Treasures @ UT Dallas
Not a member yet
7697 research outputs found
Sort by
Effective Human Allied Learning of Tree Ensembles
No full textThe current surge in machine learning is focused on automating pattern extraction from data,
primarily relying on labels, often generated by humans in the case of supervised learning.
While this approach is effective when dealing with abundant data, it is impractical when
dealing with small amounts of noisy, and structured data. Many situations involve only a
limited dataset, and obtaining more data is not sustainable.
Classically, AI methods have always used human knowledge to drive the learning process
to a better solution. In most cases, these were included as inductive biases. In many
domains, such knowledge is natural as humans have accumulated valuable knowledge over
long periods of time. Therefore, it becomes paramount to integrate human knowledge into
the model learning process, enriching and complementing the available data while mitigating
harms such as inherent biases and prejudices.
Simultaneously, machine learning models should not be treated as black boxes. The learned
models need to be easily understood by humans, expanding the knowledge possessed by
them. The transparency of models facilitates trust-building between humans and models.
Without explainability, humans may remain unaware of coincidental patterns identified by
the model within the data.
This dissertation specifically addresses these challenges within tree ensemble models, aiming
to extend the techniques to other machine learning models and enhance collaboration be-
tween humans and machine learning models for more effective problem-solving. To improve
the explainability of a robust learned model, we seek to combine tree ensembles with the
stacking method using relational templates. To enhance the interpretability of relational
tree ensembles, given the potential complexity with numerous trees, each relational tree is
first converted into a relational decision list. These decision lists are then combined and
compressed, resulting in a simplified decision list. Pre-processing and caching are employed
for efficiency, and the combination and compression are executed incrementally to save time
and space during intermediate steps. While the previous method improved the explainabil-
ity of the learned model, human knowledge is not exploited during learning. To address
this issue, we consider the injection of human knowledge from two scenarios: leveraging the
privileged information framework and directly incorporating human knowledge as first-order
logic rules. Injecting human knowledge into the learning process of tree ensembles aims to
improve model performance, fairness, and privacy
Physical Sensing of Airborne Particulates Using Complementary in-situ and Remote Sensing Approaches
No full textThe goal of this study is to physically sense and estimate the abundance of airborne particles through the combined use of in situ and remote sensing techniques. In situ detection
was carried out using a purpose-built design architecture for particulate matter (PM) detection modules powered by both solar and grid energy, integrated with LoRaWAN technology
as solution to mitigate utility costs, which has been hindered the expansion of the ground
observational network. This sensing system provides real-time data with very high temporal resolution compared to other available networking systems. This capability allows
for a deeper understanding of air quality dynamics. This sensing system was deployed at
scale in a dense urban environment in North Texas. Data, including eight highly synchronized feature variables, was used to construct models for PM concentrations in different size
fractions. The resulting models demonstrated excellent performance, indicated by correlation coefficients greater than 0.9 when tested on independent validation data. The research
outcome highlights two important factors in PM modeling. Firstly, although ground-level
particulate matter concentration depends on numerous factors, the most significant ones
are temperature, pressure, and humidity. Secondly, obtaining hyperlocal data at precisely
identical geographic locations with synchronized high temporal resolution timestamps during PM modeling is crucial. The in situ data gathered from both our own sensor network,
the OpenAQ network, and the national EPA network across the United States are then
used as ground truth for a remote sensing machine learning approach by integrating with
geostationary remote sensing observations. Machine learning models were developed for the
hourly estimate of PM2.5 concentrations in the continental United States (US). These models
incorporated data from the Geostationary Operational Environmental Satellites (GOES-16)
Aerosol Optical Depth (AOD) dataset, meteorological variables from the European Center
for Medium-Range Weather Forecasting (ECMWF), reanalysis of AOD data and air pollutant information from the MERRA-2 database, and additional data covering the period
from January 2020 to June 2023. The results of this investigation reveal that AOD data,
together with ECMWF variables and ancillary data, are effective components in modeling
PM2.5 concentrations, yielding a significant correlation coefficient of 0.849 when evaluated
on independent validation data. The reconstructed PM2.5 surfaces generated by the model
developed through this study serve as a critical data source to track pollution events and
conducting PM2.5 analyses. Furthermore, observations revealed inaccuracies in PM concentrations from our sensing system under high humidity and foggy conditions. In response,
this study implemented a humidity correction algorithm based on the hygroscopic growth
theorem for the sensing system. In addition, a computational algorithm based on aerosol
size distribution was formulated to estimate both the count and the mass concentration of
airborne nanoparticles with diameters less than 50 nm
Mechanism of Fermi Level Pinning for Metal Contacts on Transition Metal Dichalcogenides and Their Interface Thermal Stability
