Electrical and Computer Engineering

Permanent URI for this collectionhttps://uwspace.uwaterloo.ca/handle/10012/9908

This is the collection for the University of Waterloo's Department of Electrical and Computer Engineering.

Research outputs are organized by type (eg. Master Thesis, Article, Conference Paper).

Waterloo faculty, students, and staff can contact us or visit the UWSpace guide to learn more about depositing their research.

Browse

Now showing 1 - 20 of 2021

Constraining Robust Information Quantities Improves Adversarial Robustness
(University of Waterloo, 2024-12-11) Tan, Renhao
It is known that deep neural networks (DNNs) are vulnerable to imperceptible adversarial attacks, and this fact raises concerns about their safety and reliability in real-world applications. In this thesis, we aim to boost the robustness of DNNs against white-box adversarial attacks by defining three information quantities: robust conditional mutual information (CMI), robust separation, and robust normalized CMI (NCMI), which can serve as evaluation metrics of robust performance for a DNN. We then utilize these concepts to introduce a novel regularization method that constrains intra-class concentration and increases inter-class separation simultaneously among output probability distributions of attacked data. Our experimental results demonstrate that our method consistently enhances model robustness against C&W and AutoAttack on CIFAR and Tiny-ImageNet datasets, both with and without additional synthetic data. The results show that our approach enhances the robust accuracy of DNNs by up to 2.66% on CIFAR datasets and 3.49% on Tiny-ImageNet against PGD attacks, and by 1.70% on CIFAR and 1.63% on Tiny-ImageNet against AutoAttack, compared to several state-of-the-art adversarial training methods.
Towards a Robust Framework for Analyzing Random Telegraph Signals (RTS): Application to 2-level RTS in a Semiconductor Quantum Dot
(University of Waterloo, 2024-12-05) Bai, Tonghe
This thesis proposes a multi-stage, robust framework for analyzing Random Telegraph Signals (RTS), characterized by temporal fluctuations, within semiconductor research. The framework is tailored to applications across nanoscale solid-state technologies and addresses the growing need for precise and scalable RTS analysis as quantum and semiconductor technologies progress. Beyond the commonly used pulse-based measurements, this work focuses on extracting insights from steady-state measurements, enabling the study of non-equilibrium states—a critical aspect of understanding semiconductor phenomena. Specifically, the framework meets the demand of mitigating noise at high resolutions, as well as future support of real-time monitoring, and enabling automated tuning in devices such as quantum dots (QDs) and some nanoscale CMOS devices where RTS arises from single-carrier actions. Through stages of pre-processing, denoising, digitization and mean dwell time extraction, this method supports high-resolution analysis of RTS, addressing the complexities of experimental semiconductor data. The framework is validated with the real-world data of a specific QD holding 2-level RTS, demonstrating a robust 20-fold resolution increase, achieving a time bin reduction from 2 μs to 100 ns, with further explorations reaching 50 ns. This enhanced resolution uncovers hidden patterns within RTS. By accurately characterizing tunneling rates and transition dynamics, this research yields insights critical for high-fidelity quantum devices, potentially impacting applications like field-programmable gate arrays (FPGAs) and superconducting qubits, where RTS influences operational stability and performance. Beyond immediate applications, this thesis establishes a flexible RTS processing platform adaptable across various nanoscale semiconductor technologies. Future work will explore broader integration of theoretical and experimental insights to further enhance this framework, creating a versatile toolset aimed at improving robustness and adaptability in semiconductor and quantum devices operating in environments with complex noise and temporal fluctuations.
Reconfigurable Microwave/Millimeter-Wave Devices Using Liquid Crystal Technology
(University of Waterloo, 2024-10-15) Kianmehr, Hassan
In recent years, microwave liquid crystal (LC) technology has attracted significant attention from researchers due to its tunability and low-loss characteristics, extending up to terahertz frequencies. However, existing state-of-the-art LC-based devices face limitations in fabrication processes, resulting in relatively large sizes compared to other available tunable technologies. This thesis seeks to address this challenge by proposing a fabrication process for chip-scale and miniaturized LC-integrated devices. To achieve this goal, an alignment method is adopted for silicon micromachined devices, along with a comprehensive fabrication process for chip capacitors, reflective loads, and reflective-type phase shifters. Additionally, a tunable waveguide filter is designed and fabricated based on a control mechanism using a static magnetic field. The design and fabrication of silicon-micromachined variable capacitors are presented, utilizing nematic LC technology: shunt and series capacitors with and without integrated bias lines. The LC material enables electronic control over its dielectric properties, offering versatility across a broad spectrum of RF reconfigurable applications that require analog tuning. Measurement and simulation results for the chip LC shunt variable capacitor reveal a measured quality factor ranging from 44 to 123 at 1 GHz. With a biasing control voltage from 0 V to 40 V, the fabricated micromachined capacitor demonstrates an 18% capacitance shift. The LC-based series capacitor, demonstrated with an integrated bias line, isolates voltage control from RF terminals, serving a pivotal role in devices where series capacitors are essential. With a 21% shift in capacitance and a quality factor of up to 45 at 1 GHz, the capacitor’s performance is evaluated comprehensively through measurement and simulation. LC-integrated series capacitors