DSpace@RPI (Rensselaer Polytechnic Institute)
Not a member yet
6809 research outputs found
Sort by
End-to-End Table Question Answering via Retrieval-Augmented Generation
Get PDFMost existing end-to-end Table Question Answering (Table QA) models consist of a two-stage framework with a retriever to select relevant table candidates from a corpus and a reader to locate the correct answers from table candidates. Even though the accuracy of the reader models is significantly improved with the recent transformer-based approaches, the overall performance of such frameworks still suffers from the poor accuracy of using traditional information retrieval techniques as retrievers. To alleviate this problem, we introduce T-RAG, an end-to-end Table QA model, where a non-parametric dense vector index is fine-tuned jointly with BART, a parametric sequence-to-sequence model to generate answer tokens. Given any natural language question, T-RAG utilizes a unified pipeline to automatically search through a table corpus to directly locate the correct answer from the table cells. We apply T-RAG to recent open-domain Table QA benchmarks and demonstrate that the fine-tuned T-RAG model is able to achieve state-of-the-art performance in both the end-to-end Table QA and the table retrieval tasks
Damage accumulation and failure in stochastic fibrous materials
No full textDecember 2019School of EngineeringDamage accumulation and failure in random fiber networks is of importance in a variety of applications, from design of synthetic materials, such as paper and non-wovens, to accidental tearing of biological tissues. In this work we study these processes using three-dimensional models of athermal, crosslinked fiber networks, focusing attention on the modes of failure and on the relationship between network strength and network structural parameters. We consider network failure at small and large strains associated with the rupture of inter-fiber bonds. It is observed that the strength increases linearly with the bond number density, with the average distance between the bonds, and with the bond strength. Rendering the bond strength stochastic causes a reduction of the network strength. However, heterogeneity retards damage localization and increases the stretch at peak stress, therefore promoting ductility. Network strength, in general, is found to be independent of fiber material properties and fiber tortuosity. Random fiber networks, due to their inherent structural heterogeneity exhibit size effect in their strengths and we find that network strength follows Weibull statistics. We characterize the behavior and strength of random networks composed of fibers with non-circular cross-sections. Such fibers are characterized by two bending modes along different axes. For such networks, the torsional stiffness of the fibers controls the relative contribution of the two bending modes to the network stiffness at small strains. The presence of an additional bending mode does not affect the network deformation at large strains and the fiber cross-section, in general, does not affect the network strength. Using the structure-property relationships established in this work, we design a new class of materials, called the Non-Convex Voronoi networks. The Non-Convex Voronoi networks are more compliant and exhibit higher strength, rendering such networks of interest in a variety of applications, such as artificial tendons and ligaments, protective clothing etc. Finally, we also analyze network failure under multiaxial loading conditions and attempt to develop suitable failure criterion to predict network failure under generalized loading conditions, when inter-fiber bond breakage is the primary failure mechanism. The results established in this thesis can be used to design fiber networks of a specified strength and, in general, enhance the understanding of mechanical behavior of fibrous materials.Ph
High-order accurate partitioned schemes for conjugate heat transfer with advection-diffusion equations
No full textJuly 2022School of ScienceThis thesis presents a high-order accurate partitioned scheme for the solution of conjugate heat transfer (CHT) problems. The scheme is based on special second-order accurate CHAMP (Conjugate Heat transfer Advanced Multi-domain Partitioned) interface conditions and extended here to high-order accuracy and to the advection-diffusion equations. The solutions in each material domain are advanced independently with an implicit method, and domains are coupled at the interface using the new CHAMP interface conditions. These conditions are based on usual interface matching conditions involving continuity of temperature and heat flux, together with additional compatibility conditions derived from the governing equations. The new CHAMP conditions are implemented numerically using an optimized Schwarz approach, with a Taylor expansion leading to an effective domain overlap, which significantly improves the convergence rate. The partitioned time-stepping schemes are found to be stable with none or just a few sub-time-step iterations for a wide class