137091 research outputs found
Sort by
Engineering Next-Generation CRISPR Platforms for Mammalian Transcriptional Control
No full textThe synthetic control of gene expression is essential for understanding cellular function. Recently, the advent of targeted DNA-binding proteins that can be coupled with protein subdomains capable of altering local transcription (effectors) has enabled precise gene expression control strategies in any eukaryotic host, including human and other mammalian cells. Despite the vast promise of CRISPR-dCas9 technology as a broadly applicable genetic tool for highly specific gene expression control, current systems exhibit highly variable performance across both gene targets and cell lines. Furthermore, knowledge of effector behavior, and how multiple effectors work in tandem to regulate gene expression, remains elusive. To address these limitations, we systematically classified the activity of candidate effector domains, explored the effects of attaching them in various combinations, and utilized this knowledge to engineer highly potent CRISPR-based platforms.
First, we constructed a fluorescence-based assay to simultaneously quantify the expression level and activity of candidate transcription factor subdomains. After discovering a panel of high-activity repressor domains, we designed several combinatorial libraries fusing these new effectors to gold-standard systems for CRISPR-mediated gene suppression (CRISPRi) and discovered five novel domain combinations improving performance. Through robust validation, we present a next-generation CRISPR repression system, ZIM3(KRAB)-MeCP2(t), that significantly outperformed prior systems across cell lines, gene targets, and recruitment approaches. Second, we performed in-depth functional characterization of one canonical repressor protein, MeCP2, and identified two unique subdomains that individually promote effector activity, even beyond levels observed with ZIM3(KRAB)-MeCP2(t). Using confocal microscopy, we determined that nuclear localization specificity is a key determinant underlying CRISPR system performance and demonstrated further that appending nuclear localization signal (NLS) domains to a variety of transcriptional effectors vastly improves their performance. Third, we performed functional analysis of a larger panel (~80 total) of transcription factor subdomains. By quantifying activity across multiple cell lines, we construct a novel classification system and identify numerous domains that can behave both as activators and repressors depending on biological context. Using this framework, we designed combinatorial libraries to explore how specific activator and repressor subdomains work together to regulate gene expression and identified several potent activator-activator effector combinations. Finally, by incorporating several novel domains and nuclear localization optimization, we introduce a new best-in-class CRISPR repression system, ZIM3(KRAB)-NID-MXD1-BPSV40NLS, which demonstrates the strongest reported gene knockdown capabilities to-date compared to existing CRISPR tools.
This work represents a fundamental step forward in understanding the activity, context-specific behavior, and cooperativity of transcriptional protein subdomains, and how this knowledge can inform the design improved tools for CRISPR-mediated gene regulation. Furthermore, we envision that our novel, top-performing repression system, ZIM3(KRAB)-NID-MXD1-BPSV40NLS, will be a particularly valuable tool, enabling unmatched gene knockdown capability across genes of interest, cell types, and applications.Ph.D.Chemical and Biomolecular Engineerin
Machine Learning Reconstruction of Intake Activity from Historical Sr-90 Inhalation Beagle and Pu-238 Injection Mice Datasets
No full textComputational internal dosimetry is a multifaceted chapter of health physics that operates at the intersection of systemic biokinetic models that describe the intake and time-varying biodistribution of radionuclides in the body, which in turn informs internal dose estimates to assess health risks. Biokinetic models, primarily published by the International Commission on Radiological Protection (ICRP), are standard compartment models that may be encapsulated by systems of first-order differential equations, whose constants (i.e., transfer coefficients, which then give rise to dose coefficients) are published for different age groups. In the case of a radiological accident, these models are invaluable for reconstructing the intake activity for individuals and the subsequent internal dose estimate. However, the intake reconstructions and dose estimates that are made from these values may lack adequate individualization, which is a consequence of the inherently high biological variation seen among different persons. Efforts towards the stochastic sampling of biokinetic parameters aid in this shortcoming, however, the present study implements machine learning (ML) to simultaneously investigate the data that would be needed to accurately reconstruct intake activity and the limitations of machine learning in refining current and future biokinetic models. Machine learning is a subset of artificial intelligence that excels in uncovering complex, and