tailored for applications without isolating bias voltage from RF terminals are demonstrated, yielding a notable 24% shift in capacitance and achieving a quality factor of up to 105 at 1 GHz. The demonstrated series capacitors feature a 10 μm thin layer of LC material, contributing to lower control voltage requirements and faster response times. These devices are manufactured through an in-house multi-layer microfabrication process. The thesis introduces a tunable waveguide filter that integrates LC material within quartz glass tubes, actuated by a static magnetic field. This innovative filter demonstrates a 7% tuning range with minimal bandwidth variation. Experimental validation for a 7.5 GHz filter with a 2.5% bandwidth confirms the concept’s viability. The quality factor of this filter varies between 108 and 288 in the fabricated sample. Tuning of the filter is first demonstrated using both a pair of rotating magnets; then, a pair of coil magnets is used to eliminate moving parts. Finally, two monolithically integrated reflective loads and two reflective-type phase shifters are presented, employing LC material as a reconfigurable element. The LC material is confined within a micromachined space, and its dielectric properties are controlled through an applied bias voltage. The tunable reflective loads find applicability in RTPSs. Operating at frequencies of 28 GHz and 62 GHz, the reflective loads exhibit phase variations of 113◦ and 118◦, respectively, as the bias voltage ranges from 0 V to 25 V. The 28 GHz and 62 GHz devices demonstrate reflective insertion losses of 3.9 dB and 4.3 dB, respectively, indicating figures of merit of 29◦/dB and 27.5◦/dB, respectively. Employing tandem hybrids at the operating frequency alongside two identical reflective loads has led to two reflective-type phase shifters at 28 GHz and 62 GHz. While the phase shift remains the same as the corresponding reflective loads, the insertion loss increases due to the use of hybrids. The insertion loss is measured at 5.95 dB and 7 dB for the 28 GHz and 62 GHz samples, respectively. Fabrication of these devices is conducted in-house using a multi-layer microfabrication process. To the best of our knowledge, this marks the first time a fully silicon-made, chip-level LC integrated reflective load and RTPS phase shifter is presented.
From Understanding Learning Difficulties Among Students To Providing High-Quality Automated Feedback
(University of Waterloo, 2024-09-25) Chen, Huanyi
Students face various difficulties during their learning journeys. However, providing timely feedback often poses a challenge for educators due to availability constraints. Fortunately, automated feedback systems have been introduced, offering invaluable assistance. To equip instructors with a general understanding of students in their teaching activities in computing education, we conducted an analysis of students' learning analytics to gain insights. In this study, we applied clustering techniques to behavior data naturally collected within an automated feedback system. We discovered that although students spent a significant amount of time using the system, the learning outcomes were often limited. A predictive model was derived based on these observations. To assist students in their learning, we explored whether offering trivial-penalty time extensions could be beneficial and why students use them. Implementing flexible late policies was straightforward and placed minimal burden on instructors. We analyzed a fourth-year course that utilized flexible late policies and found that time conflicts and underestimation of coursework were the top two reasons for utilizing time extensions. In addition, our findings revealed a correlation between students' abilities and their usage of time extensions. This latter result was re-examined in a replication study and a reproduction study. While the automated feedback system was not initially considered in the main study, in the reproduction study, we found that even with time extensions and automated feedback systems, low/middle-performing students still could not match the performance of high-performing students. This suggests a fundamental issue: feedback from automated feedback systems may not be as effective as anticipated, which plays an essential role in assisting students' learning at scale. Consequently, the critical question arises: how to provide effective feedback from automated feedback systems. We identified two main issues in current automated feedback systems: incorrect components marked as correct and correct components marked as incorrect. To address these issues, we argue that the unit testing philosophy, widely adopted in the software industry, should not be naively applied to automated feedback systems in an educational context. We completely redesigned the procedure and proposed a novel guideline for composing automated assessments. Following this guideline, we developed an automated assessment for an entity-relationship question in a database course. Our evaluation showed that students had significantly improved their understanding of the topic.
A Quantum Repeater Sandbox with Warm Atomic Memories and Quantum Dot Photon Sources
(University of Waterloo, 2024-09-23) Li, Michael
Quantum communication is known to offer many advantages, including opportunities for distributed quantum computing and more secure information transfer through quantum key distribution. This thesis provides background on how a quantum communication network can be achieved using quantum repeaters and how these components can be constructed with a hybrid system involving a quantum dot source and warm atomic memory. It also presents three experimental projects to realize critical components to the repeater: (1) The quantum dot photon source characterizations and tuning. (2) A compact 3D printed opto-mechanical laser locking board. (3) Optical memory observed in room temperature Cs vapor cells with an EIT memory scheme. These projects have built the basic foundation to create a repeater node using room temperature vapor cells and open the doors to future investigations of warm cell experiments.
