The objective of this research is to achieve accurate and scalable bridge health monitoring (BHM) by learning, integrating, and generalizing the monitoring models derived from drive-by vehicle vibrations. Early diagnosis of bridge damage through BHM is crucial for preventing more severe damage and collapses that could lead to significant economic and human losses. Conventional BHM approaches require installing sensors directly on bridges, which are expensive, inefficient, and difficult to scale up. To address these limitations, this research uses vehicle vibration data when the vehicle passes over the bridge to infer bridge conditions. This drive-by BHM approach builds on the intuition that the recorded vehicle vibrations carry information about the vehicle-bridge interaction (VBI) and thus can indirectly inform us of the dynamic characteristics of the bridge. Advantages of this approach include the ability for each vehicle to monitor multiple bridges economically and eliminating the need for on-site maintenance of sensors and equipment on bridges. Though the drive-by BHM approach has the above benefits, implementing it in practice presents challenges due to its indirect measurement nature. In particular, this research tackles three key challenges: 1) Complex vehicle-bridge interaction. The VBI system is a complex interaction system, making mathematical modeling difficult. The analysis of vehicle vibration data to extract the desired bridge information is challenging because the data have complex noise conditions as well as many uncertainties involved. 2) Limited temporal information. The drive-by vehicle vibration data contains limited temporal information at each coordinate on the bridge, which consequently restricts the drive-by BHM's capacity to deliver fine-grained spatiotemporal assessments of the bridge's condition. 3) Heterogeneous bridge properties. The damage diagnostic model learned from vehicle vibration data collected from one bridge is hard to generalize to other bridges because bridge properties are heterogeneous. Moreover, the multi-task nature of damage diagnosis, such as detection, localization, and quantification, exacerbates the system heterogeneity issue. To address the complex vehicle-bridge interaction challenge, this research learns the BHM model through non-linear dimensionality reduction based on the insights we gained by formulating the VBI system. Many existing physics-based formulations make assumptions (e.g., ignoring non-linear dynamic terms) to simplify the drive-by BHM problem, which is inaccurate for damage diagnosis in practice. Data-driven approaches are recently introduced, but they use black-box models, which lack physical interpretation and require lots of labeled data for model training. To this end, I first characterize the non-linear relationship between bridge damage and vehicle vibrations through a new VBI formulation. This new formulation provides us with key insights to model the vehicle vibration features in a non-linear way and consider the high-frequency interactions between the bridge and vehicle dynamics. Moreover, analyzing the high-dimensional vehicle vibration response is difficult and computationally expensive because of the curse of dimensionality. Hence, I develop an algorithm to learn the low-dimensional feature embedding, also referred to as manifold, of vehicle vibration data through a non-linear and non-convex dimensionality reduction technique called stacked autoencoders. This approach provides informative features for achieving damage estimation with limited labeled data. To address the limited temporal information challenge, this research integrates multiple sensing modalities to provide complementary information about bridge health. The approach utilizes vibrations collected from both drive-by vehicles and pre-existing telecommunication (telecom) fiber-optic cables running through the bridge. In particular, my approach uses telecom fiber-optic cables as distributed acoustic sensors to continuously collect bridge dynamic strain responses at fixed locations. In addition, drive-by vehicle vibrations capture the input loading information to the bridge with a high spatial resolution. Due to extensively installed telecom fiber cables on bridges, the telecom cable-based approach also does not require on-site sensor installation and maintenance. A physics-informed system identification method is developed to estimate the bridge's natural frequencies, strain and displacement mode shapes using telecom cable responses. This method models strain mode shapes based on parametric mode shape functions derived from theoretical bridge dynamics. Moreover, I am developing a sensor fusion approach that reconstructs the dynamic responses of the bridge by modeling the vehicle-bridge-fiber interaction system that considers the drive-by vehicle and telecommunication fiber vibrations as the system input and output, respectively. To address the heterogeneous bridge properties challenge, this research generalizes the monitoring model for one bridge to monitor other bridges through a hierarchical model transfer approach. This approach learns a new manifold (or feature space) of vehicle vibration data collected from multiple bridges so that the features transferred to such manifold are sensitive to damage and invariant across multiple bridges. Specifically, the feature is modeled through domain adversarial learning that simultaneously maximizes the damage diagnosis performance for the bridge with available labeled data while minimizing the performance of classifying which bridge (including those with and without labeled data) the data came from. Moreover, to learn multiple diagnostic tasks (including damage detection, localization, and quantification) that have distinct learning difficulties, the framework formulates a feature hierarchy that allocates more learning resources to learn tasks that are hard to learn, in order to improve learning performance with limited data. A new generalization risk bound is derived to provide the theoretical foundation and insights for developing the learning algorithm and efficient optimization strategy. This approach allows a multi-task damage diagnosis model developed using labeled data from one bridge to be used for other bridges without requiring training data labels from those bridges. Overall, this research offers a new approach that can achieve accurate and scalable BHM by learning, integrating, and generalizing monitoring models learned from drive-by vehicle vibrations. The approach enables low-cost and efficient diagnosis of bridge damage before it poses a threat to the public, which could avoid the enormous loss of human lives and property

