48670 research outputs found
Sort by
Multiscale Models Of Interfacial Mechanics In Low Dimensional Systems
Get PDFCrucial thrusts in modern technology from electronic information processing to engineering cellular systems require manipulation and control of materials on smaller and smaller scales to succeed. A simple and successful way to break conventional material property limitations or design multifunctional devices is to interface two different materials together. At small length scales, the surface to bulk ratio of each component material increases, to the point that the interfacial physics can dominate the properties of the engineered system. Simultaneously, the combinatorial space of possible interfaces between materials and/or molecules is far too vast to explore by trial-and-error experimentation alone. Intuitive theoretical models can greatly improve our ability to navigate such large search spaces by providing insight on how two materials are likely to interact. The goal of this thesis is to develop predictive physical models which explain emergent phenomena at material interfaces across multiple length and time scales. A variety of state-of-the-art tools were applied to realize this goal, including analytical mathematics, quantum mechanical simulations, finite element methods, and deep neural networks. At the electron scale, a continuum model parametrized by first-principles simulations was employed to develop design criteria for confined quantum states in lateral heterostructures of two-dimensional materials. At the atomic scale, a chemo-mechanical model incorporating long-range electrostatics was developed to explain synthesizability trends in composite heterostructures of inorganic perovskites and organic molecules. A machine learning graph neural network model was developed and applied to predict the impact of general surface strains on the adsorption energy of small molecule intermediates on catalyst surfaces. Finally, at the microscale, a nonlinear kinetic model was developed to explain how cells acquire and retain memory of the mechanical properties of their surroundings across multiple timescales, which can lead to irreversible adaptation and differentiation. The methods and results presented in this thesis can improve our understanding of physical phenomena arising at interfaces and provide a blueprint for future applications of multiscale computational modeling to science and engineering problems
High Throughput Immunospecific Detection And Analysis On Subcellular Nanomaterials At Single Particle Level
No full textExtracellular vesicles (EVs) have shown great potential in diagnostics, therapeutics, and have been discovered to play a key role in intercellular communication. The study of EVs in biological fluids has proven challenging due to the nanoscale size of EVs (30 nm-1 µm diameter), the enormous quantity of EVs present in clinical samples (e.g. 1010 /mL in plasma), and the heterogeneous properties of EVs, even within those that originate from the same cell. My thesis has developed two distinct, but related, technologies to address these challenges. The first half of my thesis focuses on isolation and interpretation of specific subsets of EVs from biological samples, such as plasma, based on particular expressions of surface proteins. From these isolated EVs we have demonstrated, across multiple diseases, that there are signatures of disease states encoded in the EV RNA cargo, which we identified using supervised machine learning. To this end, building on prior work from our group, we developed a multichannel nanofluidic system that could analyze crude clinical plasma samples with nanoscale precision, which was coined Track Etched Magnetic Nanopore (TENPO). We evaluated the clinical potential TENPO by first applying it to diagnosing and staging pancreatic cancer, where current biomarkers have proven elusive to achieve sufficient sensitivity and specificity. In this work, we algorithmically combined tumor-associated EV mRNA and miRNA, isolated from plasma using TENPO, with ccfDNA levels, KRAS mutation detection, and CA19-9 via an ensemble machine learning model to form a multi-analyte panel. On an independent, blinded validation set (N = 136), we were able to distinguish patients with pancreatic cancer from those without at an accuracy of 92% (AUC=0.95). Moreover, among patients with pancreatic cancer, my model achieved significantly higher accuracy for disease staging (84%) than the current standard imaging method (64%). In addition to pancreatic cancer, I have also applied this approach to traumatic brain injury and to Alzheimer’s Disease to explored its diagnostic value in neurodegenerative diseases. Though TENPO was successful in isolating specific subsets of EVs for downstream analysis, it was not able to resolve the heterogeneity that is known to exist between individual EVs. Existing single EV analysis methods can only analyze a small number of EVs (\u3c 20,000), limiting their ability to evaluate rare EV subsets due to subsampling error when searching for these rare EVs amongst the high EV background present in plasma. To address this challenge, I have developed a high throughput, droplet based optofluidic platform to quantify specific single EVs. The key innovation of my platform is parallelization of droplet generation, processing, and analysis to