136879 research outputs found
Sort by
Vision-Language Models: The Overlooked Role of Synonyms
No full textIn the past few years, contrastively pre-trained Vision-Language Models (VLMs) such as CLIP, have significantly propelled advancements in multimodal applications. Pre-trained on internet scale images and captions VLMs learn to associate relevant text and images, making them key to downstream applications such as visual chatbots and text-to-image generation diffusion models. However, there has been limited analysis of these models��� pre-training datasets, primarily due to the challenges posed by their large scale. Moreover, given that these datasets lack human annotation, determining the presence of specific visual concepts within them poses a significant challenge.
We address this challenge by using a Large Language Model and count the pre-training texts that contain synonyms of any given visual concept. Contrary to popular belief we discover that these pre-training datasets exhibit a long-tailed concept distribution, resulting in biased performance in VLMs.
Next we propose a novel prompting strategy that leverages synonyms to improve the performance of VLMs for zero-shot recognition. Instead of prompting VLMs using original class names, we use the most frequent synonyms found in the pre-training texts. Finally we propose a novel light-weight retrieval augmented strategy, that achieves a new state-of-the-art for the zero-shot recognition
John Bickham field notebook: Herp_AK501-AK0599.pdf
No full textEach page/AK number corresponds to a karyotype slide data and/or unique specimen.Data pages for Herp_AK501-AK0599 corresponding to unique identifiers of specimens/samples examined for biological research. Specimens are primarily housed at Texas A&M University; Biodiverstiy Research and Teaching Collection
Computational Fluid Dynamics Analysis of the Blockage Accident in Wire-Wrapped Fuel Rod Bundles
No full textThe purpose for reducing CO2 emissions and enhancing the safety of nuclear reactors have led to increased interest in Liquid Metal Fast Reactors (LMFRs). These reactors offer high power density, low-pressure operation, and the ability to breed fissile material. However, LMFR fuel assemblies, comprising fuel pins enclosed in hexagonal ducts with wire-wrapped spacers, are susceptible to coolant flow blockages due to debris buildup, potentially leading to reduced heat transfer and fuel cladding damage. This PhD dissertation aims to conduct a comprehensive computational fluid dynamics (CFD) analysis of blockage accidents in wire-wrapped fuel rod bundles. The objectives include the preparation and validation of CFD models for both nominal (unblocked) conditions and various blockage scenarios, considering solid and porous blockages. Conjugate heat transfer modeling is also incorporated to simulate the cladding temperature. The proposed research activities encompass analyzing fluid flow behavior, pressure drop, velocity, turbulent structures, and temperature profiles for the different blockage configurations. Experimental data from wire-wrapped test facilities is used to validate the CFD models. These facilities have provided high-fidelity data of velocity and pressure drop for transition and turbulent flow regimes at nominal conditions, as well as for blockage scenarios with solid and porous obstructions. The CFD methodology involves solving the incompressible Navier-Stokes equations with the Reynolds Averaged Navier-Stokes (RANS) and Large Eddy Simulation (LES) methods. The results demonstrate the accuracy of the proposed methodology in predicting friction factors and velocity profiles in both unblocked and blocked bundles after the comparison with the experimental data. The findings reveal that the presence of blockages in wire-wrapped fuel rod bundles significantly impact the thermal-hydraulic performance of LMFRs. The analyses show that solid blockages cause an increase in pressure drop and a decrease in velocity, while porous blockages have a lesser impact. The turbulence analysis reveals that the blockages lead to the formation of vortices and eddies, which can further impact the flow behavior and heat transfer. This research yields valuable insights into blockage accidents and contribute to gain more reliability in the use of CFD models for safety assessments of LMFRs
Combustion Solutions for Reduced Methane Emissions from Large Bore Natural Gas Engines
No full textLegacy large bore, natural gas two-stroke engines form a vital component of the pipeline industry, but with increasing stringent pushes to reduce emissions, the necessity of improving performance of an ageing engine fleet grows more critical than ever. Precombustion chambers are frequently implemented on these engines to improve ignition stability and extend the lean limit of operation, a process which brings an additional benefit of reducing harmful emissions such as oxides of nitrogen (NOx) and hydrocarbons (HC). While prechambers reduce the carbon footprint from pipeline compressor stations, the pathway to zero emissions of the future still contains a plethora of research avenues to explore.