Proper treatment of domestic wastewater is crucial for protecting human health and the environment. However, conventional wastewater treatment processes are often high-cost, energy-intensive, and insufficient for recovering resources. Furthermore, water infrastructure in the United States is nearing the end of its intended design lifespan, posing a key opportunity for reinvention. Anaerobic secondary treatment is a promising example of next-generation water infrastructure that prioritizes resource recovery through the production of methane energy. However, anaerobic secondary effluent requires further attention due to the presence of dissolved methane, sulfide, nitrogen, and phosphorus. This dissertation explores post-treatment of anaerobic secondary effluent to maximize resource recovery from domestic wastewater. Specifically, a life cycle assessment was performed to evaluate tradeoffs between physical/chemical processes and biological processes for dissolved methane, sulfide, nitrogen, and phosphorus removal. Additionally, a membrane-aerated biofilm reactor was tested to treat anaerobic secondary effluent with high concentrations of ammonium-nitrogen and sulfide. Lastly, the use of wastewater-derived struvite as a novel fire retardant was explored to improve the profitability of phosphorus-recovery technologies. These studies serve to direct future efforts in developing complete water treatment trains with anaerobic secondary treatment.

Complex engineering design requires solving large optimization problems involving several disciplines. Collaborative Optimization is a two-level design optimization architecture that solves multidisciplinary problems as a series of single-disciplinary problems with little infrastructure overhead. It can be critical when practitioners do not have a fully coupled multidisciplinary optimizer but have access to multiple single-disciplinary optimizers. However, it comes at a steep computational cost because the limited information sharing between disciplines slows downs the progress of the optimization. Our work shows that this issue can be mitigated by exploiting the full extent of the available information. In the Collaborative Optimization framework, disciplines are repeatedly given a design point and tasked with finding the closest feasible design. These iterations are computationally expensive as they require each discipline to solve an optimization problem and report the result. To combine the information collected by all iterations, we propose to learn a predictive model of the distance to the feasible set of each discipline using datasets of feasible and infeasible design points. This requires modeling the signed distance function of each set, which is a challenging machine-learning task. Therefore, we introduce Householder networks: a new, lightweight neural network architecture that can learn distance functions more efficiently than conventional architectures. We then introduce our new method, called Bayesian Collaborative Optimization, which uses ensembles of Householder networks to represent probabilistic models of the disciplinary feasible sets. Following the Bayesian Optimization framework, these models are iteratively refined and used to find values of the design variables that improve the objective function while remaining feasible for every discipline. This method is shown to outperform previous Collaborative Optimization approaches on simple test problems. Finally, we introduce a new multi-disciplinary aircraft design problem. We optimize the airframe, propulsion system, and trajectory of an unmanned fixed-wing vehicle tasked with completing a half-marathon with a fixed battery. This problem tries to balance out realism, by including a diverse set of modeling and optimization approaches, with simplicity and low computational cost. The importance of trajectory optimization, which is efficiently solved by itself but hard to solve coupled with other disciplines, makes this problem different from those previously available. We hope that this problem will be useful to the research community as a test for multidisciplinary design optimization architectures. We have open-sourced the code that generated the results presented in this document. It can be accessed at the following links: https://github.com/jdebecdelievre/HouseholderNets.jl , https://github.com/jdebecdelievre/BayesianCollaborativeOptimizat ion.jl , and https://github.com/jdebecdelievre/ModelAirRaces.jl.

