Month: June 2021

Title: Advanced Packaging and Thermal Management of DC-DC Converters and Novel Correlations for Manifold Microchannel Heatsinks

Author: Sevket Umut Yuruker

Date/Time: July 2nd, 3:00pm – 5:00pm

Zoom link: https://umd.zoom.us/j/4052654764

Committee Members:
Professor Michael Ohadi, Chair
Professor Patrick McCluskey
Professor Jungho Kim
Professor Amir Riaz
Professor Christopher Cadou, Dean’s Representative

Abstract: An advanced packaging configuration of a dual-active-bridge 10 kW DC-DC converter module is introduced in this dissertation. Through utilization of novel heatsinks for the power switches and the transformer assembly, ~20 kW/Lit converter volumetric power density based on numerical and experimental analysis is obtained. Through a unique placement of the high power/high frequency SiC switches on the printed circuit board, many beneficial features such as double-sided cooling, complete elimination of wirebonding, circumvention of the need for TIM layers between the switches and the heatsinks, and multi functioning heatsinks as electrical busbars is achieved.

A Vertically Enhanced Manifold Microchannel System (VEMMS) cooler is developed to address the thermal challenges of a pair of power switches, simultaneously. Both air and liquid cooled versions of VEMMS cooler is presented, thermal resistances of 1.1 K/W and 0.3 K/W for the air and liquid cooled versions, respectively, at reasonable flow rates and pressure drops was obtained. Besides the power switches, thermal management of the transformer assembly is accomplished via Combined Core and Coil (C3) Coolers, where both the magnetic core and coils are liquid cooled simultaneously with electrically insulating but thermally conductive 3D printed Alumina heatsinks, where thermal resistances as low as 0.3 K/W for the magnetic core and 0.09 K/W for the transformer windings is experimentally demonstrated. Furthermore, a system level model was built to investigate the effect of various components in the cooling loop on each other, and what are the limiting factors to prevent a possible thermal runaway failure.

Lastly, using a metamodeling approach, closed form pressure drop and heat transfer correlations are developed for thermo-fluidic performance prediction of manifold microchannel heatsinks. Due to complexity and vastness of design variables present in manifold microchannel systems, adequate CFD analysis and optimization require significant computational power. Through utilization of the developed correlations, orders of magnitude reduction in computational time (from days to milliseconds) in prediction of pressure drop and heat transfer coefficient is demonstrated. Extensive mesh independence and residual convergence algorithms are developed to increase the accuracy of the created database. Between the correlation predictions and mesh independent CFD results for the entire metamodel range, a mean error of 3.9% and max error of 24% for Nusselt number, and a mean error of 4.6% and max error of 37% for Poiseuille Number are achieved.

Title: A Bayesian Network Perspective on the Elements of a Nuclear Power Plant Multi-Unit Seismic Probabilistic Risk Assessment.

Author: Jonathan DeJesus Segarra

Day/Time: Friday, July 16, 2021 at 1:00 p.m. (Eastern Daylight Time)

Examining Committee:
Dr. Michelle Bensi
Dr. Mohammad Modarres
Dr. Gregory Baecher
Dr. Katrina Groth
Dr. Robert Bunditz
Dr. Vedran Lekic

ABSTRACT

Nuclear power plants (NPPs) generated about 10% of the world’s electricity in 2019 and about 1/3 of the world’s low-carbon electricity production. Probabilistic risk assessments (PRAs) are used to estimate the risk posed by NPPs, generate insights related to strengths and vulnerabilities, and support risk-informed decisionmaking related to safety and reliability. While PRAs are typically carried out on a reactor-by-reactor basis, the Fukushima Dai-ichi accident highlighted the need to also consider multi-unit accidents. To properly characterize the risks of reactor core damage and subsequent radiation release at a multi-unit site, it is necessary to account for dependencies among reactors arising from the possibility that adverse conditions affect multiple units concurrently. For instance, the seismic hazard is one of the most critical threats to NPP structures, systems, and components (SSCs) because it affects their redundancy. Seismic PRAs are comprised of three elements: seismic hazard analysis, fragility evaluation, and systems analysis.

This dissertation presents a Bayesian network (BN) perspective on the elements of a multi-unit seismic PRA (MUSPRA) by outlining a MUSPRA approach that accounts for the dependencies across NPP reactor units. BNs offer the following advantages: graphical representation that enables transparency and facilitates communicating modeling assumptions; efficiency in modeling complex dependencies; ability to accommodate differing probability distribution assumptions; and facilitating multi-directional inference, which allows for the efficient calculation of joint and conditional probability distributions for all random variables in the BN. The proposed MUSPRA approach considers the spatial variability of the ground motions (hazard analysis), dependent seismic performance of SSCs (fragility evaluation), and efficient BN modeling of systems (systems analysis). Considering the spatial variability of ground motions represents an improvement over the typical assumption that ground motions across a NPP site are perfectly correlated. The method to model dependent seismic performance of SSCs presented is an improvement over the current “perfectly dependent or independent” approach for dependent seismic performance and provides system failure probability results that comply with theoretical bounds. Accounting for these dependencies in a systematic manner makes the MUSPRA more realistic and, therefore, should provide confidence in its results (calculated metrics) and risk insights.