achieve a throughput \u3e100x greater than typical in microfluidic systems, using only simple optics and accessible soft-lithography fabrication. I demonstrated that this improvement in throughput can be leveraged to quantify human neuron derived EVs at a limit of detection LOD = 9 EVs/µL, a \u3e100x improvement over gold standard single EV characterization methods. Additionally, I demonstrated the potential of this system for use in clinical samples by detecting EVs in a complex media, containing up to 4,000 fold more background EVs, and achieved an LOD = 11 EVs/µL. Beyond extracellular vesicles, I was also inspired to apply this immunospecific, nanoscale detection and analysis modality to other subcellular materials, namely mitochondria. I have developed a pipeline to isolate and amplify single mitochondrion DNA from individual cells with 20x higher yield than with conventional tools. With the improved yield, we were also able to reveal the pervasive single nucleotide variation on mitochondrion DNA within single cells. We also compared the genomic variation within neuron mitochondria versus that within astrocyte mitochondria, which is impossible via traditional methodology
Machine Learning Methods For The Analysis Of Single-Cell And Spatially Resolved Transcriptomics Data
Get PDFThe advent of high-throughput next-generation sequencing technologies has transformed our understanding of cell biology and human disease. It is now common for investigators to study human cell populations by profiling the transcriptomes for thousands of single cells using single-cell RNA sequencing (scRNA-seq) technologies. In addition, recent advances in spatially resolved transcriptomics (SRT) technologies have enabled gene expression profiling with spatial information in tissues. Knowledge of the relative locations of different cells in a tissue is critical for understanding disease pathology because spatial information helps in understanding how the gene expression of a cell is influenced by its surrounding environment and how neighboring regions interact at the gene expression level. In order to take full advantage of the multi-modality information when analyzing scRNA-seq and SRT data, new methods are demanded for the following challenges: (1) how to identify cell types for scRNA-seq data with closely related cell types or low sequencing depths? (2) how to jointly model gene expression, spatial location, and histology in SRT data analysis? (3) how to increase gene expression resolution in SRT to study detailed tissue structure? In this dissertation, I seek to address these various challenges and difficulties associated with scRNA-seq and SRT data analyses. To address challenge (1), I developed ItClust, a supervised machine learning method that takes advantage of cell-type-specific gene expression information learned from a well-labeled source dataset, to help cluster and classify cell types on newly generated target data. To address challenge (2), I developed SpaGCN, a graph convolutional network approach that integrates gene expression, spatial location and histology to identify spatial domains and spatially variable genes in SRT data analysis. Lastly, to address challenge (3), I developed TESLA, a machine learning framework that enhances gene expression resolution in SRT and further performs multi-level tissue annotation with pixel-level resolution. I validated the utility of each of these approaches using experimentally validated cell type labels and independent pathologists’ annotation. I also demonstrated real use cases for these methods in deciphering tumor microenvironment in various cancer types
Using Time-Resolved Electron Microscopy And Data Analytics To Quantify The Evolution Of Supported Metal Nanoparticles
Get PDFSupported precious metal nanoparticles are important heterogeneous catalysts for both industrial processes and commercial products. Their high catalytic activity stems from their high surface free energy and under-coordinated surfaces, however these same properties destabilize the particles and cause them to grow and deactivate. While research studying the degradation of supported catalysts has been undertaken for decades, the exact mechanisms at play, and how the vary with reaction conditions, are not well understood. Advances in experimental instrumentation have positioned Transmission Electron Microscopy (TEM) as an ideal tool for characterizing the dynamic evolution of these nanoscale systems with both high spatial and temporal resolution. However, the difficulty of manually analyzing large in situ datasets to quantify nanostructural evolution remains a challenge. This dissertation focuses on combining in situ experimental observations with machine learning and data analytics to quantify image data and understand nanoparticle coarsening. The first thrust of this research is developing a machine-learning pipeline for automated image segmentation. By optimizing state-of-the-art deep learning segmentation models, we were able to rapidly segment and measure particles from thousands of TEM images in a reliable and reproducible fashion. Utilizing this automated image processing pipeline, we observed the evolution of a model catalyst at high temperature and assessed the competition between coarsening by evaporation and surface diffusion as a function of particle size and temperature. After developing a physical model to describe each mechanism, we were able to characterize particle interactions along the support and to identify a critical particle size which avoids degradation. Finally, we used a combination of temperature-dependent in situ experiments and Kinetic Monte Carlo simulations to understand how the rate of nanoparticle evaporation depends on nanoparticle morphology. Our mechanistic model allows us to understand how random structural fluctuations and surface roughening contribute to the evaporation process. In all, this research aims at developing techniques and data-rich quantitative methods for understanding how supported nanocatalysts can be engineered for optimal activity and lifetime
