the expected end of Moore's Law and Dennard's Scaling, and by a tapering in computing performance and energy efficiency. Instead, the community has witnessed continuing increases in transistor density and computing efficiency. This trend has been fueled in no small part by the diaspora of accelerator architectures. In contrast to the general-purpose architectures that characterized earlier eras, accelerators are circuits specialized for particular applications. But as computing capabilities have increased, so has the size of workloads, particularly in machine learning.

To meet these demands, this thesis investigates methodologies for improving the scalability of accelerator architectures. We begin by exploring tiled Markov-Chain Monte Carlo accelerators, identify inefficiencies in prior work, resolve them, and present the BigLittleMCA architecture, which leverages burn-in to improve power efficiency by 47% relative to prior work.

While BigLittleMCA provided application-specific optimizations to improve scalability, we turn our sights to how we can facilitate scalability for large SoC architectures. To this purpose, we develop a general-purpose programming framework for multi-core, platform-agnostic accelerator development: Beethoven. We demonstrate how this work can be used to deploy accelerators through a range of approaches, from microkernel to application-scale accelerators. We begin by demonstrating the shortcomings of prior work on a 4KB memcpy benchmark. Then, we compare against High-Level Synthesis approaches using the MachSuite benchmark suite. Next, we present how Beethoven can be used to accelerate a single-core accelerator from prior work as a multi-core FPGA system, achieving 3.3 times higher throughput and 34 times lower energy per operation than a GPU, for our chosen workload, attention. And finally, we show how existing multi-core accelerator systems exhibit structural patterns that can be targeted using Beethoven's abstractions using a boosted decision-tree accelerator design from prior work.

As the final work, we present TinyProSE, an accelerator architecture for decoder-only LLM models, and its tape-out in TSMC 16N silicon using Beethoven. In this work, we identify inefficiencies in computing non-linear functions and develop an area efficient methodology, TinyAct, for computing them.

This dissertation presents two nonlinear finite element methods: one for simulating cavitation in soft isotropic tissues and the other for dynamic fracture in brittle materials. Both finite element codes implement the open-source Deal.II C++ library and derive from the same tutorial program. Here we present our computational toolbox of compatible finite element models, developed with high-performance computing (HPC) in mind, featuring technical capabilities such as adaptive mesh refinement, adaptive mesh deletion, and adaptive time stepping. These models separately address distinct nonlinear behaviors involved in the multivariate biomechanical process of injection-driven cavitation and fracture in soft isotropic tissues. Although the combined phase-field fracture-cavitation model is outside the scope of this current dissertation, it concludes by clearly outlining the path toward its forthcoming completion.

Soft isotropic tissues, e.g. brain and liver, deform in an exceedingly nonlinear manner due to their incompressible, biphasic, and viscous nature. Thus, we adapt a state-of-the-art material model which combines nonlinear poromechanics and finite viscoelasticity. This model also utilizes the vaunted Ogden constitutive relation which is the standard choice for simulating high-strain applications, e.g. deformation of soft tissues. Using this highly advanced model, we demonstrate several key experimentally observed mechanical behaviors in soft tissues: hysteresis, preconditioning, and tension-compression asymmetry. Our next numerical example simulates wetting cavitation in brain tissue, wherein fluid flows into the expanding cavity. Although this does not reflect our stated injection-driven application, wherein fluid flows out of the expanding cavity, this serves as an important proof of concept which further motivates development for the combined phase-field fracture cavitation model. We extensively detail this combined phase-field fracture cavitation model as the logical conclusion to this research in the future directions discussion.

We simulate dynamic brittle fracture in polymethyl methacrylate (PMMA) using the phase-field method, integrating time using the implicit generalized-and explicit HHT-methods for their favorable high-frequency dissipation. Using an error convergence test, we validate our elastodynamic time integrator by observing the expected quadratic spatiotemporal error convergence rate. Our first physically motivated numerical example presents a mesh convergence study, featuring a classic dynamic brittle fracture benchmark problem. To showcase the favorable high frequency dissipation, we also simulate a highly complex fracture topology, featuring multiple branching and merging events. Our next numerical example investigates delamination fracture by considering discrete elastic heterogeneity. Lastly, we produce microbranching instabilities using pre-stretch loading and a random Gaussian distribution for the critical fracture energy release rate. We also detail our phase-field fracture modeling of dynamic brittle fracture in polyacrylamide in the appendix material.

Children who survive critical illness often experience psychosocial and cognitive sequelae that persist long after hospital discharge. Across three complementary studies, this dissertation examines neurocognitive and psychosocial processes that occur on a continuum before and after pediatric critical illness. Study 1 examines pediatric intensive care unit (PICU) delirium and its relationship to cognitive dysfunction during and beyond PICU admission. The literature shows a high incidence of delirium in the pediatric intensive care unit (PICU). Children have high rates of cognitive dysfunction following PICU admission. Delirium has been linked to declines in academic performance. The primary objective of this investigation was to examine the incidence of delirium in children admitted to the PICU and the relationship between delirium and long-term cognitive dysfunction. A prospective observational design was employed at an urban, academic children’s hospital. Participants were children aged 5-17 with an expected PICU stay exceeding 24 hours. Delirium screening was conducted daily during PICU admission. Children’s cognitive functioning was assessed using the Children’s Memory Scale (CMS) during hospitalization and at three-month follow-up. Parents completed the Behavior Rating Inventory of Executive Function (BRIEF-2) to evaluate pre-hospital and post-discharge executive functioning. A total of 146 children were enrolled. A subset of 103 completed in-hospital and/or three-month CMS (mean age = 11.0 years). Approximately one-third of children exhibited cognitive dysfunction at three-months. During admission, 37.3% of children screened positive for delirium, but results demonstrated no relation between delirium burden and cognitive dysfunction at either time point. However, children exhibited sustained deficits from hospitalization to follow-up. This investigation is a novel evaluation of the influence of pediatric delirium on later cognitive function. Children admitted to the PICU demonstrated high rates of cognitive dysfunction during hospitalization and at three-month follow-up. While in-hospital cognitive dysfunction was associated with later cognitive dysfunction, delirium burden was not associated with cognitive dysfunction at either time point. Study 2 investigates the relationship between posttraumatic stress symptoms (PTS) and cognitive function from pre-hospitalization through post-discharge, examining how pre-existing trauma and in-hospital cognitive dysfunction relate to later PTS in PICU survivors. To conduct exploratory analyses of pre-existing posttraumatic stress (PTS), in-hospital cognitive function, and later PTS and cognitive function at post-hospitalization follow-up in pediatric critical care survivors. This study employed a prospective, observational design in a pediatric ICU (PICU) and general care inpatient floor at an urban, academic, children’s hospital. Children aged eight to 17 years-old without developmental delay, severe psychiatric disorder, or traumatic brain injury were included. Children’s prehospitalization trauma history was assessed with the University of California Los Angeles-Reaction Index (UCLA-RI). PTS was present if children had four of the Diagnostic and Statistical Manual of Mental Disorders-IV criteria for posttraumatic stress disorder. Chi-squared tests compared presence or absence of pre-hospitalization PTS and in-hospital cognitive function, along with PTS and cognitive function at follow-up. Over half of children reported that, prior to their PICU hospitalization, they had been exposed to trauma. Children also had high rates of PTS related to their PICU stay when assessed three-months following their PICU admission. There was a significant association between the presence of pre-hospitalization PTS and the presence of in-hospital learning/encoding deficits, χ²(1, N = 41) = 4.985, p = .026. There were no significant relations between PTS and number of cognitive deficits at either time point. However, given the limited power of the current study, findings should be interpreted with caution. These data highlight the need for increased awareness amongst mental health providers of the high incidence of PTS in children following PICU admission, and integration of proactive screening and deployment of resources for children with trauma exposure prior to hospitalization. Study 3 examined posttraumatic growth among children following PICU admission and evaluated associations with PTS and family functioning. Data from two prospective cohort studies at a large, urban children’s hospital were combined using integrative analyses. The sample had N = 121 children (ages 8-17) assessed three-months following PICU admission. PTG was measured with the Posttraumatic Growth Inventory for Children-Revised (PTGI-C-R); PTS with the UCLA PTSD Reaction Index (UCLA-RI); and family functioning (cohesion and flexibility) with the Family Adaptability and Cohesion Evaluation Scales-IV (FACES-IV). Demographic and medical covariates were collected. Group comparisons and hierarchical linear regression evaluated predictors of PTG. Children with PTS had higher PTG scores than those without PTS (t(116) = −3.61, p < .001, d = −0.67). In multivariable models, PTS remained the only significant predictor of PTG (B = 5.11, SE = 1.42, β = .33, p < .001; R² = .107). Parent education showed an inverse univariate association with PTG; family cohesion (lower) correlated with higher PTG univariately, but family cohesion/flexibility were not significant in adjusted models. Following PICU admission, children commonly report PTG, which co-occurs with distress. Findings underscore the importance of recovery approaches that both mitigate PTS and promote PTG, and they motivate research on cultural/family contexts and longitudinal trajectories beyond three months. Collectively, findings advance a multidimensional model of post–intensive care outcomes in children, emphasizing that cognitive and psychosocial functioning are interdependent domains of recovery. This work highlights the need for integrative screening and follow-up care in pediatric critical care that focus on long-term neuropsychological and emotional health beyond mere survival of life-threatening illness.