Protein function is dependent upon adoption of a native 3-D configuration whilst avoiding toxic conformations. Some proteins can fold spontaneously; however, the complexity of the proteome requires cellular cofactors called chaperones to ensure problematic proteins are natively folded. Chaperonins are a class of molecular chaperones that are universally conserved in all domains of life and contribute to the cellular proteostasis toolkit. They are 1-MDa complexes composed of two rings that undergo a dramatic conformational change with ATP hydrolysis to form an inner folding cavity. Chaperonins have significantly diverged in architecture, topology, and client repertoire so they are divided into two classes: group I and group II chaperonins. Group I chaperonins exist in prokarya whilst group II chaperonins are found in eukarya and archaea. All chaperonins help to orchestrate ATP-dependent client protein sequestration and refolding through encapsulation in the central folding cavity. This thesis is centered on enumerating the mechanism of a group II chaperonin found in M. Maripaludis (MmCpn), an archaeal methanogen. Biochemical and biophysical interrogation of structural moieties, chiefly the c-termini, were found to contribute to ATP hydrolysis, substrate re-folding, and complex integrity. New structural elements such as interfacial methionine fingers and electrostatic contacts were observed to undergo novel conformations, yielding new candidates for studying group II chaperonins in general. Electrostatic disruption of the c-termini allowed for dissection of multimeric states, including the long elusive single-ring species which could serve as a useful tool for understanding chaperonin biogenesis and function

Accurate and efficient models for the relative motion of two satellites are needed to achieve greater autonomy in an increasingly crowded orbit environment with limited computing power. Analytical models provide a computationally cheap alternative to numerical integration for relative motion propagation, typically at the expense of accuracy. Two avenues to improving relative dynamics modeling accuracy are examined herein. First, higher-order terms in the Keplerian dynamics can be included. This dissertation introduces new, second-order models that are valid for both circular and elliptical orbits in three of the most popular relative state representations: the Cartesian relative position coordinates in the rotating Hill frame, the spherical coordinates in the inertial frame, and the relative orbital elements (ROE). The performance of these models is compared with one another and with several of the most popular models from the relative dynamics literature. The second direction to improve modeling accuracy is the inclusion of non-Keplerian perturbations. A general framework for modeling arbitrary perturbations in the Hill frame is developed and used to derive the equations governing the leading-order corrections for Earth oblateness perturbation, as well as a closed-form solution for the effect of oblateness on the relative motion in near-equatorial orbits. This is compared with Keplerian models and an ROE-based perturbation model in a full-force propagation. In addition to the advancement of relative dynamics models, this dissertation examines two applications of such models. A fast and efficient method for initial relative orbit determination (IROD) from bearing-angle measurements is introduced. The second-order models developed in the course of this research provide a means resolving the range ambiguity problem that arises from linear relative motion models. The IROD problem thereby becomes one of solving a system of polynomial constraint equations linking the line-of-sight measurements to the relative state parameters. An efficient method for solving this system is developed around the insight that these parameters scale with the ratio of the inter-spacecraft separation to the orbit radius and are therefore small for most applications of interest. The method uses a truncated expansion of the quadratic formula to recursively eliminate unknowns, reduce the dimension of the system, and ultimately acquire an approximate solution. Strategies for improving robustness, efficiency, and accuracy are developed and the method is applied to general second-order systems as well as to a broad range of IROD scenarios. Modifications to the constraint equations and solution algorithm are introduced to address the challenge of bias in the bearing-angle measurements. The second application considered is that of low-thrust maneuver planning for formation reconfiguration. The adoption of fuel-efficient electric propulsion systems poses a challenge for relative orbit control schemes, which are typically based on the assumption of impulsive maneuvers. That challenge is met herein with a geometrically intuitive, semi-analytical solution to the low-thrust problem. Beginning with the equations of relative motion of two spacecraft, an unperturbed chief and a continuously-thrusting deputy, a thrust profile is constructed which transforms the equations into a form that is solved analytically. The resulting relative trajectories are the family of sinusoidal spirals, which provide diversity for design and optimization based upon a single thrust parameter. Closed-form expressions are derived for the trajectory shape and time-of-flight for two prescribed relative velocity behaviors, and used to develop a novel patched-spirals trajectory design and optimization method. The example problem of a servicer spacecraft establishing and reconfiguring a formation around a target in geostationary earth orbit is used to demonstrate the application of the patched spirals technique as well as the advantages of the relative spiral trajectories over impulsive maneuvers. The sensitivity of the trajectory solutions to deviations from the underlying assumptions, uncertainties in the state, and errors in thrust are studied through high-fidelity simulation

As artificial intelligence changes nearly every facet of modern society, we should not be surprised that it is changing how we do social science. By leveraging the power of machine learning and automated text analysis, researchers can analyze complex patterns from data and extract meaning from natural language at an unprecedented scale. However, the application of these tools to social scientific inquiry raises important issues concerning construct validity and the very nature of deductive social science. Throughout this dissertation, I examine the promises and pitfalls of applying these cutting-edge technologies specifically to the sociological study of meaning. In the first chapter, I provide a comprehensive review of popular automated text analysis methods and classify them according to the pre-analytic constructs they extract from text. In the following chapters, I present two original studies that use machine learning and automated text analysis to answer fundamental questions about culture and meaning. The first study asks: does everyday symbolic exchange contain sufficient information to effectively enculturate a tabula rasa learner? The second asks: does the way an individual understands their nation shape their immigration policy preferences? Via novel and rigorous applications of computational methods, I provide compelling evidence that supports the affirmative answers to both questions. Ultimately, this dissertation highlights the potential of machine learning and automated text analysis to produce sound social science research. However, it also underscores analytical concerns of which researchers should be mindful.