The Network Science of Distributed Representational Systems
No full textFrom brains to science itself, distributed representational systems store and process information about the world. In brains, complex cognitive functions emerge from the collective activity of billions of neurons, and in science, new knowledge is discovered by building on previous discoveries. In both systems, many small individual units—neurons and scientific concepts—interact to inform complex behaviors in the systems they comprise. The patterns in the interactions between units are telling; pairwise interactions not only trivially affect pairs of units, but they also form structural and dynamic patterns with more than just pairs, on a larger scale of the network. Recently, network science adapted methods from graph theory, statistical mechanics, information theory, algebraic topology, and dynamical systems theory to study such complex systems. In this dissertation, we use such cutting-edge methods in network science to study complex distributed representational systems in two domains: cascading neural networks in the domain of neuroscience and concept networks in the domain of science of science. In the domain of neuroscience, the brain is a system that supports complex behavior by storing and processing information from the environment on long time scales. Underlying such behavior is a network of millions of interacting neurons. Many recent studies measure neural activity on the scale of the whole brain with brain regions as units or on the scale of brain regions with individual neurons as units. While many studies have explored the neural correlates of behaviors on these scales, it is less explored how neural activity can be decomposed into low-level patterns. Network science has shown potential to advance our understanding of large-scale brain networks, and here, we apply network science to further our understanding of low-level patterns in small-scale neural networks. Specifically, we explore how the structure and dynamics of biological neural networks support information storage and computation in spontaneous neural activity in slice recordings of rodent brains. Our results illustrate the relationships between network structure, dynamics, and information processing in neural systems. In the domain of science of science, the practice of science itself is a system that discovers and curates information about the physical and social world. For centuries, philosophers, historians, and sociologists of science have theorized about the process and practice of scientific discovery. Recently, the field of science of science has emerged to use a more data-driven approach to quantify the process of science. However, it remains unclear how recent advances in science of science either support or refute the various theories from the philosophies of science. Here, we use a network science approach to operationalize theories from prominent philosophers of science, and we test those theories using networks of hyperlinked articles in Wikipedia, the largest online encyclopedia. Our results support a nuanced view of philosophies of science—that science does not grow outward, as many may intuit, but by filling in gaps in knowledge. In this dissertation, we examine cascading neural networks first in Chapters 2 through 4 and then concept networks in Chapter 5. The studies in Chapters 2 to 4 highlight the role of patterns in the connections of neural networks in storing information and performing computations. The study in Chapter 5 describes patterns in the historical growth of concept networks of scientific knowledge from Wikipedia. Together, these analyses aim to shed light on the network science of distributed representational systems that store and process information about the world
University Culture and Faculty Governance: How the Interrelationship Impacts Institutional Performance during Crisis
