Electrical and Computer Engineering, 1932-2025

Language acquisition and reasoning, two skills fundamental to human intelligence, both require internalizing and applying a collection of rules and structures derived from a limited set of examples. Crucially, this requires not only the ability to identify important structures in those examples, but also the ability to systematically combine learned structures in novel combinations, or systematicity. Interestingly, despite recent successes in the sphere of language generation, statistical learners like today's foundation models still fail to consistently reason systematically, suggesting a possible incompatibility between the current learning paradigm or model architectures and the acquisition of systematicity. We aim to develop a better understanding of the extent to which statistical learners are capable of generalizing to novel combinations, which we call compositional generalization, in systematic tasks. To that end, we identify some ways in which model architecture or optimization choices can encourage the inductive biases needed for models to generalize compositionally, as well as challenges in more realistic task settings.

Split Learning, Federated Learning, and recent combined approaches such as Split-Federated Learning represent significant advances in privacy preservation and individualization in neural network training. However, while these approaches offer promising solutions for large scale distributed learning, limitations in edge device computational power and memory have limited the practicality of biomedical and healthcare applications of neural networks on Wearable Medical Devices (WMDs). This work introduces a novel GRU-based SplitFed framework to achieve a balance of privacy, computational efficiency, and accuracy to satisfy the unique needs in the healthcare space. This approach addresses the computational limitations of edge devices by leveraging Gated Recurrent Units (GRUs) instead of traditional LSTM networks, while maintaining comparable classification accuracy. By implementing a hybrid architecture that combines the data privacy benefits of Split Learning with the collaborative advantages of Federated Learning, we demonstrate improved performance on ECG classification tasks. Experimental results on ECG arrhythmia classification show that our GRU-based SplitFed approach achieves up to 65.02% accuracy, outperforming both centralized models (58.86%) and traditional split learning approaches (50.21-50.94%). Our experimental analysis shows that both Split Learning and SplitFed approaches have comparable communication requirements for models of similar size while offering significant advantages in accuracy and privacy. Our experiments also revealed important stability challenges in the training process that must be addressed for reliable deployment. This work expands the applicability of privacy preserving distributed learning to resource-constrained medical devices, providing a framework that balances privacy, efficiency, and performance for secure processing of sensitive biomedical data.

We propose using superconducting nonlinear oscillators (SNAILs) to amplify weak signals from quantum systems, leveraging their inherent nonlinearity for reservoir computing. By modeling the SNAIL system as a neural network, we aim to improve qubit state discrimination through training, thereby enhancing the accuracy of quantum measurements. Our current work involves simulating a nonlinear system composed of a SNAIL driven by an optical pump. We generate time-series data for both the signal and idler modes, which are then used to train and perform weighted binary classification of values obtained from the SNAIL system under varying noise distributions with the same mean. Two nonlinearities are present within the SNAIL system: the third-order nonlinearity (g3), which is primarily responsible for the amplification process, and the fourth-order non-linearity (g4), or Kerr nonlinearity, which we examine as a basis for reservoir computing and classification. We seek to identify the optimal regime for classification by varying g4 and the phase of the pump.

Upon examination, we found no consistent regime with significantly improved classification performance; however, certain combinations of g4 and pump phase showed marginal gains, suggesting potential for optimization in more targeted parameter spaces.

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.

DESC is the state-of-the-art code that solves the ideal magnetohydrodynamic partial differential equations subject to a fixed boundary condition. However, to simplify the problem a standard, yet potentially inconsistent, assumption is made: nested flux surfaces. Still, this assumption is commonly used because it simplifies the partial differential equation and most desirable equilibria approximately have this property. This work extends the DESC code so that this assumption does not need to be made in vacuum fields and improves free surface equilibrium optimization. We design and implement automatically differentiable algorithms to solve the vacuum field and free surface problems. This will lead to improved study of stellarator equilibria. The boundary integral methods implemented in this work also have utility for optimization objectives to be used in the future.

Phutball is an impartial, rules-light game with an arbitrarily scalable board, making it an appealing testbed for human-inspectable multi-step reasoning. This thesis introduces PhutballEnv, a turn-based Markov game environment that is fully compatible with AlphaZero-style self-play. The system ships with a logging and visual-diagnostic stack that records every board position and action, while simultaneously producing gradient-based saliency maps that highlight the board features driving each decision. These rich traces can be automatically exported as text corpora, enabling language models to be fine-tuned on plain moves, saliency-tagged positions, or synthetic rationales generated post hoc.