Decades of advances in semiconductor technology have led to the scaling of silicon transistors towards nanometer scale dimensions, culminating in physical limitations to transistor size, and the concurrent issue of thermal degradation and excessive energy consumption. As the semiconductor industry moves towards continuously increasing computational density, for example from planar to 3-dimensional (3D) vertical stacking, the ability to effectively regulate temperature and manage heat has become a pressing bottleneck to the realization of these dense architectures. Thermal challenges in high-density memory and computation require new thermal and material considerations at the nanoscale. Accurate knowledge of temperature and heat transport within a complex system is the essential first step to successful thermal management. To this end, experimental validation of material thermal conductivity is fundamental both to understanding thermal transport and to implementing solid state thermal management approaches. The complexity of thermal transport within a dense system also necessitates novel thermal management tactics including active methods of routing heat, such as thermal switching. This dissertation evaluates the intersection of nanoscale thermal transport and nanomaterials, assessing materials that show promise for low-temperature integration in semiconductor-based systems. The studies presented in this thesis range from fundamental thermal, electrical, and material characterization thin films to the design, demonstration, and characterization of an active thermal switch. We first present thermal characterization of nanoscale boron nitride thin films transferred and deposited at low temperature (< 500 °C) for back-end of the line (BEOL) temperature compatible processes. We report the in-plane thermal conductivity measurement of large area hexagonal boron nitride thin films grown using chemical vapor deposition (CVD) and transferred at room temperature to be 55-78 W/m/K from 300-400 K. We also evaluate the complementary thermal and electrical properties of low thermal conductivity boron nitride films deposited at room temperature using electron enhanced atomic layer deposition (EE-ALD). These electrically insulating films have very low cross-plane thermal conductivities (< 0.6 W/m/K) with 10 nm thick films approaching the amorphous thermal limit of BN (~0.13 W/m/k). Finally, we present the fabrication and thermal characterization of graphene-based thermal switches for active thermal management, reporting the first thermal switch based on flexible, collapsible graphene membranes with low operating voltage (~2V)

In this work, we build up InAs as a platform for mesoscopic physics toward the eventual goal of analog quantum simulation. Specifically, we demonstrate in an InAs 2DEG quantum point contacts and quantum dots that are operational in the quantum Hall regime. Together, these serve as the basic circuit elements for controlling the flow of electrons at the nanoscale. In analogy to quantum optics, point contacts can serve as beam splitters and quantum dots can serve as artificial atoms, and having these building blocks allows us to further probe electron interactions. Our highly transparent, hybrid metal-semiconductor quantum dots have charging energies several times larger than has been achieved in other solid-state platforms, and this encourages the potential for scaling to quantum dot arrays and simulating many-body phenomena like the charge Kondo effect

Spectroscopic techniques provide us with a means of investigating a system's microscopic structure and dynamics. Accurate atomistic simulations can help us explicitly connect spectroscopic features to the underlying electronic and nuclear structure and dynamics that give rise to them. In this dissertation, I highlight my work in rendering accurate atomistic simulations of different linear and multidimensional spectroscopies more computationally tractable by leveraging semiclassical approaches for theoretically treating spectroscopies and developing machine learning (ML) models to serve as proxies for ab initio electronic structure calculations. Chapter 1 provides a quick overview of the ML approaches I employed and theoretical background for how I used molecular dynamics (MD) simulations to simulate different spectroscopies. Chapter 2 presents work I have conducted in training ML potential energy surfaces for liquid water using transfer learning to target high-level ab initio electronic structure theories in order to accurately and efficiently conduct MD simulations. In Chapters 3 and 4, I develop ML models for electronic excitation energies in order to simulate linear and 2D electronic absorption spectroscopies for various solvated chromophore systems and highlight the inability of TDDFT to treat the extent to which hydrogen-bonding affects the distribution of excitation energies. Lastly, Chapter 5 highlights my work in developing a theoretical framework to simulate novel time-resolved X-ray diffraction experiments, which can be used to probe the orientational structural dynamics of disordered condensed phase systems, and benchmarking with results for liquid chloroform

