Dissertation Defense Announcements

This qualitative study investigates the perceptions of high school-level alternative school administrators in North Carolina about the impact of performance standards (ESSA) on them and their campuses. The study interviewed four alternative high school administrators who are currently serving in alternative schools in North Carolina that were labeled Comprehensive Support and Improvement during the 2018-2019 school year due to low graduation rate and/or low performance. Analysis of the data resulted in the following findings: Alternative high school administrators perceive the ESSA guidelines for school performance as unfair and inequitable to alternative schools, and the guidelines demonstrate a lack of understanding of alternative schools and alternative education on the part of those who develop accountability guidelines.
The impact of these guidelines has resulted in changes to some practices on their campuses and receiving additional funding. Administrators perceive these accountability standards have no impact on their professional career but have increased their stress. Findings indicate that alternative school administrators perceive that the people who assist and are involved in the lives of students, their academic and social-emotional interventions, appear to be the most important strategies that can lead to successful outcomes for their students. Conclusions include a need for more awareness of the differences between traditional and alternative schools, the students served on these campuses, and more awareness of equity in education, specifically accountability and school performance. 

Deep Learning (DL) has been widely applied in both academia and industry. System innovations can continue to squeeze more efficiency out of modern hardware. Existing systems such as TensorFlow, MXNet, and PyTorch have emerged to assist researchers to train their models on a large scale. However, obtaining performant execution for different DL jobs on heterogeneous hardware platforms is notoriously difficult. We found that current solutions show relatively low scalability and inefficiencies when training neural networks on heterogeneous clusters due to stragglers and low resource utilization. Furthermore, existing strategies either require significant engineering efforts in developing hardware-specific optimization methods or result in suboptimal parallelization performance. This thesis discusses our efforts to build an efficient and scalable deep learning system when training DL jobs in heterogeneous environments. The goal of a scalable learning system is to pursue a parallel computing framework with (1) efficient parameter synchronization approaches; (2) efficient resource management techniques; (3) scalable data and model parallelism in heterogeneous environments;

In this thesis, we implement robust synchronization, efficient resource provisioning approaches, asynchronous collective communication operators, which optimize the popular learning frameworks to achieve efficient and scalable DL training. First, to avoid the "long-tail effects" for parallel tasks, we design a decentralized, relaxed, and randomized sampling approach to implement partial AllReduce operation to synchronize DL models. Second, to improve GPU memory utilization, we implement an efficient GPU memory management scheme for training nonlinear DNNs by adopting graph analysis and exploiting the layered dependency structures. Third, to train wider and deeper Deep Learning Recommendation Models (DLRMs) in heterogeneous environments, we propose an efficient collective communication operator to support hybrid embedding table placements on heterogeneous resources and a more fine-grained pipeline execution scheme to improve parallel training throughput by overlapping the communication with computation. We implement the proposed methods in several open-source learning frameworks and evaluate their performance in physical clusters with various practical DL benchmarks.

