137091 research outputs found
Sort by
Towards a Theory and Practice of Open-Ended Reasoning with Generative Models
No full textThe unreasonable effectiveness of large language modeling has enabled the rapid development of generative systems capable of increasingly sophisticated ''human-like'' reasoning. This impressive performance stems largely from two key factors: 1. A high-quality pre-training phase during which models acquire a strong prior for language, reasoning in language, and mathematics. 2. An extensive reinforcement learning phase during which the model refines knowledge and skills acquired during pre-training. This thesis proposal presents a series of theoretical analyses improving our understanding of pre-training processes coupled with practical RL-based algorithmic improvements for better generative model reasoning capability with the goal of improving generative model reasoning in both theory and practice. On the theoretical side, I establish novel generalization bounds on the performance of several generative model architectures in terms of model size and number of training samples. I then demonstrate these bounds can be used to understand commonly observed "scaling laws" during large model training. On the experimental side, I develop new RL training frameworks facilitating the open-source training of large language models (LLMs). These are then used to conduct an in-depth investigation of the factors affecting reasoning performance of LLMs after RL training. Insights from this investigation lead to the development of new RL algorithms for better LLM reasoning and self-correction.Ph.D.Machine Learnin
Defining Ballistic Experimental Methodology and Nose Performance Coefficients for Cross-Laminated Timber (CLT)
No full textProtective structures have traditionally relied on concrete and steel for ballistic resistance, but these materials are carbon-intensive and less suitable for certain military applications, such as rapid or temporary deployments. Recently, cross-laminated timber (CLT), an engineered wood product composed of stacked adhered lumber with alternating 90-degree orientations from layer to layer, has gained traction as a sustainable and rapidly assembled alternative to conventional materials. CLT systems are prime for temporary military construction due to their modular nature, so recent research has investigated the viability of CLT against extreme loading conditions, such as blast and ballistic impacts.
The current CLT ballistic testing protocol is based on the US National Institute of Justice (NIJ) Standard Test and Classification for Ballistic Resistant Materials typically applied to metallic and fiber body armor ballistic testing. Unlike metal and fiber composites, CLT exhibits more localized damage under ballistic impact, allowing for greater density of projectile impacts per surface area without shot interaction. To optimize shots per area in CLT ballistic experimentation, we subjected four Loblolly CLT target panels with varied projectile target proximities to 60 shots using a 0.5 in (1.27 cm) steel sphere projectile. Results and statistical analysis suggest that no shot path interaction occurs when impacts are spaced at least 2 in (5 cm) apart. Consequently, this study recommends reducing the standard shot spacing for partial penetration experiments from 7 in (17.8 cm) to a minimum of 2 in (5 cm) with a 0.5 in (1.27cm) steel sphere projectile. This article provides a significant contribution to the field by establishing a more efficient ballistic testing methodology tailored to CLT. The optimized approach allows for the collection of substantially more data while improving resource efficiency and reducing testing time.
Additionally, CLT’s ballistic response to varying projectile nose geometries remains unexplored. This research experimentally evaluated the penetration resistance of CLT target panels subjected to blunt (.357 and .44 Magnum) and spherical (0.5 in [1.27 cm] steel sphere) projectiles to quantify nose performance coefficients for predictive modeling within Department of Defense (DoD) design frameworks. We conducted a total of 90 ballistic experiments at the U.S. Army Engineer Research and Development Center (ERDC), generating partial and complete penetration series across a range of striking velocities. Results indicate that CLT’s ballistic resistance is highly dependent on projectile type. Blunt and spherical projectiles produced shallower penetrations and greater energy dissipation through crushing and shear plugging. The soft point projectile exhibited significant nose deformation, leading to higher energy absorption and reduced penetration efficiency. By adapting the Unified Facilities Criteria (UFC) predictive wood penetration equation with a CLT-specific nose performance coefficient, N ̅, we found that existing UFC models underpredict CLT’s resistance to nosed projectiles. These results of this research provide foundational data for developing predictive design criteria and improving CLT integration in protective infrastructure
Fluidic Diverter Controlled by Surface Jet Arrays
No full textThe surface boundary layer upstream of airframe-embedded engine inlets of modern aircraft can adversely affect the pressure recovery, distortion, and stability of the flow within the inlet systems. While boundary layer diverters that have been employed upstream of the propulsion inlets to effect thin inlet boundary layers are attractive because of their robustness, they can add significant drag and weight and are difficult to integrate. In addition, diverterless supersonic inlets (DSI) have a limited effective Mach number range. An alternative, promising approach for handling the surface boundary layer at the inlet is based on Fluidic Diverters (FDs) that use arrays of surface jets to actively transport streamwise momentum from the cross flow into the wall layer. These fluidic systems lead to minimal drag and weight penalties, as well as adjustable control throughout the flight envelope to maximize efficiency. The current investigation focused on interactions between inclined and yawed single jets and jet arrays with a subsonic turbulent boundary layer over a flat plate, with emphasis on the formation mechanisms of surface bound streamwise vortices, that lead to cross stream and transverse transport of streamwise momentum. Variation in spanwise and cross stream momentum distributions was investigated using several configurations of jet arrays with varying yaw angles. The investigations explore parameters and flow mechanisms of a Fluidic Diverter that can lay the foundation for future simulations and system level optimization