often highly non-linear, patterns in data through the use of advanced statistical methods, thus, it lends well to the highly uncertain physical system modelled by biokinetics and investigated through bioassay. The study implements two distinct datasets of internal exposures and two forms of machine learning (neural networks and random forests). The first dataset comes from historical experimentation performed by the Inhalation Toxicology Research Institute (ITRI) on beagles inhaling a soluble form of Strontium-90. The second is an aggregation of contemporary studies that all saw mice injected with known activities of Plutonium-238. Overall, the present study implements various data preprocessing techniques, augmentations, and simulations to drive cohorts of ML models which inform data quality/needs while highlighting the limitations of ML to assist in this consequential field of health physics.M.S.Medical Physic
Process Characterization of Hydrogen Direct Reduction of Stainless Steel Oxides
No full textFabrication of complex stainless-steel components using conventional melting and forming methods is often limited by geometric constraints, high energy consumption, and tooling costs. This study investigates the direct hydrogen reduction and sintering of metal–metal oxide extrusions designed to replicate the composition of 316L stainless steel. Extruded samples containing varying solids loadings and organic binders were characterized using X-ray diffraction (XRD), energy-dispersive spectroscopy (EDS), and thermogravimetric analysis (TG) to monitor thermolysis, reduction and sintering. Full reduction of iron and nickel oxides was achieved at or before 950 °C, while chromium oxide formed spinel phases, reduced to chromium metal as a member of the iron solid solution, or remained as eskolaite (a polymorph of Cr2O3). The sintering temperature and time required for densification were correlated with extrudate composition; higher solids loading require a higher hold of 1250 °C, whereas lower solids content allows for a lower 1200 °C process temperature. Higher temperatures and longer dwell times resulted in a higher proportion of austenite compared to ferrite. This process enables fabrication of complex thin-walled geometries with reduced energy consumption and lower tooling requirements compared to traditional stainless steel manufacturing techniques. These results provide a foundation for extending hydrogen-based reduction to other alloy systems, including nickel-based superalloys, and highlight the potential for increased domestic, water-based emission metal production
Ultrasound-Mediated Mild Hyperthermia Strategies for Targeting Brain Cancer
No full textUltrasound-mediated thermal stress is a promising strategy for treating brain cancer, providing a unique approach to targeted intervention. In contrast to high-temperature focal hyperthermia and ablation therapy causing direct cell destruction, the focus on mild hyperthermia (38 - 42°C) offers a spatiotemporal method for brain cancer targeting. Despite indications from studies in peripheral tumors that mild hyperthermia can modify the tumor microenvironment, its potential in brain cancer treatment remains largely unexplored due to challenges in safely applying hyperthermia in the brain. Through establishing a robust intracranial mild hyperthermia system and exploring its implications, this work demonstrates that FUS-mediated hyperthermia, combined with heat-sensitive anticancer agents (TSL-Dox and TS.BTE CAR T cells) can create unique opportunities for safer and more effective treatments against aggressive brain tumors. Moreover, this research by refining our understanding of thermal stress impact on the brain TME can support novel treatment paradigms.Ph.D.Mechanical Engineerin
Adaptable Policies and Accelerated Infrastructure for Learning for Heterogeneous Multi-Robot Coordination
No full textMany real-world challenges require heterogeneous agents to work together in a way that leverages their diverse capabilities to complete complex tasks. Heterogeneous multi-robot teams are a promising platform to address these real-world problems, and several algorithmic approaches exist to specify coordinated behavior for robots within these teams. Recently, learning-based approaches have emerged as a promising avenue for alleviating the burden of technical expertise and domain knowledge required to explicitly specify robust coordination behaviors in complex tasks. However, the adopters of learning based approaches face two significant tradeoffs when applying learning to heterogeneous robot teams:
(1) Policies for multi-robot teams can either be represented as a single set of parameters shared across all robots, a set of parameters learned for each class of robots, or fully individual parameters learned for each robot. Existing shared-parameter designs prioritize sample efficiency by enabling a single set of parameters to learn from the experience of all robots or robots within the same class using input augmentations, but tend to limit behavioral diversity. In contrast, learning separate policies for each robot enables greater diversity and expressivity at the cost of efficiency and generalization to unseen robots.