This study sought to explore two potential combustion solutions for reducing methane emissions from large bore natural gas engines. First, the Cooper-Bessemer GMV4 engine was fully simulated using Converge CFD and validated using experimental data. Before igniting the primary prechamber, radical and intermediate species were seeded throughout the main combustion chamber by use of a second, deliberately quenched prechamber. This served to boost reactivity and promote flame propagation throughout unburned regions of the chamber. Multiple temperature levels, injection timings, and chemical species compositions were investigated, for which each was then examined for early ignition limits, surviving concentration of seeded species, and overall impact on residual methane and the combustion process.
Second, the Cooper Ajax E-565 engine in open-chambered configuration was fully simulated using Converge CFD software and validated using experimental data. The open-chambered configuration was then modified to a prechambered configuration, and input parameters, such as fuel delivery and spark timing, were adjusted using best real-world design practices. This model was then used as a foundation upon which to evaluate the sensitivity of in-cylinder mixing between prechamber and main chamber gases to changes in intake manifold and port design. Eight different manifolds designs were created and analyzed for overall air flow, mixing quality, and general combustion performance. The results were then examined for an extensive investigation of factors preventing oxidation of residual methane as well as the production mechanisms of NOx emissions
Target Identification Using Single Cell RNA-Seq: Algorithms and Applications
No full textTarget identification is a crucial step in the drug development process, significantly affecting the success rate and efficiency of bringing new therapies to market. Recent advancements in single-cell RNA sequencing and computational tools have accelerated the identification and validation of therapeutic targets by enabling a deeper understanding of disease mechanisms at the cellular level. However, challenges such as inadequate understanding of the molecular basis of certain diseases and limitations in current single-cell data analysis methods, particularly in capturing gene regulatory relationships, continue to hinder the full exploitation of these technologies in precision medicine.
To this end, we aim to enhance target identification in scRNA-seq, crucial for unraveling cellular differentiation and disease mechanisms. Firstly, we develop ���scInTime���, a computational method that capitalizes on single-cell trajectory data and gene regulatory networks to accurately identify master regulators of cellular differentiation. This algorithm aims to overcome the existing challenges in mapping cell fate decisions, a critical step in advancing personalized medicine. Secondly, we propose to undertake an integrated scRNA-seq data analysis to investigate the association between pyroptosis and the severity of COVID-19. This research is expected to shed light on the immune response to SARS-CoV-2 and identify potential targets for therapeutic intervention. By focusing on the mechanisms underlying severe COVID-19 cases, we anticipate contributing to the global effort in combating the pandemic. Thirdly, we address metabolic diseases, specifically investigating the role of hepatocyte adenosine kinase in fat deposition and liver inflammation. Here, we aim to elucidate the molecular pathways that lead to excessive fat storage and inflammation in the liver, offering targets for the treatment of metabolic syndromes. Finally, we conduct a sex-based study on the role of RSPO3 in estrogen-mediated sex differences. Understanding the molecular bases of sex differences in diseases is critical for the development of gender-specific therapies and this study will contribute to that knowledge base.
Overall, our goal is to leverage scRNA-seq for precise target identification, addressing significant gaps in the understanding of cellular differentiation and disease. This thesis is designed to set the stage for a series of investigations that will collectively advance our knowledge in the field and lead to novel therapeutic strategies
An Augmented Machine Learning Approach for Real-Time Terramechanics Modeling
No full textThe purpose of this research was to develop a novel method based on Gaussian process models that can be combined with information from detailed models and field data, such that the model is capable of carrying out real-time simulation of the tire-terrain interaction. The aim of this approach is to be able to provide feedback on tire sinkage, slip, traction, and side forces to a virtual simulator for autonomous off-road vehicles. Such simulators are essential for training autonomous vehicles without the need for extensive field tests. The use of a simulator reduces the cost of autonomous vehicle training and allows for easy modification to the training/testing environment. The present approach is based using precomputed look-up tables, typically based on a Bekker terramechanics model, or a simple Coulomb friction model to determine tractive capabilities. While these provide quick information to be used in vehicle dynamics calculations, they lack the quality and quantity of information required for dynamic off-road conditions under widely different terrain conditions. Furthermore, the model is capable of accommodating disparate data sources with different ranges of validity and also take into account the uncertainties in the predictions.