No full textTransition metal dichalcogenides (TMDs) have demonstrated immense potential for application in
state-of-the-art electronic, optoelectronic, and spintronic devices because of their outstanding
electronic, optical, mechanical, and magnetic properties. However, the failure of tuning the
Schottky barrier height by the work function of metal contacts strongly limits the efficiency of
carrier injection and hence the electronic performance of TMD-based devices. This dissertation
focuses on the interface between covalent and van der Waals metal contacts and TMDs to study
the origin and mechanism of Fermi level pinning. Firstly, the interface chemistry and band
alignment of Ni and Ag contacts on MoS2 is studied. Then the mechanism of Fermi level pinning
of metal contacts on other Mo- and W-based TMDs are uncovered by considering interface
chemistry, band alignment, defects and impurities of W-TMDs, contact metal adsorption
mechanism and the resultant electronic structure. Also, the thermal stability of Ni/MoS2 systems
is investigated in the aspects of interface chemistry, elemental diffusion, and band alignment
Millennial Philanthropy: the New Voice of Philanthropy
No full textThis dissertation examines millennial philanthropy using a mixed-methods approach using data
from two nonprofits. Chapter 2 explores giving behavior and patterns at BvB Dallas, a millennial
based nonprofit, through in-depth qualitative interviews from 2017 and 2021. Chapters 3 and
Chapter 4 utilize empirical data from GRACE, a parachurch to examine how volunteer and
donations patterns differ across generations. The focus of both studies is the millennial
generation since they are the largest generation at nearly 75 million. With the Great Transfer of
Wealth coming from Baby Boomers to the younger generations including the millennial
generation, the expected transfer is expected to be up to $65 trillion dollars to the millennial
generation. How will this impact the nonprofit community and how will nonprofits continue to
motivate the millennial generation for philanthropy and fundraising
Bioinspired Surfaces for Super Liquid and Ice Repellency
No full textSurfaces with ultralow adhesion to liquids and solids are of great interest for both fundamental
research and practical applications, from passive removal of highly wetting liquids to anti-icing.
This dissertation aims to investigate and develop bioinspired methods for achieving super-
repellency and addressing issues related to the high adhesion of liquids and ice on surfaces. In
this dissertation, the limitations of the current state-of-the-art superomniphobic surfaces (rely on
air lubricant) and liquid-infused surfaces (rely on liquid lubricant) are discussed followed by the
proposal of a new design of superomniphobic surface, which mitigates the dependence on
stringed nanoparticles and can be easily converted into a liquid-infused surface with a simple
one-step process. Drawing inspiration from various bio-inspired design strategies, along with
liquid repellency anti-icing has been explored in broad aspects: (1) Delay of frost propagation
through the meniscus-mediated spontaneous movement of droplets on liquid-infused surfaces.
Surface tension forces generated by the hydrophilic oil meniscus of a large water droplet on a
hydrophilic liquid-infused surface efficiently pull neighboring tiny droplets with a diameter < 20
m from all directions, causing them to climb and coalesce. This creates a dynamic length
separation between water droplets and a neighboring frozen droplet, which eventually delays
frost bridging. This is supported by a theoretical model to characterize the dynamically changing
inter-droplet gaps. (2) A new design of quasi-liquid surfaces is demonstrated to address the
durability challenges of current state-of-the-art super-repellent surfaces in terms of liquid
repellency and ice adhesion. Inspired by cilia in human lungs, quasi-liquid lubrication is
achieved by grafting flexible polymer molecules on a flat solid substrate. The mobile polymer
chains behave like a liquid layer and significantly reduce the interfacial adhesion between the
substrate and ice, resulting in ultralow ice adhesion. (3) The adhesion mechanics of ice on
different substrates at high supercooling are studied. Due to a sudden drop in temperature from -
10 to -60ºC, ice cracking occurs and can be categorized as large, intermediate, and no cracks on
aluminum, glass, and polycarbonate respectively based on the thermal properties of the substrate.
A theoretical model is proposed to quantify the number of cracks formed at the ice-substrate
interface due to thermo-mechanical stresses. Finally, an analogy is made between the interfacial
cracks, total crack length, and the reduction in ice adhesion. Overall, this dissertation provides
insights into bioinspired design strategies for super-repellency, with potential applications in
liquid repellency and anti-icing
Learning Tractable Probabilistic Graphical Models in Continuous and Temporal Domains
No full textProbabilistic Graphical Models (PGMs) such as Bayesian networks (BNs) and Markov random
fields (MRFs) present a powerful framework for representing complex probabilistic
dependencies and reasoning about uncertainties. However, inference over these models is
typically intractable. Recently, tractable models such as cutset networks and sum-product
networks (SPNs) have become increasingly popular because they allow some inferences to be
performed in polynomial time with respect to the size of the network. However, most of this
recent work focuses on discrete domains, and the resulting models can yield poor predictive
performance in practice as a result of restrictive modeling assumptions. In temporal modeling
problems, existing probabilistic models such as dynamic Bayesian networks (DBNs),
hidden Markov networks (HMMs), dynamic sum product networks (DSPNs), and dynamic
cutset networks (DCNs) employ first-order Markov and stationary assumptions, and limit
representational power so that efficient (approximate) inference procedures can be applied.