(2) Existing platforms for training multi-robot policies often force a tradeoff between optimization for multi-agent learning, robotics relevance, and sim-to-real deployment capability. Existing multi-agent benchmark simulators are highly optimized for learning with multiple-agents, but they lack fidelity. Existing robotics simulators are high-fidelity, but are not optimized for simulating and training multiple interacting agents and do not support open-access sim-to-real deployment.
In this work, we aim to make progress on addressing these two tradeoffs, both at the architecture level and the infrastructure level. In the architecture thrust, we view shared parameters and individual parameters as two ends of a broader spectrum and propose a middle ground approach: Capability Aware Shared Hypernetworks (CASH). CASH is a soft weight sharing architecture that uses hypernetworks to efficiently learn a flexible shared policy that dynamically adapts to each robot post training. CASH outperforms baseline architectures in terms of performance and sample efficiency during both training and zero-shot generalization, all with 60%-80% fewer learnable parameters. In the infrastructure thrust, we contribute JaxRobotarium, a Jax-powered end-to-end simulation, learning, deployment, and benchmarking platform for the Robotarium. JaxRobotarium enables rapid training and deployment of multi-robot reinforcement learning (MRRL) policies with realistic robot dynamics and safety constraints, supporting both parallelization and hardware acceleration. With eight natively implemented benchmark tasks, We demonstrate that JaxRobotarium retains high simulation fidelity while achieving dramatic speedups over baseline (20x in training and 150x in simulation), and provides an open-access sim2real evaluation pipeline through the Robotarium testbed, accelerating and democratizing access to multi-robot learning research and evaluation
Efficient Inter-Process Communication Mechanisms for Data-Centric Computing
No full textModern data-intensive applications, ranging from large-scale analytics pipelines to multimodal AI inference systems, are increasingly constrained by the overhead of transferring and managing large volumes of intermediate data across complex memory hierarchies. While emerging hardware innovations, such as programmable memory engines, offer potential solutions, existing Inter-Process Communication (IPC) mechanisms remain rigid and CPU-centric, failing to adapt to dynamic workload behaviors or to leverage these architectural advances effectively. This thesis explores how IPC services can be redesigned to overcome these limitations, proposing a new class of adaptive IPC mechanisms that integrate architecture-level advances with system-level flexibility.
First, the thesis introduces \textbf{Pocket}, a resource-aware IPC system designed to enable efficient split-architecture deployments. Addressing the ``boundary tax'' inherent in decoupled components, Pocket integrates lightweight resource management directly into the messaging interface. By allowing messages to carry resource expectations, Pocket enables just-in-time resource amplification and receiver-side adaptation. This design eliminates the performance penalties typically associated with isolation, effectively bridging the gap between monolithic efficiency and microservice flexibility.
Second, the thesis presents \textbf{Rocket}, a multi-backend IPC runtime that intelligently offloads memory copy operations to hardware accelerators (e.g., Intel DSA) or optimized kernel services. Unlike naive offloading approaches, which can degrade performance due to cache interference or synchronization overheads, Rocket employs backend- and workload-aware strategies to orchestrate data movement. It provides a suite of execution modes and hybrid completion check mechanisms to determine when offloading is beneficial based on data volume, locality, and system noise, thereby maximizing computation-communication overlap in high-throughput pipelines.