    The document also describes a lightweight evaluation protocol that uses relative-Elo ladders against frozen checkpoints, along with a small user study assessing explanation clarity. Because full training was beyond the project’s time budget, the emphasis is on providing reliable implementations of the environment and interface, a data pipeline, and validation utilities that future work can build on at the intersection of reinforcement learning, language modeling, and interpretability.

Continuous device scaling and power scaling coupled with lithographic limitations have led to unacceptable levels of performance variation and an exponential increase in static power consumption. Effective measures to mitigate variation induced leakage and performance overhead are therefore greatly in demand. Previous compensation circuit solutions implore significant area-overhead solutions. In contrast, this thesis presents a low area overhead (512T) mix-signal compensation circuit, which can detect variation in performance and leakage and correct the variation to acceptable values.

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.

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.

As computing evolves and the use of cloud environments increases, there is growing concern over security and data confidentiality. Virtual machine-based trusted execution environments (VM-based TEEs) have emerged as a promising solution due to their flexibility and scalability. This thesis presents a formal verification framework for reasoning about the correctness of the Intel TDX firmware—a central component of a prominent VM-based TEE. Using CBMC, a suite of abstraction strategies, and modular verification reasoning, the framework enables tractable analysis of this complex, real-world system. The methodology is specifically applied to the TD creation sequence, demonstrating that the verification process successfully validates correct flows and reliably detects errors when deviations occur. Beyond TDX, this work contributes to the broader goal of developing a scalable and rigorous methodology for evaluating secure processor architectures and ensuring their reliability in real-world deployments.

Stealthy hyperuniform (SHU) materials are a novel class of metamaterials that combine the band gap properties of photonic crystals with spatial isotropy, enabling applications in optics and photonics such as silicon photonic waveguides, mode selection in THz quantum cascade lasers, tailored light scattering, edge detection, and mid-infrared filtering. These metamaterials often have much smaller or complete photonic band gaps (PBG) compared to their photonic crystal counterparts, without requiring the incoming light to be at normal incidence. A large PBG or gap-to-midgap ratio, balancing the width of the PBG against its central frequency, is desirable for precise control over the propagation of light or electromagnetic waves. This thesis investigated the optimization of PBGs through numerical simulation and experiments using various geometric patterns, including triangles, squares, rectangles, hexagons, circles, ellipses, and crosses. After simulating such patterns on a 1x1 unit cell in MIT Photonic Bands (MPB) software, the optimal geometric patterns were determined to be square, hexagon, and circle. For experimental validation, a circle-based SHU-patterned Indium Phosphide (InP) crystal was fabricated and analyzed using FTIR transmission spectroscopy. Compared to the unpatterned InP crystal, the patterned sample exhibited lower transmission across all mid-IR frequencies and reduced angular dependence, confirming its isotropic properties. Notably, a photonic band gap emerged around the 1000 cm⁻¹ wavenumber. These findings demonstrate the potential of SHU-based geometries in tunable, angle-independent photonic devices for mid-IR applications.

This project presents the integrated design and optimization of a personal human wing glider system augmented by electric ducted fan (EDF) propulsion. Using OpenVSP, the full 3D geometry, including a realistic human mesh, fixed-wing structure, and EDF units, is modeled to assess aerodynamic performance. The system is aerodynamically optimized for glide and cruise conditions with a span-efficient wing design that effectively integrates wing and human geometry to maximize lift-to-drag ratio while minimizing induced drag. Flight operation constraints, including stall speed, weight-to-power ratio, and structural limits, are incorporated into a wing loading optimization algorithm to extract feasible design conditions.

Propulsion is achieved through a distributed EDF system, with six ducted fans symmetrically positioned along the wingspan. Each EDF is paired with a motor selected through an efficiency-based matching process that considers thrust, RPM, and power requirements across still hover, vertical climb, and cruise conditions. XROTOR simulations inform propeller blade twist, hub sizing, and exit area ratios to optimize thrust production and minimize power draw. The system’s thrust, power, torque, and throttle demands are cross-validated with motor performance curves to ensure high efficiency under all mission segments. The final system achieves efficient low-speed flight with well-defined power and thrust requirements, supporting stable transitions across glide, cruise, hover, and climb, and demonstrating viability for compact, wearable, EDF-assisted personal flight.