Multiport Power Electronics based on Four-Quadrant Soft-Switching Current-Source Converters
No full textThe electric power sector is undergoing a fundamental transformation driven by the rapid adoption of cost-effective renewable energy generation and the emergence of high-demand applications such as EV fast charging, AI data centers, and manufacturing onshoring. These drivers have reversed a decades-long trend of declining growth in peak demand and disrupted traditional load profiles, placing unprecedented stress on the aging power infrastructure. The resulting grid instability concerns and prolonged connection delays are accelerating a shift toward distributed, flexible, and resilient microgrid deployment models where generation and consumption are co-located. However, traditional microgrids must contend with existing power electronics and use separate single-function converters for each connection point, leading to a complex, costly, and inefficient power conversion architecture.
To address these limitations, this dissertation proposes two novel families of flexible multiport power converters (MPCs) capable of interfacing and coordinating multiple sources and loads within a single-stage conversion structure. The Multiport Soft-Switching Solid-State Transformer (MS4T) realizes partially-isolated MPCs, while the Multiport Soft-Switching Current Source Inverter (MSSCSI) achieves non-isolated MPCs with increased power density. Both families support an arbitrary number of AC or DC ports with bidirectional power flow and offer high and symmetrical conversion efficiency enabled by four-quadrant Zero-Voltage Switching (ZVS).
New control techniques and modulation strategies are developed to ensure robust soft-switching operation under multiport, four-quadrant conditions. Experimental validations are provided, including demonstrations in DC-DC-DC and DC-DC-3ϕAC configurations. A comprehensive optimization framework is presented to select the Pareto-optimal resonant elements across the multiport application space. Finally, a scalable system architecture based on MS4T and MSSCSI building blocks is proposed and demonstrated through simulation of a representative 500 kW multiport system, to support the next generation of flexible, high-performance multiport power electronics for modern microgrids
On the Nature of Powder Spreading Defects and Their Detectability and Impact to Part Quality
No full textLaser powder bed fusion (PBF-LB) metal additive manufacturing allows a new paradigm for design creativity and supply chain logistics. There are many process variables that affect part quality in powder bed fusion including powder quality, layer height, laser speed and power, and hatch spacing. Spreading defects can lead to many process defects such as voids, energy density changes, and part topography variations. The overall goal of the present dissertation is to identify the critical size of notches on the recoater blade via investigation of part quality and powder bed topography. The first study was a combined simulation and experimental study investigating recoater damage and spreading defects in PBF-LB. In the experimental study, notches were machined into the recoater to investigate the effect of spreading parameters on the powder bed via a laser line scanner. A simulation model via the discrete element method (DEM) modeled the spreading conditions, resultant topography, and the distribution of particles. This investigation will aid in understanding the spreading behavior at the particle level. The second study investigated the ability of optical based methods to predict recoater damage from in-process signals. A range of machine learning methods were employed to study this behavior. Key features from the optical images were used to identify height and width variations within the powder bed. The third study investigated the effect of recoater defects on part quality characteristics of PBF-LB printed parts. Witness specimens were analyzed via computed tomography and optical profilometry for their roughness and porosity content. This investigation will aid in determining the criticality of the size of the recoater damage and its influence on porosity distribution. Spreading deviations of size 0.0241 mm were not shown to cause swelling at the surface, whereas spreading defects of size 0.0707 mm, which was almost two times the layer thickness, showed noticeable signs of swelling. Altogether, these studies will inform a comprehensive understanding of powder spreading with a damaged recoater and its subsequent effects on powder bed and part level defects.Ph.D.Mechanical Engineerin
Smart In-Process Inspection with Human-Automation Symbiosis in Industry 5.0 Manufacturing Systems
No full textSince the advent of Industry 4.0, manufacturing industries have been continually integrating cutting-edge technologies into different aspects of manufacturing systems, creating a more complex and dynamic production environment. To maintain high product quality throughout the manufacturing processes, in-process inspection (IPI) becomes an efficient strategy, enabling prompt identification of defective parts and real-time process control through defect mitigation. Consequently, smart in-process inspection (s-IPI) has emerged as a critical research area in manufacturing.
