137091 research outputs found
Sort by
Efficient AI Stack: Deployment-Aware Neural Architecture Search and Serving of Deep Neural Networks
No full textThe increasing deployment of Deep Neural Networks (DNNs) on the critical path of production applications in both datacenter and the edge require production systems to serve these DNNs under unpredictable and bursty request arrival rates. Serving models under such conditions requires these systems to strike a careful balance between the latency (R1) and accuracy (R2) requirements of the application and the overall efficiency of utilization of scarce resources (R3). To efficiently balance trade-offs in R1-R3, production systems need to navigate between models, choices of hardware, and application contexts. This thesis proposes an efficient AI stack to solve this tension in the R1-R3 trade-off space. The key idea in the efficient AI stack is to produce and consume Pareto-Optimal (w.r.t latency/accuracy) DNNs. On the production side, the thesis proposes several neural architecture search algorithms namely CompOFA, DES, and SuperFedNAS that automatically specialize DNNs to produce the highest accuracy under different hardware and latency targets in centralized and federated data environments. On the consumption side, the thesis proposes a) an inference serving system SuperServe that consumes these DNNs and schedules them under bursty workloads with resource efficiency, and b) DSched that schedules data pipelines to DNNs in a timely and cost-efficient manner. Overall, the proposed efficient stack co-optimizes R1-R2 under dynamic workloads with resource efficiency R3.Ph.D.Computer Scienc
The Effects of Controlled Porosity on the Dynamic Compression and Tensile Failure Strength of Additively Manufactured 316L Stainless Steel
No full textProcess-inherent microstructural heterogeneities in materials can have marked effects on the shock wave propagation and the resulting tension-induced spall failure. Incorporating intentional heterogeneities in the form of micro-scale pores of controlled size and distribution can allow for a deeper investigation of their effects on shock wave motion and interactions. The research was conducted on 316L SS printed using Powder Bed Fusion (PBF) to generate a better understanding of the effects of size, fraction, and location of pre-existing pores on dynamic mechanical properties. A high-throughput experimental method involving multiple samples simultaneously impacted in each experiment employing the 80-mm diameter single-stage gas gun and multiplexed PDV diagnostics was utilized. The resulting analysis revealed shock-wave mitigation caused by the collapse of pre-existing pores that varied dependent on the pore size and location within the sample leading to a decrease in damage in the spall plane. When many pores are present at a lower peak stress, a higher number of pores and a larger pore size led to an increase in the dissipation of the shock wave as well as a dispersion of the shock wave. An investigation into isolated pores at higher peak stresses revealed that a single pore most effectively disrupts the shock wave and limits the spall damage experienced by the material, with larger pores having a more exaggerated effect. However, the presence of multiple pores at a higher peak stress, both in the plane of the impact direction as well as perpendicular to it, does not dissipate the shock wave as effectively, and more damage is observed. This work provides insights into the shock mitigation mechanisms of alloys containing small volume fractions of pores, and furthers our understanding of the resulting microstructural deformation processes dependent on the pore size and locationPh.D.Materials Science and Engineerin
Autonomous Transfer Hub Networks for Self-Driving Trucks
No full textThe emergence of autonomous driving technology is expected to fundamentally transform the future of the transportation industry.
In the domain of freight transportation, ATHN, deploying autonomous trucks for the middle mile and traditional trucks on the first and last miles, is perceived as the most likely adoption of this technology.
This thesis develops a scalable and comprehensive framework for optimizing the operations and managing hub utilization within large-scale ATHN systems.
Ultimately, the framework is applied to the entire U.S. freight market, envisioning the future freight transportation and its impact on cost savings and the labor market dynamics.
The first part of the thesis presents a scalable flow-based MIP model designed to select loads for the ATHN and establish delivery schedules.
By exploiting the problem structure, this model achieves optimal ATHN design within hours and facilitates data-driven analysis for large-scale systems.
To demonstrate the MIP model's performance, a case study utilizing real data from a dedicated trucking company is conducted, representing a freight system spanning the entire United States over a one-month horizon.
This study quantifies the potential cost-saving benefits of using autonomous trucks and presents an extensive sensitivity analysis on various factors of the ATHN system.
The second part of the thesis addresses limitations in the MIP model regarding hub operations.
Despite its efficacy, the MIP model does not account for operations within the hub.
Optimizing hub utilization with MIP methodologies requires adjusting the MIP model, resulting in a complex framework where scaling becomes nontrivial.