Accurate medical image classification using computer vision remains a challenge in clinical radiology, particularly in low-data settings where labelled examples are scarce or expensive to obtain. This thesis evaluates 20 separate model configurations across 3 binary chest X-ray classification tasks, to determine the impact of pre-training, base architecture, and fine-tuning strategies on diagnostic accuracy. The models' variations include convolutional (ResNet-18, ResNet-50) and transformer base models (ViT, DINOv2), application of pre-training or not on a large corpus of healthy chest X-rays images, and method of fine-tuning (transfer learning, full fine-tuning, Low-Rank Adaptation, Chain of Low-Rank Adaptation, and Weight-Decomposed Low-Rank Adaptation).

The findings prove that the effect of domain-specific pre-training significantly boosts downstream performance of CNNs and models trained on small datasets (< 2,500 diseased images) by an average of 2.55% and 3.9% respectively. Convolutional architectures consistently outperform the top transformer-based models by an average of 9.5%. Almost all pre-trained ResNet models match or exceeded benchmark standards for public datasets, achieving up to 96.1% accuracy on the largest dataset (8,716 labelled examples) and up to 79.4% average accuracy across the 3 tasks, with a mean dataset size of 5,209 labelled images. As a result, this work comprehensively shows CNN-based models in computer vision-based medical diagnostics and pre-training on a large, related, healthy corpora improves downstream classification accuracy. The two of which should be adopted into routine use in critical fields such as radiology, where high-quality data is scarce and accuracy is paramount.

We introduce Pok'eChamp, a minimax agent powered by Large Language Models (LLMs) for Pok'emon battles. Built on a general framework for two-player competitive games, Pok'eChamp leverages the generalist capabilities of LLMs to enhance minimax tree search. Specifically, LLMs replace three key components: (1) player action sampling, (2) opponent modeling, and (3) value function estimation, enabling the agent to effectively utilize game play history and human knowledge to reduce the search space and address partial observability. In the second phase of our research, we develop a ReAct-like framework and incorporate retrieval-augmented generation (RAG) to evaluate the efficacy of LLMs in the specialized task of competitive team generation. Notably, our frameworks requires no additional LLM training. We evaluate Pok'eChamp in the popular Gen 9 OU format. For battling, the battling agent achieves a win rate of 76% against the best existing LLM-based bot and 84% against the strongest rule-based bot when powered by GPT-4o, demonstrating its superior performance. Even with an open-source 8-billion-parameter Llama 3.1 model, Pok'eChamp consistently outperforms the previous best LLM-based bot, Pok'eLLMon powered by GPT-4o, with a 64% win rate. For the team generation task, the LLM agent was able to achieve high performing teams on par with a heuristic approach that specifically utilized statistical metagame usage data. These specialized tasks show the efficacy of LLMs trained only on generalized prior data, especially when given the same tools as current heuristic-based approaches and real human players. This work here led to publication (Karten et al., 2025)

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.

Large-scale fault-tolerant quantum computers and processors have the potential to significantly impact fields of cryptography, physics simulation, optimization, and more. Superconducting qubits are a leading implementation platform for quantum computers. The 2D transmon is a particular type of superconducting qubit that is widely used due to its charge noise resilience. Still, one of the key challenges in superconducting quantum computing is to improve 2D transmon relaxation (T1) and coherence (T2) times, for which the highest reported values have been shown to be up to 2.5 milliseconds. There is a demonstrated inverse relationship between the the relaxation time (T1) and the qubit operating frequency (ω_q) represented by T1 = Q/ω_q, where qubit quality factor is represented by Q. Recent work on tantalum-on-silicon-based 2D transmons demonstrates a Q = 2.5 × 10^7, attaining a maximum qubit lifetime of 1.6 milliseconds with a qubit frequency of 2.4 GHz. This work investigates the potential of lowering ω_q on tantalum on-silicon-based 2D transmons to increase qubit relaxation times, and provides an overview of the process of design, simulation, fabrication, and measurement required to do so.