While scientists and practitioners call for increasing ambition in environmental objectives, state capacity to meet existing goals is in many places weak. I examine potential to improve government agencies' implementation of environmental public policy through more strategic resource deployment. I do so through a mixed-methods study of personnel management and deforestation control in Brazil's federal protected areas agency, the Chico Mendes Institute for Biodiversity Conservation (ICMBio), which manages one of the world's largest systems of protected areas. First, I examine the subjurisdiction conditions under which additional public employees will most positively impact deforestation outcomes. To do so, I combine a comparative case study of six ICMBio management teams operating in the Amazon region with panel regression analysis covering 322 federal protected areas over ten years. I find that the Chico Mendes Institute could have prevented on the order of 1,700 km2 of deforestation over this period through more strategic allocation of its personnel. I then examine the geographic, political, and organizational factors that produced a misalignment between subjurisdiction personnel needs and actual personnel allocation in the agency's first decade. This chapter draws on historical institutional analysis and descriptive analysis of quantitative data related to personnel, protected areas, and socioeconomic characteristics of the regions in which Brazil's federal protected areas are located. Altogether, I demonstrate the value of "looking inside" of public environmental organizations to understand environmental outcomes. In addition, I generate concrete policy recommendations for the Chico Mendes Institute itself: the principal guardian of some of the most important ecosystems on Earth

As brain imaging technologies measure increasingly higher spatial resolutions and faster time scales, complementary advances in analyzing whole-brain activation time-series data are necessary. Topological models offer a powerful framework for this analysis by describing the dynamical organization of the brain as a graph and can effectively capture the underlying shape of the space explored by the brain, for example, during ongoing cognition. A recently established approach using the Mapper algorithm from topological data analysis (TDA) now enables the construction of these graphs from whole-brain functional imaging data. The work described in this thesis advances that approach in three ways. First, we provide new open-source tools for visualizing and extracting insights from shape graph representations of neuroscientific data learned by Mapper. Second, we introduce a new Mapper algorithm inspired by the high dimensionality of brain imaging data designed to reduce both information loss and computational cost. Third, we extend the Mapper-based approach to naturalistic fMRI data analysis, quantifying more ecologically valid transitions in the unstructured data and leveraging annotations provided by the paradigm (e.g., tasks, stimuli-derived features). By simultaneously addressing usability, scalability, and ecological validity, this dissertation takes us three steps closer to translational applications of our Mapper-based approach, and ultimately, to realizing the promise of precision neuroimaging. Along the way, we introduce a new fMRI dataset collected using a naturalistic self-viewing paradigm and describe a novel link between brain dynamics and behavior

This dissertation presents a new interpretation of American responsibility and reveals overlooked contingencies in the acceleration of key drivers of anthropogenic climate change in the mid-twentieth century. Historians have recognized links between the ascent of the United States to superpower status in the 1940s and the launching of the Great Acceleration--a global skyrocketing of greenhouse gas emissions and resource consumption starting soon after 1945. Scholarship has focused largely on longer-term climate effects of American actions and domestic consumption. Less attention has been paid to another form of American contribution: choices by U.S. foreign policymakers as they shaped transnational patterns of industrialization and resource use across a pivotal decade for global politics and environmental change. More fully understanding the making of the Great Acceleration requires investigating the degree of choice and knowledge of risk among U.S. policymakers from the late 1930s to the early 1950s. Engaging U.S. military, diplomatic, and intelligence archives, this dissertation investigates how American policymakers developed ideas that stressed equality over competition in postwar industrial geography and resource use in Europe and Asia--and why they ultimately discarded these ideas. Each chapter focuses where competing choices and risks appeared sharply: debates regarding Germany and Japan during war and occupation. While U.S. policymakers did not understand the climate impacts of their choices, they knowingly turned from plans that supported more balanced and sustainable transnational patterns of industrialization and resource use in Europe and Asia to accept risks of global inequality, resource exhaustion, and preparation for future world war. This work bridges historiographies of the American empire, U.S. foreign relations, and critical Anthropocene studies to expand historical insights on the deepening climate crisis.

