137091 research outputs found
Sort by
Indigenous-Led Transportation Solutions: A Pasefika Framework of Transportation Planning in the Koʻolauloa Moku of Oʻahu, Hawaiʻi
No full textIn this research, I employ indigenous, pasefika methods to create a framework for transportation planning. Koʻolauloa, a rural community home to Native Hawaiian and other Pacific Islanders indigenous to other United States territories, has long standing transportation justice and equity issues surrounding safety and resilience. Using talanoa, a pasefika engagement method of inclusive, transparent and cultural dialogue-based meetings, and talk-story, a Hawaiian storytelling method used to exchange knowledge, culture, beliefs and values between two parties, I engage the indigenous and pasefika community, government agencies, and professional planners and engineers in Koʻolauloa. Complementing these storytelling methods, I use indigenous statistics in the form of a community preference survey to understand safety and resiliency for Koʻolauloa.
Using these indigenous, pasefika methods I can reframe safety and resiliency around the culture, values and beliefs of Koʻolauloa. From the talk-story with government agencies and professional planners and engineers, I can identify how their data-driven approach is leading to gaps in engagement and projects that don’t match the community. There is a need to go beyond participation and inclusion from the stakeholders and for them to build deeper long-lasting relationships with the community. Using methods like talanoa and talk-story should be utilize by planners and engineers, and the stories and conversations should be viewed as legitimate data in the planning process.
For the community of Koʻolauloa, the talanoa revealed how these stories and conversations can lead to rich planning solutions that is supported by the community. The relationship between the people, the ʻāina, kai, and their ancestors and descendants is important to understand for planners, as it influences the solutions being made by the community. For example, the community saw that a contributing factor to their fatalities and collisions was because the spirit of aloha was missing. This belief and the important cultural relevance of aloha led to them wanting safety solutions such as gateways, signs and RRFBs that remind people to “Drive with Aloha.
New Directions in Multi-Objective Optimization with Applications
No full textThis thesis introduces the portfolio framework for optimization with multiple objectives. A portfolio is a small set of solutions that approximately optimizes every objective under consideration. This approach recognizes the inherent plurality of objectives and provides a structured way to navigate competing goals. Instead of insisting on a single `best' solution, portfolios offer a small number of high-quality solutions that together span the space of possible preferences. This work discusses the theoretical foundations, algorithmic techniques, and practical applications of this framework across various problems in machine learning and combinatorial optimization
Deep Learning Based Manufacturing Capability Modeling for Process Planning Automation
No full textRealizing the vision of a generative manufacturing process planning system requires a set of efficient computational tools that enable automated decision making at different stages of the design-to-product pipeline. While efforts have been made over the years to develop automated systems for feature recognition, shape retrieval, process selection, and manufacturability assessment of part designs, such methods are hampered by the lack of a systematic, realistic, and scalable approach to model and integrate the knowledge required to enable decision-making at different granularity levels of a manufacturing system (e.g., process level, shop level, and online marketplace level). Specifically, methods to date have relied on simplistic and sometimes ad-hoc classification and/or encoding of manufacturing process capability knowledge, which are incomplete or too simplistic, difficult to scale, and heavily reliant on human expertise. Recent advances in machine learning and especially deep learning have shed light on potential pathways for data-driven inference of manufacturing process capabilities from existing design and manufacturing data of successfully produced parts. This dissertation presents a coherent set of research efforts to develop deep learning-based computational tools to enable data-driven decision making at the process, shop, and marketplace levels. Specifically, the dissertation addresses the following research questions: (1) Can we learn and represent the shape, material, and part quality transformation capabilities of discrete manufacturing processes from design and manufacturing data as a latent probability distribution? (2) Can we develop data-driven computational tools to enable decision-making in key process planning steps ranging from process/operation selection to operations sequencing? (3) Can we enable efficient automated search of manufacturers based on a computed score for the match between the desired part and the manufacturer’s capabilities? To answer these research questions, this dissertation presents (1) a deep embedding modeling approach for inferring the shape, material properties, and part quality transformation capabilities of machining processes/operations from historical design and manufacturing data as a latent probability distribution to enable automated process/operation selection and manufacturability assessment at the process level; (2) a deep-learning based part geometry segmentation approach