Emerging threats in modern processor design underscore the need for robust hardware security measures, particularly in scenarios where unintended information leakage can compromise sensitive data (Kocher et al.; Lipp et al., 2020). Information Flow Tracking (IFT) has become a critical technique to enforce confidentiality and integrity requirements by identifying unintended data propagation paths (Hu et al., 2022). However, existing IFT approaches often introduce considerable overhead, limiting their applicability in performance- and cost-sensitive domains.

In this work, we propose a custom IFT solution integrated into a constant-time multiplier for the Rocket Core. By engineering the taint propagation logic to address data correlations and reconvergence conditions, our approach reduces unnecessary complexity while preserving security guarantees. We compare our implementation to both CellIFT (Solt et al.) and a self-composition method (SPV, JasperGold) to assess relative trade-offs in precision, performance, and resource utilization. Formal verification results demonstrate a remarkable improvement in tracking precision, reducing false positives compared to prior approaches. Furthermore, our module design yields approximately 24% as many gates as CellIFT, offering tangible cost and area benefits. Compared to SPV, we observe 21% fewer overall flip-flops, making our solution attractive for resource-constrained hardware applications. This work presents an approach for reconvergence-informed IFT within an open source RISC-V processor, providing new insights into efficient hardware taint tracking and reinforcing the viability of IFT for secure processor architectures.

Drone-based item delivery has thus far in its short but innovative history revolved around a multi-rotor, miniature aircraft, performing a full landing to deliver packages. This system carries a range of technical and structural problems, including relatively poor flight dynamics, high battery power consumption, and range limitations. This thesis explores the deployment of electronically guided Smart Parachutes for drone-based package delivery, allowing the drone to perform a ‘fly-by’ during which it ejects the target’s package instead of the delivery requiring a full landing. Successfully designing and building a Smart Parachute delivery system repairs a wide range of social ills by adding yet more efficiency to the drone delivery network. Modern society suffers severe automotive congestion, pollution, theft, and latency between customer order and customer product receipt. A scaled drone delivery system mitigates the vast majority of these ills. As such, the project is worth pursuing.

The thesis specifically will explore a number of electromechanical solutions, highly reliant on granular, software-driven, location-sensing (via camera feed), as well as optimizing GPS coordination. These measurements will then feed into a rapid system state update algorithm to plan an optimal path for the Smart Parachute. Lastly, the system state algorithm will control a servo motor to steer the rigging of the parachute and update the feedback loop.

This project involves constructing an apparatus that mounts to a standard player piano mechanism in place of the paper piano roll and allows MIDI files to be played on the instrument. While modern player retrofit kits exist using solenoids, these need to be directly installed onto the piano action by an experienced technician. Additionally, installing such a kit on a player piano necessitates removing most, if not all of the original player mechanism. For my project, I aim to construct a portable device that can be easily installed by any individual on any functioning player piano. This will not only reduce the cost and effort required to update the piano, but it will also preserve its originality whilst giving it the ability to play MIDI files.

In the Radio Frequency Integrated Circuit design industry today, the design process is both complex and labor-intensive, demanding deep domain expertise and significant time investment. A designer first starts with target performance specifications. After establishing a general architecture, the designer then chooses topologies for each gain stage, iteratively adjusting parameters until the active portion of a circuit is produced. Then, the designer must match the impedances of each stage, utilizing predefined parameter sweeps and heuristic techniques to adjust and optimize their passive component designs. This process is extremely tedious, taking anywhere from a few weeks to several months depending on the complexity of a design. To address this issue and expedite the design process, this thesis tackles the development of both passive and active components by utilizing machine learning methods. Specifically, we utilize inverse design methods for passive structures and reinforcement learning for active components to synthesize power amplifier circuits from end-to-end algorithmically. Moreover, we consolidate these tools into graphical user interfaces to provide a ready-to-use product for RFIC design engineers anywhere.

Though hardware accelerators decrease the operational carbon emissions of contemporary systems-on-chip (SoCs), their significant area, and thus embodied carbon cost, makes their impact on the overall lifetime sustainability of a chip less clear. As opposed to instantiating separate hardware for each accelerated application kernel, time sharing hardware resources for multiple kernels can decrease the total chip area required for acceleration, at the cost of increased energy consumption. This paper examines a spectrum of methods for this hardware sharing, in order to find the architectures which optimally trade off operational and embodied carbon emissions.