Furthermore, as manufacturing technologies advance, the relationship between humans and automation agents evolves. Industry 5.0, as the next phase in manufacturing, differs from Industry 4.0 by shifting its focus from economic gains to human social value and well-being, emphasizing a human-centric philosophy. This underscores the importance of implementing human-automation symbiosis (HAS), which fosters closer partnership and mutually beneficial collaboration between humans and automation agents.
In this regard, this dissertation envisions smart in-process inspection with human-automation symbiosis in Industry 5.0 manufacturing systems (I5MS) as an emergent research paradigm. It explores the integration of smart in-process inspection into manufacturing systems and the implementation of HAS for I5MS from the task allocation and task execution perspectives, which builds upon multiple disciplines, including advanced manufacturing systems, cognitive engineering, and human-automation interaction.
With s-IPI and HAS as the two technical pillars, four fundamental issues are identified, including in-process inspection, defect mitigation planning, dynamic and adaptive task allocation, and behavioral intervention design for human-automation collaboration. Corresponding technical approaches are proposed for each issue.
Given the interdisciplinary nature, the research findings of this dissertation are poised to have a broad impact within and beyond manufacturing systems, such as service systems or operational systems. They have the potential to influence various sectors by providing a framework for integrating advanced inspection and human-automation symbiosis into diverse industrial applications. This research sets the stage for future innovations in manufacturing, ultimately contributing to the development of more intelligent, efficient, and human-centric environments in the era of Industry 5.0.Ph.D.Mechanical Engineerin
Navigating Mixed Traffic: Lateral and Longitudinal Control for Connected and Autonomous Vehicles with a Human-Centric Approach
No full textWith the emergence of connected and autonomous vehicles (CAVs), the transportation system is undergoing a revolution driven by advanced vehicle automation and connectivity technologies. CAVs, with enhanced situational awareness and advanced automation capabilities, promise to improve traffic safety, smoothness, and efficiency. However, the coexistence of CAVs with human-driven vehicles (HDVs) in mixed-traffic environments presents significant challenges, particularly in the context of lane changes. These challenges include: (i) disruptions to CAV platoons caused by HDV lane changes, (ii) mental discomfort experienced by HDV drivers and CAV users during HDV lane changes, (iii) uncertainties and compromised safety performance in HDV lane-change behavior in mixed traffic, and (iv) diminished control performance of CAVs and comfort for HDV drivers during CAV lane changes. To address these challenges, this dissertation develops human-centric control strategies for CAVs to improve interactions with HDVs in the context of both HDV and CAV lane changes. The contributions are reflected over several interconnected chapters, each building upon the insights and methods introduced in prior chapters.
First, the dissertation proposes a deep reinforcement learning-based proactive longitudinal control strategy for CAVs to preclude disruptive HDV lane-change behaviors that can induce disturbances, and to preserve the smoothness of traffic flow in the CAV platooning control process. In it, a Transformer-based lane-change traffic condition predictor is constructed to predict whether an HDV will likely perform a disruptive lane change under ambient traffic conditions. If no disruptive lane change is predicted, an extended intelligent driver model is activated for the CAV to perform smooth car-following behavior under cooperative CAV platooning control. If a disruptive lane change is predicted, a rainbow deep Q-Network-based lane-change preclusion model is proposed through which the CAV can alter the lane-change traffic condition to preclude the HDV’s lane change. The proposed control strategy can effectively reduce disruptive lane-change maneuvers by a HDV in the vicinity of the CAV and improve string stability performance, and serves as a building block for proactive control in mixed-traffic environments.