to segment and label machinable volumes with candidate machining operations; (3) a deep sequence learning approach to automatically learn precedence relations among machining operations to enable automated operations sequencing at the shop level; and, (4) at the online marketplace level, a deep unsupervised learning-based manufacturing capability model to enable efficient process-aware part retrieval, which, when combined with federated learning, enables an implicit and secure manufacturer search without the need for a centralized parts database. From these studies, it has been demonstrated that (1) a three dimensional variational autoencoder generative adversarial network can model machining process shape transformation capabilities as latent probability distributions, which, when combined with a Siamese neural network, achieved 100% accuracy and area under the curve (AUC) of 1 in manufacturability analysis, as well as 89% class-average accuracy in automated process selection, outperforming baseline discriminative models; (2) a semantic segmentation method using generative pre-trained neural networks achieved over 96.8% intersection over union (IoU) for machinable volume identification in complex machined parts produced in a lathe; (3) a three-dimensional convolutional recurrent neural network model can be used to learn precedence relations in machining operations sequencing, outperforming a baseline binary classifier with 97.6% validation accuracy, and demonstrating practical applicability in validating operations sequences for realistic machined parts; (4) the proposed deep unsupervised part retrieval model improved part retrieval precision by incorporating manufacturing capability information, achieving a combined process and function class precision at 1 of 93.0%, which, when combined with federated learning, demonstrated an accuracy of 89% in identifying suppliers with non-overlapping manufacturing capabilities and a 87.8% accuracy when suppliers’ capabilities overlap. The findings and contributions of this research serve as key technology enablers for future generative manufacturing process planning systems.Ph.D.Mechanical Engineerin
Using the TIN-Based Real-Time Integrated Basin Simulator (tRIBS) to Model Streamflow and Terrain Processes in the Rio Grande de Añasco, Puerto Rico
No full textRainfall–runoff models support decisions in water resources management, such as hydraulic design, planning, and flood control, especially where observations are limited in space and time. These models exist in several forms, and each has been developed for various purposes. Physically based distributed models, which solve governing physical equations, are particularly useful for representing hydrological processes in space and time, but they require greater expertise. The Triangulated Irregular Network (TIN)-based Real- time Integrated Basin Simulator (tRIBS) is one such model, and although it has been applied in various settings, additional case studies are valuable for informing its practical use. This study applies tRIBS to simulate streamflow, terrain processes, and hydrological dynamics in a clay soil-dominated, predominantly evergreen-forested subbasin of the Río Grande de Añasco in western Puerto Rico. The goals are to establish and calibrate the model, evaluate its performance across an event-based window and an extended period, and document insights on initialization, sensitivity, and calibration strategies to guide future tRIBS applications and water resources studies in the region. The model was successfully calibrated for August 2000–June 2001 and produced reasonable streamflow simulations during both the event-based (November 4–17, 2003) and long-term (September 1990–July 1991) validation periods. Performance depended strongly on precipitation patterns: watershed-wide, uniform rainfall events were captured well, whereas localized storms were more difficult to reproduce due to limitations in gauge- based rainfall representation. For this study, spin-up tests indicated that recycling the xiv hydrometeorological forcing was the most effective approach for initializing the groundwater table. The sensitivity analysis showed that saturated hydraulic conductivity in the clay- dominated study area played a critical role in controlling infiltration, groundwater, and baseflow. The pore-size distribution index was also important because it strongly affected the model response and baseflow. Lower values of this parameter substantially decreased baseflow and amplified the impact of changes in air entry pressure. Other influential soil parameters included the hydraulic decay parameter, anisotropy ratio, and saturated soil moisture. Land use parameters were less influential, although the optical transmission coefficient and stress thresholds for evaporation and transpiration had noticeable effects on model response. Model performance results are summarized as follows. For the 11-month calibration (August 2000–June 2001), hourly performance had a correlation coefficient (CC) of 0.68, a Nash–Sutcliffe efficiency (NSE) of 0.39, and a root-mean-square error (RMSE) of 11.54 m³/s. The model was validated with both a short high-rainfall window where precipitation was relatively uniform across the watershed (November 4–17, 2003) and an independent period equal in length to the 11-month calibration (September 1990 to July 1991). In the short, high-rainfall period, when rainfall was more uniform across the watershed, hourly performance showed CC 0.91, NSE 0.69, and RMSE 47.85. For the 11-month period that includes wet and dry seasons and a mix of uniform and non-uniform events, hourly performance showed CC 0.58, NSE 0.24, and RMSE 13.55. Performance improved at the daily time step: NSE and CC increased, while RMSE decreased