To overcome this difficulty and optimze hub utilization, a CP model is developed to minimize the hub capacities by shifting the start times in the initial schedule from the MIP model.
The power of combining the strengths of MIP and CP is demonstrated using the same case study data, which shows that the CP model can find optimal schedules efficiently and reduce the necessary hub capacities substantially.
Furthermore, the sensitivity analysis centered on hub operations provides insights for improving the ATHN operations, laying the foundation to achieve the ultimate goal of assessing the economic and labor impact of the ATHN on the entire U.S. freight market.
The final part of the thesis envisions the impact of autonomous trucking on the U.S. freight market by applying the proposed framework to all freight loads.
Two ATHN systems are introduced based on the availability of autonomous trucking.
Due to the absence of a representative dataset for total U.S. freight movement, synthetic load data is generated by processing publicly available freight datasets.
This application constructs an ATHN system capable of accommodating national freight demands while offering insights into labor dynamics.
Additionally, a thorough economic and labor study is conducted using the collaborative market framework as the U.S. freight market gradually adopts the ATHN system.
Analyzing these results is crucial for understanding the landscape of the future freight market, designing labor policies, and effectively implementing the national-scale ATHN system in the years to come.Ph.D.Operations Researc
Compound Flood Analysis with GIS-Integrated Dynamic Flood Models for Coastal Urban Areas
No full textThis study addresses the complex challenges of compound flooding in coastal urban areas, specially focusing on the City of Savannah and the Savannah River Basin, which are highly vulnerable to flooding risk. Compound flooding in these areas arises from the interplay of fluvial, coastal, and flash flooding, which are further exacerbated by climate change and urbanization. To tackle these risks, this research conducts three main investigations: (1) fluvial flood analysis using a physical model integrated with geospatial data, (2) development of a high-resolution hydrology-hydraulics urban flood model, and (3) flood driver analysis based on scenario simulations. The first investigation introduces the Physics-Informed Geostatistical Approach (PIGA), which leverages spatial correlations and real -time data to enhance flood prediction. PIGA demonstrates enhanced accuracy and computational efficiency during hurricane events, particularly in data-sparse regions, compared to traditional models. The second investigation develops a high resolution urban flood model that captures the dynamics of surface runoff and drainage systems through a multi-scale, 2D nested simulation process. This model, supported by GIS, achieves a balance between computational efficiency and the detailed accuracy required for complex urban environments. The third investigation quantifies the contributions of key flood drivers— precipitation, riverine inflow, and coastal surge—through advanced statistical analysis using the developed integrated model. Findings reveal that precipitation and downstream conditions are the primary contributors to urban flooding in the study area, with coastal surge significantly increasing flood risk in the area. These spatially distributed insights offer critical information for resilience planning against compound flooding in coastal urban areas. Collectively, this research establishes a comprehensive framework for real-time flood prediction, high-resolution urban modeling, and flood driver analysis. The results provide practical tools for improving flood management strategies and enhancing resilience in vulnerable coastal urban communities.Ph.D.Civil Engineerin
Optimization and Evaluation of Wearable ERG Prototype Designs
No full textThe field of electrophysiological monitoring has seen significant advances with the development of wearable devices designed for continuous health monitoring. This thesis focuses on the optimization and evaluation of wearable electroretinography (ERG) prototypes, aimed at improving the reliability and accuracy of oscillatory potential (OP) recordings for diabetic retinopathy assessment. The wearable ERG device integrates electrodes mounted on a lightweight frame designed for prolonged use, offering enhanced comfort and mobility compared to traditional static systems. Through iterative design and optimization, key factors such as consistency, signal-to-noise ratio, and repeatability were assessed. Experimental evaluations comparing the wearable ERG prototype to the gold-standard ERG method demonstrated similar a-wave amplitudes and OP1 implicit times, indicating a potential for more reliable and consistent retinal measurements. The outcomes of this research contribute to the development of wearable, user-friendly devices that can transform the monitoring and diagnosis of retinal diseases, with particular relevance to diabetic retinopathy.M.S.Biomedical Engineerin
Distributed and Time-Varying Optimization for Autonomy and Decision-Making
No full textAs the use of autonomous agents expands, agents are being tasked with completing
more complex tasks in increasingly challenging environments. To complete these tasks
agents must make decisions with limited information and onboard resources such as low
computational power and inexpensive measurement units. One prevailing method to mitigate these challenges is to generate a solution, i.e., a reference trajectory, beforehand and