Our analysis quantifies the carbon impact of sharing for both compute and memory hardware. We show that the decreased area and power consumption which comes with highly specialized compute hardware outweighs the benefits of sharing compute hardware with a reconfigurable fabric. Importantly, however, the power and area overhead required to share memory hardware is low, and so sharing memory comes with a substantial carbon-efficiency improvement. Hence, the best option is fixed-function accelerators which keep compute separate but maximally share on-chip memory hardware. Our analysis also shows sustainable sizing options for tailoring these shared memory pools to the memory needs of an application's accelerator collection.

The rapid emergence of advanced therapeutics has highlighted the need for innovative drug delivery systems (DDSs) capable of overcoming biological barriers and ensuring precise spatial and temporal control. This work focuses on the design and simulation of a dual-mode microfluidic DDS that combines electrochemical and magnetic actuation mechanisms to enable reliable and minimally invasive therapeutic delivery. Computational simulations demonstrate the system’s ability to maintain laminar flow under varying conditions, including different Reynolds numbers, fluid viscosities, and multi-input configurations, ensuring robust performance across diverse scenarios. Guidelines for fabrication have been established, detailing the integration of electrochemical drug release via gold membrane dissolution and a magnetic fallback system using iron-doped PDMS membranes for power-independent operation. The front-end fluid control system has been physically implemented on the macro-scale and addresses critical challenges in localized drug delivery by combining spatiotemporal precision, adaptability, and robustness.

In light of the tremendous possibilities offered by waveform design in the near field, we develop simulations of a sub-THz radar system that senses using Airy beams. Remarkably, this class of waveforms is “self-accelerating”: they trace out parabolic trajectories in free space in the absence of external forces. As such, a radar equipped with such a beam should be capable of around-the-corner sensing. This newfound capability opens entirely new dimensions in traffic sensing (among others), and its deployment in traffic radar may lead to significantly improved safety outcomes. Our central contribution lies in determining that conjugating the complex E-field incident on an obstacle (via an active surface) generates an Airy beam that returns to the transceiver plane by tracing the path taken by the original. This is true regardless of obstacle orientation. We also investigated the effect of (vertical) obstacle length on the link-budget. As expected, the percentage of returned power increases with obstacle length, reaching a plateau at ~ 80%. We determine that an ideal obstacle length might be in the vicinity of 0.4 (twice the aperture size of 0.2), since it provides a robust middle ground between returned power and size. In the same line of inquiry, we also found that the smallest obstacle length should be about 0.05 (a quarter of the aperture size). At this limit, an (unconjugated) specular reflection might, in fact, yield greater power returns in the primary lobe than conjugation.

Perovskite solar cells, or PSCs, especially those based on a formamidinium lead iodide (FAPbI3) chemistry, have emerged as a promising technology for next-generation photovoltaics, achieving record power conversion efficiencies exceeding 25%1, 2. However, the standard use of methylammonium chloride (MACl) as a processing additive presents significant challenges, including irreversible reactions with formamidinium (FA), production of volatile byproducts, and limited long-term stability under operational conditions3. This study aims to address these issues by systematically investigating alternative additives to replace MACl in FAPbI3-based perovskite solar cells. The research explores a series of ammonia-based, sulfur-based, and organic acid-based additives, including but not limited to NH4Cl, NH4F, iPACl, thiourea, and NH4SCN, focusing on their reactivity with FA, impact on perovskite formation, and potential to improve or replicate FAPBI3 film uniformity and stability. Through in-solution testing and evaluation of the additives' effects on crystallization, film morphology, and optical properties, this study seeks to identify possible optimal additives or combinations to replace MACl and gain an improved understanding of additive-perovskite interactions. The research employs X-ray diffraction to gain insight into crystallization dynamics, optical microscopy for morphological analysis, and UV-Vis spectroscopy to assess optical properties across various lengths of sample aging. Through structural, optical, and morphological characterization, we demonstrate potential combinations of NH4Cl, NH4PF6, and thiourea, as well as pre-aged equimolar combination substitutes of FACl and thiourea as promising substitutes with band gap, morphology, phase stability, and performance characteristics similar to that of and exceeding MACl-based FAPbI3 films and devices.