Robust Reinforcement Learning for Linear Temporal Logic Specifications with Finite Trajectory Duration
(University of Waterloo, 2024-09-23) MortazaviMoghaddam, SeyyedSoroush
Linear Temporal Logic (LTL) is a formal behavioral specification language that offers a mathematically unambiguous and succinct way to represent operating requirements for a wide variety of systems, including autonomous and robotic systems. Traditional methods in this domain rely on model-checking approaches to ensure that a devised policy adheres to the provided specification. However, these methods are limited in the scope of problems they can solve and often lack generalizability to novel specifications and environments. Despite progress in synthesizing satisfying policies for LTL specifications under different operating conditions, learning policies that reliably satisfy complex LTL specifications in challenging environments remains an open problem. With the emergence of Machine Learning (ML) approaches, researchers have explored the use of ML-based techniques with LTL policy synthesis. Among the various approaches investigated, Reinforcement Learning (RL) has garnered particular attention for this task. While LTL specifications are evaluated over infinite-length trajectories, this work focuses on satisfying a class of specifications within a finite number of steps, as is to be expected in most real-world applications involving robotic or autonomous systems where the run-time of the robot is limited before it needs to recharge itself, e.g., a robot vacuum which has to perform certain cleaning tasks before recharge. Therefore, in this work, an RL-based technique is developed for the problem of generating trajectories of a system that satisfy a given LTLf specification in a system with finite (discrete) states and actions and a priori unknown transition probabilities modeled as a Markov Decision Process (MDP). The proposed approach builds upon the popular AlphaGo Zero Reinforcement Learning (RL) framework, which has found great success in the two-player game of Go, to learn policies that can satisfy an LTLf specification given a limit on the trajectory duration. In this thesis, first the motivation and the necessary background on the problem are provided, followed by a brief overview of existing methods. Then the problem statement is introduced, the proposed methodology and its variants are presented, and extensive simulations of complex robot motion planning problems are conducted and their results are explained. These simulations demonstrate how the approach achieves higher success rates under time constraints compared to state-of-the-art methods. The thesis concludes with a section discussing potential directions for future work and examining the results and their implications for the work completed.
Digital Twin-Enhanced Radar and Joint Communication-Sensing Systems: Application in Accurate Fall Severity Classification and Beyond
(University of Waterloo, 2024-09-20) Elbadrawy, Abdelrahman
The growing population of old seniors presents a significant challenge to healthcare systems worldwide. According to the United Nations Population Fund (UNFPA), the number of people aged 65 and older is 10.3% of the global population and is expected to reach 20.7% by 2074. The World Health Organization (WHO) reports that in the near future, by 2023, one in 6 people will be over the age of 60. This increase in the elderly population poses a serious challenge to healthcare providers at retirement homes because of the need to provide individual care to their residents. Falls stand out as the predominant cause of injury and death among the elderly. As stated by the National Council on Aging (NCOA), 1 in 4 Americans aged 65 and older fall each year, which equates to 14 million people. The NCOA also reports that the cost of treating injuries resulting from falls is expected to reach $101 billion by 2030. Moreover, the Center for Disease Control and Prevention (CDC) reports that repeated fall incidents double after falling once. This sets a significant burden on healthcare providers to assess the severity of falls and provide immediate care for those who are in need. In this study, we present a novel approach for fall detection, leveraging radar-based sensing systems as well as joint communication-sensing systems and advanced digital twin simulations. The choice of radar technology is rooted in its capability for high-resolution detection of micro-movements and its inherent respect for individual privacy, as it does not require visual imaging. Moreover, the choice of joint communication-sensing systems is motivated by the growing potential of 5G technology in enabling real-time sensing along with communication. Both systems have the capability for utilizing more physical resources, enabling greater resolution enhancement and more accurate detection. Both systems offer a non-intrusive and privacy-preserving solution for fall detection, ensuring the safety and dignity of the elderly. The integration of digital twins, replicating a diverse array of human physiology and fall dynamics, allows for extensive, varied, and ethical training of sophisticated machine learning algorithms without the constraints and ethical concerns of using human subjects. Our proposed methodology has led to significant advancements in the accuracy and sensitivity of detecting and assessing fall severity, especially in diverse populations and scenarios. We observed notable improvements in the system’s ability to discern subtle variations in falls, a critical factor in elderly care where such incidents can have serious health implications. Our approach not only sets a new benchmark in fall detection technology but also demonstrates the vast potential of combining radar and joint communication-sensing technology with digital simulations in medical research. This research paves the way for innovative patient monitoring solutions, offering a beacon of hope in improving senior care and proactive health management. In this study, the digital twin environment was created for both systems, radar and 5G, to simulate various fall scenarios under different conditions. For both systems, the simulated data was used to train machine learning models to detect the severity of falls, verifying the proposed methodology for severity of fall classification in an ideal environment. Furthermore, the correlation between the simulation and measurement results is presented. Measurement campaigns were conducted for both systems to validate the simulation results and to demonstrate the feasibility of the proposed methodology in real-world scenarios. Employing convolutional neural networks for the radar system, we obtained an accuracy of 99.45% using simulated data and 81.25% using measured data in detecting the severity of falls. The analysis addressed various parameters distinguishing different scenarios, including fall speed and the participant’s body size. On the other hand, for the 5G system, we achieved an accuracy of 92.46% using simulated data and 88.9% using measured data in detecting the severity of falls.