Conservation capacity development has been widely promoted as a key mechanism to achieve sustainable natural resource management, yet persistent inequities, unclear assumptions, and limited evidence of long-term impacts raise questions about its effectiveness and fairness. This dissertation addresses these gaps by examining the context, processes, and outcomes of conservation capacity development through the lens of social equity in the study of the World Wildlife Fund for Nature’s (WWF) Rufiji-Mafia-Kilwa (RUMAKI) Seascape program’s nearly two decades of collaboration with community-based organizations in the coastal fishing community of Somanga, Tanzania. This case study research investigates how capacity development has been implemented, experienced, and applied by local actors and the implications it holds for implementing equitable, sustainable coastal governance.To explore these issues, I employed a qualitative, co-developed case study design to holistically reflect on conservation capacity development experiences. I collected data through key informant interviews, semi-structured interviews, focus group discussions, and validation workshops in 2023 and 2024, which I complemented with a review of gray and scientific literature on Tanzania’s coastal governance and policy context. My data analysis combined thematic inductive and deductive coding, guided by literature on capacity development and theories of social learning and multidimensional social equity (i.e. recognitional, procedural, distributional, and contextual equity). The study findings reveal that WWF’s long-term engagement fostered significant learning that community participants leveraged in their community-based organizations for coastal resource management and livelihoods, and, ultimately, improved acceptance of fisheries co-management. Community views shifted from resistance to active participation in coastal governance. Conservation capacity development contributed to these processes by increasing environmental awareness, leadership, entrepreneurship, and community inclusion in natural resource management. However, inequities remain in access, participation, and leadership for the community-based organizations and their engagement in externally led opportunities like capacity development. The findings of the case study demonstrate that conservation capacity development in a coastal fishing community involves a wide diversity of processes, learning can ripple out from individuals to community organizations and wider social units towards adopting new beliefs and practices that result in social and environmental change. These change processes are dependent on trusting relationships, opportunities, and motivations to engage in and apply learning from the capacity development that is deeply influenced by historical, political, social, and economic structures. While case study findings are limited to a singular program and village in Tanzania, other tropical, coastal fishing case studies could also apply the case study methods and approaches to analysis to assess broader applicability. Based on the case study findings, future efforts to strengthen conservation capacity could advance from ad-hoc, brief technical trainings toward more transformative, equity-centered approaches that recognize community agency, redistribute design or decision-making power, and address the broader systemic factors shaping local coastal resource management. Theories or expertise from behavioral science and educational fields could inform future capacity development research and practice in addressing how adults learn and apply their learning in coastal resource management. Further equity analyses could improve the design and evaluation of conservation capacity development to identify barriers and opportunities to fair practices and wider acceptance of sustainable natural resource governance practices.

Adolescent and young adult (AYA) cancer survivors face unique post-treatment challenges, including disrupted developmental trajectories and limited engagement in survivorship care. While self-management is essential for long-term wellness, it remains ambiguously defined and inconsistently supported in this population. This dissertation explores self-management as a developmental construct shaped by identity, autonomy, and evolving goals, and examines how peer support contributes to that process. Across four integrated studies, this work clarifies the relationship between self-management and patient activation, identifies the self-management roles and responsibilities and their alignment with developmental indicators of Positive Youth Development (PYD), and analyzes goal pursuit and coaching interactions within a peer support intervention. Findings demonstrate that self-management and patient activation are interrelated processes, with survivors’ roles and responsibilities aligning closely with Positive Youth Development indicators. AYA cancer survivors pursued layered, personally meaningful goals that reflected broader developmental aims and identity formation. Peer support interactions varied in style and type over time, responding to survivors’ evolving needs and developmental context, and were associated with outcomes in self-management and mental health-related quality of life. These findings contribute to a more nuanced understanding of self-management as a developmental achievement and underscore the value of developmentally responsive, survivor-led peer interventions. They also inform future research and practice aimed at optimizing peer support models and promoting resilient survivorship in AYA populations.

Tropical coastal marine ecosystems (TCMEs), including coral reefs, mangroves, and seagrass beds, are rich biodiversity hotspots which provide many ecosystem services for local human populations. TCMEs face increasing threats from local drivers of overexploitation and development and global drivers of anthropogenic climate change. Global concern for the health of TCMEs has led to an increase in both marine conservation interventions and climate mechanism research on TCMEs. Local management and conservation initiatives are important for addressing climate change, and understanding community-level perspectives is increasingly seen as important for successful implementation of those initiatives. In this dissertation, I use guiding frameworks identifying both social and ecological impacts of conservation interventions on TCMEs and anthropogenic climate change on coral reef systems. I approach these tropical systems as social-ecological systems (SES), and thus use systems approaches to understand systems-level impacts. The collective aim of my dissertation chapters is to inform science and practice where social-ecological lenses of both research and perception focus on these systems, versus where there are gaps in both research and understanding.In Chapter 1, I apply a systematic mapping approach to describe the literature on social and ecological outcomes associated with conservation interventions in TCMEs and show the extent, occurrence, and characteristics of evidence related to conservation interventions and outcomes. I find research gaps on common interventions (e.g., enforcement & prosecution), important social outcomes (e.g., knowledge & behavior), habitats (e.g., mangroves), and geography (e.g., the Middle East). In Chapter 2, I focus on coral reef social-ecological systems and apply network approaches to scientific literature reviews to code qualitative connections from climate threats to subsequent outcomes as nodes in a social-ecological network (SEN). I find connections generally focused on certain climate threats (e.g., sea surface temperature) and the most-connected literature nodes related to coral reefs and reef fish while impacts to shellfish and human communities were less connected in these generalized reviews. In Chapter 3, I assess stakeholder perceptions of climate threats impacts in the Caribbean, utilizing the same qualitative coding as Chapter 2 to create both aggregate and country-level perception SENs. I find that coastal stakeholders exhibit climate threat perceptions in the form of complex SES. I also find that individual country networks showcase nuance in climate perceptions, while an aggregate Caribbean SEN can help distill common climate threat perception pathways throughout all country sites. This dissertation closes in a Conclusion section where I summarize key dissertation findings and suggest areas where further research could continue to build upon this dissertation.This dissertation contributes to the field by characterizing tropical social-ecological systems via impacts from both conservation and climate change. Conceptualizing these systems as social-ecological networks adds nuanced perspective to the pathways by which both scientific reviews and local perception describe coral reef SES

Purpose:

To develop, validate, and demonstrate a customizable photon-counting detector computed tomography (PCD-CT) model for optimizing energy settings and generating harmonized virtual monoenergetic images to enhance spectral separation, inter- and intra-scanner consistency, and material imaging accuracy.

