80264 research outputs found
Sort by
Quantifying Power Loading of Self-Supported Carbon Nanotube Fiber Devices for Gas Heating
No full textAs industries seek more energy-efficient and sustainable technologies, the electrification of heating processes has gained focus, driving the demand for advanced materials that can operate reliably with high power loading at various temperature ranges. Many electrical heating devices utilize Joule heating of metal alloys (e.g. Kanthal or Nichrome) to transfer heat to flowing gases. Carbon nanotube fibers (CNTFs) are promising options to replace metal alloys as next-generation electrothermal heaters because the CNTF have similar electrical properties while also offering higher strengths, lower densities, and the ability to be processed using textile manufacturing due to their flexibility and cut resistance. However, prior measurements have not reported the achievable ranges of power loadings and specific power loadings for self-supported CNTF intended for flexible, lightweight gas electric heating applications. Here, we use Joule heating experiments supported by thermal modeling to quantify the convection heat exchange between flowing gases and CNTF electrothermal devices made entirely of CNTF monofilaments and CNTF textiles. Material characterization shows that moderate-temperature annealing of CNTF to remove remnant acid improves the specific strength and enhances the resistivity compared to as-spun fibers, as desired for Joule heating. Single-filament heating experiments in quiescent fluids show that CNTF monofilaments with 20 µm diameters can achieve specific power loadings ~ 8*10^8 W kg^(-1) in inert gases and >1*10^8 W kg^(-1) in air compared to 25 µm Nichrome, whose power loading is >14*10^6 W kg^(-1) and >28*10^6 W kg^(-1) in inert gases and air respectively. Similar power loading was observed when Joule-heated fiber array devices, used in our study to demonstrate the air heating. Experiments on CNT fabric devices in flowing air achieve power loadings >17*10^6 W kg^(-1) compared to >1*10^6 W kg^(-1) for textiles/meshes made of multifilament CNTF yarns and Nichrome wire, respectively. Thus, this work quantifies the ranges of achievable power loadings for high-strength CNTF devices and motivates future work optimizing the design of electrothermal CNTF textiles, arrays, and meshes for gas heating applications
Birth-Death Process Models for Clonal Hematopoiesis of Indeterminate Potential (CHIP)
No full textClonal Hematopoiesis of Indeterminate Potential (CHIP) is a condition marked by the expansion of hematopoietic stem cell (HSC) clones carrying somatic mutations, in the absence of overt hematologic malignancy. Although CHIP is associated with an increased risk of hematologic cancers and cardiovascular diseases, the clonal dynamics that govern its progression remain to be understood further.
This thesis develops and applies Birth-Death Process (BDP) models to study CHIP progression, with an emphasis on computationally efficient and statistically rigorous inference methods. The first part focuses on single time-point observations using the Linear Birth-Death Process (LBDP), modeled via negative binomial distributions for recurrent mutations and logarithmic distributions for single mutations. These models are validated through extensive simulation studies and applied to the ARIC dataset, enabling gene- and mutation-specific growth rate estimation.
To extend the analysis to longitudinal data, an approximate likelihood framework for LBDP is introduced, replacing the full likelihood with a proxy to enable faster computation. This method is analytically shown to be statistically consistent, and its performance is benchmarked against traditional approaches, including Galton-Watson estimators, saddlepoint approximations, and maximum likelihood estimation. Simulations and real-data applications from ARIC datasets demonstrate the accuracy and efficiency of the proposed method.
To further accommodate real-world variability in blood sampling, a hierarchical LBDP model is developed. Five estimation techniques are evaluated within this framework: the approximate estimator, Laplace approximation (mean trajectory approximation), Hamiltonian Monte Carlo, EM algorithm, and ODE-based inference. Simulation studies using ARIC-like data reveal that the mean trajectory approximation achieves the highest accuracy, while the approximate estimator offers a computationally efficient alternative. Additionally, the mean trajectory method is shown to extend naturally to more complex dynamics, including logistic growth.