Finally, the thesis proposes \textbf{SkyRocket}, a runtime-adaptive framework that optimizes IPC for heterogeneous data flows. Unlike static or naive configurations that apply a uniform strategy regardless of payload characteristics, SkyRocket leverages lightweight workload signatures to drive real-time adaptation. By dynamically switching between execution modes and backend strategies, it achieves adaptive control, ensuring that the IPC mechanism evolves in lockstep with the varying demands of multimodal applications. SkyRocket effectively narrows the performance gap between general-purpose IPC stacks and finely tuned, task-specific solutions.
Together, Pocket, Rocket, and SkyRocket demonstrate a practical and scalable path toward performance-aware IPC systems. By aligning IPC logic with modern hardware capabilities and application-level variability, this dissertation presents a comprehensive design space for efficient data movement in the era of hardware-accelerated, data-centric computing
Incorporating Geometric and Consistency Constraints into Deep Models for Robust Phase Reconstruction and Speech Enhancement
No full textPhase reconstruction of short-term Fourier transform (STFT) spectra remains a central challenge in speech generation tasks.
Although early studies emphasized magnitude over phase, recent work has shown that accurate phase reconstruction is crucial for reducing artifacts and distortions that degrade speech quality in enhancement (SE).
With the rapid advancement of deep neural networks (DNNs), recent research has aimed to estimate both the magnitude and phase of clean spectrograms simultaneously.
However, most approaches directly predict the phase spectrogram—a task made difficult by the phase’s unstructured nature, wrapping ambiguities, and extreme sensitivity to time shifts.
Such direct estimation also overlooks alternative phase configurations that can yield perceptually valid speech.
This dissertation proposes a new framework for phase estimation that overcomes these limitations and demonstrates its effectiveness across multiple DNN-based SE models.
We begin by introducing the first deep state-space-based SE model operating on complex-valued spectrograms.
While it surpasses baseline models with a compact U-Net architecture, its estimated phase offers limited improvement over the noisy phase, underscoring the difficulty of direct phase prediction.
To address this, we develop a novel explicit consistency-preserving loss that leverages the observation that perceptually high-quality speech arises when magnitude and phase are mutually consistent.
Building on this insight, we integrate geometric constraints under additive noise conditions with the consistency principle, resulting in the Multi-Sourced Griffin-Lim Algorithm (MSGLA).
MSGLA jointly refines speech and noise phases through iterative updates guided by DNN-estimated magnitudes and geometric relationships, outperforming direct phase estimation and prior geometric methods.
Finally, we extend these ideas to a large-scale generative pretraining framework that models the distribution of clean speech spectrograms and incorporates the consistency-based phase loss during training
Mechanism-Based Study of the Performance of Printed RF Elements under Stretching and Bending
No full textFlexible hybrid electronics (FHE) has attracted substantial interest by combining the high performance of silicon integrated circuits (ICs) and mechanical versatility of flexible printed circuits. FHE systems are being increasingly used in health, automotive, and Internet-of-Thing (IoT) applications, where Radio Frequency (RF) communication is an important element for the wireless operation of the FHE devices. The interconnect is a key component in the RF module of the FHE, as it connects the ICs, antennas, and other passive components together. Thus, understanding the RF performance of printed interconnects under various extents of deformation in their operating conditions, such as bending and stretching, is crucial to the reliable design of FHE devices. A more accurate evaluation of RF performance requires the characterization of scattering parameters (S-parameters). The scarcity of experimental data and numerical models for FHE relevant components in the RF regime remains a major obstacle to improving the reliability of FHE devices. This dissertation explores the multi-physics characterization and mechanism-based modeling of RF performance changes in printed interconnects subjected to mechanical deformation. Through a combination of empirical testing, computational simulations, and modeling improvements, this work addresses critical gaps in understanding how stretching and bending impact the S-parameters of screen-printed coplanar waveguides (CPWs) and inkjet-printed microstrip lines.