Methods:

For the development, validation and utility-demonstration, a customizable simulation model, DukeCounter, was developed to replicate real PCD-CT systems. Photon transport and crosstalk in PCDs were modeled using Monte Carlo simulations, and charge sharing was implemented using an analytical Gaussian charge cloud model. The fundamental interactions in PCDs including photoelectric absorption, Compton and fluorescence x-ray scatterings, charge cloud formation, and charge diffusion and repulsion were modeled. Spatio-energetic detector responses were generated for face-on CdTe-, CZT-, GaAs-, and edge-on Si-based PCDs. These responses, combined with standardized scanner parameters, were integrated into a CT simulator to create virtual DukeCounter PCD-CT scanners. The framework was benchmarked against experimental data from a clinical CdTe-based PCD-CT scanner across three dose levels. To demonstrate its utility, three pilot studies were conducted using a computational ACR phantom for task-generic image quality assessment, an XCAT model with bronchitis and emphysema for COPD biomarker extraction, and an XCAT with liver lesions for lesion detectability analysis.

For the PCD-CT optimization, CdTe-, CZT-, and Si-based PCD-CT scanners were modeled using an in-silico imaging framework by incorporating scanner geometries and validated spatio-energetic detector responses of DukeCounter to account for inter-pixel and inter-energy crosstalk and noise correlation. Using these models, two energy settings (thresholding and binning), two dose levels, and two sizes of a cylindrical phantom containing inserts with various concentrations of calcium, iodine, and gadolinium were simulated. The tube voltage (120 kV), pitch (1), gantry rotation speed (0.5 rot/second), and reconstruction settings were held constant across all acquisitions. Quantitative metrics including the separability index (s’) and contrast-to-noise ratio (CNR) were determined. Energy settings were ranked based on the s’, and the rank-sum method with a Friedman test was used to determine the optimal setting.

For the PCD-CT harmonization, in-silico models of CdTe-, CZT-, and Si-based PCD-CT scanners were simulated to generate four energy-bin CT images for both multi-material cylindrical and anthropomorphic (XCAT) phantoms. Images were acquired across four exposure levels (50, 100, 200, 400 mAs), three helical pitches (0.8, 1.0, 1.2), and two reconstruction kernels (Hann and Ram-Lak), while keeping other parameters constant. Ground-truth VMIs (in HU) were generated from the known linear attenuation coefficients at energies 30 to 90 keV (5 keV intervals). A conditional U-Net architecture was trained to predict a target-energy VMI given the four energy-bin images and an encoded energy level as inputs. The model was trained on 75% of data, validated on 15%, and tested on 10%. Quantitative performance was assessed using root mean square error (RMSE), mean absolute percentage difference (MAPD), structural similarity index measure (SSIM), and voxel-wise Bland-Altman analyses across scanners and acquisition conditions.

Results:

From the PCD modeling study, the simulated charge cloud size increased with energy and was more pronounced in Si due to its low atomic number. The detector response across a 3×3-pixel neighborhood varied with PCD material, design, and energy threshold settings. Validation results demonstrated strong agreement between simulated and real ACR images. For the 20-keV-threshold images, the mean relative difference (MRD) in f50 of MTF was 4.15%1.21 for air and 2.54%2.08 for bone, and the MRD in fav of NPS was 0.83%0.97. The MRDs in noise magnitude were 2.65%1.68, 3.05%1.97, and 2.78%1.79 for the 20-keV-threshold, 65-keV-threshold, and 70-keV-VMI images, respectively. The MRDs in CT number for the same image types were 0.03%0.03, 0.11%0.09, 0.11%0.05 for air, and 1.85%0.20, 1.84%0.55, 0.50%0.36 for polyethylene. DukeCounter-generated images showed that task-generic and task-specific image qualities were influenced by PCD materials, designs, and energy threshold settings. GaAs-based DukeCounter exhibited the highest image noise, the largest error in COPD biomarker quantification, and the lowest performance in liver lesion detection, under consistent acquisition and reconstruction settings.

From the PCD-CT optimization study, the optimal energy settings varied primarily with phantom size and material pairs to separate (Ca-I, Ca-Gd, I-Gd) rather than dose level. When aggregated across all imaging tasks and conditions, the optimal energy settings were 30-65 keV and 20-35-50-70 keV for two- and four-threshold CdTe-, 20-35-50-70 keV for four-threshold CZT-, and 5-35-50-80-120 keV for four-bin Si-based PCD-CT systems. Four-threshold CdTe showed significantly higher s’ performance than two-threshold (mean difference of 0.87 ± 1.57, p < 0.001). Four-threshold CZT performed slightly but consistently better than CdTe (mean difference of 0.11  0.28, p < 0.001).

From the PCD-CT harmonization study, the conditional U-Net model achieved high quantitative and structural accuracy in generating harmonized VMIs across all scanners and imaging conditions for both cylindrical and XCAT phantoms. For cylindrical phantoms, the mean RMSE and MAPD were 8.5  4.4 HU and 11.1  4.3 % for CdTe-, 9.4  3.8 HU and 11.2  3.6 % for CZT-, and 12.2  6.9 HU and 16.4  8.0 % for Si-based PCD-CT scanners. For the XCAT phantom, performance was comparable (RMSE = 12.1-13.2 HU and MAPD = 11.8-16.0 %), improving at higher exposure levels (RMSE = 11.9  5.3 HU, MAPD = 11.2  5.2 % at 400 mAs). Voxel-wise Bland–Altman analysis revealed minimal bias (< 3 HU) and tight 95% limits of agreement ( 17 HU) between scanner pairs, confirming inter-scanner harmonization. Structural similarity between predicted and ground-truth VMIs was high (SSIM = 0.96-0.99), demonstrating high spatial and spectral fidelity across all imaging conditions.

Conclusion:

A customizable, modular simulation framework was developed to model spatio-energetic detector responses for various PCD materials and designs. The detector responses were integrated into a CT simulation pipeline to build DukeCounter PCD-CT systems. The framework’s utility was demonstrated through task-specific assessments of image quality and clinical performance of DukeCounter systems using XCAT phantoms. This approach enables systematic PCD-CT design evaluation and optimization, supporting translational research in medical imaging by reducing the cost, time, and radiation burden of physical experiments.

The s-prime optimization work presented a simulation-based framework for optimizing PCD-CT scanners and provided clinically translatable energy settings to advance the clinical adoption of PCD-CT and support accurate, standardized spectral imaging.

The proposed conditional U-Net effectively learns the nonlinear mapping between multi-energy PCD-CT data and corresponding VMIs across various scanners and acquisition protocols. By leveraging realistic, physics-informed in-silico datasets that provide ground-truth mappings, this framework offers a computationally efficient, data-driven pipeline for harmonized VMI generation. The approach eliminates the need for scanner-specific calibrations or material-basis selection and provides standardized, scanner-agnostic spectral imaging, advancing the clinical translation and quantitative reliability of PCD-CT.