of CHT problems. The first part of this thesis focuses on the conjugate heat transfer problem for the diffusion equations and extends the current second-order accurate CHAMP schemes to higher-order accuracy. A detailed fourth-order accurate derivation is given to demonstrate the approach, with a general discussion on deriving a pth-order accurate method. The scheme is then analyzed to determine the optimal coupling coefficients in the CHAMP condition based on solving an optimization problem. The CHAMP iteration is studied in isolation while keeping the time-step fixed, and the iteration amplification factor of the CHAMP scheme is compared to the optimized Schwarz scheme with different overlap widths. The CHAMP time-stepping scheme is then studied for the case where no sub-iteration is taken. The un-iterated CHAMP time-stepping scheme is analyzed to show the overall fourth-order accuracy when a fourth-order accurate CHAMP interface condition is applied. The CHAMP condition is also derived for general curvilinear grids. Numerical results using manufactured solutions on curvilinear grids are given showing the accuracy and stability of the un-iterated CHAMP schemes. To solve problems with large time steps, an adaptive variable sub-iteration CHAMP algorithm is proposed. The new algorithm chooses the number of sub-iterations adaptively based on a measure of the residual for the CHAMP conditions at each time step. A large time step study using the new adaptive algorithm is given to show the robustness of the method. The second part of this thesis generalizes the approach to solving the conjugate heat transfer problems with advection-diffusion equations, also to higher-order. Involving the advective terms in the governing equations and interface conditions complicates the analysis. The pth-order accurate method is derived first, followed by a presentation of the complete CHAMP time-stepping algorithm. A detailed second-order accurate analysis is presented to describe the approach, with a general discussion on analyzing a pth-order accurate method. The CHAMP iteration amplification factor for the advection-diffusion equations is computed numerically using some numerical software packages. The coupling parameters in the CHAMP condition are calculated by solving an optimization problem based on the convergence factor of the sub-iterations. The iteration amplification factors for the CHAMP schemes using the optimal coupling parameters with different orders of accuracy are presented and compared. The convergence factor of a pth-order accurate un-iterated CHAMP time-stepping scheme is also derived and analyzed. Finally, the accuracy of the scheme is verified using several numerical examples.Ph
Augmenting collaborative immersive systems through cognitive technologies for educational scenarios
No full textAugust 2022School of ArchitectureVirtual reality continues to proliferate in the classroom. Human-scale immersive virtual environments provide natural benefits and assuage many of the issues raised with head-mounted displays. While a headset removes a user from his or her physical environment, these spaces allow users to experience virtual realities in groups, maintaining authentic communication between peers. This work contributes the creation and utilization of necessary cognitive technologies and services for educational scenarios as well as conducts the first controlled quantitative study on learning within these environments this author is aware of. A believed benefit of learning within these environments is the ability to enter virtual spaces without abandoning physical and social connections. In a successful educational environment, students and instructors must be able simultaneously to see content on the screen and each other. A designed intelligent, steerable lighting system illuminates the workspace for the task at hand while reducing illumination at the screen by over 97 %. Students and instructors can affect the system from their personal devices, while a network layer maintains input from spatial tracking and user-developed systems. This combination of experienced simplicity and retained flexibility for a lighting system within a human-scale immersive environment is an important contribution to educational scenarios. The irregular geometry of the CRAIVE-Lab screen introduces distortion to panoramic imagery, which is counteractable via a geometric transformation. Whereas for traditional cube-shaped CAVE systems, this transformation is linear and can be described with simple calculations, that required for the CRAIVE-Lab is more complex due to the nonlinearity introduced by the rounded corners. This work determines the mathematically-derived solution for describing this necessary transformation. This generated image warping solution is adaptable to other environments of irregular projections. It is executed via a Python script and performed for users without needing any additional software or files or knowledge of the underlying mathematics. The script does so 100 times faster than the previous method, in a fraction of a second, and