Projection-based model order reduction is a technology for dramatically reducing the cost of expensive computational physics simulations by approximately solving the governing equations in a data-driven approximation subspace. This dissertation proposes a set of techniques that exploit the concepts of locality and clustering to further accelerate reduced-order models on challenging problems. The first exploits the fact that many physical systems exhibit spatially localized features such as shocks or wakes to construct space-local reduced-order models that are more efficient than existing model-reduction schemes. The second takes advantage of locality in the simulation's state-space to construct smaller bases composed of a few nearby solution snapshots. Both novel techniques proposed here are applied to multiple systems drawn from computational fluid dynamics, including a challenging large-scale turbulent flow problem. In all applications, multiple orders of magnitude of speedup are observed relative to the original high-dimensional computational model, and substantial speedups are also observed relative to previously-existing model-order reduction techniques

Healthy coastal waters are critical for public well-being, thriving ocean economies, and functioning ecosystems. Yet threats to coastal health are ubiquitous: pollution, invasive species, overfishing, warming, and acidification harm much of the planet's coastlines and the human and natural communities in their proximity. In order to identify and mitigate these threats, it is important to keep a pulse on the state of coastal health over time by monitoring representative parameters. If anomalous measurements are observed, coastal managers (public health agencies, fishery managers, etc.) can take preventative or remediative action. In practice, technical and operational restrictions limit coastal monitoring to coarse spatial and temporal resolutions. This is an issue because the variability of environmental parameters of interest often occurs on finer spatial and temporal scales. As a result, public health and natural resource managers have limited information about the state of coastal health because sparse measurements are rarely representative of entire systems. While there has been an increase in coastal monitoring ability in recent times, translation of these growing environmental datasets into meaning and decision support is still lacking. To make coastal monitoring more useful, it is necessary to better understand how environmental drivers influence coastal health. Knowledge of how oceanic, hydrological, and meteorological processes modulate coastal health can be leveraged for better planning, monitoring, and management of coastal systems. Through a combination of observational and data science techniques, the studies documented in this dissertation elucidate how various environmental processes affect various facets of coastal health. Moreover, this dissertation provides insight into how these relationships can be leveraged to improve coastal health monitoring and management. Environmental processes are typically unsteady (i.e. time-variant) and indicators of coastal health can vary across multiple temporal scales. It is known that fecal indicator bacteria (FIB) -- a commonly-used proxy for water quality - can vary substantially at subhourly frequencies, yet little previous work had studied natural variation of FIB in enclosed (i.e. sheltered from the open ocean) beaches. Chapter 2 presents data from a high-frequency water quality sampling (HFS) event conducted over two days at a beach within a harbor. Results show that the temporal variability of FIB at enclosed beaches can be higher than at open beaches, and that environmental drivers such as chlorophyll concentration, turbidity, and tide level partially explain this dynamic. Chapter 3 leverages the HFS methodology to yield data sufficient to calibrate predictive beach water quality models. While such models have typically been developed at beaches where extensive historical FIB and environmental datasets exist, we show that effective models can be developed with data collected at HFS events. This study thus provides a new technique that can be used to rapidly develop predictive modeling systems for 'data-poor' beaches (which make up the majority of the planet's coastlines). Chapter 4 investigates the feasibility of beach water quality forecasting (as opposed to 'nowcasting' same-day water quality, the predominant technique used in operational prediction systems). We successfully develop a novel framework that can provide FIB forecasts at beaches with up to three days lead time. This framework can be applied to provide beach managers more time to allocate resources and beachgoers more time to make decisions on where to recreate based on health risk. Chapter 5 applies similar methodologies to a different facet of coastal health. We present a high-frequency, long-term environmental DNA (eDNA) dataset collected in a coastal California stream. Along with auxiliary environmental data, this eDNA dataset is used to model temporal trends in endangered fish abundance in the stream. Results show the efficacy of eDNA as a tool for long-term biodiversity monitoring. Overall, this dissertation provides a variety of techniques to gage the health of coastal systems and its environmental drivers. Results from these field-scale and modeling studies contribute to the growing field of environmental data science, providing tools to better translate environmental data into meaning and decision support

My dissertation is focused on studying the assembly of both ecological and social communities in the life sciences. First, I aimed to reimagine the current scientific enterprise within its own terms. By publishing scientific papers that center scientific art (Ch. 1), intergenerational mentorship (Ch. 2), and celebration of diverse lived experiences within academic science (Ch. 3), I show that science can flourish from diverse, equitable, and interdisciplinary groups. The topic of this research focuses on the assembly of microbial communities in the nectar of Diplacus aurantiacus (Ch. 1). By applying population genetics, functional genomics, and experimental evolution to this wild microbiome, I identify mechanisms that connect genetic variation to community-level processes such as priority effects (Ch. 2) and show that population-level variation can alter molecular traits associated with community assembly (Ch. 3). Second, I question the academic structure itself as the central nexus for scientific discovery. By dissolving the duality between science and art (Ch. 4), creating new science spaces that center culture and lived experience (Ch. 5-6) and reimagining entire educational ecosystems outside of "traditional" scientific venues (Ch. 7), I propose new frameworks for how science can be conducted and perhaps, what we consider science to be. This work is centered in community-centered art/science programming in the Greater Bay Area of California. In total, I hope this dissertation highlights how the life sciences can flourish when we celebrate a diversity of perspectives and approaches