This dissertation investigates data-driven and simulation-based approaches for quantifying the resilience and sustainability of wastewater treatment operations under increasingly variable climatic conditions. The growing frequency and intensity of rainfall events threaten the continuity of biological nutrient removal processes, leading to hydraulic overloads, biomass washout, and effluent quality deterioration. To address these challenges, this work integrates machine learning, resilience quantification, and life-cycle assessment (LCA) to better understand and manage wet-weather impacts at wastewater resource recovery facilities (WRRFs).In the first chapter, we develop and implement machine-learning models for predicting wet-weather flows. Because extreme inflow events are rare and difficult to capture in conventional datasets, the study applies data-level resampling methods such as Synthetic Minority Oversampling (SMOTE) and Adaptive Synthetic Sampling (ADASYN) to improve the representation of infrequent but operationally critical high-flow conditions. Testing across two full-scale WRRFs, one served by a combined sewer system and the other by a separate sanitary system, demonstrates that the ADASYN–Random Forest configuration provides the most reliable early warnings of extreme hydraulic loads, offering actionable lead time for operators to implement flow-management strategies. Building on predictive modeling, the second component introduces a data driven framework for quantifying process resilience directly from routine operational data. Using Statistical Process Control (SPC) to identify wet-weather disturbances, three complementary metrics are derived to capture the magnitude, duration, and cumulative effect of performance degradation. Application of this framework to two nitrogen-removal systems reveals distinct differences in resilience between storm and non-storm disturbances, with integral performance loss emerging as a robust indicator of process resilience. The results from this chapter demonstrate that empirical resilience analysis can be achieved using existing monitoring data, providing utilities with an efficient diagnostic tool for assessing process resilience without the need for additional sensors or external hydrologic predictors. The final component integrates dynamic plant-wide simulation with LCA to evaluate both plant resilience and sustainability of different wet-weather management strategies, including direct wet-flow treatment, flow equalization, and targeted chemical dosing. Lifecycle inventory derived from the process model quantifies energy use, chemical consumption, and greenhouse-gas emissions under each scenario. The analysis shows that nitrous oxide (N2O) dominates the carbon footprint of the liquid treatment line, while flow equalization significantly reduces both emission variability and eutrophication potential by buffering hydraulic peaks. Importantly, the results reveal that environmental burdens are more related to cumulative performance loss (rppr) than with performance degradation (mpr), underscoring the link between operational resilience and sustainability. Collectively, this work advances our understanding of wet weather and quantifying both resilience and sustainability of plant operations for various wastewater systems. By combining predictive analytics, empirical resilience metrics, and sustainability assessment, the dissertation provides new methodological tools for operating treatment plants that are both robust to hydraulic disturbances and aligned with broader decarbonization goals.

Genome-wide association studies (GWAS) have identified hundreds of thousands of genetic variants associated with human traits and diseases, yet the vast majority of those variants lie in non-coding regions of the genome, and their molecular consequences remain largely unknown. A major mechanism through which non-coding variants act is by altering transcription factor (TF) binding. TFs are sequences-specific DNA-binding proteins that orchestrate gene regulation by recognizing specific DNA sequences or motifs within regulatory elements, reshaping chromatin environments, and guiding the activity of RNA polymerase. They are central to almost all aspects of cellular functions: controlling the cell cycle, modulating responses to environmental stimuli, guiding cell development and differentiation, and enabling cellular reprogramming. Since TFs are essential for those critical functions, disruptions of TF binding caused by non-coding variants can propagate through gene regulatory network and lead to change in phenotype or contribute to disease risk. Understanding how non-coding variants affect TF binding therefore remains a central challenge in functional genomics. This dissertation addresses this challenge in four parts. Chapter 1 (introduction) provides biological and computational background for this problem. It reviews the roles of TFs in gene regulation, the importance of modeling the effects of non-coding variants on TF binding, and major experimental platform for study TF binding, including both in vivo and in vitro technologies which map TF occupancy in cellular contexts and characterize intrinsic binding preferences respectively. The chapter also summarizes existing computational approaches, highlighting both their contributions and their limitations. Chapter 2 introduces QBiC-SELEX, a modular computational method designed to predict the effects of genetic variants on TF binding using HT-SELEX data. QBiC-SELEX addresses systematic biases across SELEX-based platforms by explicitly modeling the experimental artifacts. Trained on over 4,000 HT-SELEX experiments covering more than 1,000 human TFs with a robust model curation strategy, QBiC-SELEX consistently outperforms existing methods across in vitro and in vivo benchmarks. Importantly, QBiC-SELEX not only provides a large-scale resource of curated variant-effect models for more than 1,000 TFs but also offers new instructive insights for how SELEX-based experiments can be more effectively designed and computational methods can utilize this type of data. Chapter 3 presents the second method, which addresses the complementary problem of calling TF binding sites across a broad affinity range. Unlike traditional methods such as position weight matrices (PWMs), which prioritize high affinity binding sites, CtrlF-TF uses k-mer data derived from universal protein binding microarray and processed HT-SELEX data generated by QBiC-SELEX to better capture medium- and low-affinity binding sites which are functionally important for regulatory activity. Benchmarks with PBM, ChIP-seq and genomic footprinting data demonstrates that CtrlF-TF calls substantially more TF binding sites than PWM while preserving lower false positive rates. The Chapter 4 summarizes these contributions and outlines a path forward. A potential opportunity is to integrate curated in vitro TF-DNA binding data into a small TF-specialized pretrained genomic language model, which can then be fine-tuned with in vivo datasets such as ChIP-seq and ATAC-seq. This integrative approach has the potential to capture both intrinsic binding preferences of TFs and context-specific occupancy with high specificities. Together, QBiC-SELEX and CtrlF-TF provide valuable community resources for variants effect prediction and TF binding discovery. They deepen our understanding of into how SELEX-based data can be effectively modeled and demonstrate the principles of building robust and interpretable methods in functional genomics, laying the foundation for future integrative approaches such as genomic language model that unifies all TF data.