in a live classroom setting. To generate spatialized audio for headsets, developers are presented with APIs. However the application of this model to a human-scale immersive environment is a necessary development. The produced Spatial Audio Worker changes the way developers and users consider the environments’ audio arrays. It presents developers with standardized inputs to the loudspeaker arrays, requiring no knowledge of underlying hardware or software. Developers of educational applications (and beyond) can spatialize audio and create congruent content with a few lines of familiar code. This method has been adopted by multiple applications, which now require no tailoring to be used across multiple arrays and environments. Many internet ``cloud’’ services enable new functions and accelerate content creation. A produced navigation interface utilizes the Street View Static API to retrieve and properly format for the screen the necessary imagery for panoramic scene generation. This navigator, the only such known for human-scale immersive environments, instantly transports classes to locations around the globe. This is combined with additional cloud services such as natural language processing and object detection for scenarios such as the Language Learning Environment.Additionally, as these environments proliferate, their limited available time is most valuable for final use cases and courses. Therefore, a virtual testing environment is produced which can be experienced at scale in a head-mounted display. This testing environment enables parallel and remote testing while keeping the physical environment free for educational purposes, providing instructors and students alike with the first known headset-based testing application of a human-scale environment. In a holistic presentation of these cognitive technologies and services, the Language Learning Environment at Rensselaer transports students to virtual- and real-world scenes in which to practice their foreign language skills. This experience is made possible directly through the discussed cognitive technologies and services. It is assessed in both a qualitative and quantitative study. The quantitative study is believed to be the first of its kind. Results indicate that learning in the environment is effective, knowledge is retained over time, and students rate the experience as enjoyable and engaging. This work has profoundly altered interaction with the environment. The ability for parallel contributions of content by students and instructors to a visual and aural human-scale immersive environment via personal devices for the purposes of educational scenarios is a contribution of this work. Users engage directly with the panoramic display via the Pin-Up application, and tasks that previously consumed large portions of class are now completed quickly, asynchronously, and remotely. Multiple novel courses are held in the environments, including Aural Architecture taught by Jonas Braasch in the Acoustics Department and Nuclear Phenomena for Engineering Applications taught by Emily Liu in the Department of Mechanical, Aerospace, & Nuclear Engineering. The wide variety of use cases which have already adopted these cognitive technologies and services indicate much more future research is available for pursuit.Ph
Multi-scale computational modeling of coupled chemical, physical and mechanical phenomena in cementitious materials
No full textAugust 2021School of EngineeringWith the increasingly popular use of cement-based materials like concrete in buildings and infrastructure, it has become crucially important to reliably predict the aging and deterioration of cement-based materials. One of the major phenomena that cause this deterioration is the ingress of corrosive agents into concrete materials, e.g., chloride penetration and sulfate attack. This is a delicate yet challenging task to tackle because of the following nature of the problem to be considered (1) the mass transport in concrete is a phenomenon related to various length scales. First, the microstructure of the cement-based material significantly affects the penetration process. Second, the corrosion induced degradation initiates at the mesoscale level. Finally, the overall transport problem needs to be solved at a macro-scale level to consider the internal and external conditions like temperature, relative humidity, and fracture. (2) The cementitious matrix is a chemical compound with ongoing hydration and membrane-like properties. These features make the mass transport involved in multiple chemical and physical reactions. (3) Evolution of mechanical properties including strength gain due to hydration and strength loss due to degradation, plays an important role in the formation of fracture in cement-based materials and thus in turn, affecting the mass transport in concrete. In an attempt to address the problem described above, a comprehensive multi-scale computational framework is proposed in this research and is applied to the mass transport problem in cementitious materials. The research covers four main projects. In