Title: Towards Trust and Transparency in Deep Learning Systems through Behavior Introspection & Online Competency Prediction

Author: Julia Filiberti Allen

Date/Time: Monday, June 28th at 1pm

Committee Members:
Professor Steven A. Gabriel, Chair, Dept. of Mechanical Engineering
Professor Dinesh Manocha, Dean’s Representative, Dept. of Computer Science
Professor Jin-Oh Hahn, Dept. of Mechanical Engineering
Professor Shapour Azarm, Dept. of Mechanical Engineering
Professor Jeffrey Herrmann, Dept. of Mechanical Engineering

Abstract:
Deep neural networks are naturally “black boxes”, offering little insight into how or why they make decisions.  These limitations diminish the adoption likelihood of such systems for important tasks and as trusted teammates.  We employ introspective techniques to abstract machine activation patterns into human-interpretable strategies and identify relationships between environmental conditions (why), strategies (how), and performance (result) on both a deep reinforcement learning two-dimensional pursuit game application and image-based deep supervised learning obstacle recognition application. Pursuit-evasion games have been studied for decades under perfect information and analytically-derived policies for static environments.  We incorporate uncertainty in a target’s position via simulated measurements and demonstrate a novel continuous deep reinforcement learning approach against speed-advantaged targets.  The resulting approach was tested under many scenarios and performance exceeded that of a baseline course-aligned strategy.  We manually observed separation of learned pursuit behaviors into strategy groups and manually hypothesized environmental conditions that affected performance.  These manual observations motivated automation and abstraction of conditions, performance and strategy relationships.  Next, we found that deep network activation patterns could be abstracted into human-interpretable strategies for two separate deep learning approaches.  We characterized machine commitment by the introduction of a novel measure and revealed significant correlations between machine commitment, strategies, environmental conditions, and task performance.  As such, we motivated online exploitation of machine behavior estimation for competency-aware intelligent systems.  And finally, we realized online prediction capabilities for conditions, strategies, and performance.  Our competency-aware machine learning approach is easily portable to new applications due to its Bayesian nonparametric foundation, wherein all inputs are compactly transformed into the same compact data representation.   In particular, image data is transformed into a probability distribution over features extracted from the data. The resulting transformation forms a common representation for comparing two images, possibly from different types of sensors.  By uncovering relationships between environmental conditions (why), machine strategies (how), & performance (result) and by giving rise to online estimation of machine competency, we increase transparency and trust in machine learning systems, contributing to the overarching explainable artificial intelligence initiative.

Title: INTELLIGENT INTERSECTION MANAGEMENT THROUGH GRADIENT-BASED MULTI-AGENT COORDINATION OF TRAFFIC LIGHTS AND VEHICLES

Author: Manuel Aurelio Rodriguez

Date/Time: July 7, 2021 1pm-3pm

Zoom Meeting: https://umd.zoom.us/j/5586796034

List of committee members
Dr. Hosam Fathy (Chair)
Dr. Yancy Diaz-Mercado
Dr. Shapour Azarm
Dr. Jin-Oh Hahn
Dr. Derek Paley (Dean’s Representative)

Abstract
This dissertation examines the problem of coordinating both traffic lights plus connected and autonomous (CAVs) vehicles in urban traffic. We propose a coordination strategy for CAVs and smart traffic lights based on combining ideas from gradient-based multi-agent control, trajectory planning and control barrier functions.

The work is motivated by an extensive previous literature showing that traffic network synchronization has the potential to alleviate congestion while reducing fuel consumption and delay. The literature presents many algorithms for coordinating the traversal of intersections by connected and automated vehicles (CAVs), as well as the synchronization of traffic lights. However, the integrated solution of these two synchronization problems remains relatively unexplored. One of the main challenges of any algorithm proposed in this area consists of managing the trade-off between computational efficiency, communication requirements, and performance.

In this dissertation, the overall proposed control framework consists of coordinating the timing of the agents through decentralized gradient-based control, using novel potential energy functions that encode the desired interactions between the timing of connected agents. The negotiated timing is then achieved through acceleration minimizing trajectory planning. Finally, feasibility and safety constraints are handled by a regulator that uses control barrier functions. The strategy is validated in simulation for different types of intersections and traffic scenarios. We show potential savings in fuel and time of up 15% and 70% respectively.