Building on this foundation, the dissertation introduces an innovative human-emulation-based proactive longitudinal control strategy for CAVs, to assist HDV lane changes and counteract their negative effects on CAV platoon smoothness. It constructs a Transformer-based behavior predictor to predict the HDV lane change behavior. If a lane change is anticipated, a lane-change assistance model is developed using the proximal policy optimization, which enables the CAV to emulate the lane-change assistance maneuver of human drivers. If no lane change by the HDV is anticipated, a multi-anticipative car-following model is adopted, through which the CAV executes cooperative platooning control. This strategy emulates the courteous lane-change assistance maneuver of human drivers in cooperative lane changes, where the drivers yield the lane-change vehicle through decelerations. By executing legible motions, CAVs can communicate their intentions in advance, assisting HDV drivers in making safe lane changes and promoting smooth CAV platoon operation.
Third, this dissertation expands the previous two strategies by proposing a proactive human-centric control strategy that enhances CAV control performance and mitigates the mental discomfort of both HDV drivers and CAV users. It operates in a hierarchical reinforcement learning framework and comprises a two-level solution to the intricate lane-change management problem. The upper-level task determines whether to assist or preclude the HDV lane change, and the lower-level task executes the corresponding CAV control actions. For situations without potential lane changes, a proximal policy optimization-based car-following control model is developed for the CAV to perform smooth car-following behavior. An adversarial inverse reinforcement learning-based behavior planner is proposed to regulate the lane-change management and car-following control models. This study emphasizes the potential of incorporating human factors into CAV operations to improve overall traffic performance, while ensuring the mental comfort of both HDV drivers and CAV users.
Fourth, to validate the aforementioned strategies, the dissertation applies driving simulator experiments to investigate HDV drivers’ experiences and safety during lane changes in mixed traffic. It first examines both the perceived and objective complexity experienced by HDV drivers during lane changes and explores their behavior evolution in repeated interactions with CAVs. It emphasizes how HDV drivers may be surprised by the way CAVs behave on roads at their first encounter with CAVs, but then can learn to understand CAV driving behavior after sufficient experience and adapt their lane-change behavior accordingly. Next, the dissertation introduces a comprehensive safety performance framework that combines multiple surrogate safety metrics to analyze the safety performance of HDV lane changes in both HDV-only traffic and mixed-traffic conditions. These investigations highlight the dynamic interactions between CAVs and HDVs in mixed-traffic, evolving nature of HDV behavior, and potential targeted interventions to shape HDV lane-change behavior in mixed-traffic environments.
Last, the dissertation proposes a human-like lane-change control strategy for CAVs to produce human-like lane-change behavior to improve CAV-HDV interactions. Operating within a theory of mind framework, the strategy introduces human-like reasoning into CAV operations. It first develops a human-like lane-change model that combines the enhanced control enabled by robot driving behavior and the human expertise of natural human driving behavior. The model produces human-like lane-change maneuvers to improve CAV control performance and mitigate HDV drivers’ mental discomfort. Further, this strategy proposes a HDV behavior predictor that learns human driving patterns to model human decision-making. The predictor anticipates potential responses of HDVs to the CAV lane change, enabling the CAV to perform informed and adaptive lane-change maneuvers that mitigate disruptions and maintain traffic smoothness. This human-like control strategy aims to integrate human intelligence and cognition with advanced CAV capabilities (in situational awareness and reaction times) to promote the smooth integration of CAVs into society.
In summary, this dissertation enhances the understanding of challenges in CAV-HDV interactions during lane changes in mixed traffic, which can guide CAV manufacturers in designing human-centric and adaptive control systems, while informing transportation planners on strategies to support safe and effective CAV deployment. The proposed human-centric control strategies demonstrate significant potential for improving CAV reliability and safety under diverse real-world conditions. By enabling smoother, more intuitive interactions between CAVs and HDVs, these strategies support a smooth transition to a more automated transportation environment, facilitating the measured and scalable adoption of CAV technologies as they continue to evolve.Ph.D.Civil Engineerin
Biologically Inspired Visual Communication System for Improving Multiple Robot System Operation and Resiliency