This study extends the CONFINE algorithm as a framework for uncertainty quantification onto transformer-based language models. CONFIDE (CONformal prediction for FIne-tuned DEep language models) applies conformal prediction to the internal embeddings of BERT and RoBERTa architectures, introducing new hyperparameters such as distance metrics and PCA. CONFIDE uses either [CLS] token embeddings or flattened hidden states to construct class-conditional nonconformity scores, enabling statistically valid prediction sets with instance-level explanations.

Empirically, CONFIDE improves test accuracy by up to 4.09% on BERT-TINY and achieves greater correct efficiency compared to prior methods, including NM2 and VanillaNN. We show that early and intermediate transformer layers often yield better-calibrated and more semantically meaningful representations for conformal prediction. In resource-constrained models and high-stakes tasks with ambiguous labels, CONFIDE offers robustness and interpretability where softmax-based uncertainty fails.

This project combines the Boston Dynamics Spot robotic platform with virtual reality to develop a more intuitive remote inspection experience. Two means of visualizing camera data from Spot in real-time in a VR interface were developed and tested. The first approach uses point cloud data from the robot’s stereo cameras. Because of a bottleneck in data transmission latency, the second approach focuses on improved latency at the tradeoff of a slightly less natural two-dimensional display of data. This approach surrounds the human user in six planes that show the live JPEG-compressed image feed from each of the robot’s cameras: gripper, front right, front left, right, left, and back. The location of the user is tracked and sent back as movement commands to Spot; as the user moves in their space, the robot moves analogously in its space. The user can send additional commands, such as sitting and standing, via handheld controller buttons. Considerations for privacy in a remote inspection were also implemented, including features that protect sensitive information from being transmitted and ensure the robot can not approach certain areas. Possible applications for this project lie in its ability to facilitate more natural feeling remote inspections, including those of dangerous areas, spaces where direct access is difficult, and private or secure locations.

As robots are increasingly integrated into human-centric environments, equipping them with emotional intelligence is essential for safe, ethical, and meaningful interactions. This thesis explores the intersection of emotional intelligence and Human-Robot Interaction (HRI) by leveraging Boston Dynamics’ quadruped robot, Spot, to detect one user’s emotional state, specifically happiness, and respond physically in real time to the detected emotion. The project uses a machine learning model with tools from the Deepface Python library for facial emotion recognition (FER) to interpret the user’s emotion, which is captured by the camera in Spot’s gripper arm. When happiness is detected, a servo powers a flag to wave left and right to provide a clear, expressive response.

This project serves as a proof-of-concept for real-time emotional responsiveness in non-humanoid robots and demonstrates the potential for future quantitative and qualitative studies in affective HRI. By focusing on a single emotion and a single user, the project also prioritizes controlled experimentation and ethical design, avoiding implementing generalized facial recognition. The research contributes to the emerging field of emotionally expressive robotics and highlights the potential for enhancing human trust and engagement of useful robots in human environments.

This thesis presents the design and implementation of a magnetically actuated Kresling robot capable of rolling locomotion, bistable folding transitions, and modular self-assembly. Leveraging the geometric properties of the Kresling structure and inplane magnetized plates, the system responds to uniform magnetic fields generated by a triple-axis Helmholtz coil. Two design iterations were developed—one using silicone-neodymium composites, and another with permanent magnets for improved control. Real-time tracking via ArUco markers and color segmentation enables visionbased pose estimation. Analytical models identify optimal torque conditions for state transitions, and experiments validate consistent actuation and successful magnetic docking between units. This work demonstrates the feasibility of scalable, untethered origami robots, with future potential for autonomous control and reconfigurable soft robotic systems.