Type 2 diabetes (T2D) and its precursor states, including impaired fasting glucose (IFG) and impaired glucose tolerance (IGT), develop gradually over years. During this early phase, individuals may experience intermittent dysglycemia, altered autonomic regulation, and subtle cardiometabolic stress, often without meeting clinical thresholds for diagnosis. Current frontline screening tools -- fasting glucose, hemoglobin A1c (HbA1c), and, less commonly, the oral glucose tolerance test (OGTT) - are limited in their ability to capture this early biology. HbA1c and spot glucose tests provide static, low-resolution snapshots of a highly dynamic system, and they can fail in both directions: they can classify individuals with ongoing physiologic dysregulation as “normal,” and they can label others as high risk even when physiology is improving. OGTT, which probes post-challenge glucose handling over time and can detect IGT before overt T2D, is more sensitive but remains underused in practice because it is time-intensive, clinic-bound, and logistically burdensome. As a result, many people at greatest risk for progression to T2D are either not identified or not stratified in a way that supports targeted intervention.This thesis addresses that gap by developing and validating digital biomarkers of dysglycemia derived from passively and semi-passively collected physiological data in free-living settings from individuals with normoglycemia, prediabetes (PD), and type 2 diabetes (T2D) in the DiabetesWatch study. We designed and execute a novel fully remote, sensor-rich digital health study to create and characterize a benchmark dataset that captures both dynamic glucose responses and continuous cardiometabolic physiology outside the clinic. We then leverage two complementary data streams collected in DiabetesWatch: (1) CGM-based meal challenge data enabling an at-home mobile OGTT (mOGTT), and (2) continuous multimodal physiological signals measured from wrist-worn wearable devices, including heart rate, heart rate variability (HRV), peripheral oxygen saturation (SpO₂), electrodermal activity, skin temperature, sleep, and activity to model autonomic, respiratory, and metabolic stress physiology. Across these datasets, we design computational methods to extract physiologically interpretable features, quantify individual metabolic responses, and map those responses to established clinical constructs. We present the DiabetesWatch dataset as a resource for discovering digital biomarkers of glycemic health. The cohort is deeply phenotyped with 14 days of free-living continuous glucose monitoring, commercial and research-grade wearable data, detailed food logs, HbA1c, and an oral glucose tolerance test. To our knowledge, this is the largest dataset of its kind spanning normoglycemia, prediabetes, and unmedicated type 2 diabetes with this level of multimodal categorization. Using DiabetesWatch, we demonstrate the feasibility of fully remote digital health studies for glycemic health that achieve high participant satisfaction, demographic diversity, broad geographic reach, and strong interest in follow-up, while capturing rich, heterogeneous sensor data - including relatively underexplored modalities such as SpO₂ and electrodermal activity. The study design and data collection pipeline provide a replicable template for future glycemic health research. We also introduce the mobile oral glucose tolerance test (mOGTT), a CGM-enabled alternative to the traditional OGTT that can be performed outside the clinic. In mOGTT, participants ingest a standardized glucose challenge while wearing a CGM, allowing us to reconstruct classic OGTT readouts - fasting glucose, post-challenge glucose, total glycemic exposure over two hours (AUC), and recovery back toward baseline - without venipuncture and without requiring repeated in-person sampling. Using data collected from adults spanning normoglycemia, PD, and T2D, we show that mOGTT metrics reproduce clinically meaningful patterns observed in traditional OGTT. These metrics align with diagnostic thresholds for IFG and IGT and capture a wide range of physiologic phenotypes, including elevated post-load excursions, delayed recovery, large total glycemic burden, and persistent fasting elevation. Importantly, we observe that individuals with “normal” HbA1c can still exhibit multi-metric dysglycemia on mOGTT, while some individuals with HbA1c in the T2D and PD range show physiologic improvement (e.g., rapid recovery, low excursions), suggesting that HbA1c alone can both under-call emerging risk and over-call people who may be reversing disease through lifestyle change. We then formalize these glucose dynamics into a composite glucose challenge response index (GCRI) score. This score is constructed by quantifying, for each participant, how far their fasting, peak, post-challenge, and cumulative exposure metrics deviate from clinically used OGTT thresholds, and then aggregating those deviations into a single continuous index of dysglycemia. Methodologically, this treats glycemic regulation as a multidimensional control problem rather than a single cutoff problem. We show that the GCRI increases monotonically with clinical risk and generalizes to an independent external cohort. Further, we evaluate whether chronic, passively acquired physiology from wearables can provide a scalable, fully noninvasive screen for dysglycemia. From continuous wrist-worn data streams, we derive autonomic and cardiometabolic features including resting heart rate, HRV, oxygen saturation, thermoregulatory fluctuation, and sleep-linked recovery metrics. We train and evaluate multiple predictive model families - regularized logistic regression, random forest, gradient boosting, and XGBoost, along with simple ensembles - using an analysis pipeline with strict separation between model selection and testing. These models distinguish PD/T2D from normoglycemia with moderate-to-high discrimination (area under the ROC curve up to 0.90 and area under the precision–recall curve up to ~0.96 in held-out evaluation), driven by physiologic signatures consistent with known early cardio–metabolic stress: elevated resting rate, autonomic imbalance, oxygenation instability, and blunted nighttime recovery. The features that repeatedly emerge as most informative are physiologically coherent and reproducible across model classes, suggesting that subclinical metabolic dysfunction is already imprinted in cardiovascular and autonomic control before overt diabetes. Taken together, this thesis demonstrates an end-to-end framework for digital glycemic phenotyping: from at-home, CGM-based dynamic stress testing (mOGTT), to composite dysglycemia scoring tied to clinical thresholds, to passive cardiometabolic risk screening from wearables. The central conclusion is that continuous and near-continuous biosensing can surface early metabolic dysfunction in a way that is richer, more personalized, and more scalable than traditional single-point clinical tests. By translating raw signals into quantitative digital biomarkers, this work lays the groundwork for proactive, physiology-aware monitoring of PD and emerging T2D. It supports a shift from episodic detection to continuous risk assessment, and from binary diagnosis to mechanism-informed stratification, with the longer-term goal of enabling earlier intervention, individualized prevention, and remote metabolic care at population scale.

This dissertation presents the implementation and characterization of a cryogenic trapped-ion quantum computing platform developed as part of the Software-Tailored Architecture for Quantum Co-design (STAQ) project. The system utilizes $^{171}$Yb$^+$ ions confined in a microfabricated surface-electrode trap and manipulated using Raman and microwave-driven gates to realize high-fidelity quantum operations. Operating at cryogenic temperatures ($\approx4$K) within an ultra-high vacuum environment, the computer integrates advanced optical, mechanical, and electronic subsystems to achieve improved motional coherence and scalable ion control. A key focus of this work is the design of the system’s architecture—spanning trap packaging, laser beam delivery, RF drive circuitry, and cryostat layout—to support long-range programmable spin-spin interactions via Mølmer–Sørensen gates. Special emphasis is placed on the optimization of voltage delivery for improved motional mode stability, as well as the development of custom hardware and software tools for system automation. The thesis includes detailed studies of experimental protocols for entanglement generation and characterization. Algorithmic applications, including quantum simulations of spin models and ergodic protocols for frustrated ground state preparation are also explored. Numerical simulations complement the experimental work, providing insight into entanglement generation via engineered dissipation, and thermal state preparation. This work contributes a fully integrated experimental framework for trapped-ion quantum computing. It offers both a reproducible model and a testbed for future advances in quantum control and hybrid quantum-classical computation.

Period product insecurity, defined as the lack of enough period products due to income, remains an understudied aspect of material hardship in the United States. Despite increasing public awareness, few tools exist to measure this type of hardship, which limits efforts to understand its prevalence, outcomes, and policy responses. This dissertation focuses on four connected issues: (1) the lack of a validated measure for period product insecurity, (2) the limited understanding of its link to mental health, and (3) the scarce empirical evidence on the social and political factors influencing state period product policies and (4) the consequences of experiencing period product insecurity and what period supply banks are doing to address them. To address these gaps, I created the Van Ness Period Product Insecurity Scale (VNPPIS), based on cognitive interviews, pilot tests, and survey methods adapted from the U.S. Household Food Security Survey Module. I estimated that roughly 8 percent of the U.S. adult population experiences period product insecurity and its associations with depressive and anxiety symptoms. At the same time, I compiled a new dataset of over 1,300 state bills related to period product introduced between 2009 and 2024. I use multilevel regression with poststratification (MRP) to model how public attitudes toward groups correlate with the likelihood of adopting strong policies. I then explore experiences of individuals experiencing PPI in the absence of public policies to support them. Findings indicate that period product insecurity is widespread, especially among low-income groups. State-level analysis shows that more comprehensive, inclusive policies are more likely to pass in states where public opinion supports marginalized communities and considers them deserving of aid. Overall, these findings emphasize that period product insecurity is both measurable and impactful, and that meaningful solutions can be implemented through public policy.

