137091 research outputs found
Sort by
Novel Machine Learning Approaches with Applications in Healthcare and Social Welfare
No full textAs the world evolves rapidly, the use of machine learning (ML) and artificial intelligence (AI) in healthcare and social welfare has become crucial. These fields are fundamental to our society, and machine learning offers powerful tools to tackle prevalent issues such as handling high-dimensional and sparse data, improving data quality and hence, predictive accuracy for chronic disease management, and optimizing decision-making under uncertainty in these fields. Similarly, social welfare professionals and policy makers urgently need targeted interventions and equitable resource allocation. This thesis covers three studies, each of which targets critical aspects of healthcare and social welfare.Ph.D.Machine Learnin
Atmospheric Boundary Layer Effects on Ship Airwakes Using Synthetic Eddy Method in Lattice-Boltzmann Simulations
No full textShipdeck landing is amongst the most challenging operations a modern navy deals with. To ensure such operations are safely carried out, extensive training of prospective pilots is required. One of the first steps of such training is exercises with flight simulators. To realistically simulate the conditions in the wake of a ship, Computational Fluid Dynamics (CFD) simulations should be conducted. There are different CFD methodologies to obtain an accurate ship airwake simulation, one of them is the Lattice-Boltzmann Method (LBM). It has been proven that the LBM is a computationally efficient mid-fidelity method to solve various flow cases. Previous studies showed that the accuracy of the LBM ship airwake predictions was acceptable, hence results were deemed suitable to be used in flight simulators.
Many aspects need to be considered in a ship airwake simulation to represent real-world conditions as closely as possible. Ship geometry modeled in simulations, boundary conditions used to model ship, and inflow modeling are among those important aspects. Ship geometry fundamentally affects the flow field over the landing deck, hence, choice of an appropriate ship model is crucial. NATO Generic Destroyer (NATO-GD) is a more accurate representation of modern naval combatant designs in comparison to widely-adopted Simple Frigate Shape 2 (SFS2). Therefore, it is important to understand the difference between the airwakes generated by different ship geometries. To be able to properly model curved surfaces of NATO-GD, simple boundary conditions like bounce-back fall short, as it transforms geometries into voxelized Cartesian approximation. Rather, boundary conditions that can model non-Cartesian aligned surfaces in the LBM simulations would be preferred such as Mei-Luo-Shyy (MLS) and Grad immersed boundary conditions.
The Atmospheric Boundary Layer (ABL), being a natural phenomenon, needs to be modeled in such simulations. It is relatively straightforward to model the steady part of the ABL as it can be modeled by power law or logarithmic law profiles. However, inherent turbulence of the ABL is not represented in such mean velocity profiles. There are numerous methods to simulate the inherent turbulence in numerical simulations. While some of those methods have been proven to be ineffective, some others have been proven to be computationally expensive. With the Synthetic Eddy Method (SEM) it is possible to retain the computational efficiency of the LBM, while accurately simulating the inherent turbulence.
While it was not the primary objective of this study, a new method was proposed to evaluate pilot workload for ship deck landing. Using solely the vertical flow fluctuation statistics obtained from the LBM simulations, this new method eliminates the subjectivity inherent in pilot workload evaluation. The comparisons with pilot workload ratings available for SFS2 in the literature demonstrated good correlation, indicating the potential of this new method.
This thesis aims to inspect the effects of the ship geometry and the ABL on the ship airwake by simulating a realistic ABL by using the Synthetic Eddy Method. It also evaluates the solid wall boundary conditions to model ship geometry accurately in the LBM simulations. Additionally, the effects of floating-point precision and Lattice-Boltzmann velocity sets on ship airwake simulation accuracy were evaluated, allowing for the determination of minimum viable computational resources. The accuracy of simulation results was evaluated by comparisons with multiple experimental measurement data sets and high-fidelity CFD results
Biocapacitance Sensors for In-Line Viable Cell Density Monitoring in Complex Bioreactors
No full textThe precise monitoring of Viable Cell Density (VCD) in bioreactors is crucial for optimizing bioprocesses, improving yield, and maintaining product quality. VCD can be effectively indicated by biocapacitance, a capacitance associated with the viable cell membrane, which can be detected through Impedance/Dielectric Spectroscopy (IS/DS) The advancements of high-performance computing and Artificial Intelligence (AI) in recent decades have expedited the measuring and analysis of IS/DS enabling real-time monitoring of biocapacitance. In-line biocapacitance sensors has been widely applied for VCD monitoring in lab-scale adherent cell cultures and in high-throughput suspension cell cultures. However, this promising technology has not been effectively applied to high-throughput adherent cell cultures and lab-scale suspension cell cultures due to the complexity of the bioreactors. The Hollow Fiber Bioreactor (HFBR) is a high-efficiency bioreactor for high-throughput adherent cell culture thanks to its high surface-to-volume ratio. However, the packed interior of the HFBR presents unique challenges in achieving in-line biocapacitance sensing.