Airy beams have numerous imaging applications due to their ability to follow different propagation paths, including parabolic trajectories as well as being self-healing. As it travels it also does not spread out, retaining its intensity, and it exhibits resilience to obstacles. Ultrasound has had many imaging applications over the years but the use of ultrasonic transducers for the production of airy beams is still novel. The use of a phased array to do so is especially useful because by carefully arranging and modulating an array of signals, typically through phase and amplitude manipulation, it is possible to synthesize the spatial structure required to produce Airy beams. The interference of these mixed signals allows engineers and scientists to shape the beam’s energy distribution, creating the characteristic Airy waveform with a main lobe and trailing sidelobes.

This thesis analyzes the performance of photovoltaic (PV) systems across four solar fields within a 1.3 km radius at Princeton University to explore how hyperlocal environmental factors affect energy yield. Despite similar irradiance and ambient conditions, differences in surrounding surfaces (grassy vs. concrete) led to notable variations in module temperature. Heavy snow fall events were identified as resulting in different power output behavior even in fields with the same geometric configuration. A custom power output model, accounting for tilt and temperature, flagged unexpected power output anomalies, leveraging a combination of partial on field irradiance measurements with online irradiance data. Results show that environmental factors do affect the power yield of solar fields in the same region, underscoring the value in further examining such factors to enhance the performance of solar installations in the future.

This project aims to develop an understanding of network attacks on Ethereum Proof-of-Stake and the tools that are available study them. Our goal is to explore facets of Ethereum’s behavior under adversarial conditions that cannot be probed in a non-emulated environment. After a brief review of Ethereum’s protocol, we walk through a sequence of malicious behaviors that threaten Ethereum’s safety and liveness. We discuss a set of emulators that aid our study of Ethereum, as emulation provides a safe testing grounds for attacks. After selecting an emulator and upgrading its capabilities through a series of functional improvements, we launch a set of experiments that measure Ethereum’s resilience in the face of network-level disruptions. First, we consider the impact of network delay on liveness by measuring fork rate. We show that adding delay between ASes causes fork rate of up to 6.9% with a uniform distribution of validators on the network and up to 16.7% with a concentrated distribution of validators. Then, we illuminate the danger of concentrating validators within nodes on the network by instigating packet loss for clusters of validators. We examine validator revenue, validator balance, and blockchain finality as evidence of the seriousness of this attack. Our results suggest that targeting clusters of validators allows us to cause an inactivity leak such that validators lose up to 93.8% of their original stake. As a result, we advocate for a more uniform distribution of validators across the network. Finally, we provide an outline for future work that should address the limitations of the research in this report and push to explore ways we can make Ethereum more safe and secure.

Thin-film semiconductors have been essential to modern technologies. But as devices become smaller and demand more power, traditional materials like silicon (Si) are no longer sufficient. Two-dimensional (2D) semiconductors, especially transition metal dichalcogenides (TMDs), are gaining interest. TMDs have a direct bandgap as monolayers, and past research shows that strain can shift this bandgap. Stacking TMDs also enables complex structures for advanced device applications. This work focuses on producing clean, intrinsic TMD films for 3D structures and introducing engineered strain. WS2 was grown, and various topographies were etched into SiO2. Two transfer methods were tested: traditional exfoliation and a cyclododecane-based approach. Raman spectroscopy, photoluminescence, SEM, and AFM were used to study the resulting devices. The CDD-based transfer showed higher yield and more consistent results than exfoliation. Drying the TMD before transfer reduced strain after placement on the patterned substrate. CDD was also tested to improve coupling between multiple layers of TMDs. While some transfers succeeded, reproducibility is still uncertain. CDD was further used to roll TMD flakes by placing it between the flakes and substrate. Future work will explore using CDD to fabricate suspended films and improve interfaces between stacked TMD layers.

When working with biological cells, two key principles should guide the process: acceleration and accuracy. Estimating cell viability is a task that greatly benefits from methods that embody these characteristics. Traditionally, viable and dead cells have been classified manually using techniques such as cell staining with chemical reagents or fluorescence. However, these processes are limited by their reliance on manual intervention and the inherent fragility of cells. deep Through the methods introduced in this report, a third key principle is introduced: automation. Combining imaging techniques with deep learning algorithms presents a promising alternative, enabling the automation of single-cell viability classification. By leveraging deep learning and dimensionality reduction, systems can be developed to streamline and improve cell classification. To validate these systems, it is assumed that stained and unstained cells can be analyzed in a similar manner. This research aims to establish that no meaningful difference exists between stained and unstained cell images, laying a solid ground truth for the development of automated classification systems. These advancements hold significant potential for applications in cancer research, drug development, and stem cell studies.