Layered transition metal dichalcogenides are an emerging class of atomically-thin, layered materials for use in solar cells, transistors, LEDs, and other optoelectronic devices. These materials are flexible and durable, with bandgaps that shift from indirect as a bulk film to direct as a monolayer, allowing for finely tunable devices. Two of these materials, WSe2 and WS2, have bandgaps around 2 eV, making them more versatile than other 2D materials such as graphene, which lacks a bandgap. This project focused on mono- and multilayer WSe2 and WS2 films, and determined their conductivity, established the position of the Fermi level, and introduced [RuCpMes]2, an air stable dimer, as an n-type dopant to determine shifts in Fermi level and changes in conductivity with doping. The introduction of the Ru dimer increased the conductivity of monolayer WSe2 films by 3 orders of magnitude and increased the conductivity of monolayer WS2 films by 4 orders of magnitude, indicating successful doping. The Fermi level of undoped monolayer WSe2 was found to be close to the conduction band minimum and the Fermi level of monolayer WS2 was found to be slightly below the conduction band minimum. After doping with [RuCpMes]2, the Fermi level of monolayer WSe2 did not shift significantly due to its proximity to the valence band minimum, but the Fermi level of monolayer WS2 shifted upward. Due to charges on the surface, the work function of both monolayer films decreased with doping, with a greater change in monolayer WS2.

Unmanned Surface Vehicles (USV) and Unmanned Underwater Vehicles (UUV) are essential for a wide range of applications, including deep-sea exploration, environmental monitoring, and sea-bed surveying for valuable minerals such as rare earth elements. Understanding fish locomotion is crucial for designing energy-efficient marine robotic vehicles that can travel either on the surface of or under the water without any human operators on board. This work introduces a low-cost miniature fish-like robot, BlueGuppy, that can swim up to approximately three body lengths per second and can turn with its sole actuator. First, we detail its hardware design, emphasizing the rationale behind specific design choices. Second, we perform both straight-line and turning experiments demonstrating how swimming speed scales with flapping frequency and how temporally asymmetric flapping enables controlled turning. Finally, we present ongoing efforts to develop a Proportional-Integral-Derivative (PID) controller framework for the robotic fish using Bluetooth Low Energy (BLE) and WiFi as a means of communication. This BlueGuppy project highlights the potential of free-swimming robotic models in advancing our understanding of fluid-swarm interactions within biological collectives and inspiring the development of next-generation marine robotic vehicles.

Reforestation efforts, in particular those led by NGOs and nonprofits, often struggle in maintaining efficient and reliable workforces, relying heavily upon volunteers and part time contributors. This is to the detriment of meeting project timelines in the face of urgency as the climate crisis grows stronger despite community efforts to improve environmental resilience through reforestation. This project has developed a prototype of an autonomous robot that can reliably dig 500 holes in under six hours in muddy environments where traditional tools are less effective, thereby increasing productivity, freeing up bottlenecks, and reducing reliance on skilled labor or limited workforces. The robot is equipped with a robust power supply, a custom linear actuator, easy to repair modularity, custom digging mechanisms and GNSS sensing capabilities to determine location. The approach was informed by Pro Eco Azuero, a NGO in Panama that does community reforestation work. This is a scalable, manageable, and safe technology equipped with electrical and logical safeguards that could increase any organization’s efficiency in a cost-effective manner. Keywords: robotics, reforestation, Latin America, sustainability, autonomous systems.

Experimenting with tax policy in the real world can be prohibitively expensive and politically infeasible. Governments need innovative simulation and modeling techniques to evaluate policy impacts before deployment. Existing approaches in optimal income taxation theory create sup-optimal policies by relying on economic models that make simplifying assumptions about human behavior. This thesis argues that large language models [LLMs] learn tax policies that result in higher social welfare than the tax policies produced by existing economic models by providing a scalable, affordable method to model societal behavior and optimize for social welfare. We model policy decisions as an infinite dynamic game between a tax planner (leader) and workers (followers), optimizing for Stackelberg equilibria that maximize social welfare. We use LLMs to generate synthetic human data facilitating policy mechanism design, testing, and optimization. To increase realism, we implement simulation scenarios where the tax planner is elected by worker agents. We validate our LLM-based approach by comparing our results in a two worker agent, one tax planner simulation to a Stackelberg equilibria that we calculate through backwards induction. We investigate the effect of different simulation scenarios and skill distributions on social welfare. We find that our LLM-based approach achieves higher social welfare than the tax policy calculated according to economist Emmanuel Saez's optimal income taxation formulas. Future work could implement extensions to Saez's formulas that incorporate more elements of human economic activity with the goal of achieving higher social welfare with learned policies in these more complicated scenarios.