Glass is one of the most important and frequently used materials due to its special properties: its hardness and transparency makes it an ideal material for windows, and its stability makes it a great candidate for immobilizing radioactive nuclear waste, etc. As a result, almost all aspects of glass ranging from its fabrication, properties and application, characterization, to stability and destruction have been hot topics of material science research for a long time. Among all researching methods, molecular dynamics (MD) simulation is a new emerging technique that has been applied more and more to glass research in modern years thanks to its advantages over conventional experimental methods such as high efficiency, high accuracy and low cost. This thesis focuses on using MD simulations to evaluate glass properties from two main aspects: 1) the equivalence of glasses produced from modern methods such as vapor deposition, sol-gel condensation and irradiation and those fabricated from conventional melt-quenching, and 2) the effects of temperature, pH and glass composition on zeolite precipitation during the nuclear waste immobilization glass dissolution process. The thesis is thus divided based on these two topics. In Chapter I, we aim to compare the equivalence of SiO2 glasses obtained from MD-simulated vapor deposition, sol-gel condensation and irradiation processes and the melt-quenched glasses. That is, to evaluate whether these glasses can (available) or cannot (forbidden) be obtained by using the traditional melt-quenching method by changing the cooling rate. We will show that the availability of glasses can be determined and explained by observing the medium-range structural features and the energy landscape of the atoms. In Chapter II, we explore another important field of glass application: the nuclear waste-immobilization glass dissolution. Though vitrification: the process of melting and mixing radioactive nuclear waste and glassy materials, is widely considered the best way of treating nuclear waste due to the extraordinary stability of glasses, observation of continuous glass dissolution (alteration resumption process, or stage III of the nuclear waste-immobilization glasses dissolution process) is reported in many experimental cases which will lead to nuclear waste leakage. Though the exact reason of alteration resumption is still being discussed, it is generally believed that this phenomenon is closely related to the precipitation of secondary phases like zeolites. Based on ab initio MD simulations, we first construct a complete methodology to calculate the thermodynamic properties (enthalpy and entropy of formation, heat capacity, etc.) of any zeolites given its composition and lattice structure. Moreover, with these thermodynamic data of zeolites, we use Gibbs free Energy Minimization (GEM) simulation to build a database of zeolite precipitation under various temperatures, pHs and initial glass compositions. Finally, we manage to train a machine learning (ML) model using the precipitation database that can predict zeolite formation given the aforementioned conditions as inputs.

Accurate estimation of 3D human pose/shape from a single image remains a challenging task under occlusion or domain shift. We try to solve this problem by investigating three different geometric priors: camera pose priors, scene geometry priors, and parametric body-model priors. The first part of the dissertation focuses on analyzing the difference in the popular 3d human pose datasets and proposes a plug-in camera pose module to improve cross-dataset generalization. In the second part, we evaluate the usefulness of scene geometry in helping improve 3d pose estimators. Finally, we build strong pose/shape estimators from a single image by leveraging the best of the statistical model-based methods and nonparametric methods.

Cardiac magnetic resonance imaging (MRI) is a proven technique for the evaluation of myocardial structure and function. An introduction to cardiac physiology is presented in Chapter 1, wherein a measure of LV rotational mechanics called LV twist is defined. LV twist is the apparent rotation of the LV apex relative to the LV base and provides insight into LV function beyond that traditionally reported in measures like ejection fraction. Importantly, cardiac MRI techniques are very well suited to evaluating cardiac structure and function and these methods are outlined in Chapter 2. In particular, MRI tagging can be used to non-invasively generate myocardial tissue landmarks that facilitate the qualitative and quantitative assessment of left ventricular (LV) myocardial deformation in both research and clinical settings. This thesis focuses on several developments. First, we developed a new tagging technique called complementary radial tagging (CRT) that generates a tagging pattern that better matches the annular shaped LV myocardium in the short-axis view. CRT also has better tag contrast during late diastolic phases and can be used to accurately measure the LV mechanics during all cardiac phases (Chapter 3).