This dissertation presents the development and characterization of biocapacitance sensors tailored for HFBRs. The study encompasses theoretical analysis, modeling, design, fabrication, and data analysis to create and optimize these sensors. The technology developed in this study is the first to achieve direct in-line VCD monitoring for high-throughput adherent cell culture. Key findings demonstrate that the sensors exhibit high sensitivity and potential to achieve spatial resolution, effectively monitoring VCD in real-time. Comprehensive data analysis and theoretical insights provides deeper understanding and optimization of sensor performance. Additionally, the work includes the development and assessment of two other biocapacitance sensors for lab-scale suspension cell cultures, which are also insufficiently addressed in current market and research. This work highlights the critical role of sensor technology, theoretical analysis, and data processing in advancing bioreactor systems.Ph.D.Materials Science and Engineerin
Low-Gravity Sloshing in Spherical Tanks: Experimental Investigation Toward Numerical and Analytical Modeling
No full textUncontrolled sloshing in spacecraft propellant tanks can severely disturb vehicle attitude and dynamics, jeopardizing critical proximity maneuvers. Despite decades of research, the dynamics of low-gravity sloshing remain insufficiently understood and difficult to model with reliability, also due to limited and incomplete experimental data. In this context, this thesis presents: (i) a new experimental dataset on low-gravity sloshing in a spherical tank; (ii) a numerical CFD framework and an analytical spherical pendulum model for simulating low-gravity free contact line fluid motion; and (iii) the development of numerical and analytical models for contact angle hysteresis. The experimental dataset, originating from the SILA (Sloshing Imaging and Load Analysis) payload, shows low- to large-amplitude sloshing in a spherical tank for a range of Bond numbers below 50. The force exerted on the tank walls is recorded, and the position of the center of mass is reconstructed from free surface tracking to analyze the fluid’s response in both time and frequency domains. Video imaging and Bond number evolution are reported to characterize the low-gravity environment and transitions between high and low gravity. A numerical setup is implemented in ANSYS Fluent and validated in microgravity for constant contact angles. The analytical model of the spherical pendulum is implemented in MATLAB, and the results for the center of mass displacement qualitatively match the numerical ones. Still, both deviate from the experimental data due to overly simplified free-contact-line assumptions. For the analyzed low-gravity range (Bo <50), the contact angle hysteresis is revealed as a fundamental physical factor that simplified pendulum analogies fail to simulate, thus invalidating their use in widespread spacecraft guidance, navigation, and control algorithms. The development of a User Defined Function for ANSYS Fluent modeling the contact angle hysteresis in both two and three dimensions is initiated, valid for structured grids and for surfaces of any shape. The development of a two-pendulum model is proposed to capture the dynamics given by the contact angle hysteresis
A Systematic Methodology for Generating Effective Thermo-Physical and Thermomechanical Properties of TRISO Fuel Pellets
No full textThis work presents a detailed methodology for generating effective
material properties of tri-structural isotropic (TRISO) fuel pellets for use in
reduced-order homogenized thermal and/or mechanical calculations. This is
achieved by use of the finite-element (FE) method to perform simulations of
explicitly modeled TRISO fuel pellets, in conjunction with derivation of
analytical equations for analogous homogenous pellets which preserve a
particular parameter of interest. This method is applied to several properties
including thermal conductivity, specific heat capacity, thermal expansion
coefficient, Young’s Modulus and Poisson’s Ratio, and the properties
generated are a function of temperature and packing fraction.
It is demonstrated that the state-of-the-art for homogenizing the
thermal conductivity is inadequate for realistic TRISO pellets, while for the
other properties no existing standard is found in the literature for
homogenizing. Additionally, a multi-physics application of the properties is
demonstrated, which involves coupling neutronics with thermo-mechanical
calculations. The fission power distribution is obtained from the neutronic
analysis, and then the temperature, displacement and stress-strain
distributions from the explicit model are compared with those from the
homogenous calculations. For the finite-element simulations, the commercial
software Abaqus is used, and the Monte Carlo code MCNP is used for the
neutronics to obtain the power distributions.M.S.Nuclear and Radiological Engineerin
Advancing Expressibility and Scalability in Graph Learning
No full textGraph learning models have become the dominant approach for learning from graph-structured data, yet two fundamental challenges persist: expressiveness and scalability. Advances in model expressiveness often compromise scalability, underscoring the need for holistic approaches that address both objectives simultaneously. This thesis tackles these interrelated challenges through a model co-design perspective, developing methods that jointly optimize for expressiveness and scalability.
We present four key contributions. First, MG-GCN, a multi-GPU framework for efficient full-batch Graph Convolutional Network training, establishes foundational strategies for scaling Graph Neural Networks through 1D row-wise partitioning, memory optimization via buffer reuse, and latency hiding through overlapped communication and computation.
Second, we introduce the Graph-Enhanced Contextual Operator (GECO), an architectural innovation that integrates neighborhood propagation with global convolutions to capture long-range dependencies in quasilinear time, enabling scalable Graph Transformers without sacrificing quality.
Third, we propose Global Neighbor Sampling (GNS) and Global Layer Sampling (GLS), efficient sampling strategies that overcome the quadratic complexity of dense Graph Transformers by distinguishing local from global neighbors, scaling these models to graphs with millions of nodes.