Integration of Green Ammonia into Smart Grids: Neural Network-Based Modeling and Control for Direct Ammonia Synthesis and Fuel Cells
(University of Waterloo, 2024-09-18) Syed, Miswar
The current green ammonia production method involves generating green hydrogen via an electrolyzer and combining it with nitrogen through the Haber Bosch (HB) process to produce ammonia. A newer method, Direct Ammonia Synthesis (DAS), is gaining attention as it can produce green ammonia directly using an Electrochemical Ammonia Synthesizer (EAS) without the electrolyzer and HB system, significantly reducing costs and energy consumption. The produced ammonia can be directly converted to electricity using Direct Ammonia Fuel Cells (DAFC). Additionally, ammonia addresses hydrogen-related issues such as high flammability, poor volumetric density, and high storage costs. First, the thesis focuses on the DAS approach. It explores the integration of EAS and DAFC into the grid as a means to provide stable power through the utilization of various power smoothing filters. The EAS converts excess wind/solar power into green ammonia, which is then used by DAFC to produce electricity during power deficits. Second, a novel neural network (NN) model for EAS is developed to simplify the traditionally complex and sensor-intensive modeling of electrochemical systems. This NN model accurately predicts ammonia production based on solar power, nitrogen, and water inputs. Third, an NN model for DAFC is created to output electrical power from ammonia. Both EAS and DAFC NN models can be integrated into the existing microgrid system models in MATLAB-Simulink and Python. Finally, the thesis introduces a Neural Network-based Model Predictive Control (NNMPC) approach for regulating EAS output and meeting the ammonia demand, which demonstrates superior accuracy and efficiency compared to the traditional fuzzy logic control method. Unlike a traditional MPC, which uses a mathematical plant model for predictive optimization, an NN model demonstrates superior accuracy in encapsulating plant dynamics. The NNMPC addresses mathematical intricacies in MPC models, especially as plant complexities increase. Simulation results confirm the effectiveness of the NN models and NNMPC in practical applications. The research conducted in this thesis has resulted in journal and conference research publications as well as a collaborative project with a Waterloo-based company.
Quantum Dot Mediated Bead-Based Assays for Continuous Biomarker Monitoring in Diabetes and Ex Vivo Lung Perfusion Systems
(University of Waterloo, 2024-09-17) Srikant, Sanjana
The continuous monitoring of biomolecules such as hormones, polypeptides, and monosaccharides can provide crucial information about biological systems, leading to a better understanding of human physiology and an improved ability to diagnose and treat complex, multisystemic illnesses. There is a lack of technological advancements that can measure a diverse range of biomarkers with the necessary sensitivity, specificity, and temporal resolution at physiologically relevant concentrations, in a multiplexed manner. Traditionally, enzyme-linked immunosorbent assays (ELISAs) have been used to detect protein biomarkers, while small carbohydrate biomarkers like glucose and lactate levels are measured enzymatically. However, these methods are time consuming, do not offer real-time or multiplexed detection, and require substantial amounts of reagents, sophisticated detection instrumentation, and trained personnel to perform. This study focuses on developing assays for use in two diseases with significant clinical impact: diabetes, where the detection of hormones insulin and glucagon is demonstrated, and ex vivo lung perfusion, where the detection of small molecules lactate and glucose is shown. The system utilizes Quantum Dot mediated Bead-Based Assays (BQA) with microfluidic modules to create a real-time ELISA (QIRT-ELISA) platform for continuous and multiplexed detection of diverse biomarkers in heterogeneous biological matrices. This system surpasses traditional ELISAs by significantly improving sensitivity, specificity, and measurement times, while also enabling multiplexing through the use of quantum dots and their unique optical properties to extract multiple discrete emission spectra signals from targets of interest with a single UV excitation laser. Validation experiments demonstrate that QIRT-ELISA can detect insulin and glucagon in the low picomolar range in whole blood, and lactate and glucose in the low millimolar range in ex vivo lung perfusate. Continuous measurements of insulin and glucagon in vivo in rats undergoing glucose tolerance tests further validate the platform, showing comparable results to conventional ELISA. Thus, the platform presents unique advancements in continuous monitoring technology that can be applied clinically to improve personalized precision medicine and enhance patient outcomes. We believe that the platform can be further developed to encompass a wider range of biomarkers, making it universally applicable across various clinical settings.
Designing a Tightly Coupled Ultra-Wideband Millimeter-Wave Phased Array Antenna
(University of Waterloo, 2024-09-17) Rahmati, Nazanin
The advent of 5G and 6G wireless communication networks has significantly increased interest in Ultra-Wideband (UWB) phased array antennas operating in the millimeter-wave (mm-wave) spectrum, particularly within the 24-71 GHz range. Tightly Coupled Arrays (TCAs) have emerged as a promising candidate for planar, low-profile UWB phased array implementations. Despite the wideband capabilities of Tightly Coupled Dipole Arrays (TCDA), a primary challenge remains designing compact, low-profile baluns suitable for integration within small unit cells at high mm-wave frequencies. This thesis addresses this challenge by intro- ducing a novel, planar Marchand balun-based antenna element that achieves an acceptable Voltage Standing Wave Ratio (VSWR) across the 22.6 to 72.6 GHz band while maintaining a thin profile suitable for mobile devices. In addition, a wideband bowtie element, utilizing the planar Marchand balun, is developed to cover two key 5G mm-wave frequency bands (24-29 GHz and 37-43 GHz). The performance of these elements is evaluated through the design, fabrication, and measurement of a small 2x2 antenna array. In this project, in order to provide more space for accommodating the complex feeding network and reduce the cost of the array by reducing the number of required transceiver (T/R) modules, a thinned 8x8 TCDA is proposed in which only 22 elements are excited. By strategically terminating a part of unexcited elements (20 elements) and leaving the rest as open-circuit (22 elements), the array leverages mutual coupling to induce appropriate currents and maintain acceptable radiation characteristics. The scheme of excitation is obtained by investigating the current distribution among elements and with the help of optimization. Genetic Algorithm (GA) optimization is employed to determine the optimal excitation voltages and terminating impedances for broadside radiation and scanning up to 45 degrees in the E-plane. An innovative, low-profile distributed circuit, comprising a cascade of striplines ter- minated to a resistance, is designed to realize the required terminating impedances. Ad- ditionally, a wideband matching network is developed for each excited port using the Real Frequency Technique (RFT) to ensure impedance matching across the operating fre- quency range and steering angles. The thinned 8x8 TCA, including the antenna array, terminating impedances, matching networks, and feedlines, is simulated using Ansys High Frequency Structure Simulator (HFSS). Simulation results demonstrate the effectiveness of the thinned TCDA concept, achieving acceptable input reflection coefficients at all ex- cited ports and maintaining desired Side Lobe Level (SLL) and directivity across the entire frequency range and steering angles.