In this thesis, the efficacy of the hybrid approach is demonstrated by combining a modified Bekker terramechanics model with finite element analysis (FEA) simulations using a Gaussian process regression (GPR) model that is capable of predicting tire forces and sinkage in off-road conditions. The Bekker-style model is valuable due to its simplicity, but it lacks dynamic loading information. While the FEA model provides valuable information about the full tire-terrain interaction it is computationally very expensive and unsuitable for real-time simulations. Therefore, a GPR model is trained to predict the difference between the steady-state information provided by the Bekker-style model and the full response provided by the FEA. Then the Bekker-style model and trained GPR model can be deployed in tandem to predict full response (forces) in a real-time applications such as off-road autonomous driving
Biogas Processing and Safety Challenges: Improving Process Components Functionality Through Reliability Engineering Approach
No full textTo achieve the objectives of carbon neutrality, energy security, and sustainability, biogas is increasingly used as fuel in the transportation and power generation sectors. Due to the support strategies for the reduction of greenhouse gases and air pollution implemented by several governments around the world, biogas production has tripled recently, leading to an increase in the number of related facilities. The forecast indicates a further increase in biogas facilities up to 2035. The number of incidents in the biogas industry has increased more than five times from 2001-2012; this trend outpaces the number of biogas facilities. The following categories of causes can be categorized (in order of quantitative relevance) in the analysis of accident causes and consequences: Failures of the entire piece of equipment, component failures, operational errors, design flaws, and maintenance errors.
The aim of this study is to address safety issues of biogas processing activities based on the process components��� failure analysis and reliability studies. Because of the scarce number of documented biogas processing components��� failure rates and the limited details available, biogas processing plants can benefit from the technologies already used for the processing of conventional natural gas, but due to the modest amounts of biogas produced by production and upgrading plants, the size of biogas processing plants must be substantially smaller. On this basis, this work sourced the existing reliability data from conventional natural gas components��� failure rates databases such as: Industrial database (e.g., Offshore and Onshore Reliability Equipment Database-OREDA), manufacturer���s field failure studies and end-user field failure studies. Selected biogas processing components��� unit(s) from active biogas piping and instrumentation diagram P&ID were utilized and data sourced from industrial database was used to perform series of failure analysis and reliability studies; results obtained from this study can be used to analyze the reliability of biogas processing units.
Enhancing the quality of failure data is critical to decreasing the uncertainties associated with biogas processing system reliability estimates to improve the system availability and maintainability and reduce the increasing number of accidents related to biogas processing. Consequentially, improve system availability and maintainability substantially, hence leading to improved productivit
Ultra-High Strain Rate Impact Behavior in High Molecular Weight Thermoplastics
No full textThe advent of commercial and military hypersonic vehicles introduces formidable challenges to existing thermal protection systems and ballistic armor. Concurrently, all spacecraft face escalating threats from hypervelocity impacts (HVIs) by micrometeoroids and orbital debris (MMOD). Overcoming these obstacles requires materials/structures capable of withstanding extreme conditions. Yet, a clear knowledge gap exists in the fundamental understanding of the complex multiphysics phenomena generated when materials are subjected to ultra-high strain rate (>10^6 s^ ���1 ) collisions. This is especially true for polymers, despite their capacity for energy absorption and tailorability. This dissertation endeavors to bridge this knowledge gap by exploring the ultra-high strain rate impact behavior of three high molecular weight thermoplastics, ultra-high molecular weight polyethylene (UHMWPE), high-density polyethylene (HDPE), and polycarbonate (PC). The primary objective is to help answer fundamental questions about their behavior and illuminate their applications. This goal necessitated a four-pronged strategy: (i) the establishment of an impact testing facility; (ii) the execution of HVI experiments; (iii) the application of numerical simulations to infer impacted target conditions; and (iv) the integration of these findings with existing literature to explain the role a polymer���s microstructure plays in its macroscopic energy absorption. The ensuing findings (i) revealed the difference in UHMWPE and HDPE���s HVI deformation, failure, and energy absorption was governed by an interplay between strain rates and rates of polymer chain disentanglement and reorientation and (ii) demonstrated the