Doctor of Philosophy / Department of Computer Science / William H. Hsu / Since the introduction of new curriculum standards at K-12 schools, computational thinking has become a major research area. Creating and delivering content to enhance these skills, as well as evaluation, remain open problems. This work describes different interventions based on the Scratch programming language aimed toward improving student self-efficacy in computer science and computational thinking. These interventions were applied at a STEM outreach program for 5th-9th grade students. Previous experience in STEM-related activities and subjects, as well as student self-efficacy, were surveyed using a developed pre- and post-survey. The impact of these interventions on student performance and confidence, as well as the validity of the instrument are discussed. To complement attitude surveys, a translation of Scratch to Blockly is proposed. This will record student programming behaviors for quantitative analysis of computational thinking in support of student self-efficacy. Outreach work with Kansas Starbase, as well as the Girl Scouts of the USA, is also described and evaluated. A key goal for computational thinking in the past 10 years has been to bring computer science to other disciplines. To test the gap from computer science to STEM, computational thinking exercises were embedded in an electromagnetic fields course. Integrating computation into theory courses in physics has been a curricular need, yet there are many difficulties and obstacles to overcome in integrating with existing curricula and programs. Recommendations from this experimental study are given towards integrating CT into physics a reality. As part of a continuing collaboration with physics, a comprehensive system for automated extraction of assessment data for descriptive analytics and visualization is also described.

Graphene is an atomically thin two-dimensional (2-D) crystal with unique thermal, mechanical, and electronic transport properties such as the high mobility of carriers, perfect 2- D confinement and linear dispersion, etc., has been attracted many interest as a promising candidate for nano-scale devices over the past decades. Multilayer stacks of graphene and other stable, atomically thin, 2-D materials offer the prospect of creating a new class of heterostructure materials. Hexagonal boron- nitride (hBN), is a great candidate to be stacked with graphene due to an atomically 2-D layered structure with a lattice constant very similar to graphene (1.8% mismatch), large electrical band gap (∼4.7eV), and excellent thermal and chemical stability. The graphene/hBN based tunneling transistors show the resonant tunneling and strong negative differential resistance (NDR). These devices which have potential for future high-frequency and logic applications such as high-speed IC circuits, signal generators, data storage, etc., has been studied both theoretically and experimentally recently. The aim in this dissertation has been to study the effect of the uniaxial strain on the graphene nanoribbon resonant tunneling transistors (RTTs). The uniaxial strain may be induced either by an external stress applied to the graphene in a particular direction or by a substrate due to deposition of graphene on top of the other materials. The strain modifies distances between carbon atoms which leading to different hopping amplitudes among neighboring sites. A resonant tunneling transistor consisting of armchair graphene nanoribbon (AGNR) electrodes with three layers of hBN tunnel barrier between them has been considered. By using the nearest-neighbor tight-bind (TB) method and the nonequilibrium Green function (NEGF) formalism, the electronic transport characteristics of a RTT is calculated. In this work, we focus on how the strain affects the current-voltage characteristics of AGNR/hBN RTT.

The pricing of exotic derivatives continues to attract much attention from academics and practitioners alike. Despite the overwhelming interest, the task of finding a robust methodology that could derive closed-form solutions for exotic derivatives remains a difficult challenge. In addition, the level of sophistication is greatly enhanced when options are priced in a more realistic framework. This includes, but not limited to, utilising jump-diffusion models with mean-reversion, stochastic volatility, and/or stochastic jump intensity. More pertinently, these inclusions allow the resulting asset price process to capture the various empirical features, such as heavy tails and asymmetry, commonly observed in financial data. However, under such a framework, the density function governing the underlying asset price process is generally not available. This leads to a breakdown of the classical risk-neutral option valuation method via the discounted expectation of the final payoff. Furthermore, when an analytical expression for the option pricing formula becomes available, the solution is often complex and in semi closed-form. Hence, a substantial amount of computational time is required to obtain the value of the option, which may not satisfy the efficiency demanded in practice. Such drawbacks may be remedied by utilising numerical integration techniques to price options more efficiently in the Fourier domain instead, since the associated characteristic functions are more readily available. This thesis is concerned primarily with the efficient and accurate pricing of exotic derivatives under the aforementioned framework. We address the research opportunity by exploring the valuation of exotic options with numerical integration techniques once the associated characteristic functions are developed. In particular, we advocate the use of the novel Fourier-cosine (COS) expansions, and the more recent Shannon wavelet inverse Fourier technique (SWIFT). Once the option prices are obtained, the efficiency of the two techniques are benchmarked against the widely-acclaimed fast Fourier transform (FFT) method. More importantly, we perform extensive numerical experiments and error analyses to show that, under our proposed framework, not only is the COS and SWIFT methods more efficient, but are also highly accurate with exponential rate of error convergence. Finally, we conduct a set of sensitivity analyses to evaluate the models’ consistency and robustness under different market conditions