Fabrication of atomic force microscope probes with high aspect ratio silicon tips, silicon/silicon nitride cantilevers and stair-shaped handles
(University of Waterloo, 2024-09-10) Pan, Aixi
The atomic force microscope (AFM) is a versatile tool with promising applications in biomedical detection, optical spectroscopy, and material characterization. However, its widespread utilization faces challenges due to limitations in conventional fabrication methods of commercially available probes. Specifically, the standard tips may not meet the requirements for scanning deep or narrow structures accurately, and the rectangular handle design can block a portion of the reflected laser signal, leading to inconsistent feedback. Overcoming these limitations is crucial for expanding the utility and unlocking the full potential of AFM in diverse scientific and technological applications. In this thesis, we propose innovative strategies to enhance the scanning performance of AFM probes with high aspect ratio (HAR) tips and stair-shaped handles. Chapter 3 explores four distinct silicon (Si) AFM tip fabrication methods, each offering unique advantages and contributions to the field. The first method (Section 3.1) integrates non-switching pseudo-Bosch etching with wet sharpening techniques to achieve an exceptional aspect ratio of 1:135, marking a significant advancement in tip fabrication. The second method (Section 3.2) introduces an innovative approach utilizing tapered silicon oxide (SiO2) masks to fabricate Si nanocones with extraordinary apex measuring just 3.54 nm. The third method (Section 3.3) explores a novel two-step cryogenic etching process to yield a controllable tip profile with a slight taper angle of 2.2°. The fourth method (Section 3.4) combines the Bosch process with the pseudo-Bosch process, incorporating periodic oxygen (O2) shrinking steps. This approach achieves a remarkable tip apex sharpness of 20 nm, pushing the boundaries of nanofabrication capabilities. Chapter 4 details the fabrication of a stair-shaped Si handle to mitigate laser blocking. Two techniques are described: one leveraging the loading effect and RIE-lag to attain stage heights of 71 μm, 151 μm, 168 μm, and 287 μm, while the other employs pseudo-grayscale lithography with a titanium (Ti) mask, yielding final stages of 52 μm, 83 μm, 161 μm, and 211 μm. Both these methods are applicable for practical AFM probe fabrication without laser blocking. Chapter 5 delves into the mass fabrication of all-Si HAR AFM probes, merging tips fabricated by the O2 shrinking method with handles fabricated using the loading effect and RIE-lag. Furthermore, Chapter 6 explores the adoption of silicon nitride (SiNx) as an alternative to Si for cantilevers. Amorphous low-pressure chemical vapor deposition (LPCVD) SiNx offers benefits such as low spring constants and precise deposition control, resulting in thin and low-spring constant cantilevers. This configuration minimizes damage to the sample and tip, making it ideal for delicate samples. By combining a SiNx cantilever with a Si tip, the probe capitalizes on both tip apex and thin cantilever advantages, facilitating accurate AFM imaging with high resolution while preventing false images on fragile samples.
Fabrication of Silicon Out-of-Plane Microneedles for Potential Drug Delivery and Interstitial Fluid Extraction
(University of Waterloo, 2024-09-10) Hu, Wenhan
Microneedles represent a successful application of MEMS technology, forming minimally invasive platforms for transdermal drug delivery, body fluid sampling, and diagnostics. Silicon microneedles, in particular, are favored due to their exceptional mechanical strength and biocompatibility. This thesis focuses on the three different fabrication methods of silicon microneedles using MEMS techniques. Initially, we fabricated silicon out-of-plane cone-shaped hollow microneedles with sharp apexes and off-axis pores. This process involved backside hole etching and frontside pillar etching via the Bosch process, followed by pillar sharpening using a HF-HNO3 mixed solution. The resulting microneedles were 160 μm high. However, to penetrate the epidermis and access abundant body fluids for health monitoring systems, taller microneedles longer than 500 μm are required. Fabricating these higher microneedles proved challenging due to difficulties in achieving uniform sharpening through wet etching. To address this, we developed a novel method for fabricating silicon out-of-plane hollow microneedles with beveled tips. This method included frontside slope etching, backside hole etching, and frontside pillar etching, combining anisotropic wet etching and dry etching (Bosch process). The resulting microneedles were approximately 600 μm tall with beveled sharp tips. We tested various fundamental functions of these microneedles by connecting the chip to a syringe using a 3D-printed applicator, successfully demonstrating liquid extraction, liquid injection, and simulated drug delivery process. To minimize the impact of inevitable lateral etching during frontside pillar etching in the Bosch process, we proposed sacrificial structures surrounding the pillars to shield them from lateral etching. Testing two types of sacrificial structures, we found both structures could effectively reduce lateral etching, enabling the fabrication of 370 μm high ring pillars with vertical sidewalls. Additionally, grayscale lithography combined with subsequent Bosch processing presents an effective and flexible method for fabricating complex 3D structures like bevels. We first acquired contrast curve for the photoresist before grayscale lithography. Then we used this technique integrating frontside slope etching and frontside pillar etching into a single step, resulting in the fabrication of hollow microneedles measuring 325 μm in height.