the first project, a coupled multi-physics model is proposed to simulate chloride penetration in saturated and unsaturated concrete. This transport model considers different chemical and physical phenomena encountered in the diffusion process of a concentrated solution, and meanwhile characterizing the effect of microstructure as well as the mesoscale tortuosity. In the second project, moisture diffusion is taken as an example to study how the existence of cracks affects the mass transport. The water transport in cracked concrete is successfully simulated with the modified diffusivity and the consideration of disequilibrium condition between adsorption and desorption. The first two projects are both developed at mesoscale. To accurately characterize the effect of underlying mechanism at the microscale, a chemo-mechanical model, μLDPM, is developed in the third project by making use of the advantages of two state-of-the-art models: (1) microscale chemical reactions modeled by using μic, a vector-based model that simulates the formation of reaction products around idealized spherical particles, and (2) microstructural elastic and damage behavior modeled by a variant of the Lattice Discrete Particle Model (LDPM) known for its wide successes in modeling cementitious materials failure under various stress states. In the last project, a multi-scale coupling method is proposed to connect the microscale model and the transport models proposed in the first two projects. With this scheme, the values of parameters that depend on underlying microscale phenomena can be directly identified once the mixture is provided, thus avoiding empirical estimation due to lack of sufficient experimental data.Ph
Hempwerks : non-corroding concrete reinforcing made with natural fiber and thermoplastics
No full textAugust 2021School of ArchitectureInfrastructure has been deteriorating for decades and climate change and urbanization are acceleratingtheir degradation. Much of our infrastructure is built with reinforced concrete, whose steel
reinforcement can become a liability if allowed to corrode. This thesis explores the development of a
non-corroding reinforcing alternative made from natural fibers and thermoplastic. The impact of
natural fiber composite (NFC) reinforced concrete has been investigated at multiple scales of
performance: structural, processing, and environmental. Structural and environmental performance has
been calculated based upon material database values, while processing performance has been observed
though the production of multiple NFC samples. The structural performance calculations reveal that
flax reinforced composites can match the tensile strength of steel with a fiber volume ratio between
44% and 50% and match the elastic modulus of GFRP with a fiber volume ratio between 46% and
49%. The processing performance experiments reveal the “jacket” commingling method results in
better fiber saturation than the “parallel” or “twisted” methods. The environmental performance results
reveal that given a constant fiber volume ratio, PLA matrix composites demand only 50%-51% the
embodied energy of GFRP, regardless of the selected natural fiber reinforcing.M
Non-destructive platform for morphologic and cell density characterization of developing and drugged multicellular tumor spheroids
No full textMay 2022School of EngineeringIn vitro tumor models are important tools for exploring cancer progression and novel therapeutic strategies. These models are often oversimplified and lack many features of the 3D tumor environment, limiting their ability to simulate in vivo tumor behavior. Recapitulating key features of the tumor microenvironment, particularly 3D structure, is critical for the next generation of in vitro cancer diagnostic tools. Emerging 3D tissue-engineered in vitro models, such as multicellular tumor spheroids (MCTSs) and organoids, are a promising solution to this challenge. These cell aggregates have the ability to mimic several key aspects of in vivo tumors, such as 3D structure and pathophysiological gradients, and possess great promise for high-throughput in vitro testing. However, lack of standardization of MCTS fabrication has led to a large spectrum of “spheroid” and ”organoid” models, few of which are able to replicate the necessary sphericity for physiologically-representative behavior. This discrepancy is thought to be a major contributor to the high failure rate in drug discovery, where only a low percentage of drugs investigated in vitro succeed in clinical trials. As an additional challenge, the required size and shape of these in vitro tumor models precludes them from conventional microscopy, thus limiting our ability to use these models to collect data. Advancement of these MCTS models relies heavily on our ability to characterize and assess them. Herein, I propose an image-based analytical tool capable of characterizing mesoscopic (∼ 3mm thick) samples. Based on Optical Coherence Tomography (OCT), a structural imaging modality, this approach will be non-destructive and maintain cell-scale resolution while also boasting