Fourth, we introduce Neighbor Aware Skip Connections (NASC), a lightweight mechanism that adaptively balances node and neighborhood information, serving as a versatile plug-in for both GNNs and Graph Transformers with minimal overhead.
Together, these contributions demonstrate that achieving scalability and expressiveness in graph learning does not necessitate increasingly complex architectures. Rather, a model co-design approach—integrating system-level optimizations, architectural innovations, sampling strategies, and simplified components—can deliver high performance efficiently
Graphene Oxide-based Catalysts for Waste Lignin Valorization to Value-added Biochemicals
No full textAs the most abundant aromatic polymer in nature, lignin is considered a promising resource for producing value-added biochemicals, such as monophenols and biofuels. However, achieving high conversion and selectivity in lignin valorization remains challenging. Among various valorization methods, oxidative depolymerization can selectively cleave the C-C and C-O bonds between the fundamental units of lignin, producing monophenolic products. However, this method faces limitations due to low conversion and yield, the need for harsh conditions (e.g., high temperature, high pressure, or toxic solvents), the use of expensive catalysts, and the complexity of the products. Therefore, it is crucial to design cost-effective catalysts that are efficient and selective. Graphene oxide (GO) is a good candidate as the catalyst for lignin depolymerization for lignin valorization due to its large surface area, abundant functional groups, and excellent dispersibility in various solvents. Furthermore, modifying GO with metal nanoparticles, such as copper oxide (CuO), enhances its catalytic activity and selectivity. This thesis aims to develop GO and metal-modified GO catalysts for efficient and selective lignin depolymerization.Ph.D.Chemical and Biomolecular Engineerin
Improving energy efficiency in secondary aluminum melting furnaces through waste heat recovery
No full textApproximately 80% of aluminum production in the United States uses the secondary process, which recycles scrap metal, instead of the primary ore refining process. The melting step is most energy-intensive, requiring temperatures of over 700°C and accounting for nearly 50% of total energy (natural gas) usage. Waste heat recovery technologies present an opportunity to enhance process efficiency by capturing energy from flue gas instead of rejecting it to the ambient. Two stages of waste heat recovery can be implemented for aluminum melting furnaces: (i) a high-temperature heat exchanger (400–900°C) located directly downstream of the melting furnace, and (ii) a medium-temperature recovery system (100–300°C) located further downstream. The first part of this study investigates the performance of two different high-temperature heat exchangers (recuperator and regenerator) for preheating combustion air. For this, an effectiveness – NTU method was used to analyze performance and to monitor the degradation of a recuperator over one year of operation. A one-dimensional, two-phase numerical model was developed to represent heat transfer and air flow through a packed-bed regenerative burner subject to clogging of the media particles (degradation). A parametric study was performed to examine the effect of key inputs (bed porosity, air flow rate, particle size, etc.) on the heat recovery performance over time. Both models were compared and validated with facility-specific data from secondary aluminum melting furnaces.
The second part of this study explores the use of medium-temperature recovery technologies, including for on-site power generation, load-preheating, and thermal energy storage. Specifically, Organic Rankine Cycles, Kalina Cycles, and waste heat boilers were evaluated for power generation. In addition, the use of sensible thermal storage (1D packed bed) was analyzed for offsetting load-preheating demands. Overall, this work provides recommendations for improving existing heat recovery technologies and adopting new practices for enhancing energy efficiency in secondary aluminum melting furnaces
A General Framework Linking Adsorbent Characterization and Process Simulation: Kinetics, Isotherms, and Adsorption Bed Modeling
No full textThis dissertation presents an integrated framework that combines material characterization,
numerical modeling, data-driven analysis, and adsorption-bed simulation to evaluate novel
adsorbents for adsorption-based separation processes. A modeling approach was developed to
extract adsorption kinetic parameters directly from dynamic gravimetric and volumetric
experiments, eliminating restrictive assumptions of traditional analytical solutions and providing
accurate diffusivities and mass transfer coefficients across diverse materials. To incorporate
complex mixture adsorption behavior into process simulations, a symbolic regression method was
created to generate empirical gas mixture isotherm equations from discrete experimental or
molecular simulation data, with numerical stability filtering enabling reliable implementation
during adsorption bed modeling. These equations were successfully used in breakthrough
simulations to predict mixture adsorption dynamics in systems where classical models or IAST
are insufficient. An efficient adsorption bed simulation toolbox was further developed using
numerical quadrature for discrete isotherms, a high-order WENO scheme for breakthrough
modeling, and pre-generated isotherm grids with spline interpolation to eliminate repeated IAST
calculations and improve computational speed. The framework was demonstrated through a case
study on atmospheric water harvesting with LiCl-impregnated MIL-101 analogs, showing that
although higher salt loadings increase water uptake, LiCl induced mass transfer limitations slow
kinetics and constrain rapid cycling, while higher gas flow velocities partially mitigate these
effects. Overall, this work provides a flexible and computationally efficient pathway for translating
laboratory-scale adsorption measurements into process-level performance predictions