Verifying Explanations for Neural Networks
(University of Waterloo, 2024-09-05) Le, Nham Van
Deep neural networks (DNNs) have been applied in solving many complex tasks across multiple domains, many of which have direct effects on our daily lives: generative models are replacing traditional search engines for answering questions, cars are being driven by neural networks, doctors and radiologists are using neural nets to diagnose patients more efficiently, financial systems are run by automated trading bots, etc. Coupled with the ever-increasing of DNNs' complexity, the need for explaining their predictions and verifying their safety is clear. Generally speaking, verifying a DNN involves checking if it behaves as expected for unseen inputs in a particular region, and explaining a DNN involves interpreting the network's prediction on a given input. Both approaches have their own pros and cons: the output of any input in a verified region is proven to be correct (with respect to a spec), but such regions are minuscule compared to the whole input space, not just because of the performance of the tools, but because of the inherent limits in e-robustness -- the commonly used verification specifications; and while explanation methods can be applied to explain the output given any input, they are post-hoc and hard to judge: does an explanation make sense because the DNN is working close to how a human being process the same input, or because the explanation visualizes the input itself without taking the model in consideration? Our main insight: we can combine both verification and explanation, resulting in novel verification problems towards a robust explanation for neural networks. However, any verification problem (or specification) can not exist in isolation, but in a symbiosis relationship with the tools solving it. When we propose a new spec, it is expected that existing tools cannot solve it effectively, or may not work at all. Interesting problems push developers to improve the tools, and better tools widen the design space for researchers to come up with even more interesting specs. Thus, in this proposal, we are introducing not just novel specifications, but how to solve them by building better tools. This thesis presents a series of results and research ideas based on that insight. First, we show that by extending e-robustness with an explanation function (the activation pattern of the DNN), we can verify a bigger region of the input space using existing verification tools. Second, by verifying the explanation functions, we provide a robust way to compare different explanation methods. Finally, even when the combination of existing DNNs' verification specs and explanation functions is friendlier to existing verification tools, we still run into scalability issues as we increase the size of the networks. Thus, in this thesis we also present our results on building a distributed SMT solver, which lies at the heart of many neural network verification tools.
Improving Knowledge Distillation by Training Teachers to maximize Their Conditional Mutual Information
(University of Waterloo, 2024-09-03) Ye, Linfeng
Knowledge distillation (KD) and its variants, as effective model compression methods, have attracted tremendous attention from both academia and industry. These methods usually use the pretrained teacher models' outputs and the ground truth labels as supervision signals to train the lightweight student model, which can improve student performance in terms of accuracy. One aspect of KD that has rarely been explored in the literature is how the behavior of the teacher models affects the students' performance. Specifically, in most existing KD projects, teacher models are usually trained to optimize their own accuracy. However, recent studies have shown that a teacher with higher accuracy does not always lead to a student with higher accuracy \cite{cho2019efficacy, stanton2021does}. To explain the aforementioned counter-intuitive observations and advance the understanding of the role of teacher models in KD, the following research problem naturally arises: \textit{How can a teacher model be trained to further improve student's accuracy in scenarios where the teacher is willing to allow its knowledge to be transferred to the student in whatever form?} In this thesis, we assert that the role of the teacher model is to provide contextual information to the student model during the KD process. In order to increase the contextual information captured by the teacher model, this thesis proposes a novel regularization term called Maximum Conditional Mutual Information (MCMI). Specifically, when a teacher model is trained by conventional cross-entropy loss plus MCMI, its log-likelihood and conditional mutual information (CMI) are simultaneously maximized. The new Class Activation Mapping (CAM) algorithm further verified that maximizing the teacher’s CMI value allows it to capture more contextual information in an image cluster. Via conducting a thorough set of experiments, we show that by employing a teacher trained by CE plus MCMI rather than one trained CE in various state-of-the-art KD frameworks, student's classification accuracy consistently increases, with a gain of up to 3.32\%. In addition, we show that such improvements in the student's accuracy are more drastic in zero-shot and few-shot settings. Notably, when 5\% of the training samples are available to the student (few-shot), the student's accuracy increases with the gain of up to 5.72\%, and increases from 0\% to as high as 84\% for an omitted class (zero-shot).