In this work, we aim to design a general probabilistic framework for continuous, temporal
modeling that allows efficient and accurate approximate inference. To this end, we first
extend the idea of cutset conditioning into continuous domains and propose a probabilistic
model that encodes the joint distribution as the product of a local, complex distribution
over a small subset of variables and a fully tractable conditional distribution whose pa
rameters are controlled by a neural network. This model admits exact inference when all
variables in the local distribution are observed, otherwise we show that “cutset” sampling
can be employed to efficiently generate accurate predictions in practice. We then extend our
framework into temporal domains and model the full transition distribution as a tractable
continuous density over the variables at the current time slice only, while the parameters are
controlled using a Recurrent Neural Network (RNN) that takes all previous observations as
input. We show that, in this model, various inference tasks can be efficiently implemented
using forward filtering with simple gradient ascent. Lastly, we enhance the robustness of
our continuous tractable model against distribution shifts using the Distributionally Robust
Supervised Learning (DRSL) framework. We demonstrate the approach of applying DRSL
in learning robust generative probabilistic models and develop an efficient linearithmic algorithm
for addressing the adversarial risk minimization problem. We evaluated our models’
predictive performance and robustness through various tasks on real-world datasets, and the
experimental results demonstrate the superior performance of our models against existing
competitors
Controllability, Reachability, and Inference for Complex Dynamic Systems
No full textThe focus of this dissertation is on developing a data-driven framework to infer interconnections in networks of dynamic agents, development of theoretical foundations for the controllability of bilinear systems, and reachability analysis of linear time invariant systems.
Inspired by some existing methodologies on discovering the governing dynamics using time-
series data, here we introduce two new data manipulation techniques that can be applied to
infer interconnections when the agents are of the second-order dynamics and have hidden
states, the states not available for direct measurements. In the next step, the controllability
of single-input bilinear systems is investigated. It will be shown that some of the conditions
required for the controllability of such systems can be relaxed at the expense of losing control over regions with zero Lebesgue measures in the state space. We then show how these
relaxations can open path to achieve conditions on the near controllability of multi-input
bilinear systems. Lastly, the reachability problem is studied for discrete-time linear systems
where we propose new techniques for verifying the inclusion and exclusion of given sets with
applications to the security of cyber-physical systems
Active Load Control of Wind Turbines Using Plasma Actuation
No full textWind turbines are progressively moving toward larger rotor diameters, hub heights, and
power ratings to lower the Levelized Cost of Energy (LCOE) by increasing the Annual Energy
Production (AEP). However, as wind turbines get larger and larger, simply scaling the rotor
diameter and rated power comes into conflict with the ‘square-cube law’, where the amount of
material used in the blade scales with its volume (the cube), while the energy capture scales
with the area of the rotor (the square). Thus, the capital costs can grow faster than the gains
in energy production. One approach to disrupt the square-cube law and reduce the LCOE
further is to mitigate aerodynamic loads by employing on-blade Active Flow Control (AFC)
devices. By reducing aerodynamic loads, with no adverse effects on turbine performance, this
approach allows longer blades with less material and, thus, lower capital costs.
The research in this dissertation seeks to quantify the performance of plasma-based active
load control under different setups and wind conditions. It aims to explore the potential of
using plasma-based actuators for active load control through detailed modeling, validation,
simulation, and analysis processes.