Content Addressable Memories for In-Memory Search to Enable AI Applications
No full textIn this thesis we address the demand for fast, accurate and compact similarity search to enable data-intensive and edge AI applications using our novel content addressable memory (CAM) designs. CAMs, with their ability to perform parallel in-memory searches, offer the potential to replace traditional search schemes. Despite these benefits there are many challenges for using CAMs due to scaling issues at advanced nodes and limited search resolution at larger Hamming Distances. In this thesis we highlight these challenges by analyzing CAMs at the 7nm technology node using layout extracted parasitics and considering various variability sources. We study these effects at the application-level using a CAM-based recommendation system. We suggest design modifications at the algorithmic, schematic and layout-levels to solve these issues to enable the use of CAMs in data-intensive AI applications. To enhance the use of hyperdimensional computing (HDC) for edge applications, we develop novel energy-efficient and variation-tolerant CAM and in-memory encoding designs using resistive memory devices at the 22nm node. We use an HDC-based speech recognition system to evaluate and benchmark our designs against other technologies. Overall, in this thesis, we focus on developing efficient CAM designs tailored to meet the varying requirements of diverse AI systems.Ph.D.Electrical and Computer Engineerin
On k-Winners-Take-All as a Model of Neuron Communication
No full textUnderstanding how local interactions between neurons give rise to high-level structures in the brain is one of the key open questions in neuroscience. This thesis investigates discrete-time, weighted digraph models of the connectome, which not only provide insight into biological neural networks, but also form the backbone of modern artificial neural networks. The primary model of interest is the k-cap process, a recurrent neural network employing k-winners-take-all (k-WTA) as a nonlinear gating function; this mechanism models firing rate regulation via global lateral inhibition. By exploring the computational properties arising from this model, this work aims to offer insights into the emergence of structures in the connectome
Data-Driven Frameworks for Predictive and Prescriptive Control of Incremental Manufacturing Processes
No full textModel-based approaches for manufacturing processes play a critical role in enabling first part correct production frameworks. Deep learning (DL) methods offer significant opportunities for online control and offline process design; however, there remains limited understanding regarding the fundamental applicability of such methods. This dissertation seeks to explore the suitability of deep learning-based approaches for online control and offline process design for additive manufacturing, with a particular focus on directed energy deposition and wire arc additive based processes. This work is organized in three complementary studies that explore applicability of deep learning for (1) online process monitoring and control and (2) offline process design. The first study showed that both SVR and CNN models trained on melt pool data collected from a coaxial optical camera can accurately predict the bead height with a mean absolute percentage error of 3.67% and 3.68%, respectively. The second study explores a novel data-driven and physics-informed framework proved successful in scenarios with parameters outside of the training dataset as well as in unstable process settings such as the beginning and end of the deposition as well as during the transition between laser powers and standoff distances achieving a mean absolute percentage error of 6.51% for height and 14.89% for width. The final study creates a model that can predict layer height with an error of 0.446 mm for a test part with a wire feed speed and geometry not included in the training data. A key factor in these studies that is considered includes understanding of feasibility to be integrated onto edge devices, with implications for feedback and feed-forward control of machine platforms.Ph.D.Computational Science and Engineerin
The Dynamic Interplay of Ice Mélange and Calving Glacier Interfaces
No full textFuture sea level projections are subject to uncertainties in ice sheet processes, including slow variations in ice flow speed and fast variations in melting and calving, which are driven by increasing oceanic and atmospheric temperatures. Recent studies suggest that the contribution of ice sheets to sea level rise over coming centuries is likely to be higher than previously projected due to the potential for rapid ice fracture and iceberg calving from the edge of ice sheets. This thesis seeks to address this issue by investigating the dynamic feedbacks between glaciers and the icebergs they generate through calving, with a particular emphasis on the role of ice mélange. Ice mélange is a slushy amalgamation of icebergs and sea ice, which is thought to function like an ice shelf by slowing glacier flow from the ice sheet interior and preventing fracturing and calving of new icebergs. Although mélange currently persists seasonally or throughout the year at numerous glaciers in Greenland and Antarctica, its importance may rise in conjunction with escalated calving rates. The principal aim of this research is to resolve open scientific questions concerning the role of ice mélange in the mechanics of iceberg calving and the retreat of glaciers.