Emerging deep learning applications require unprecedented computation and memory capacity. To accelerate these applications, novel processing systems such as dataflow accelerators strive to exploit multiple dimensions of parallelism within deep learning models, e.g., tensor and pipeline parallelism. Although these systems provide ultra-high performance when fully utilized, compiling deep learning applications to harness their computation capability remains a challenging problem. With recent advances in domain-specific programming language, accelerator design, and machine learning, we now have the potential to better serve the needs of training and evaluating large deep learning applications on dataflow accelerators through algorithm, software, and hardware co-design. In this dissertation, I present the design and development of efficient deep learning optimizations and programming frameworks. I present two frameworks: SpatialRNN for accelerating recurrent neural network language models on spatial accelerators and Sigma for expressing and accelerating high-data-reuse deep learning kernels using reconfigurable dataflow accelerators. Our end-to-end evaluation using Sigma demonstrates a 5.4x speedup on kernels encompassing financial applications, traditional machine learning, language modeling and computer vision tasks over an Nvidia V100 GPU accelerator

Treating wastewater is a resourceful and economical approach to providing potable water to our growing population. Because chlorination causes carcinogenic byproducts, and wastewater can be heavily contaminated, other purification methods are needed. One method involves using UV-light and hydrogen peroxide (H2O2) to oxidize contaminants and has already been implemented around world. However, H2O2 is often produced unsustainably via the anthraquinone process. An alternative is to produce H2O2 via a two-electron water oxidation reaction (2e-WOR) at the anode of an electrochemical cell. Despite this proposal, much research is needed before this technology can be implemented, hence the work of this thesis. Herein, we optimize parameters of various components in the 2e-WOR -- the electrode, electrolyte, and electrochemical cell -- to improve H2O2 production. First, we introduce a stable anode of manganese (Mn) doped titanium dioxide (TiO2), in which increasing the amount of Mn increases its catalytic activity. Through a facile solution gelation process, we synthesize TiO2 with various amounts of Mn in its crystal lattice. In a 0.5 M H2SO4 electrolyte, we found the onset potential to decrease by 370 mV with the addition of Mn, while in a 2 M KHCO3 electrolyte, the onset decreased by 260 mV. It was discovered that the selectivity towards H2O2, oxygen, and peroxysulfate, also depends on the amount of Mn and the electrolyte. Next, we moved to optimizing components of the electrolyte and fixing the anode material to be commercial fluorine doped tin oxide (FTO) on glass. After varying the applied voltage, HCO3-- to CO32-- ratio, total dissolved inorganic carbon concentration, and cationic species, we present an optimized electrolyte mixture of 0.5 M KHCO3 and 3.5 M K2CO3 under 3.25 V vs. RHE. We compare our optimized electrolyte to the field standard 2 M KHCO3 electrolyte and find a twelve-fold increase in the selectivity and H2O2 production rate. The optimized electrolyte also led to a 20-time higher concentration of accumulated H2O2 over 3 hours. We then implemented a machine learning (ML) model to aid in the optimization of the electrochemical H2O2 production process and to map out various performance parameters of the system. From the model, we discovered a further optimized electrolyte composition of 0.92 M KHCO3 and 3.08 M K2CO3, which has become the new standard electrolyte for the field. Our results also demonstrate the efficacy of using ML when experimental data and mechanistic knowledge is limited. Finally, we turn to the electrochemical cell/setup. We optimize the electrolyte flow rate to maximize efficiency on an FTO anode. We then switch to a new and improved indium tin oxide on platinum on titanium (ITO/Pt/Ti) layered anode, which triples the H2O2 production rate. With this anode, we were able to achieve the highest production rate of ~ 40 µmoles/min/cm2 from a voltage of 3.4 V vs. RHE. We also achieved a continuous production rate of ~ 23 µmoles/min/cm2 from a voltage of 3 V vs. RHE over four hours. With the stated improvements in H2O2 production via the 2e-WOR, we hope to bring this technology closer to implementation for clean and sustainable H2O2 production