In parallel with the development of the CRT technique we used conventional clinical tagging protocols to measure LV rotational mechanics quantitatively. Myocardial fibrosis is known to frequently occur in patients with Duchenne Muscular Dystrophy (DMD). The consequences of myocardial fibrosis on LV mechanics, however, are incompletely understood. In Chapter 4 we report on the LV rotational mechanics in patients with DMD with or without myocardial fibrosis. The results suggested that both DMD and the presence of myocardial fibrosis independently worsen LV rotational mechanics.

In Chapter 5 we report on the effects of conventional cardiac MRI exams, which require repeated breath holding and place a burden on some clinical patients, on measure of LV rotational mechanics. This breath hold paradigm presents two problems for patients with DMD who typically develop progressive respiratory impairment and the signs and symptoms of cardiac involvement at an early age. Currently, clinical protocols use both breath-hold and free breathing techniques, as needed, to acquire data. Chapter 5 compared the LV rotational mechanics between these techniques in healthy subjects and patients with DMD. It was found that free-breathing significantly decreases estimates of LV twist compared to breath hold measures. The results suggest that when using quantitative imaging biomarkers of LV rotational mechanics to monitoring disease progression or the response to therapy, especially in patients with DMD for whom decline in respiratory function is certain, it is important to use a free-breathing strategy for all studies to facilitate intra-subject longitudinal comparisons.

In Chapter 6, LV rotational mechanics were further evaluated in patients with mitral regurgitation. In patients with mitral regurgitation LV twist decreases, while CL-shear angle (an alternate measure of LV rotational mechanics) pseudo-normalized due to subtle changes in the heart’s geometry.

Lastly, in Chapter 7, a technique for acquiring two slices simultaneously – Controlled Aliasing In Parallel Imaging Results In Higher Acceleration (CAIPIRINHA) was implemented in a cardiac MRI tagging sequence. The application of CAIPIRINHA to a LV tagging sequence is shown to achieve similar estimates of peak LV twist in a single breath hold, which simplifies the exam and avoids measurement differences that may arise from data acquired in different breath holds.

In conclusion, this thesis reports on several technical developments and clinical applications related to estimating quantitatively the function of the left ventricle. Taken together these developments can be combined to provide fast and accurate estimates of LV rotational mechanics that provide insight to LV function beyond traditional measures.

During the past few years, the number of applications that need to process large-scale data has grown remarkably. The data driving these applications are often uncertain, as is the analysis, which often involves probabilistic models and statistical inference. Examples include sensor-based monitoring, information extraction, and online advertising. Such applications require probabilistic data analysis (PDA), which is a family of queries over data, uncertainties, and probabilistic models that involve relational operators from database literature, as well as inference operators from statistical machine learning (SML) literature. Prior to our work, probabilistic database research advocated an approach in which uncertainty is modeled by attaching probabilities to data items. However, such systems do not and cannot take advantage of the wealth of SML research, because they are unable to represent and reason the pervasive probabilistic correlations in the data.

In this thesis, we propose, build, and evaluate BayesStore, a probabilistic database system that natively supports SML models and various inference algorithms to perform advanced data analysis. This marriage of database and SML technologies creates a declarative and efficient probabilistic processing framework for applications dealing with large-scale uncertain data. We use sensor-based monitoring and information extraction over text as the two driving applications. Sensor network applications generate noisy sensor readings, on top of which a first-order Bayesian network model is used to capture the probability distribution. Information extraction applications generate uncertain entities from text using linear-chain conditional random fields. We explore a variety of research challenges, including extending the relational data model with probabilistic data and statistical models, efficiently implementing statistical inference algorithms in a database, defining relational operators (e.g., select, project, join) over probabilistic data and models, developing joint optimization of inference operators and the relational algebra, and devising novel query execution plans. The experimental results show: (1) statistical inference algorithms over probabilistic models can be efficiently implemented in the set-oriented programming framework in databases; (2) optimizations for query-driven SML inference lead to orders-of-magnitude speed-up on large corpora; and (3) using in-database SML methods to extract and query probabilistic information can significantly improve answer quality.