then design a control law for the agents to track the a priori generated solution. Typically,
an optimization algorithm uses a priori known data and dynamic models to generate the offline solution. However, in some environments it may not be possible to generate a solution
offline. For example, we cannot generate offline optimal trajectories for agents to navigate
an unmapped cave system in a search and rescue mission. One approach that has been proposed to enable agents to make decisions online is to use an optimization algorithm within
the control loop. However, vital questions arise regarding stability, performance, and implementability when proposing to use optimization in-the-loop. Therefore, this dissertation
addresses some of these challenges that arise when considering optimization in-the-loop
onboard agents with limited information and onboard resources.Ph.D.Electrical and Computer Engineerin
NorSand Implementation in Additive and Multiplicative Elastoplasticity
No full textProblems characterized by large deformations are ubiquitous throughout the field of geotechnical engineering. From the cone penetration test to geohazards such as landslides, understanding soil behavior within the regime of large strains is essential for both effective design and risk mitigation with broad economic and humanitarian implications. The NorSand constitutive model, grounded in critical state soil mechanics, is a state parameter-based model capable of capturing realistic soil behavior of dilative and contractive sands. Modern implementations of NorSand typically assume small-strain kinematics, which become inaccurate when applied to problems involving substantial strains and rotations. In this thesis, we seek to enhance the predictive capabilities of NorSand, with a focus on model performance under large deformations. To this end, NorSand elastoplasticity is formulated according to both the additive decomposition of strain, which aligns with small strains, and the multiplicative decomposition of strain, which aligns with finite strains. Full descriptions of kinematics in the two regimes are given, with more rigorous mechanics emphasizing rotational invariance incorporated for the multiplicative formulation. A performance evaluation shows that both formulations agree with a benchmark model in the small- and medium-strain regime, with the multiplicative formulation capable of accurately describing soil response in the large-strain regime. Once verified, the multiplicative formulation of NorSand is implemented in Kratos Multiphysics, specifically its Geotechnical Particle Finite Element Method (G-PFEM) module, to simulate cone penetration. Cone penetration simulations give classical readings of tip resistance, sleeve friction, and excess pore pressures. Extension of NorSand into the finite strain regime represents an advancement in computational granular mechanics for large deformation problems.M.S.Civil Engineerin
Development of New Heme Sensors and Novel Insights into Heme Homeostasis
No full textHeme is an essential iron-containing nutrient important for a myriad of biological processes and defects in heme homeostasis have been implicated in a variety of disorders. However, heme is potentially cytotoxic, suggesting that cells must have evolved mechanisms to limit the deleterious effects heme may have on membranes and proteins. While some heme processes are well characterized, many aspects of heme biology remain unclear. To enable the development of better medical solutions for heme-related conditions, there is a need to improve our understanding of heme homeostasis mechanisms. Previous work had developed genetically encoded heme sensors that have vastly increased our understanding of heme homeostasis across several different organisms, and to expand the ability of these tools to probe heme in multiple contexts, this work sought to target the heme sensors to the ER and diversify the color palette of heme sensors for multi–compartmental and –cellular imaging. Additionally, the novel phenomenon of intercellular heme transfer in Baker’s yeast was preliminarily characterized.
Attempts to target the prototype heme sensor (HS1) to the S. cerevisiae ER were unsuccessful due to an incompatibility between the integral fusion sensor design and yeast import and folding capacity. While ER heme sensors redesigned into a linear domain structure overcame fluorescence and localization challenges, they failed to recapitulate HS1 functionality. Interestingly, a humanized ER-HS1 was correctly localized and functional, validating its utility in expanding our understanding of heme biology, and highlighting species differences in heterologous protein expression.
The diversification of the HS1 color palette involved replacing the green fluorescence module with a cyan variant. The S. cerevisiae ECFP-HS1 functioned similarly to the prototype sensor, responding to titrations of labile heme levels across different strains. Importantly, the ECFP-HS1 signal was distinctly differentiable from HS1, allowing both constructs to be simultaneously utilized for multiplexing applications.
Finally, cell-cell communication and nutrient exchange in the context of heme was explored in S. cerevisiae. Heme-deficient strains were found to uptake heme from heme-replete donor strains in a manner that involved cell-cell contact. This finding has implications not only for our understanding of heme homeostasis, but also general nutrient regulation for yeast population control.
Taken together, this work has expanded the toolkit for probing heme trafficking and signaling in biological systems, as well as identified a novel process of heme regulation in Baker’s yeast.Ph.D.Chemistry and Biochemistr