Dielectric Barrier Discharge (DBD) plasma actuators have several advantages over other
AFC devices, including having no moving parts, being lightweight, having a high bandwidth
response, and being easy to integrate on blades. These devices can be used to modulate the
local lift along the blade span. Therefore, the first part of this dissertation research focuses
on the modeling and validation of the DBD plasma-based lift actuator. A detailed modeling
process of the lift actuator in the NREL FAST simulation tool is presented. The results of a
preliminary wind tunnel experimental validation for the lift actuator are also presented and
discussed. Then, an investigation of using multiple lift actuators compared to using a single
lift actuator on each blade for dynamic load control is conducted. Open-loop and closed-loop
feedback control systems are designed to investigate the performance difference between using
single and multiple actuators. Next, inspired by a previously published bound on achievable
load reduction, parametric and sensitivity studies of typical blade design parameters are
carried out to explore and evaluate the use of the bound to better understand the relation
between actuator effectiveness and blade geometry. After that, a methodology for designing
and evaluating a feedback control system using lift actuators to alleviate the extreme loads
under gust wind is proposed. A switching scheme is described to integrate the gust load
controller into turbines with an existing dynamic load controller to achieve extreme loads and
deflections reduction while maintaining the fatigue load reduction capability. Finally, as wind
turbines in the field can be exposed to off-design conditions intentionally or unintentionally,
which could adversely affect the wind turbine loading condition and power production,
the combined plasma-based fatigue and gust load controller are evaluated under off-design
conditions. The performance of the load controller on fatigue loads, turbine performance,
and extreme loads is investigated under two off-design conditions—yaw misalignment and
blade contamination
Chemically Tuned Virus Like Particles: From Cancer Therapy to Targeted Delivery
No full textIn recent years, nanoparticle-based therapeutics have been increasingly applied in broad range of
clinical applications from diagnosis to treatment of many diseases such as cancer, diabetes and
neurodegenerative disorders. A wide range of synthetic and naturally occurring materials such as
polymers, metal oxides, silicate, liposome, and carbon nanotubes have been developed to
overcome some of the key barriers in free therapeutics including intracellular trafficking,
cell/tissue targeting, poor biodistribution, and low efficiency. However, despite all achievements
in creating these nanomaterials with different chemical and physical properties such as size, shape
and surface properties, developing a nanoparticle to surmount these limitations all in one is a big
challenge. Virus like particle (VLP) as protein-based nanomaterials that closely mimic the highly
symmetrical and polyvalent conformation of viruses while lack the viral genomes have emerged
as a solution for these limitations. Their unique features such as high biocompatibility,
biodegradability, monodisperisty, intrinsic immunogenicity, and safety combined with interior and
exterior modification capability offer new tool to scientists for careful design and engineering of
multi-component therapeutic agent with intended biological behavior and pharmacological profiles. Herein, various chemistry strategies are introduced in combination with biology and
immunology to turn virus like particle to a favorable engineered biomaterial for several functions
such as cancer therapy and intracellular delivery. We showed how by modifying surface of VLP
Qβ with NIR organic molecule we can make a highly efficient and stable photothermal agent that
can cause thermal ablation of tumor while simultaneously activating the immune response. We
found this immunophotothermal agent, suppress primary tumor, control metastasis, and prolong
survival time in mice bearing breast cancer. We also addressed one of the biggest challenges in
biologic delivery which is direct delivery of therapeutic cargo into cell cytoplasm. Using organic
chemistry we designed a cytosolic targeting linker that when attached to surface of VLP Qβ, helps
to escape endosomes and be released into cytoplasm, Moreover, this proteinaceous material is
shown to have a great potential in combination with other materials such as metal organic
framework to construct a multimodal cancer therapeutic agent enabling delivering mulit
therapeutic agents such as immunotherapeutic drugs while taking advantage of all unique features
of virus like particles. These works clearly show the significant potential of VLP in design and
modification of new therapeutic platform
Computational Analysis of Laser Impact Welding Processes Using Eulerian Formulation
No full textOver the past 70 years, impact welding has been used to join metallic substrates with the
advantage of not exposing the materials involved to significant amounts of heat. This allows
the direct joining of dissimilar or heat-sensitive alloys in ways not considered practical using
fusion welding. Most impact welding techniques create a weld between large parts by applying a very large impulse to a flyer part within a short period, and cause the surfaces to collide
at a particular range of angles. A number of analytical and, more recently, computational
approaches have been developed to investigate the phenomena that occur at rapidly forming collision interfaces, as direct, in situ observation of such processes are experimentally
challenging. This challenge is only compounded when the parts to be joined are very small,
such as in laser impact welding, which welds flyer foils of 0.1 mm thickness or less to various
geometries of substrates.
As a prerequisite to success, impact welding requires the ablation of thin layers of material
from both parts’ contact surfaces in a high velocity, high temperature jet. This process
removes oxides and other contaminants, while indenting the surface asperities to promote
mutual contact at high pressure. Extreme shear stresses occur within both the flyer and its
target, temperatures increase dramatically at the collision point, and extreme plastic strains
result from the concentration of shock stresses. Larger scale impact weld techniques, such as
those involving the use of explosives, magnetic discharge, or the vaporization of an electrically
conductive foil, are energetic enough such that meso- or micro-scale characteristics such the
metallic grains’ microstructure, or the contact surface roughness, may be neglected when
performing numerical or analytical investigations. However, the foils generally used for
flyers in laser impact welding are thin enough such that these effects can become significant.
Presented is a comprehensive analysis of laser impact welding using computational methodology based on the Eulerian finite element formulation. Frameworks including the inhomogeneous microstructure and measured rough surfaces of metallic foils are developed to predict
the evolution of transient phenomena throughout the formation of laser impact welded joints,
using material definitions appropriate for welds between dissimilar, as well as similar, alloys.
Correlations are found between the weld morphologies numerically predicted and found experimentally; additionally, a basis to use the Eulerian formulation in laser impact welding
modeling is established to help promote further development and adoption of the technique