a field of view sufficient for full aggregate viewing. Using this tool, we can characterize morphology, cellular density, and cell viability of MCTSs, providing key longitudinal information on tumor model development and response to drug. The non-destructive nature of this technique uniquely positions us for longitudinal investigations within singular aggregates. One such application we sought to explore is the acquired resistance of cancer cells to pharmacologic anti-cancer treatments, a complex issue that can add substantial challenges to treatment and eradication of solid tumors. Recentstudies have shown that following initial treatment, indications of sustained health and increased aggressiveness have been observed in cells, which may serve to promote cell proliferation/migration rather than eliminating the cells as intended. Utilizing OCT/Imaris to assess aggregates made from cells that have been exposed to a drug and later treated again with the same drug, we can non-destructively assess changes in cell density/viability in response to different priming concentrations/drugs. The goal of this study is to establish a platform for studying the effects of drug-exposure on subsequent drug response. Overall, this research developed a non-destructive tool for analysis of tumor spheroid morphology, cell density, and cell viability within dense cellular aggregates. When applied to aggregates that were primed and subsequently treated with the same drug, the tool provided evidence of cell density spikes/dips that mirrored clinical acquired resistance observations. In summary, this tool holds great promise for advancing multicellular tumor spheroid use, especially toward drug screening applications.Ph
Designing a strong test for measuring true common-sense reasoning
No full textCommon-sense reasoning has recently emerged as an important test for artificial general intelligence, especially given the much-publicized successes of language representation models such as T5, BERT and GPT-3. Currently, typical benchmarks involve question answering tasks, but to test the full complexity of common-sense reasoning, more comprehensive evaluation methods that are grounded in theory should be developed
Studying tertiary structural transitions of tRNALys3 and Arf1 protein with pressure
No full textMay 2022School of ScienceRNA is believed to be one of the original molecular components present in the beginning of life about 3.5 billion years ago. RNA structure comprises both double helical secondary structures involving canonical nucleic acid base pairing and stacking, as well as tertiary structures in which more sequence distant interactions fold the molecules into a more compact structure. The molecular determinants of folded RNA structures and the pathways of their folding remain poorly understood. First half of this dissertation focuses on the effects of pressure perturbation on structural transitions of RNA molecules. According to Le Chatelier’s principle, pressure shifts (bio)chemical equilibria towards states of lower molar volume. The differences in volume between biomolecular conformations can arise from differences in packing and cavities, differences in interaction with solvent molecules, including electrostriction for charged molecules such as RNA. The effect of pressure on the structure of a folded RNA molecule, tRNALys3, was investigated using a combination of Nuclear Magnetic Resonance (NMR) and Small Angle X-ray Scattering (SAXS). Pressure leads to significant perturbation of the local interactions that define the tRNA tertiary structure, with little effect on secondary structure. Moreover, the global shape of the molecule was not strongly affected by pressure. The effects of divalent cation and water activity on these transitions were explored as well. In the second half of this dissertation, study model has changed from RNA to protein that we studied on one of the small GTPases - ADP ribosylation factors (Arfs). Arfs function in regulation of vesicular transport with lipids and protein trafficking in eukaryotic cells. GTPase activity of Arfs is activated by Guanine nucleotide Exchange Factors (GEFs); then inactivated upon binding GTPase Activating Proteins (GAPs). As a GDP/GTP switch, massive conformational differences between the GDP- and GTP-bound forms are observed in the N-terminal and switch regions of Arfs. In addition, previous studies suggest the N-terminal helix controls the conformational transitions of GDP/GTP switch that involves local unfolding. Using pressure perturbation coupled with NMR and SAXS, we have characterized the excited conformational states likely populated during the switch transitions. In addition, we have investigated the effects of GDP and Mg2+ on the Arf1 conformational landscape. These results helped to map a plausible GDP/GTP switch pathway and better understand the allosteric mechanism of Arfs family of small GTPases.Ph
Loss and modulation of bone matrix and fragility during diabetes mellitus