Optimization is a fundamental part of robotics and can be seen in various aspects of the field, such as control and simulation. Both of these areas involve finding the best solutions to various optimization problems to achieve desired outcomes. Efficiency is key when it comes to solving these optimization problems. By finding solutions quickly and reliably, we can execute optimization-based controllers in real-time on hardware. The ability to quickly generate large amounts of simulation data is also valuable for offline optimization tasks such as policy optimization, co-design optimization, and system identification. Oftentimes, the optimization problems arising in robotics control and simulation have structure. Some problems directly fit into well-studied categories, for instance, the Linear Quadratic Regulator (LQR), other control problems can be cast as Linear Programs (LP), or Quadratic Programs (QP). For each of these categories there exist efficient and reliable solvers. Fitting your problem into one of these categories is often a safe strategy. However, there exist control and simulation tasks that involve complex optimization problems that do not fit these categories and for which there are currently no satisfactory solvers. In this dissertation, we focus on such problems. We are particularly interested in coupled optimization problems where the solution of one optimization problem is a parameter of another one. These coupled optimization problems can naturally arise in robotic simulation. For instance, the simulation of contact physics requires solving the least action principle and the maximum dissipation principle. We can solve these two optimization problems jointly. Coupled optimization problems also frequently arise in autonomous driving scenarios where agents are interacting. Indeed, each vehicle or pedestrian in the scene is optimizing its path to rally its destination as fast as possible while avoiding collisions. Conversely, we can deliberately choose to decompose a single complex optimization problem into a set of coupled optimization problems. Problem decomposition is a strategy that can yield significant benefits in terms of the speed and reliability of the solver. In this context, optimization problems exchange gradient information by leveraging differentiable optimization. The strategy behind these choices is what we call Composable Optimization. In this dissertation, we focus on a few applications in robotics control and simulation namely game-theoretic control, control through contact, physics simulation, and collision detection. For these problems, we leverage composable optimization to exploit problem structure and devise efficient solvers. In some cases, we may combine multiple problems into a single optimization problem, while in other cases we may decompose the problems into simpler chunks. This approach allows us to tackle more complex optimization problems in a structured and efficient manner

One of the most profound discoveries of modern science thus far is that the breathtakingly beautiful diversity of form and function in life is encoded by the genome. Despite major advances in genome sequencing and functional testing, there is yet much to learn about how traits are specified by underlying DNA sequence. For vertebrate genomes, linking genotypic variation to phenotypic outcome is especially challenging given that their typical size spans hundreds of megabases to several gigabases in length. In my graduate work, I contributed to efforts addressing this challenge. Each study described in this thesis involved computational analysis of vertebrate genome sequence, either in the form of whole nuclear genome assemblies or of high-throughput short-read sequencing data, to identify candidate regions associated with and hypothesized to control traits of interest. For two of the studies, including my main thesis project described in chapter 6, I also performed functional testing to evaluate the extent to which candidate regions actually contribute to the associated phenotype. Together, the efforts described in this thesis represents genotype-phenotype mapping methods to identify the protein-coding and cis-regulatory sequence basis of disease-related and naturally evolved traits that arose over the last 10s to 10s of millions of years in humans, other placental mammals, and fishes

This thesis focuses on the intersection of materials screening, machine learning, and ab-initio calculations, with regards to low dimensional materials - mainly 1D dimensional van der Waals wires and materials with 1D and 2D structures that are particularly suitable for phase change applications. 1D van der Waals materials are a less studied counterpart to their 2D dimensional cousins, but there exists a wide range of such 1D prototypes, including conducting wires that may be suitable as consideration for interconnect replacement candidates. We find exfoliation energies can be reasonable and of the same magnitude as their 2D layers, and evaluate their electronic property changes as we thin these wires to the 1D limit. Due to the fact there is a limited number of such materials, only a few hundred, we machine-learn to expand this composition space, training using the known compositions and using input features based on the elemental-makeup. We particularly target conductive and magnetic materials, as well as those containing transition metals and members of the chalcogen, pnictogen, and halogen families, as those are often particularly suited for chemical vapour decomposition (CVD) and chemical vapour transport (CVT) synthesis mechanisms. Previous work in the group has examined the phase changes in the transition metal dichalcogenides (TMD), and we extend this to a wider range of materials as well as changes where only the relative stacking between low-dimensional components differ. The multiple phases present can display unique dynamics, including switching mechanisms potentially relevant for phase change memory applications. We present a bottom-up screening across all crystalline material to evaluate their phase change potential. For the stacking changes, we study the 1T' and the $T_d$ transition present in select members of the transition metal dichalcogenide family. Past research has focused on the 1T, 1T' and the 2H phase transitions, where the individual 2D layers support differing structures. In contrast, the 1T' and the $T_d$ phases have the same individual layers, but a differing relative orientation that is present between the two. This type of sliding phase transition offers the potential to further reduce switching timescales. We present an analysis on the few-layer behavior of the transition metal dichalcogenide structures of 1T' and T$_d$ WTe$_2$ and MoTe$_2$ with the observation that the energy differences are far more subtle