The coal-fired steam electricity plants are interested in finding efficient ways to manage by-products from the combustion process out of environmental and regulatory considerations. As one of the major solid by-products, the coal fly ash (CFA) is required by the Coal Combustion Residuals (CCRs) rules to be disposed of in an engineered landfill to protect groundwater. While the disposal of the CFA in the landfill needs water for moisture conditioning and dust control measures, it is convenient to use liquid by-products as alternative moisture sources. The concentrates (brines) generated from the volume reduction of the flue gas desulfurization (FGD) wastewater, such as reverses osmosis and evaporation treatment, can be an alternative liquid source to achieve zero-liquid-discharge (ZLD) for its economic benefit and environmental responsibility considerations. It is reasonable to investigate the potential methods to co-dispose the CFA and FGD brine in the ash landfill. In this study, chloride was the dominant anion with a significant presence of sulfate and bromide in the hypersaline FGD brine, and the cations were mainly calcium and magnesium. The class F CFA used in this study was acquired from an electric plant in the southeast U.S. and did not possess cementitious properties needed for stabilization/solidification (S/S) of co-disposal material. Methods investigated were the co-disposal through compaction and paste encapsulation technology. Instrumented testbeds with leachate and runoff collection systems for each co-disposal method were used to study their field behaviors under the weather conditions of Charlotte, North Carolina.
The chemical analysis of leachate and runoff samples from the compacted testbed found that the method released 79.1% of chloride and 88.6% of bromide in added FGD brine due to the absence of solidification/stabilization of the material. While the electrical conductivity (EC) was used as an indicator of the pore solution’s salinity, the chemical compositions of the fluid could vary as observed in the shifting of dominant anion from chloride to sulfate in the leachate. This study established a set of empirical equations to translate the permittivity to volumetric water content (VWC) for the pore solution's of a known EC (salinity). The low intensity, high-frequency precipitation provided high infiltration during the winter, resulting leachate generation by the testbed with a little amount of runoff. In contrast, the summer's high intensity, low-frequency rains resulted in a high runoff with little infiltration, coupled with extensive evaporation, causing a pause in leachate generation.
Compared to the compacted method, the paste encapsulation method successfully sequestered the halides mainly through the solidification process of the material, as indicated by the leaching test. Further analysis of the chemical composition of inner and annulus leachate coupled with the low hydraulic conductivity (1.44×10-8) of parallelly tested laboratory samples and the negligible leachate volume collected from the inner section of the leachate collection system suggests the leachate collected in the annulus section originated from side leakage. The chemical analysis of leachate and runoff showed on average 80% of retention of chloride and bromide during the experiment period and 97% retention if the side leakage could have been eliminated. The relatively impermeable paste suggests storm management of a paste landfill should expect runoff quantity approximately equivalent to the local precipitation. The surface temperature of the paste was elevated by solar radiation during the summer, which indicated the paste landfill could serve as a heat source that could impact the local microclimate. The mineralogy study of different samples showed the formation of poorly structured minerals which caused interpretation challenges of XRD results. The anticipated halide stabilization pathway through the precipitation of Fridel’s salt and Kuzel’s salt was complicated by the significant presence of magnesium in the brine.
Although the compacted method failed to retain halides under current weather conditions (Charlotte, NC, US), it does not necessarily disqualify its use in different environments. Therefore, a physics-based COMSOL-MATLAB (CM) model was established to simulate the field behaviors of the compacted co-disposal material, which was validated with the field data. The CM model consisted of three main components: heat transfer (HT), unsaturated flow (UF), and solute transport (ST) processes. The model also simulated the runoff, evaporation, and solar heating at the surface of the testbed. The CM model could appropriately reproduce the field leachate/runoff generation pattern, moisture content variation, temperature profile, and the change of chloride and bromide concentrations in the leachate during the washoff stage. The accuracy of simulation results could be improved with a better estimation of the conditions on the testbed surface.
While the validated physics-based model could be used to explore potential management methods for the compacted landfill and its behaviors under different weather conditions, the abundance of data spurred the interest in developing data-driven models. Since the bulk dielectric permittivity, which could be translated into VWC, was the measured property, a data-driven model simulating the change of permittivity in the compacted testbed was developed. The data-driven model was structured as three layers of material stacked in spatial order to address the standard operation of implementing new layers on top of old materials during landfill operation. With a forecast interval of 24 hours, the prediction over time of three years had an average R2 of 97.6% with the data-driven model trained with the first-year data, and R2 of 99.5% if two years of data were used in the training. The scenario studies showed that the data-driven model could only accurately predict permittivity values included in the training dataset, which indicates that a failure to predict could happen when unprecedented values occurred. Further investigation showed the data-driven model could simulate processes that would have conventionally required additional physics-related information through unique pattern recognition in the training dataset.

Due to its reduced dimensionality, monolayer molybdenum disulfide (MoS2) exhibits many unique optical properties, making it an excellent candidate for future optoelectronic devices. Given the multitude of applications, understanding the optical limitations of MoS2 under intense excitation is essential to optimize its performance. To that end, we investigate the femtosecond laser-induced breakdown of monolayer MoS2 with a variety of techniques. In this study, the substrate is discovered to have a profound effect where the ablation threshold itself can vary by more than one order of magnitude due to a simple interference phenomenon within the monolayer. Via substrate engineering, the ablation threshold can be reduced such that laser patterning using pulse energies less than 100 pJ is possible. Similar to many other optical nonlinearities, absorption measurements and theoretical modeling reveal that avalanche ionization is also enhanced where more than 75% of the generated free carriers at breakdown are due to avalanche ionization alone. Finally, multi-shot studies demonstrate that MoS2 is one of the most optically robust materials with very weak incubation effects. Notably, the onset of optical damage results in the formation of nano-voids where clusters of atoms are removed while the overall integrity of the monolayer remains intact. All these findings help establish MoS2 as a promising candidate for strong field devices and provides foundational knowledge regarding the strong field physics of two-dimensional materials.