No full textMay 2022School of EngineeringBone is a complex hierarchical material that depends on both its quantity and quality to provide the mechanical, endocrinological and immunological functions in vivo. As such, deleterious conditions leading to impaired bone quality and quantity stemming from age or associated disease states have a severe impact on individual health. Diabetes mellitus (both the type 1 and 2 variants) is a disease state in which there is an increase in skeletal fragility. Notably, bone mineral density (BMD) a marker for clinical diagnosis of osteoporosis and incorporated into the fracture risk assessment tool (FRAX) to predict risk, is unchanged in type 2 diabetes mellitus (T2D). Consequently, in the absence of decreased bone quantity, the quality and composition of the bone matrix may explain the poorly understood mechanisms underlying diabetic skeletal fragility. The three major constituents of bone are the mineral phase (hydroxyapatite), the organic phase (collagen, non-collagenous proteins) and water. As such, the loss and/or modifications in these submicron components may explain and contribute to our understanding of the causes of skeletal fragility. Type 1 diabetes mellitus (T1D) and T2D, an autoimmune disease and metabolic disorder respectively, are both characterized by hyperglycemia. Arising from different etiology, determining the variation in the alterations in the organic and inorganic phases on bone which contribute to strength is integral to understanding diabetic fracture. To this end, three animal models reflecting the T1D, T2D and the effect of select therapeutics were investigated here: (a) A non-obese transgenic murine model of T2D provides a vehicle of assessing diabetic skeletal fragility in the absence of the confounding influence of obesity; (b) a drug-induced rat model of T1D treated with a sodium-glucose cotransporter-2 inhibitor provides a mean to understand the role of stabilizing glycemic control on the bone structure and matrix composition with T1D; and (c) an ovariectomized rat model with both metformin and strength exercise regimens yields insight into the effect of common diabetic interventions on bone matrix quality. Each of these models was evaluated for skeletal structure, fracture resistance and matrix composition through mechanical (strength and toughness), biochemical protocols (“in bulk” fluorescent advanced glycation end-products (AGEs)), and imaging techniques (microcomputed tomography, small-angle x-ray scattering, confocal Raman spectroscopy).
To applying the findings observed in the animal models of diabetes to the human diabetic condition, a retrospective cohort study was performed on a T1D population. Extracted from the OptumLabs Data Warehouse, a T1D population was garnered to assess the efficacy of longitudinal HbA1c as means of assessing fracture risk when accounting for various covariates, including medications, comorbidities, and patient demographics. To this end, univariate Kaplan-Meier survival modeling was used to estimate fracture risk and multivariate Cox proportional hazards modeling was used to assess the independent hazard ratios of longitudinal HbA1c and medications when adjusting for the covariates. Notably, elevated longitudinal HbA1c was associated with a significant increase in T1D fracture risk, though it was not an independent predictor of fracture risk when accounting for other covariates. Furthermore, metformin and bisphosphonates, both medications affecting bone turnover, were independent predictors of T1D fracture, imparting beneficial and deleterious effects, respectively. The findings from human T1D cohort are consistent with the rodent findings describing T1D as skeletal fragility in T1D occurred through changes in bone quantity rather than the changes in bone quality observed in the rodent models of T2D.
The evaluation of these models illuminated variations in the impact of the disease states on the bone quality and structure. The non-obese murine model revealed altered mineralization correlating with glycoxidation products leading to reduced resistance to fracture within the bone. The rat model of T1D treated with SGLT-2 inhibitor demonstrated that structural changes observed with T1D occur despite reducing diabetic hyperglycemia and that the reduction in blood glucose improves bone strength and alters the accumulation of AGEs without impacting the mineral phase. The rodent model treated with metformin and exercise saw the greatest improvement in fracture resistance with dual intervention when compared to other ovariectomized groups. Ovariectomy provided a more significant change in matrix turnover than any intervention groups without causing changes in bone mineral or modifications of collagen by AGEs – possibly due to changes in non-collagenous proteins or enzymatic processes – requiring further investigation. The combination of these studies provides insight into how the diabetic condition and ovariectomy impacts bone health leading to the observed fragility characteristics. Identifying aspects of bone quality affected by the disease state creates an avenue to evaluate the efficacy and potential targets for therapeutic interventions.Ph