With increasing numbers of common diseases affecting the general population, diagnoses are becoming more time-consuming, expensive, and stressful for patients, requiring frequent laboratory tests and specialized equipment to track disease progression. Recent innovations in wearable medical sensing and machine learning have enabled portable smartwatch-based \textit{health decision support systems} (HDSSs). HDSSs use real-time samples of physiological signals to proactively diagnose diseases through deep neural network classifiers, then alert medical professionals so that further laboratory tests can be administered once a detection result is positive. However, these models have a large parameter space, placing a latency and energy burden on battery-powered smartwatches and causing high model variance (i.e., sensitivity to the specific training and validation datasets used before deployment). Neural network classifier models are also generally poor at uncertainty estimation, making them difficult for detecting disease comorbidities or distinguishing between mutually exclusive disease outcomes. This work uses dataset preprocessing, along with state-of-the-art energy-based model frameworks including Joint Energy-Based Models (JEM) and Variational Entropy Regularized Approximate (VERA) maximum likelihood generation, to optimize the accuracy, variance, size, latency, and confidence calibration of existing detection model frameworks across multiple disease categories, helping to enhance their usefulness in pervasive healthcare applications.

Coordinating autonomous drone swarms in decentralized environments presents a significant challenge, especially when designing strategies that scale effectively. In this thesis, we propose a graph-based multi-agent reinforcement learning (MARL) framework that enables drone swarms to autonomously learn cooperative interception behaviors in pursuit-evasion scenarios. Our approach employs graph neural networks (GNNs) to enforce permutation invariance, accommodate varying team sizes, and support decentralized decision-making under limited observability. Each agent operates without access to global state information, relying solely on local observations and limited-range communication. We begin by outlining the relevant background concepts, then detail our proposed methodology. We demonstrate generalization to unseen team sizes and the emergence of decentralized strategies by evaluating our model in benchmark scenarios.

This project builds upon prior work into the use of image tactile sensors to improve gripper performance through the adjustment of marker identifying algorithms and the initial construction of a physical gripping system. I focused on improving marker tracking algorithms, refining the physical design of sensor modules, and enabling real-time data processing. The previous image tactile sensing system has been enhanced through the development of new sensor designs, integration of a multi-camera system, and incorporation of a MySQL database for real-time data storage and analysis. Marker tracking accuracy and processing speed were significantly improved by introducing techniques like k-means clustering, watershed segmentation, and frame-skipping, alongside the use of diffused lighting and optimized illumination parameters. Three novel sensor modules were fabricated, including a black rubber module with manually placed markers, a clear silicone pad for small object interactions, and a gel-tip-inspired design for dynamic sensing. These modules demonstrated improved adaptability, scalability, and accuracy for tactile sensing. MySQL integration facilitated long-term data aggregation and enabled retrospective analysis of sensor performance, paving the way for machine learning applications in predictive force estimation. The database allows for the categorization of force measurements by sensor type, object size, and timestamp, enabling continuous system refinement. Real-time feedback was achieved by addressing prior computational challenges, reducing noise, and eliminating reliance on external lighting conditions. Testing with multiple gripper models validated the system's capability for precise tactile sensing in robotic manipulation tasks, allowing for accurate handling of a larger array of object sizes. The updated system supports continuous force tracking and object recognition during gripper operation. Future work will focus on miniaturizing the sensors, further optimizing image analysis algorithms, and leveraging accumulated data for enhanced system performance. This semester's advancements mark a significant step toward developing low-cost, efficient tactile sensing systems for applications in robotics, prosthetics, and human-robot interaction. Documenting these aspects of the implementation would make improving robotic gripping and navigation more accessible for a wider group of users. Because of the limitation of time, the physical calibration of the grippers and experiments regarding improvement to object manipulation performance are future considerations.