Neurons derived from induced pluripotent stem cells (iPSCs) are powerful tools for modeling neurological diseases, but current methods often produce cells with an immature, fetal-like phenotype, limiting their relevance. Here, we developed a high-throughput CRISPR activation (CRISPRa) screening platform to systematically identify transcription factor (TF) combinations that synergistically enhance neuronal maturation. Using a SYN1 promoter-driven fluorescent reporter to enrich for mature neurons, we screened a library of 1,920 TFs and epigenetic modifiers paired with the neurogenic factor NEUROG3 (NGN3). The screen identified multiple novel TF combinations that significantly improved neuronal differentiation. We focused on a top hit, the epigenetic modifier KMT2B, and demonstrated that co-activation of KMT2B and NGN3 generates neurons with significantly enhanced transcriptional, morphological, and functional maturity compared to NGN3 alone. KMT2B+NGN3-derived neurons exhibited gene expression profiles more closely correlated with adult human brain tissue and showed enrichment for GWAS risk genes associated with major psychiatric disorders. These neurons also displayed longer neurites, higher synaptic puncta density, and more robust, coordinated network activity on multi-electrode arrays. Mechanistically, ATAC-seq revealed that KMT2B+NGN3 co-activation remodels the chromatin landscape to increase accessibility at loci critical for axonal guidance and synaptic plasticity while suppressing pathways associated with pluripotency. This work establishes a scalable platform for dissecting the combinatorial logic of neuronal maturation and provides an optimized protocol for rapidly generating more physiologically relevant human neurons for in vitro disease modeling.

Irregular migration is one of the most controversial topics of our time. While natural disasters, war, poverty, and inequality are driving people to move at record numbers, wealthy countries are devoting every effort to keeping unwanted migrants out. One of the most overlooked ways that wealthy countries do this is by using other countries that migrants pass through on their way to the wealthy destination. This third category of country is called a transit state. While their use to repel migrants has not gone unnoticed by the leaders of the Global North, researchers have not fully explained how wealthy states convince/coerce them to help and the tools transit states use to stop these migration flows. This dissertation begins to answer this question by looking at how aid, refugee policy, and data capacity could be used to leverage transit states against unwanted, irregular migration.I use observational data from the US, EU, UNHCR, World Bank, and more to answer my question. In the first paper I look at migrants apprehended at the southwest US border and construct migration routes across the Western Hemisphere. I then look at how the use of these transit routes influences the amount of foreign aid a country on that route receives from the US. The second paper analyzes how policy pertaining to refugees changes in the Middle East and North Africa before and after the Syrian Civil War and refugee crisis. I use the new DWRAP database to compare policy changes across time in the relevant countries. I also use expenditure data from UNHCR to see how this kind of aid might influence such policy. The third paper uses metadata on FDP datasets from the World Bank and UNHCR to see which countries have the most data availability and capacity on the forcibly displaced. I find that transit countries are more likely to receive certain types of aid from the US than non-transit countries. Additionally, transit countries are not more likely to open their borders during a refugee crisis but are more likely to receive assistance from UNHCR. Finally, I find that transit states tend to have slightly more datasets and data availability than non-transit states.

Lung cancer remains one of the leading causes of cancer-related mortality worldwide, primarily due to late-stage detection and the complexities of accurate diagnosis. Recent advances in artificial intelligence (AI) and medical imaging have created new avenues for improving early detection and diagnostic accuracy.This dissertation explores how clinical knowledge and virtual imaging trials can be integrated to enhance AI-driven lung cancer diagnosis. The research leverages weakly supervised deep learning models trained on large-scale body CT datasets, linking textual radiology reports with corresponding image data. By incorporating simulated imaging environments, known as virtual imaging trials, the work systematically evaluates algorithmic performance under controlled but realistic conditions. This approach not only bridges the gap between experimental validation and clinical applicability but also ensures reproducibility and standardization. The results demonstrate that incorporating domain-specific priors and synthetic datasets significantly improves detection sensitivity while reducing false positives. Furthermore, the investigation clarifies how representational biases in both data and model training can be mitigated through hybrid strategies that reflect authentic clinical decision-making. Overall, the dissertation presents a novel framework that unites computational modeling and radiological expertise, providing a robust foundation for the next generation of AI systems in medical imaging. The findings have broad implications for translational healthcare, regulatory assessment of AI tools, and clinical adoption pathways.

Delayed diagnosis of Mendelian disease substantially prevents early therapeutic intervention that could improve symptoms and prognosis. Although clinical genetics has historically focused on identifying protein-coding variants that cause disease, growing evidence indicates a substantial role for non-coding variants, especially those that alter splicing and/or gene expression. This likely contributes to the incomplete diagnostic yield of exome sequencing in patients with suspected genetic disease. However, non-coding variants present additional challenges in detection, functional interpretation, and classification as pathogenic or benign. In a cohort of genetically undiagnosed Mendelian disease patients, we identified two siblings with glycogen storage disease (GSD) type IX γ2, both of whom had a classic clinical presentation, enzyme deficiency, and a known pathogenic splice acceptor variant on one allele of PHKG2. Despite the autosomal recessive nature of that disease, no variant on the second allele was identified by gene panel sequencing. To identify a potential missing second pathogenic variant, a combination of genome sequencing, short-read RNA-seq, and long-read RNA-seq detected putative deep intronic splicing variant and a corresponding pseudoexon inclusion in both siblings. We confirmed the functional splicing effects of this variant by generating a HEK293T cell model in which we installed the variant using CRISPR editing. That cellular model has genetic, biochemical, and cellular impacts that are consistent with GSD IX γ2 and allow us to screen candidate antisense therapeutics that target aberrant splicing, optimizing a potential personalized therapeutic prior to evaluation in the patient’s own cells. This demonstrates a robust pathway for detecting, validating, and reversing the impacts of novel non-coding causes of rare disease. Expanding the use of this approach could markedly improve the diagnostic yield of genetic sequencing and advance precision medicine.

Polydimethylsiloxane (silicone) wristbands have gained popularity over the last decade as low-cost tools to assess personal exposure to a wide range of semi-volatile organic compounds (SVOCs). The potential of these wristbands has been indicated by prior research demonstrating that concentrations of SVOCs sorbed to the wristbands correlate with paired biomarker concentrations. Therefore, these devices could offer a more convenient strategy for environmental exposure assessment. However, several unknowns remain which temper the full realization of the wristband’s capability. Specifically, although we know that the wristband concentrations correlate with biomarker concentrations, we do not understand the mechanisms and factors influencing these correlations. Another uncertainty is the wristband’s ability to sample metals such as lead (Pb) which are also ubiquitous in our environment and can cause health effects that overlap with SVOCs. This work reviews the motivation for the use of silicone wristbands as passive samplers in chapter 1 and addresses these uncertainties which are outlined in the following specific chapters:2. Measurement of the effect of movement speed on SVOC uptake to unworn silicone wristbands. 3. Quantification of SVOC uptake from gas and particle-phase sorption to worn silicone wristbands. 4. Modification of the wristbands for improved metal sampling and evaluation of their performance compared to other exposure assessment samples.

In chapter 2, we suspended silicone wristbands under static conditions on a ring stand or attached them to centrifuge tube rotators, which were spun at different speeds ranging from 0.05 to 1.1 m/s. The results of this experiment showed that faster moving wristbands had greater uptake rates of SVOCs. The fastest speed (~1.1 m/s), equivalent to a leisurely walking pace, displayed uptake rates for most compounds that were 4 to 5 times higher than the uptake rate of static wristbands. We also observed that the magnitude of increase with speed for a given compound was negatively correlated with its diffusivity in air and positively correlated with its octanol-air partition coefficient (log KOA), potentially indicating particle uptake or sampler side limiting mass transfer. When compared to worn wristbands, unworn wristbands moving at 1.1 m/s exhibited lower uptake rates for most compounds. This disparity became more pronounced as the SVOCs increased in KOA indicating an additional mechanism other than diffusion (e.g. particle adsorption or direct surface contact) may be driving uptake of the compounds.