In the first project described in this thesis, I employ the Helsinki Discrete Element Model (HiDEM) to simulate calving-mélange feedbacks down to spatial scales of meters and temporal scales of seconds. This high-fidelity model facilitates understanding of how mélange buttressing influences the stress state at the glacier terminus and enables the identification of variations in calving rates and styles. I analyze bulk calving statistics from this model, including calving event size and recurrence time, which provide insights into short time-scale events that regulate glacier calving. Though such a high-fidelity model is useful for conducting process studies of glacier calving, its high computational expense prevent its use for projecting ice sheet behavior over climate-relevant time scales of decades and longer. Consequently, I conduct a case study of Sermeq Kujalleq, the fastest flowing glacier in Greenland, using the Ice-Sheet and Sea-Level System Model (ISSM) to discern the relative impacts of submarine melting and a weakened ice mélange on the glacier's recent retreat. By adjusting the sensitivity of melt rates and a calving stress threshold parameter, I conclude that Sermeq Kujalleq's ocean-induced retreat starting in the late 1990's was predominantly driven by a weakened ice mélange following an influx of warm ocean waters. However, parameterized representation of mélange variations hinder the ability of such models to project future coupled evolution of mélange and glacier calving. To remedy this shortcoming of current models, I develop a two-dimensional continuum model of ice mélange, GLACIOME2D, which integrates the granular physics of mélange and its mechanical interactions with the ice-ocean environment over climate-relevant timescales. Collectively, my findings integrate insights from high-fidelity discrete element modeling with efficient continuum modeling to provide advances in long-term projections of ice sheet mass loss and contributions to sea level rise
Methods for Analysis of High-Resolution Muscle Recordings During Dynamic Movement and Generation of Realistic Motor Unit Datasets for Ground Truth
No full textEvery movement an animal makes is dependent on sophisticated coordination across thousands of motor neurons controlling hundreds of thousands of muscle fibers. Each motor neuron can control dozens to hundreds of muscle fibers, which forms a functional unit of the motor system, called a motor unit. From over 60 years of research, we have learned a lot about individual motor unit physiology and function, but we still have a limited understanding of large-scale motor unit population coordination. By leveraging novel, flexible Myomatrix arrays (Chung et al., 2023), we were able to record from many motor units during multiple dynamic behaviors in rats. This project focused on producing new methods for spike sorting, that is, robustly identifying the activation times and identities of many simultaneously recorded motor units across a range of force conditions during dynamic behavior. The methods we developed were shown to provide a substantial performance boost in the accuracy of identifying when each motor unit became active during behavior. In addition to this, we developed new methods for generating simulated motor unit datasets that closely replicate the properties of the real recorded data we collected. These new datasets enabled the development of our novel methods to match candidate motor units with ground truth units to be able to robustly compute the accuracy of motor unit spike sorting. Overall, this work provides improved methods for spike sorting motor unit datasets, unlocking the potential for discovery of new diagnostic techniques for motor coordination disabilities and for addressing gaps in our scientific understanding of motor unit coordination.Ph.D.Biomedical Engineerin
Terahertz Nondestructive evaluation Techniques for Industrial Applications and Imaging
No full textThis thesis is dedicated to advancing Terahertz (THz) technology for nondestructive evaluation and imaging, with specific applications in industrial manufacturing and cultural heritage conservation. The core of this work lies in the development and application of signal processing and machine learning techniques to overcome the limitations of conventional THz analysis.
In the industrial applications, the research first establishes a framework for comparing deconvolution methods to precisely measure mill scale thickness on steel, providing guidance for method selection under different conditions. Furthermore, a neural network model that accurately estimates thickness is trained, and a novel filtered deconvolution technique that improves signal clarity and simplifies analysis is proposed.
For electronics inspection, THz imaging demonstrates high sensitivity in mapping the thickness of conformal coatings on circuit boards and in identifying hidden defects within a complex multi-layer interposer. By combining deconvolution with polarization analysis and unsupervised machine learning, the study successfully locates and characterizes various defect types, validated by X-ray imaging.
Addressing a challenge in cultural heritage, the thesis reformulates the problem of detecting iron gall ink on multi-layer documents from a low-contrast imaging task into a classification problem. A convolutional neural network, trained with co-teaching which is a strategy to handle noisy labels, is developed to reliably identify ink patterns on single and multi-layer paper stacks, revealing features not discernible in traditional time or frequency domain images.
Collectively, this research underscores the significant potential of integrating advanced algorithmic approaches with THz technology to push the boundaries of precision and reliability in nondestructive testing across diverse fields