137091 research outputs found
Sort by
Hardware-Algorithm Co-Design for Energy-Efficient and Low-Latency Domain-Specific Machine Learning Systems
No full textThis dissertation develops hardware–software co-design methodologies across multiple stages of the machine-learning (ML) system stack, spanning sensor-to-model interfaces, neural network architectures, and hardware accelerators. Modern ML workloads must balance competing demands of latency, energy, throughput, and accuracy, creating a multidimensional design space where no single configuration suffices. By jointly optimizing algorithms and hardware, this research demonstrates how tailored co-design strategies can deliver substantial performance and efficiency gains.
At the sensor-ML interface, we introduce mixed-signal compute-in-memory accelerators for efficient feature extraction and employ Quantization-Aware Training (QAT) with low-Hamming-weight binary representations, significantly reducing inference energy costs. At the neural network level, we design ChirpNet, a radar-specific architecture that sequentially processes Frequency Modulated Continuous Wave (FMCW) chirps, thereby lowering memory and compute requirements. We also propose LUGA, an adaptive ML-sensor feedback framework that dynamically tunes sensing resolution and model complexity based on real-time confidence estimates, improving energy efficiency. At the accelerator level, we develop hardware-algorithm co-design techniques for distributed SAT solving using stochastic recurrent neural networks (S-RNNs). Leveraging low-precision weights and tensor-parallel execution with Reduce-Scatter communication, we achieve notable speedups and reduced latency.Ph.D.Electrical and Computer Engineerin
Development of Lipid Nanoparticles for RNA Delivery to Hematopoietic Stem Cells and Development of Alternative Model Systems for Nanoparticle Discovery
No full textLipid nanoparticles (LNPs) have been used to deliver RNA in several FDA-approved drugs, including treatments for genetic liver diseases (e.g., ONPATTRO) and as vaccines for COVID-19 (e.g., SPIKEVAX, COMIRNATY). The current LNP discovery pipeline selects for nanoparticles with high liver biodistribution, yet development of next generation RNA therapies that treat diseases outside of the liver will require LNPs that de-target the liver and deliver RNA to other therapeutically relevant cell types.
Hematopoietic stem cells (HSCs) represent a clinically attractive target for RNA therapies as dysfunction in this single cell type drives multiple different pathologies including sickle cell disease, ß-thalassemia, anemias, immunodeficiencies, and metabolic disorders. Despite the potential impact RNA therapies can have on the treatment of blood disorders, there are few delivery vehicles that can deliver RNA to these cells. In my work, I sought to identify LNPs that escape the liver and functionally deliver RNA to HSCs. Firstly, I worked to identify lead candidate HSC-LNPs using high-throughput LNP screens at single-cell resolution. Next, I worked to characterize delivery profiles of lead HSC-LNPs across pre-clinical models to investigate the translational potential of identified particles. Next, I identify formulation and post-processing parameters that improve LNP potency and explore biological responses to LNP addition in the bone marrow.
Finally, I explored alternative model systems in the nanoparticle discovery pipeline. Current approaches to discovering LNPs for use in human RNA therapies relies on identifying lead candidates in cell culture or animal models and evaluating delivery profiles of these lead candidates across model species which have either underexplored or poorly predictive relationships. Typically, nonhuman primates (NHPs) are used as the final stage animal model to validate LNP-RNA drugs before human trials due to their physiological and immunological similarity to humans. Many initial LNP formulations fail to translate across model systems. Performing initial LNP screens directly in healthy NHPs to circumvent these translational challenges is not feasible due to ethical, practical, and economic concerns. Here, I performed initial LNP screening directly in NHPs with spontaneous disease. These animals are independently identified for euthanasia, such that characterizing LNP delivery in these animals adds no additional loss of animal life and may more closely represent true clinical situations where LNP-RNA drugs are used to treat sick patients.Ph.D.Biomedical Engineerin
Characterizing Lipid Nanoparticle mRNA Delivery to the Central Nervous System