Low Power TFT Logic Implementation on Display Backplane, Using Unipolar TFTs
(University of Waterloo, 2024-08-30) Kapar, Sparsh
Active-Matrix (AM) displays are key components in devices such as computer monitors, smartphones, laptops, portable gaming consoles, and wearable devices. These displays are trending towards becoming the primary display technology in today’s market. The demand for current displays has surpassed 4K (4096 rows by 2160 columns) and have even reached 8K Ultra-High-Definition (UHD, 7680 rows by 4320 columns). The display panel is generally fabricated on low-cost thin-film transistors (TFT), while the driver and control circuits are built using conventional Complementary-Metal-Oxide-Semiconductor (CMOS) circuits. A bonding pad serves as interface between CMOS circuits and TFT display panel. For each row and column of pixels added to the display backplane, an additional bonding pad needs to be added to properly interface the off-panel peripheral row and column control circuit with TFT and OLED pixel array. As the pixel density of the display increases, the bonding pad pitch must decrease, to accommodate. However, the pitch can only reduce finitely, and this imposes a bottleneck on achieving high-density large-scale displays. Additionally, with each row-line connected to a gate of thousands of pixels, there is a high capacitive load, which contributes to a high dynamic power consumption. Recent research has been investigating the use of TFTs to change the off-panel integrated-circuit (IC) and integrate it onto the display backplane, eliminating the need for bonding pads, while also reducing the overall power consumption. Amorphous silicon (a-Si:H) TFTs have good uniformity and low mask count fabrication process, making it suitable for large-scale displays, in comparison to Low-Temperature-Polysilicon (LTPS) TFTs. However, a-Si:H TFTs are naturally unipolar, which make it hard to replicate the CMOS like logic that the off-panel ICs have. This thesis aims at tackling the bottlenecks of large-scale displays, that is, the high dynamic power consumption, and the limited space from the bonding pads. The proposed row driver circuit presented in this thesis can be used to eliminate the off-panel row IC, and be integrated into the display backplane, while reducing the dynamic power consumption, with a-Si:H TFT technology.
Fast and Efficient Calibration for Phased Arrays and Beamforming Circuits using Novel Constellation Characterization Method and Channel Transformation Technique
(University of Waterloo, 2024-08-30) Chen, Yuxuan
Traditional phased array calibration methods, like exhaustive search, are time-consuming. This thesis proposes a novel, efficient technique that significantly reduces calibration time without sacrificing performance. The proposed approach utilizes a constellation characterization method. It extracts and describes element responses using a set of mapping functions based on a strategically chosen data subset. A custom solver then generates control codes for desired beam patterns. For robustness, a closed-loop calibration routine is introduced to verify the solutions. Additionally, a taper awareness scheme is incorporated to optimize the output power by accounting for element variations and the desired tapering profile. The proposed method demonstrates significant speed-up compared to the exhaustive search method. On two beamforming integrated circuits (BFICs) and two test arrays, it achieves improvements of up to 1100 times while maintaining performance. Furthermore, a channel transformation technique is proposed to leverage similarities between array elements. This technique reduces measurements by transferring mapping functions between array elements, avoiding full characterization for each element. A sequencing technique is also introduced to optimize the transformation route, maximizing success rates and further minimizing measurements. Experimental validation shows significant reductions in addition to the savings achieved via the proposed characterization method. Compared to the exhaustive search method, reductions of up to 3000 times are achieved. This thesis presents significant advancements to phased array calibration, paving the way for efficient and scalable solutions in future high-resolution massive multiple input and multiple output (MIMO) systems.
Resource Management for Vehicular See-through Service Provisioning
(University of Waterloo, 2024-08-29) Wang, Ruoxu
Vehicular see-through service is one of the driver assistance services expected to be provided by smart vehicles in the future. It is envisioned that this service will expand the vision of a smart-vehicle driver by timely alerting the driver to obstacles in the front blind spots. In this study, we investigate cooperative perception and resource allocation in the provision of the vehicular see-through service to ensure that drivers are able to obtain accurate and timely environment information about their front blind spots. To cope with the consumption of communication and computing resources associated with the transmission and processing of multi-view point clouds, we propose a cooperative sensor data collection and processing scheme. In this scheme, smart vehicles cooperatively collect point cloud data for each object and complete the point cloud processing at the on-board computing unit. For each object, we select a set of collectors to collect the point cloud data and an analyzer to process the data. To address the tradeoff between classification accuracy and resource efficiency, we investigate the impact of the location of cooperative vehicles on object classification accuracy and propose an indicator to assess the quality of the collector sets. We develop a collector set pre-selection algorithm that identifies all collector sets for each object, to satisfy the classification accuracy and have minimal data redundancy. To address the tradeoff between the service requirements and the overall resource consumption of the system, we treat each potential collector set and its analyzer of an object as a worker pair and select the worker pair for each object based on its resource consumption. Taking into account the resource competition among worker pairs, we develop a joint vehicle selection and resource allocation algorithm based on ant colony optimization to minimize the overall resource consumption, while satisfying the delay requirements of the service. Simulation results demonstrate that our proposed service provisioning scheme outperforms benchmark schemes in terms of resource efficiency, task completion rate and request completion rate.