No full textThis study focused on the interrelationship of organizational culture and faculty governance at public research universities. The purpose of this study was to explore how universities responded in organizationally culturally appropriate ways to crisis situations that had the potential to strain shared decision-making processes. Through a qualitative study using case studies of two institutions—the University of Michigan (U-M) and the University of California Los Angeles (UCLA)—this study investigated the conditions under which university stakeholders made decisions during the COVID-19 global pandemic and explored why decisions may have been consistent with institutional culture or violated cultural norms. The two institutions in this study were chosen because they reflected two common types of institutional responses to COVID-19: those that responded in ways that appeared to violate cultural norms (U-M), and universities with COVID-19 decisions consistent with cultural norms and expectations (UCLA). Data were gathered through document analysis, observations, and interviews with senior and midlevel administrators, academic administrators, faculty senate leaders, and other faculty and staff who were part of task forces or committees formed around COVID-19. This study found faculty governance plays an important role in how universities effectively respond during times of crisis and makes multiple contributions to institutional activity. Both informal and formal faculty governance advances a practice of consultation in institutional decision making, resulting in better decisions and outcomes. Academic leaders (i.e., deans) play an important role in institutional shared governance and, as a group, strengthen a sense of collective responsibility for the university. Universities that (a) demonstrate strength in the degree to which norms are held, (b) have congruent cultural elements, and (c) have norms with contents that evoke attitudes and beliefs have more stable and effective cultures, especially when responding to crises. Higher education institutions will continue to face challenges and crises even if they are not at the scale of COVID-19. Understanding how organizational culture affects faculty governance and how faculty governance affects culture with respect to decisions and outcomes is important, because both are tied to how a university operates and perform
Deep Basis Fitting for Depth Completion
No full textRecovering depth information from a single image is a challenging task. It is a fundamentally ill-posed problem since there exists an infinite number of scene geometries that could give rise to a given image. However, knowing the depths for a few pixels can significantly constrain the set of solutions. Recovering a plausible depth map from an image with sparse depth measurements is referred to as image-guided depth completion and is the focus of this thesis. We first developed a novel approach called Deep Basis Fitting (DBF) that builds upon the strengths of modern deep learning techniques and classical optimization algorithms which significantly improves performance. The proposed method replaces the final 1 × 1 convolutional layer used in most depth completion networks with a least-squares fitting module which computes weights by fitting the implicit depth bases to the given sparse depth measurements. In addition, we show how our method can be naturally extended to a multi-scale formulation for improved self-supervised training. We then extend DBF for depth completion within a Bayesian evidence framework to provide calibrated per-pixel variance. The DBF approach falls short when the underlying least-squares problem is under-determined, i.e. the number of sparse depths is smaller than the dimension of the basis. By adopting a Bayesian treatment, our Bayesian Deep Basis Fitting (BDBF) approach is able to 1) predict high-quality uncertainty estimates and 2) enable depth completion with very few or even no sparse measurements. While many depth completion methods rely on an external 3D sensor to produce accurate sparse measurements, it is still possible, albeit much more challenging, to generate dense depth from a single camera. Structure-from-motion algorithms, such as visual odometry or visual SLAM, solve for both camera motion and scene structures which can be used for depth completion. To this end, we developed a visual odometry system named Direct Sparse Odometry Lite (DSOL), which builds upon the original Direct Sparse Odometry (DSO).DSOL adopts several algorithmic and implementation enhancements that speed up computation by an order of magnitude compared to the baseline. We follow the data-oriented design philosophy and layout data contiguously in memory, which improves cache-locality and allows for easy parallelization. The increase in speed allows us to process images at higher frame rates, which in turn provides better results on rapid motions. Finally, we show that the two systems developed above can be integrated together, where sparse points from the monocular visual odometry can be used for depth completion and the completed depth can in turn be used to initialize odometry keyframes
Fermionic Diagonal Coinvariants
Get PDFLet be a complex reflection group of rank acting on its reflection representation V \cong \mb{C}^n. The doubly graded action of on the exterior algebra induces an action on the quotient by the ideal generate by -invariants with vanishing constant term \FDR_W = \wedge (V \oplus V^*) / \langle \wedge (V \oplus V^*)^W_{+} \rangle. We describe the bi-graded -module structure of \FDR_W and introduce a variant of Motzkin paths that descends to the standard monomial basis of \FDR_W with respect to certain term order. The top degree of \FDR_W exhibits the Narayana refinement of Catalan numbers. When , the symmetric group, \FDR_{S_n} \cong R_{n,0,2}, where is the special case of the Boson-Fermionic diagonal coinvariants with two sets of Fermionic variables. In this case, the -th degree component is a difference of Kronecker product of two hook Schur functions.