No full textLipid nanoparticles (LNPs) have demonstrated safety, versatility, and clinical relevance as vehicles for RNA delivery. Currently, there are three FDA-approved LNP-RNA drugs: ONPATTRO®, the systemically administrated siRNA treatment for hereditary liver disease from Alnylam; and two intramuscular mRNA vaccines against COVID-19: SPIKEVAX® produced by Moderna and by Pfizer-BioNTech’s COMIRNATY®. However, beyond vaccination and liver targeting, LNP-RNA drugs are far from reaching their full potential. Research has shown that LNPs can deliver RNA effectively to non-liver tissues, such as lung, spleen, solid tumors, and bone marrow. For the treatment and management of nervous system disorders, gene therapies hold great promise but require safe and effective delivery vehicles. Therefore, a need exists to design LNPs that transfect cells in the central nervous system (CNS). Presently, LNPs with CNS tropism either i) carry ligands, ii) rely on blood-brain barrier disruption, or iii) are administrated locally. In this work, we have investigated mRNA delivery to the CNS upon systemic administration without targeting ligands. Firstly, we optimized the isolation of various cell types from the brain. Secondly, we have studied mRNA delivery readouts for liver de-targeted LNPs, and identified nanoparticle characteristics for subsequent high-throughput LNP screening. Thirdly, we formulated and screened LNPs with these chosen characteristics that we administrated intravenously, and characterized the in vivo tropism of brain-delivering LNPs at the cellular level. Lastly, we used spatial transcriptomics to understand where in the brain the best-performing LNP achieved functional delivery and how mRNA was distributed within the tissue. Through this work, we have shown CNS delivery by liver de-targeted LNPs, and demonstrated the feasibility of systemic delivery of therapeutic mRNA to the cells of the blood-brain barrier without the use of targeting ligands.Ph.D.Bioengineerin
From Hand-Picking and Conveyor Belts to Optical Scanners and Robotic Sorting: Data-Driven Municipal Solid Waste Recovery Models for Regional Secondary Manufacturing Feedstock - A Case Study in Kent County, Michigan, USA
No full textThis thesis explores the transition from linear to circular waste management models, emphasizing the environmental, economic, and equity benefits of regional circular municipal solid waste (MSW) infrastructure. Using a case study of Kent County, Michigan, where the ambitious goal is to divert 90% of waste from landfills by 2030, the study evaluates the potential of a proposed Sustainable Business Park (SBP) to optimize material recovery, reduce greenhouse gas (GHG) emissions, and support local manufacturing supply chains. A hybridized methodology involving Material Flow Analysis (MFA) and Life Cycle Assessment (LCA) assesses six scenarios ranging from traditional landfilling to advanced circularity practices incorporating SBPs.
Results demonstrate that more circular scenarios, particularly those with high recovery rates, significantly outperform linear approaches in reducing GHG emissions, conserving energy, and creating economic opportunities from recycled materials. The study also introduces a circular revenue capture metric to evaluate the economic value of recovered materials, highlighting the transformative potential of SBPs to foster sustainable industrial ecosystems. Equity considerations are integrated to address historical injustices in waste management, advocating for inclusive decision-making and equitable resource distribution. Limitations and future research directions are discussed, including the need for continuous material composition analysis and the exploration of social lifecycle impacts. This research contributes a comprehensive framework for advancing circularity in regional waste management systems, aligning economic, environmental, and social priorities to achieve a circular economy.M.S.Environmental Engineerin
“Listening to Shakira is an important part of the process”: On the importance of “non-productive” labor in the university library
No full textBook chapter: Library Jackassery: The silly ideas that work and embracing working silly (2026
Tests of General Relativity and Investigating Gravitational-Wave Emission from Binary Black Holes in Highly Eccentric Orbits and their Prospects for Black Hole Populations
No full textThis thesis investigates the profound influence of gravitational wave (GW) detection on both astrophysics and fundamental physics, a field revolutionized by LIGO's pioneering observations. I begin with the theoretical underpinnings of GWs, derived from Einstein's field equations, and describe the functioning of current detectors, such as LIGO and Virgo. A key emphasis is placed on the computational techniques for analyzing GW data, particularly the novel Bayesian wavelet-based method that facilitates model-independent analysis. My research also examines various consistency tests applied to GW signals, providing insights into their effectiveness in detecting deviations from General Relativity. Additionally, the thesis presents a Bayesian analysis of numerical waveforms from hyperbolic encounters of binary black holes, estimating detection rates for upcoming observatories. I conclude with a study of stellar black holes within active galactic nucleus disks, highlighting how high-eccentricity encounters can yield detectable GW emissions.Ph.D.Physic
Challenging the Benefit of Anthropomorphism on Human-AI Collaboration with AI Voice Agents