Nonclassicality of Propagating States From 3-Photon Interactions in a Superconducting Parametric Cavity
(University of Waterloo, 2024-08-28) Jarvis-Frain, Benjamin
The multimode superconducting parametric cavity has proven to be a powerful and versatile system for producing nonclassical states of light in the microwave regime. Utilizing its ability to realize nonlinear and multimode Hamiltonians, we can produce strongly correlated propagating signals from the cavity at different frequencies including entangled photons. In this thesis, we study the generation of photon triplets using a cubic Hamiltonian in the cavity under a parametric drive. We demonstrate the implementation of 3-photon Spontaneous Parametric Down-Conversion (SPDC) into different frequency modes of the cavity and study the non-Gaussian statistics of the outputted photon triplets through purely linear detection. We detail our methodology for performing absolutely calibrated measurements of the cavity output using a Shot Noise Tunnel Junction (SNTJ) as well as our use of a near quantum-limited Travelling-Wave Parametric Amplifier (TWPA). In addition to the primary results of this thesis, we present calibrated measurements of the noise temperature of the Crescendo TWPA from QuantWare. Through the use of this TWPA and SNTJ, we are able to obtain the correlations between frequencies with low uncertainty up to the 4th moments. From these moments, we can compute a nonlinear entanglement witness on the propagating triplets and demonstrate the non-Gaussian genuine tripartite entanglement between photons with over 15 sigmas of certainty.
Modeling and Simulation of Multilayer MoS2 Schottky Barrier Field-Effect Transistors
(University of Waterloo, 2024-08-28) He, Zhuoyang
As a member of the transition metal dichalcogenide (TMDCs) family, Molybdenum Disulfide (MoS2) exhibits a layered 2D structure with exceptional electronic and optoelectronic properties, making MoS2 a promising candidate for next-generation nano-devices. However, despite numerous efforts in synthesis, fabrication process, and device structure for few-layer MoS2. A key challenge that remains is the Fermi-pinning effect. Due to defects distributed at the MoS2/metal interface, the Schottky-mott rule fails to predict the Schottky barrier height based on the Fermi-level difference between MoS2 and the metal. Instead, the Fermi level of the metal is pinned near the conduction band edge regardless of the work function of the metal used. This phenomenon results in an inevitable Schottky barrier, which must be recognized in device simulations. In this thesis, the electrical and optoelectronic performance of multilayer MoS2 field-effect transistors is predicted and analyzed through simulation techniques. Drift-diffusion equations are employed to model electronic properties, utilizing finite element methods (FEM) to solve the corresponding partial differential equations. FEM discretizes space and divides the solution domain into finite elements. For optoelectronic simulations, When solving Maxwell's equation for optical absorption and carrier generation rates. Finite-Difference Time Domain (FDTD) methods are applied, which discretize both space and time, representing fields as discrete values on a grid in both dimensions. We examine the effect of Schottky barrier height on MoS2-based devices and complete the missing p-branch of MoS2 SBFETs due to the Fermi-pinning effect. Initially, we verify our model against experimental data, results prove the capability of our model to predict the electrical performance of Schottky barrier FETs. By varying the Schottky barrier height from 0 eV to 1.3 eV across the MoS2 bandgap, a transition from n-type transport to p-type transport is observed. However, the ambipolar transport is limited by the relatively large bandgap. Ambipolarity is enabled through asymmetric metal configurations, as evidenced by simulations showing that the Pd-Au configuration can achieve ambipolar behavior with currents comparable to symmetric MoS2-based FETs. Furthermore, we simulate dual-gated FETs by incorporating an additional gate, allowing for reconfigurability between NP and PN configurations. These devices exhibit an outstanding rectification ratio that can be optimized under low gate voltage conditions. Nevertheless, the asymmetry in performance between PN and NP configurations indicates the significant impact of metal choices. The successful establishment of p-n junction in the dual-gated devices illustrates their potential as photodetectors. Simulation results indicate a photoresponsivity of 24.2 mA/W for PN configuration.
Predictable Cache Subsystem Design for Real-time Multicores
(University of Waterloo, 2024-08-23) Wang, Xinzhe
This thesis focuses on the design of the cache subsystem for multicore real-time systems. Specifically, it looks at the two common contention sources in the cache subsystem that undermine the system's timing predictability: temporal contention (queuing delays resulting from simultaneous shared resource accesses) and spatial contention (conflicts in cache line allocations). Accordingly, this thesis proposes two solutions to address the two problems. First, to address temporal contention induced by the shared last-level cache (LLC), we propose PECC, a predictable exclusive cache coherence mechanism. Unlike the common choice of inclusive cache hierarchies, PECC incorporates an exclusive LLC free of back invalidation, which is a significant contributor to the per-request worst-case latency (PR-WCL) in inclusive hierarchies. As a result, PECC reduces the bound of PR-WCL by 6% and improves the average performance by 2.33 times over the predictable solution with an inclusive LLC. Second, to address spatial contention in the shared LLC, we propose ParRP, a novel cache partitioning scheme that provides cache space isolation for shared data. Conventionally, achieving cache space isolation for shared data is challenging due to its intrinsic sharing nature, which violates the resource isolation principle. ParRP overcomes this by partitioning the replacement policy instead of cache entries. Consequently, ParRP guarantees isolation property even with shared data, enabling isolated worst-case execution time (WCET) analysis without interference from other cores. Our evaluation shows that this leads to a 2.4 times reduction in the WCET of multi-threaded tasks with shared data at the cost of a 16.5% decrease in average performance.

Browse

Recent Submissions