Includes bibliographical references. / The South Africa Coloured population of the Western Cape is the result of unions between Europeans, Africans (Bantu and Khoisan), and various other populations (Malaysian or Indonesian descent). The world-wide burden of tuberculosis remains an enormous problem, and is particularly severe in this population. In general, admixed populations that have arisen in historical times can make an important contribution to the discovery of disease susceptibility genes if the parental populations exhibit substantial variation in susceptibility. Despite numerous successful genome-wide association studies, detecting variants that have low disease risk still poses a challenge. Furthermore, admixture association studies for multi-way admixed populations pose constant challenges, including the choice of an accurate ancestral panel to infer ancestry and for imputing missing genotypes to identify possible genetic variants causing susceptibility to disease. This thesis addresses some of these challenges. We first developed PROXYANC, an approach to select the best proxy ancestral populations for admixed populations. From the simulation of a multi-way admixed population, we demonstrated the ability and accuracy of PROXYANC in selecting the best proxy ancestry and illustrated the importance of the choice of ancestries in both estimating admixture proportions and imputing missing genotypes. We applied this approach to the South African Coloured population, to refine both the choice of ancestral populations and their genetic contributions. We also demonstrated that the ancestral allele frequency differences correlated with increased linkage disequilibrium in the SAC, and that the increased LD originates from admixture events rather than population bottlenecks. Secondly, we conducted a study to determine whether ancestry-specific genetic contributions affect tuberculosis risk. We additionally conducted imputation genome-wide association studies and a meta-analysis incorporating previous genome-wide association studies of tuberculosis.

Includes abstract. / Includes bibliographical references (leaves 126-150). / Mycobacterium tuberculosis is a bacterial pathogen that causes tuberculosis, a leading cause of human death worldwide from infectious diseases, especially in Africa. Despite enormous advances achieved in recent years in controlling the disease, tuberculosis remains a public health challenge. The contribution of existing drugs is of immense value, but the deadly synergy of the disease with Human Immunodeficiency Virus (HIV) or Acquired Immunodeficiency Syndrome (AIDS) and the emergence of drug resistant strains are threatening to compromise gains in tuberculosis control. In fact, the development of active tuberculosis is the outcome of the delicate balance between bacterial virulence and host resistance, which constitute two distinct and independent components. Significant progress has been made in understanding the evolution of the bacterial pathogen and its interaction with the host. The end point of these efforts is the identification of virulence factors and drug targets within the bacterium in order to develop new drugs and vaccines for the eradication of the disease.

Emphysema is a type of Chronic Obstructive Pulmonary Disease (COPD) characterized by breathing difficulties due to airflow obstruction, and results in structural and functional changes of the lungs. Structural changes include alveolar wall destruction and the formation of enlarged alveoli, or bullae, which appear as low attenuation areas in the CT image of emphysematous lungs. Functional changes include increased lung compliance and decreased bulk modulus in emphysematous lungs. Previous mathematical and computational models have attempted to explain either general lung structure or function, but have not linked the two to explore patient-specific lung mechanics. We propose that we can link the structure and function by creating CT-based spring network models of the lung parenchyma and manipulating these networks to predict the regional tissue stiffness and global pressure-volume relationship of the lung during disease progression. The goal of this thesis is to predict these patient-specific changes during emphysema progression by approximating the lung tissue stiffness distribution from CT densities and predicting parenchymal destruction over time from high-strain regions of a non-linear elastic spring network representing lung tissue. First, we used simple spring network models to determine the appropriate non-linear spring force-extension equation to implement into the full lung network. We then mapped a spring network onto a CT image to create a lung network, applied the non-linear force-extension equation to the network springs, and developed a lung deflation model to capture the quasi-static pressure-volume curve of the lung. Finally, we reduced the stiffness of high-strain regions of the lung network and deflated the model to predict the loss of tissue elastance and the reduced bulk modulus over time. Our method shows evidence of a reduced bulk modulus and similar tissue destruction between predicted and actual lung networks, but further development and testing are necessary to create more accurate prediction network models.

A wide variety of phenomena in condensed matter systems is driven primarily by interactions between electrons in the system. This work is concerned with the application of functional renormalization group (fRG) as a generalized solver for the multi-band Hamiltonians that describe these systems. We consider a decoupled formulation of the fRG equations that is optimized in the frequency and momentum domains and retains the flow of relevant modes in the system. Approximate truncations that extend the scheme to arbitrary multiband systems are addressed. This optimized decoupling is then used to derive the flow equations that describe fluctuations in model Hamiltonians for Cuprate and Pnictide superconductors. We construct a full phase diagram of the systems studied as a function of doping, temperature and coupling. Access to the frequency modes in the system allows us to explore the impact of coupling phonons to these model Hamiltonians. Alterations to the diagram due to electron-phonon coupling is derived. The results of the decoupled formulation is in agreement with results in the literature for many of the models considered. Further the fRG captures the sensitivity of susceptibilities of Cuprate Hamiltonians to band structure, the enhanced role of Hunds coupling in Pnictide systems and the impact of phonons in multiband Hamiltonians.