In addition we consider a module spanned by -ary strings of length . When , as a vector space, M_{n,2} \cong \mb{C}[X_n] / \langle x_1^2, \ldots, x_n^2 \rangle. The trivial component of \dr_n \otimes M_{n,2} is a weighted sum of -Narayana numbers which is a different -Catalan number than the alternant of \dr_n. At , the trivial component equals the inversion generating function for -avoiding permutations
Synthesis And Characterization Of Functional Materials Using Silica Colloidial Crystals, Their Inverse Replicas, And Layered Double Hydroxides
Get PDFThe design of materials with tunable properties is at the forefront of material-based applications. The key to materials design is understanding their fundamental characteristics and establishing a structure-property correlation. This dissertation explores fundamental aspects of synthesis and characterization of functional materials designed using colloidal crystals, inverse replicas, and layered materials for electronics and energy devices applications. We have combined particle assembly and High Pressure confined Chemical Vapor Deposition (HPcCVD) to create ordered and electrically continuous 3D nanostructures of metals and semiconductors, defined as metalattices. These nanostructures have crystalline arrays of uniform particles in which the period of the crystal is close to the characteristic physical length scale of the material, for example, exciton Bohr radius in semiconductors, making them tunable for electronic, plasmonic, thermoelectric and spintronics applications. Silica nanoparticles in the range of 20-120 nm, assembled as micron thick films using vertical deposition technique, were used as templates for metalattice design. The interstices in the colloidal crystal films were infiltrated with polycrystalline semiconductors (Ge/Si/ZnSe) and metals (Ni/Pt/Ag/Pd/Au) using HPcCVD to obtain corresponding metalattices.We have developed a core-shell chemical passivation strategy for Ge metalattice prepared by infiltration of ~70 nm silica colloidal crystal using HPCVD. The oxide-free Ge core shows quantum confinement which depends on the void size in the silica template. The size of Ge sites dictated by the voids in the template and core-shell interdiffusion of Si and Ge can, in principle, be tuned to modify the electronic properties of the Ge metalattice. We have also investigated the structures of colloidal crystalline films and germanium metalattice in detail by scanning electron microscopy (SEM) and small angle x-ray scattering (SAXS). Particles smaller than ~32 nm diameter assemble into body centered cubic, whereas particles larger than 32 nm assemble into random hexagonal close pack structures with 2D hexatic phase. Polycrystalline films of these materials retain their structure, and long-range order upon infiltration at high temperature and pressure, and the structure is preserved in Ge metalattice. This detailed understanding of particle arrangements in the template can help in establishing structure-property relationships in the metalattices. We also explore material design made from layered materials for application in energy systems. We discuss method for controlled assembly of oppositely charged nanosheets using tri-block co-polymer F127 to tune their interactions and study the synthesis and anion exchange of Mg-Al, Zn-Al and Co-Al layered double hydroxides. We characterize their structural, thermochemical, and ionic conduction properties to understand their fundamental behavior for applications as anionic conductors in electrochemical systems operating between 100-250 °C
Virtualizing Reconfigurable Architectures: From Fpgas To Beyond
Get PDFWith field-programmable gate arrays (FPGAs) being widely deployed in data centers to enhance the computing performance, an efficient virtualization support is required to fully unleash the potential of cloud FPGAs. However, the system support for FPGAs in the context of the cloud environment is still in its infancy, which leads to a low resource utilization due to the tight coupling between compilation and resource allocation. Moreover, the system support proposed in existing works is limited to a homogeneous FPGA cluster comprising identical FPGA devices, which is hard to be extended to a heterogeneous FPGA cluster that comprises multiple types of FPGAs. As the FPGA cloud is expected to become increasingly heterogeneous due to the hardware rolling upgrade strategy, it is necessary to provide efficient virtualization support for the heterogeneous FPGA cluster.
In this dissertation, we first identify three pairs of conflicting requirements from runtime management and offline compilation, which are related to the tradeoff between flexibility and efficiency. These conflicting requirements are the fundamental reason why the single-level abstraction proposed in prior works for the homogeneous FPGA cluster cannot be trivially extended to the heterogeneous cluster. To decouple these conflicting requirements, we provide a two-level system abstraction. Specifically, the high-level abstraction is FPGA-agnostic and provides a simple and homogeneous view of the FPGA resources to simplify the runtime management and maximize the flexibility. On the contrary, the low-level abstraction is FPGA-specific and exposes sufficient low-level hardware details to the compilation framework to ensure the mapping quality and maximize the efficiency. This generic two-level system abstraction can also be specialized to the homogeneous FPGA cluster and/or be extended to leverage application-specific information to further improve the efficiency. We also develop a compilation framework and a modular runtime system with a heuristic-based runtime management policy to support this two-level system abstraction. By enabling a dynamic FPGA sharing at the sub-FPGA granularity, the proposed virtualization solution can deploy 1.62x more applications using the same amount of FPGA resources and reduce the compilation time by 22.6% (perform as many compilation tasks in parallel as possible) with an acceptable virtualization overhead, i.e.,
Finally, we use Liquid Silicon as a case study to show that the proposed virtualization solution can be extended to other spatial reconfigurable architectures. Liquid Silicon is a homogeneous reconfigurable architecture enabled by the non-volatile memory technology (i.e., RRAM). It extends the configuration capability of existing FPGAs from computation to the whole spectrum ranging from computation to data storage. It allows users to better customize hardware by flexibly partitioning hardware resources between computation and memory based on the actual usage. Instead of naively applying the proposed virtualization solution onto Liquid Silicon, we co-optimize the system abstraction and Liquid Silicon architecture to improve the performance