No full textThis research considers the problem of signal loss on radio frequency communication within Multiple Robot Systems. As more consumer devices utilize wireless communication technology, there are fewer open radio frequencies available. This dissertation will investigate biological visual communication methods and develop an active robotic visual communication system suitable for use within a Multiple Robot System. Humans employ a few different methods of active visual communication, but none have been applied to robotic communication. There are numerous active visual communication systems within biology that may be better suited for robotic applications as compared to current human systems. A key component of this research will be investigating and decomposing these communication systems. The decomposition will be used to develop the foundation of the robotic visual communication system. The development of the visual communication system will then be detailed and will include the software, the visual communication receiver, and the visual communication transmitter. The functionality of the visual communication prototype will then be discussed in both a laboratory setting and for practical deployment on a robotic platform for real-world application testing. The completion of this work will result in establishing a communication protocol that can operate in areas with heavy radio frequency interference or where radio communications are not desired. This thesis will support the view that multi robot systems and other collaborative robotic systems do not need to be reliant on radio frequency communication to reliably operate.Ph.D.Mechanical Engineerin
In-slot cooling for high power density electric motor
No full textElectric motor is a key component of the electric vehicle (EV) drivetrain as it converts electrical energy into mechanical energy. Therefore, a traction motor with high power density is essential to improve the overall performance of the EV drivetrain. Permanent magnet synchronous motors (PMSM) are commonly used in EV powertrains because of their high power and torque densities. In a high-power-density electric motor, major heat loss occurs in the windings in the form of resistive heating. In a typical liquid-jacket cooled motor, high thermal resistance between the winding and the cooling medium limits heat extraction. Therefore, by moving the coolant closer to the winding, thermal resistance between the winding and coolant can be significantly reduced, and hence, the winding temperature can be maintained below the thermal limit. This dissertation numerically compares the electro-thermal performance of three embedded cooling techniques, namely direct winding heat exchanger (DWHX), embedded circular, and rectangular cooling channels within stator core, with conventional jacket cooling (JC). The results confirm that the DWHX approach provides the best performance compared to the other cooling techniques. However, DWHX suffers from high winding-liner contact resistance. To enhance heat extraction by eliminating this contact resistance, this research examines a novel concept of capillary flow-assisted evaporative cooling (EC) confined between the slot liner and active winding. Electro-thermal performance of the EC has been assessed and compared with JC using a combination of modeling and motorette testing. Results show that the EC can provide 106.2% higher steady-state rms current density compared to the JC and enable 30 s peak rms current density of 40.8 A/mm2 within the thermal limit of class H insulation of 180oC. Leveraging the enhanced electro-thermal performance of EC, this dissertation also presents an optimized design of a high-speed heavy rare-earth-free PMSM to meet the U.S. Department of Energy (DOE)’s 2025 power density target of 50 kW/L for EV motors. The optimized motor layout can provide a 10 sec. peak power density of 86.59 kW/L, which is ~73% higher than the DOE’s peak power density target. Furthermore, based on the EC and DWHX designs, this dissertation proposes a new concept of EC from side and middle liner (EC-SML) and the effectiveness of EC, DWHX, and EC-SML has been assessed for different motor topologies and slot sizes. The results confirm that EC and EC-SML provide the best electro-thermal performance for small-slot and large-slot motors, respectively. Finally, to overcome the high leakage risk of EC and EC-SML, a novel 3D-printable hollow liner cooling concept has been proposed, and the thermal performance of a 3D-printed hollow liner has been experimentally evaluated using an in-house motorette test bench.Ph.D.Mechanical Engineerin
In Silico Discovery of New Materials for the Separation of Linear and Cyclic Siloxanes
No full textThis thesis presents the use molecular modeling to inform and accelerate the implementation of adsorption-based separation strategies for the removal of cyclic siloxanes from silicone fluid. This work focuses on a back end adsorption system to separate cyclic and linear siloxane oligomers. We first introduced a newly developed transferable molecular force field (FF) for cyclic and linear siloxanes. This force field is used to calculate adsorption of siloxanes in Metal-Organic Frameworks (MOFs) using molecular simulations. We then used this FF to explore the kinetic separation of cyclic siloxanes from linear siloxanes using MOFs by detailed molecular dynamics simulations. We determined a range of pore-limiting diameters that only allows the diffusion of linear siloxanes. We developed a strategy and workflow to investigate the equilibrium separation of linear and cyclic siloxanes in MOFs. We showed that configurational entropy effects drive the preferential adsorption of linear siloxanes over
cyclic siloxanes. We also explored vacuum-temperature swing adsorption (VTSA) supercritical CO2 as a potential method for the recovery of adsorbed linear siloxanes. Finally, we extended the work to investigate the potential of pure-silica zeolites for separating linear and cyclic siloxanes. We developed a new FF for siloxanes in pure-silica zeolites. This FF is used to calculate adsorption of siloxanes in pure-silica zeolites using molecular simulations. Overall, the developed geometric principles and screening strategies efficiently narrowed down a list of candidate MOFs and pure-silica zeolites, and identified promising adsorbents for further experimentation in collaboration with experimental partners.Ph.D.Chemical and Biomolecular Engineerin