In chapter 3, we explored these mechanisms further by recruiting participants to wear two watchband samplers, each containing a silicone segment that can accumulate SVOCs. One watchband was covered with a 10 micron stainless steel screen to filter out particles and while the other watchband was unscreened. The fraction of each SVOC passing the screen and accumulating on the silicone represents the fraction of exposure from the gas-phase. This data revealed that many compounds experienced noticeable uptake on the silicone sampling surface from both gas-phase and suspended particulates. Additionally, when comparing the concentration on the unscreened silicone in the watchband to the concentration on a paired standard wristband, we witnessed a similar trend as in aim 1 in which high KOA compounds had significantly greater concentrations on standard wristbands. Combined, these results point to direct contact with surfaces and skin as major mechanisms of SVOC uptake by standard wristbands.

In chapter 4, having known the evidence of particulate SVOC uptake, we investigated the effectiveness of a modified wristband as an exposure assessment tool for Pb and other metals. The modification consisted of placing a polyurethane foam substrate in the cavity of a fitness style watchband. The polyurethane foam contains higher specific surface area than silicone for improved particle uptake and retention. We recruited participants to wear these modified wristbands for five days and to provide dust wipe, blood, and water samples. In some cases additional vacuumed dust and soil samples were also collected. We found that Pb and other metals (Cu, Fe, Mn, Cd, Ca, Na) on PUF were positively correlated with paired dust wipes samples. However, Pb on wristbands was not strongly correlated with Pb in blood. Additionally, we compared ratios of Pb:Fe, Cu:Fe, Mn:Fe, etc. across sample types, which revealed that the particle composition on the PUF was more similar to that of dust wipes and vacuumed dust than outdoor soil. Overall, these modified wristbands can be paired with silicone wristbands to expand the range of chemicals and provide a more comprehensive assessment of exposure.

Combined, these three chapters demonstrate that wristband passive samplers capture valuable exposure information by integrating multiple exposure routes and diverse chemical classes, much like the human body itself. Collectively, these results lay the groundwork for mechanistically linking wristband data to health outcomes and for extending the wristband approach to broader exposome assessment.

This dissertation showcases the capability of machine learning within the computer vision domain for applications among environmental and manufacturing domains. Improvements in the accessibility and resolution of imaging now allow vast image streams to be converted into reliable signals for decision making across surveillance, monitoring, and anomaly/defect detection. However, the evolution of image data in the age of machine learning exposes a pivotal requirement with their use: methods that remain applicable across domains, resolutions, and sensors that can scale smoothly from limited data. In practice, real-world datasets are heterogeneous in size, exhibit realistic structured missingness (i.e., cloud cover, human error), and vary in spatial resolution. This work emphasizes the ability to use machine learning among different image types and among different use cases in different domains.Many popular state-of-the-art models are trained on web-scale corpora, and excel at canonical benchmarks, yet the data they learn from often differ markedly from the imagery encountered in remote sensing-based tasks, or manufacturing microscopy tasks. This domain dependence is well known among the computer vision domain. Convolutional Neural Networks pretrained on RGB-banded images can degrade performance when used out of the box on single channel, or hyperspectral data applications, despite the use of images as inputs. It is imperative to approach the application of real-world data for real world tasks (not everything is about classification) with the use of machine learning with a careful and calculated approach. With these real-world datasets comes the reality of small, task specific datasets. Recognizing this reality, the work presented here leverages appropriate pretrained networks to translate large-model strengths into reliable performers for daily environmental monitoring and manufacturing inspection. In the era of large foundation models (LLMs, and diffusion models), memory efficient convolutional architectures remain highly effective for targeted image analysis, especially in small-sample, domain-shifted settings. This work demonstrates two complementary applications on a curated set of 613 high-resolution (3m/pixel) satellite images over Thilafushi Island, a manmade landfill in the Maldives historically known for consistent waste burning. Firstly, a pretrained CNN was finetuned to classify images as containing or not containing a waste burning plume. Secondly, a pretrained-UNet was adapted for binary semantic segmentation to localize plume pixels within the images—a task unforeseen in previous moderate and low-resolution datasets. Leveraging transfer learning and careful regularization enables robust detection despite limited images and hence limited positively class images. As of 2021, a governmental ban on open waste burning was placed on the Thilafushi Island, and this developed model on archived data from 2016-2021 provided support for image-drive compliance monitoring after the ban. A second case study applies high-resolution satellite imagery to surveil and track harmful algal blooms along the Chowan River of North Carolina (USA). We employ commercially available 3m/pixel imagery, contrasted against coarser products (~300 m/pixel), as input to a pretrained CNN for image classification. The higher spatial resolution provides finer, attuned detail of coastline and shoreline blooms, that tend to be sub-pixel or unrepresented at all at 300m resolution. Real-world riverine scenes introduce unique variability and noise via sunglint, detritus, aquatic vegetation, and seasonal changes to the watercolor, while the blooms themselves exhibit inconsistent and amorphous shapes, providing a unique challenge for state-of-the-art labeling methods. This study demonstrated the practical use of high-resolution, learning-based bloom monitoring for public-health management via support to rural sampling teams. The last use case regarding remote sensing data that this dissertation addresses regards spatiotemporal gaps in a new generation of air quality monitoring. NASA’s TEMPO instrument delivers hourly, neighborhood scale retrievals of tropospheric trace gases, enabling diurnal analysis unforeseen with legacy low-Earth-orbit sensors. However, TEMPO is not exempt of the caveats of missing data that plague real world remote sensing datasets; daylight-only sampling produces nocturnal gaps, while cloud coverage and quality-flag filtering create spatially irregular missing pixels. We treat these challenges separately. For temporal gaps, we train classical regression models to forecast overnight NO₂ from evening TEMPO snapshots and auxiliary drivers (meteorology, boundary-layer indices, emissions proxies), supported when available by lunar Pandora spectrometer observations. For spatial gaps, we develop a mask-aware Partial Convolutional U-Net that conditions on the missing data mask, which learns to preserve edges as well as learns spatial gradient trends without ground truth pixels. This adapted U-Net shows improved spatial coherence and structure when compared to baseline image inpainting techniques. This study highlights the versatile use cases of machine learning, traditional and deep learning, for use on novel, real-world datasets to produce robust data-efficient methods. On the extreme other end of the resolution spectrum, this dissertation examines the microscopic limit of the same anomaly-detection problem. In laser powder bed fusion metal additive manufacturing, microscopic pores and lack-of-fusion defects can compromise performance of parts meant for high stress and high pressure uses. This study aims to bridge the modalities of thermal tomography (TT) and scanning electron microscopy (SEM) with machine learning to enhance detection efforts. A multi-task U-Net with a shared encoder takes defect relevant information from datasets from both quality control datasets, and performs their respective tasks; segmentation on SEM images, and classification of defect parameters within TT images. Despite the disjoint nature of the datasets themselves, the shared encoder design yielded improved performance over the separate tasks over the separate datasets. This work, while distant in domain, complements the central theme of this dissertation: machine-learning-driven image analysis that translates pixels into patterns across scales—from satellite-based scenes to microstructures. The results of this dissertation show the versatility of machine learning across visual domains and datasets, emphasizing how practical, image-based machine learning methods can work among a variety of resolutions for a variety of real-world use cases. These results offer a straightforward way to utilize images and transform them into trustworthy signals to rely on for public-health decisions, and manufacturing quality checks, spotlighting the versatility of machine learning for use cases we encounter every day.