viability of threat-optimized protective structures. The dissertation also includes a novel study on similar impact phenomena, generated by decreasing spatial scale instead of increasing velocity. A Laser Induced Projectile Impact Test (LIPIT) apparatus and gas gun were used to launch alumina spheres ranging five orders of magnitude in diameter (3 ��m���10 mm) into scaled PC targets at 550 m/s. Length scale reduction set in motion a remarkable 230% amplification in specific energy absorption and a 240% increase in relative impact deformation area. In all, the dissertation leads to an irrefutable conclusion: while the behavior of polymers at lower rates (<10^3 s^ ���1 ) is well understood, there remains much to discover about their behavior as loading conditions become extreme. Data from lower rates and smaller scales cannot be extrapolated to higher rates and larger scales
Megahertz Rate Spectroscopic Investigation of Hypervelocity Impact Flash
No full textUnderstanding hypervelocity impacts (HVI) has become crucial to several fields, such as space exploration, planetary science, aerospace engineering, and defense-related applications. HVIs are highly dynamic and extreme phenomena characterized by the immense and rapid transfer of energy between colliding bodies. One manifestation of this transformation of kinetic energy during and immediately following an HVI is the emission of an intense flash of light. Researchers investigating these optical emissions emanating from HVIs refer to this phenomenon as the ���impact flash.��� The impact flash can be noticed across a wide range of impact scenarios: both high and low impact velocities; in the vacuum of outer space, as well as the atmosphere of the earth; with both solid and liquid objects; and across metallic, ceramic, and polymer materials. Investigations into the impact flash have sought to establish connections between impact conditions, temporal evolution of different spectroscopic features, and underlying material failure mechanisms. Characterization of this phenomenon can also act as an indicator for a spatial and temporal evolution of material damage. Typical studies of the impact flash make use of a diverse array of optical diagnostics tools and methods. These include high-speed cameras, photodiodes and photomultiplier tubes, pyrometers, laser interferometers, and spectrometers. Stop-motion or relatively low-rate spectral analysis systems have been employed to measure HVI-induced light emission, providing some spectral information related to the dynamic material behavior and projectile-to-target energy transfer. The relatively short time scales (microseconds) and rapidly evolving impacted material response characteristics of HVI events necessitate the use of highspeed detectors with superior temporal resolution to capture the impact flash evolution with sufficient temporal resolution.
Hence, the objective of this thesis research is to develop two high-speed transient spectral analysis systems for investigating the evolution of the impact flash. The first system, an ultrahigh-speed spectrometer (UHSS), has been developed to capture and analyze the transient light emission induced by HVIs in a two-stage light gas gun (2SLGG) facility, with sub-microsecond (i.e., megahertz-rate) temporal resolution. The UHSS makes use of a MHz-rate complementary metal��� oxide���semiconductor (CMOS) camera coupled to an imaging spectrometer, resulting in MHz-rate spectral imaging characterization. This system was used to record time-resolved spectral images of HVIs at speeds up to 6 km/s from spherical aluminum projectiles impacting both aluminum and stainless-steel targets. The high-speed spectral images recorded allowed us to compare the evolutionary behavior of energized metallic species and combustion byproducts in the nano- and microsecond time scales following the impacts. This study found that even at similar impact conditions, the light emission behavior of aluminum-on-aluminum and aluminum-on-stainless-steel impacts can be substantially different. Aluminum target impacts showed one temporal peak in light intensity, whereas the stainless-steel targets showed an additional secondary peak of emitted light. Additionally, in stainless-steel impacts, combustion byproducts were seen to feature a stronger second peak than pure metallic species.
The second high-speed spectral analysis system, a fiber-based multi-spectral diagnostics (FMSD) system, makes use of a fiber bundle to direct light from the impact flash into a collection of highspeed silicon photodiodes. A gigahertz-rate oscilloscope receives the output voltage from these photodiodes as light reaches each photodiode. By placing unique spectral filters in front of each photodiode, these photodiodes are used to capture signals from a specific species. The design of the FMSD was validated using laser-induced plasma experiments. Engineering improvements and subsequent implementation in future HVI studies are discussed
John Bickham field notebook: AK3001-AK3500.pdf
No full textBound book, each page corresponds to a karyotype slide data.Data pages for AK3501-AK4000 corresponding to unique identifiers of specimens/samples examined for biological research. Specimens are primarily housed at Texas A&M University; Biodiverstiy Research and Teaching Collection