No full textContrary to conventional wisdom, theoretical frameworks, and current trends in AI design, for human collaboration, simple robotic-sounding agents may be better than more complex anthropomorphic, or human-like, agents. This dissertation tested two extremes of anthropomorphism and social intelligence in an AI voice agent across four studies that examined various types of social influence. The results uncovered a consistent discrepancy between the subjective ratings of the agent and the social behavior. In the trivia task in Study 1, participants conformed less when they perceived the AI agent as more anthropomorphic, despite viewing the more anthropomorphic agent as more likable. In the moral judgment task in Study 2, participants conformed less to the anthropomorphic agent than the robotic agent, regardless of the agent’s morality, which, again, contrasted with the subjective ratings. In the prisoner’s dilemma task in Study 3, participants cooperated less with the anthropomorphic agent as they applied human social behaviors to the AI (e.g., retaliating to the degree of lowering their game score) that were not found in interactions with the robotic agent. In the automated vehicle task, compliance varied by the agent type, agent driving style, and driving scenario despite the anthropomorphic agent being consistently preferred. Evidently, the implementation of human qualities in an AI agent does not guarantee more conformity, cooperation, or compliance to the agent. A possible theoretical explanation, garnered from these four studies, is that automation bias amplifies the effects predicted by the Computers are Social Actors theory, leading people to have higher subconscious social performance expectations of an anthropomorphic AI agent in interactive tasks than a nonanthropomorphic agent or other humans. Developers should consider the desired human behavior, contextual factors, performance of the technology, and social influence type before applying human-like features to AI technology
Advanced Disaster Relief and Aid Planning Tool
No full textPresented at AIAA SciTech 2026Defense acquisition and science and technology (S&T) investment decisions are inherently uncertain, requiring a careful balance between short-term capability procurement and long-term innovation. To explore this trade-space in a transparent, data-rich environment, the Advanced Disaster Relief and Aid Planning Tool (ADAPT) leverages Humanitarian Aid and Disaster Relief (HADR) operations as an unclassified analog for military missions. ADAPT integrated agent-based and discrete-event modeling with an interactive dashboard to assess how alternative fleet compositions, technology infusions, and acquisition strategies affect mission outcomes such as amount of aid delivered, aid delivery time, and asset utilization. The 2025 iteration of ADAPT expands the framework by introducing three major advancements: (1) expanded quantification of risk within the dashboard; (2) a redesigned Tabletop Exercise (TTX) format featuring structured participant roles, defined objectives, and dynamic scenario vignettes; and (3) integration of uncertainty and variability in risk, cost, and vignettes to mirror real-world acquisition environments. These enhancements enable participants to explore procurement and S&T investment tradeoffs interactively, observing how decision strategies evolve under uncertainty. Through multiple TTX campaigns based on Australia’s 2016 Cyclone Winston response, participants developed adaptive acquisition strategies, most notably a 25/75 allocation between investment and procurement, emerging from live negotiation and risk-balancing. The results demonstrate how risk-informed, role-driven experimentation can illuminate decision dynamics in complex acquisition environments and support data-driven recommendations for future defense and humanitarian operations
Stable Routing and Control System Requirements for Lunar Optical Communication Networks
No full textFuture lunar missions will require high-capacity, delay-intolerant communication networks between the Moon and Earth, pushing the need for optical satellite constellations. This study challenges the assumption of perfect line-of-sight communication by accounting for the physical limitations of satellite control systems required to maintain optical links. A routing strategy is proposed that prioritizes stable, longer lasting paths over traditional shortest path solutions, reducing the frequency of satellite slews. A time-varying graph-based framework is used to compare routing methods and estimate the resulting angular velocity and torque requirements imposed to the spacecraft. Results suggests that the stability-aware routing method can reduce the number of reorientation events and lower average control torque requirements, despite a moderate increase in path length. These findings offer a more physically grounded perspective on the feasibility and sustainability of large-scale cislunar optical networks