Historically, polymer network fracture was studied primarily or solely through the lens of an engineering perspective. Recent investigation has broadened this perspective into one that considers the chemical reality of such materials, as covalent scission is necessary for polymer network failure to occur. For over half a century, Lake-Thomas theory has remained the dominant theory connecting molecular details to the critical failure of a network. Prior advances from our group include adjustments of the molecular energy parameter and modifications of the theory to account for the complex, molecular details of a polymer network. Chemical reactivity which can be mapped onto macromolecular properties is therefore desirable.Reactive Strand Extension (RSE) has been established as a stress-relieving, mechanochemical response wherein polymer strands elongate under force rather than undergoing scission. The response has been documented on the single polymer strand level by Atomic Force Microscopy (AFM) and by pulsed ultrasonication, and incorporated into complex multinetwork architectures. Since the energy dissipation per mechanophore unit can be estimated, RSE-reinforced polymers hold the potential to assist in bridging the quantitative gap between molecular and mechanical frameworks. Chapter 2 discusses the synthesis of end-linked networks containing RSE units and the consequence on their resultant toughness. We perform ROMP of cyclooctene and a cyclooctene-derivative containing RSE to construct two highly similar polyolefins, P1 and P2, differing only in the content of RSE loading, with P1 at 0% RSE and P2 at 20% RSE. We then exploit activated alkyne-hydroxyl “click chemistry” to crosslink them into networks, and prepare them as organogels N1 and N2. Network characterizations via oscillatory rheology and comparison of swelling ratios and sol fractions suggest the networks differ only in the latent mechanochemical ability of the networks. N2 exhibits a tearing energy of 9.6 ± 0.7 J·m-2, and N1 a tearing energy of 6.9 ± 1.1 J·m-2, (p = 0.01, t-test). We attribute the difference of 30% tearing to the installation of RSE. Chapter 3 describes a workflow leveraging advances in machine learning to identify useful scissile mechanophores that can be adapted into tougher polymer networks. We report that the installation of trimethysilyl groups in the meta position to the polymerization/crosslinking handles (NUSZEG) decreases the mechanochemical stability of the resultant metallocene. We synthesize copolymers of similar size containing ferrocene + gDCC and NUSZEG + gDCC and sonicate them, observing ring opening values of ~20% and ~5% respectively. The lowered ring opening is consistent with competition from a weaker bond, one of the lowest phi values we have reported for this sort of competition experiment. We also report that installation of NUSZEG as a crosslinker into acrylate-based networks enhances the tearing energy by to four times. Chapter 4 reviews the mechanochemistry of the four-membered, nitrogen containing heterocycle azetidine. Azetidine undergoes a force-mediated [2+2] cycloelimination to reveal a linear oxime and alkene, and in the presence of water, creates complex hydrolysis products. Its potential in dynamic networks and future RSE-related systems is also discussed.

Research into high temperature photocatalysis has revealed advantages to using light to augment or replace thermal energy for the synthesis of feedstock chemicals. However, poor penetration of light into a typical powder catalyst bed poses a challenge for efficient high temperature photo-driven heterogenous catalysis. To address this issue, we present a 3D porous optical diffuser loaded with plasmonic Rh nanoparticles enabling volumetric illumination of the nanoparticle catalysts. Within the field of plasmonic catalysis, much attention has been given to the tuning of the catalysts optical properties by manipulating of the plasmonic nanoparticles itself. However, much less attention has been given to the impact which the plasmonic particles dielectric supporting substrate has upon the system’s optical profile. This dissertation presents a monolithic plasmonic Rh/SiO2 structure which exhibits dramatically increased response to light compared to a powdered catalyst due to the control exerted over the dielectric substrate scattering profile. Chapter 1 presents an introduction to plasmonic catalysis. Chapter 2 shows how thermal non-thermal mechanisms can be distinguished with accessible and common place laboratory tools and procedures. Two of these methods, the mass-dependent method and the cover/uncover method revolve around embracing the poor penetration of light into a powder catalyst bed compared to the deeper penetration of photothermal heat. In addition to describing accessible means to differentiate between bulk thermal and local thermal, and nonthermal effects, Chapter 2 also highlights some previous misconception and misused terminology regarding quantum efficiency claims in plasmonic catalysis. Chapter 3 shows how changing from a powder supported catalyst to a monolithic aerogel supported catalyst homogenizes the optical environment, makes the catalyst more responsive to light, and drives selectivity toward a single product. In a powdered photocatalyst the intensely illuminated surface is followed by an abrupt transition to a dark but thermally active subsurface, creating a dichotomy of dueling thermal and non-thermal reaction paradigms. However, in an aerogel supported catalyst a broad gradient of optical intensity extending throughout the entirety of the catalyst bed was achieved. This broadening of the optical gradient between catalyst bed surface and subsurface improves the selectivity of the CO2 reduction reaction. The creation of a continuous and more uniform optical environment also provides greater reaction efficiency compared to the powdered catalyst with a sharper illumination gradient. The broad illumination gradient can be further tuned by augmenting the loading scheme to broaden or sharpen the optical distribution. This dynamic illumination paradigm is achieved through minimizing the supports optical scattering by employing an aerogel where the particle size is reduced to a size where only weak Rayliegh scattering is relevant. Support optical scattering can then be reintroduced in a controlled manner through the use of a sacrificial template composed of fused zinc oxide tetrapods. The pores added by this template-based approach provide interfaces upon which light scatters to create a highly effective optical diffuser. These pores also improve mass transport over conventional aerogel supported catalysts. The ZnO scaffolding implemented to improve mass flow also provides easier nanoparticle loading as well as the potential for more facile adjustments to the aerogel's surface chemistry compared to an unmodified aerogel were tight native pores heavily restrict surface accessibility. This approach to supported photocatalyst design provides exceptional flexibility for tuning the balance between optical properties, mass flow considerations, and available metal surface area. Chapter 4 presents work which chronologically preceded Chapter 3 and established the potential influence of dielectric scattering on the plasmonic particles’ local electric field strength. To probe this interaction silica microlenses and Mie resonators were synthesized and used to combine these two types of light matter interactions to further concentrate light at the nanoscale beyond what the plasmonic and dielectric scattering modes achieved separately. While the local field was enhanced in line with simulations, it was concluded that altering the dielectric support to be more diffusive as opposed to more concentrating would provide greater benefit. This directly informed the approach of Chapter 3. Chapter 5 focuses on a second challenge facing plasmonic catalysis, the short charge separated lifetime of hot carriers produced by plasmon dephasing, typically on the order of femtoseconds. Previous works have leveraged charge injection into the supporting material to prevent recombination due to the barrier posed by the Schottky junction The final portion of this dissertation explores the introduction of a type-II heterojunction to the supporting structure to further separate carriers and prolong excited lifetimes. The junction is achieved by oxidizing ZnS to form a layer of ZnO on the surface and then vacuum annealing to convert both phases to wurtzite, and to promote interfacial connectivity. Materials characterization including TEM, XRD, Raman, and uv-vis reflectance measurements support the formation of ZnO layer on the ZnS surface. Differences in the photoluminescent spectra and lifetimes and EPR spectra of oxidized Zns and unoxidized ZnS suggest that photoexcited electrons reside in ZnO phase for an extended period before recombining, compared to ZnS samples where photoexcited electrons persist in sulfur vacancy traps. Chapter 6 presents the conclusion from this body of research into designing the supporting material architecture to better complement the plasmonic catalyst in terms of both optical and electronic properties.