Genetic variation is shaped by admixture between populations in an evolutionary process. The mixture dynamic between groups of populations results in a mosaic of chromosomal segments inherited from multiple ancestral populations. The distribution of ancestral chromosomal segments and the recombination breakpoints in an admixed genome provide information about the time of admixture. Studying populations with particular ancestries has become a major interest in population genetics because of medical and evolutionary impacts of the patterns of single nucleotide polymorphisms. It provides a better understanding of the impact of population migrations and helps us uncover interactions between several populations. Most of the research on admixed population dating has focused on a single interaction between two populations using various approaches. Some have extended this to mixing of three populations based on assumptions and approaches which differ from one tool to another. However, the inference of distinct ancestral proportions along the genome of an admixed individual and plausible dates of admixture, still remain a challenge in the case of multi-way admixed populations. This dissertation consists of three research initiatives. First, provide a succinct review of current methods for dating the admixture events. We accomplish this by providing a comprehensive review and comparison of current methods pertinent to date admixture event. Second, we assess various admixture dating tools which estimate the time of admixture between two parental populations. We do so by performing various simulations assuming a particular number of generations and use these to evaluate the tools. Third, we apply the top three assessed methods to some admixed populations from the 1000 Genomes project. Despite MALDER shows improvement and produces reasonable date estimates over other current methods, the results from both simulation and real data suggest that dating ancient admixture events accounting for the effect of other admixtures remains a challenge. Our results suggest the need for developing a new approach to date ancient and complex admixture events in multi-way admixed populations.

The cellular function of proteins, and their targeting by drug applications, are both governed by biomolecular thermodynamics and kinetics. In order to make meaningful and efficient predictions of these mechanisms, molecular simulations must be able to estimate the binding affinity and rates of association and dissociation of a protein-ligand complex, or the populations and rates of exchange between distinct conformational states (i.e. folding and unfolding, binding and unbinding). The above studies are typically done using different, but complementary approaches. Alchemical methods, including free energy perturbation (FEP) and thermodynamic integration (TI), have become the dominant method for computing high-quality estimates of protein-ligand binding free energies. In particular, the widely-used approach of relative binding free energy calculation can deliver accuracies within 1 kcal mol−1. However, detailed physical pathways and kinetics are missing from these calculations. In principle, all-atom molecular dynamics (MD) simulation, with the help of Markov State Models (MSMs), can be used to obtain this information, yet finite sampling error still limits MSM approaches from making accurate predictions for very slow unfolding or unbinding processes. To overcome these issues, a new approach called multiensemble Markov models (MEMMs) have been developed, in which sampling from biased thermodynamic ensembles can be used to infer states populations and transition rates in unbiased ensembles. In this dissertation, two distinct biophysical problems are investigated. In the first part, we apply expanded ensemble (EE) methods to accurately predict relative binding free energies for a series of protein-ligand systems. Moreover, we propose a simple optimization scheme for choosing alchemical intermediates in free energy simulations. In the second part, we employ MEMMs to estimate the free energies and kinetics of protein folding and ligand binding, to achieve greatly improved predictions. Finally, we combine the above EE method and a maximum-caliber algorithm to study how sequence mutations perturb protein stability and folding kinetics. In summary, this work comprises a wide range of current methodology in biophysical simulation, complementing and improving upon existing approaches. / Chemistry

In recent years, new Artificial Intelligence technologies have mimicked examples of collective intelligence occurring in the natural world including flocks of birds, schools of fish, and swarms of bees. One company in particular, Unanimous AI, built a platform (UNU Swarm) that enables a group of humans to make decisions as a single mind by forming a real-time closed-loop feedback system for individuals. This platform has proven the ability to amplify the predictive ability of groups of humans in realms including sports, medicine, politics, finance, and entertainment. Previous research has demonstrated it is possible to further enhance knowledge accumulation within a crowd through curation and bias methods applied to individuals in the crowd.\newline This study explores the efficacy of applying a machine learning pipeline to identify the top performing individuals in the crowd based on a structural profile of survey responses. The ultimate goal is to select these users as Swarm participants to improve the accuracy of the overall system. Unanimous AI provided 24 weeks of survey data collection consisting of 1,139 users from the NHL 2017-2018 season. By applying a machine learning pipeline, this study able to curate a crowd consisting of users that had an average z-score 0.309 and Wisdom of the Crowd prediction accuracy of 61.5%, which is 4.1% higher than a randomly selected crowd and 1.4% lower than Vegas favorite picks.