The study begins by thoroughly characterizing the material sets, analyzing their mechanical and electrical properties to establish a foundation for design and fabrication of high-quality printed interconnects. Once the fabrication is completed, it then investigates changes in S-parameters through stretching and bending experiments, defining failure criteria and benchmarking results for simulation validation. This includes developing testing methodologies compatible with S-parameter measurements. Numerical models are then first developed in CSTTM with the assumptions of changed geometry and conductivity degradation due to mechanical strain in three cases: in-situ stretching of CPWs, in-situ bending of microstrip lines, and post bending up to 12500 cycles of microstrip lines. With the omission of other potential strain-induced mechanisms, the predictive models consistently underestimate the extent of RF degradation observed in measurements. To address these shortcomings, other experimentally observed mechanisms, including strain-induced changes in dielectric properties and surface roughness of printed conductors are systematically integrated into refined CSTTM simulation models for all three cases. By incorporating all deformation-dependent material parameters, these models demonstrate significantly improved prediction accuracy across a range of strain and frequency levels. Validation against measured S-parameters confirms the necessity of incorporating multiphysics-informed adjustments of dielectric constant of and conductor surface roughness, beyond traditional geometry and conductivity assumptions.
Ultimately, by integrating material characterization, experimental data, and simulation insights, this dissertation establishes a mechanism-based modeling strategy that bridges empirical characterization and predictive simulation, contributing critical insights into the design and reliability of FHE. The methodology developed here serves as a valuable tool for future research and device development in FHE for wearable, automotive, and IoT applications requiring mechanically robust RF performance
Will AI Change Cartography, Or Will Cartographers Change AI?
No full textPresented on January 15, 2026 at 12:00 p.m. in the Centergy One Building, Hodges Room.Dr. Anthony Robinson is E. Willard & Ruby S. Miller Professor of Geography, Director of Online Geospatial Education Programs, and Director of the GeoGraphics Lab at Penn State. His research focuses broadly on designing and evaluating geovisualization tools to improve geographic information utility and usability.Runtime: 62:06 minutesAt its heart, making maps involves simplifying reality to help us navigate and to explain our world. Cartographers are trained in the art and science of shaping spatial data to support a wide range of user needs. The rise of AI begs us to consider how the process of designing geovisualizations should change. AI also offers the potential to serve as an interactive assistant to help people understand what they see in a geographic visualization. In this talk, I highlights results from recent efforts to envision how cartographers might use AI, as well as ongoing work to leverage AI approaches for improving map accessibility for people who are blind or visually-impaired. I conclude with reflections on why automating cartographic design is hard, and why I think it might remain that way despite the promises of AI
Enhancing Realism in Indoor Navigation
No full textThis body of work addresses key challenges in indoor navigation for embodied AI by enhancing simulation realism and developing adaptive training strategies to relax assumptions towards realism in policies. We introduce innovative methodologies in scene reconstruction and policy learning to advance goal-conditioned navigation in complex environments. The first of the works presents a novel framework that integrates 3D Gaussian Splatting with smartphone-based captures to create high-fidelity simulation environments. Our approach enables faster scene capture while maintaining high-quality performance, overcoming limitations of traditional, costly methods. Through extensive evaluations, we demonstrate the relationship between scene reconstruction quality, measured by Peak Signal-to-Noise Ratio (PSNR), and navigation performance, achieving successful policy transfer and improved performance in real-world scenarios. The second work tackles GPS dependence in indoor social navigation tasks, proposing curriculum learning strategies that progressively reduce GPS reliance. Our approach demonstrates significant performance gains, challenging conventional assumptions about GPS-based localization. Together, these contributions improve simulation fidelity, real-to-sim-to-real gap, and localization capabilities in agents, advancing the state-of-the-art in embodied AI navigation systems.M.S.Computer Scienc