By refining inference techniques and validating them across simulated and real-world datasets, this work contributes a robust statistical framework for analyzing CHIP dynamics. The proposed models enhance our understanding of clonal evolution in hematopoiesis and support the development of predictive tools for clinical and research applications
Efficient and Faithful Algorithms for Interpretable Machine Learning
No full textAs deep learning models continue to grow in complexity and scale, the demand for interpretable machine learning (ML) methods becomes increasingly critical across a wide range of applications. This thesis addresses the challenges of interpreting deep neural networks (DNNs) by designing efficient and faithful algorithms tailored to existing mainstream models: multilayer perceptrons (MLPs), graph neural networks (GNNs), vision transformers (ViTs), and large language models (LLMs). My ultimate goal is to develop frameworks that are not only theoretically grounded but also computationally efficient for interpenetrating DNN models.
For tabular data modeled by MLPs, we focus on accelerating Shapley value computation, a widely used method rooted in cooperative game theory. While Shapley values provide theoretically sound feature attributions, their computation is NP-hard due to the exponential number of input coalitions. To address this, we propose SHEAR (Shapley Explanation Acceleration Rule), a novel approach that leverages a theoretical chain rule to identify a small set of contributive cooperators that preserve attribution
accuracy while significantly reducing computation. SHEAR achieves substantial speed-up without degradation of fidelity across several benchmark datasets. For graph-structured data and GNNs, we propose LARA (Local Attribution via Removal-based Amortization), a fidelity-oriented framework for node attribution. Traditional GNN explanation methods often struggle with high computational cost and low faithfulness, especially on large-scale graphs. LARA addresses these limitations by introducing a bidirectional attribution mechanism that produces explanation-oriented node embeddings. It also incorporates subgraph sampling to enhance scalability and amortized training mechanism to generalize explanations across unseen nodes.
In the domain of vision models, particularly ViTs, we present the TVE (Transferable Vision Explainer) to enable efficient and reusable explanations. While existing vision explainers require retraining for each model and task, TVE introduces the concept of meta-attribution, a generalized, pre-trained attribution representation that can be adapted to diverse downstream tasks without further training. By pertaining TVE on large-scale image datasets, we demonstrate that it can generate faithful and transferable explanations for multiple vision architectures, such as ViT, Swin, and DeiT, across several datasets. This pretrain-once, explain-everywhere mechanism offers a scalable solution for vision interpretability in real-world deployments.
Finally, for interpreting LLMs, we focus on improving the faithfulness of natural language explanations. Existing approaches frequently yield inconsistent or non- faithful outputs due to the intrinsic complexity of LLMs and their one-pass generation
style. To address this, we propose a novel fidelity metric based on contrary explanations and introduce FaithLM, a self-consistency-based framework that iteratively refines natural language explanations using in-context learning. FaithLM leverages feedback from fidelity evaluations to optimize explanation prompts, which achieves significant alignment of explanations with model decisions.
Together, these contributions provide a comprehensive study for advancing the interpretability of modern AI systems across multiple data modalities. This thesis cannot only improve the transparency and trustworthiness of ML models but also act as a groundwork for safe and responsible AI deployment in high-stakes domains
Towards the Controlled Synthesis and Industrial Applications of Two-Dimensional (2D) Materials Guided by Machine Learning
No full textSince the first exfoliation of graphene from graphite in 2004, atomically thin two-dimensional (2D) materials have gained significant attention due to their unique properties that emerge during the transition from bulk to monolayer form. These characteristics have enabled a broad range of applications spanning nanoelectronics, optoelectronics, and energy systems. While substantial progress has been made in the discovery and synthesis of 2D materials, challenges remain in achieving controlled growth and scalable, cost-effective production.
Recent advances in integrating machine learning (ML) into practical engineering processes have accelerated its adoption in 2D materials research by reducing the labor required for large-scale data analysis. This thesis addresses the pressing challenges in 2D material development by establishing a data-driven framework to investigate both the top-down exfoliation and bottom-up synthesis mechanisms, while also exploring their potential for industrial applications.
A customized miniature CVD platform was developed to facilitate real-time optical monitoring of MoS₂ monolayer growth. Through image processing techniques, real-time growth footage was digitized, enabling the extraction of key morphological parameters such as nucleation density, crystal coverage, and growth rate. Machine learning algorithms were employed to correlate these parameters with process conditions, enabling predictive modeling and CVD optimization. This closed-loop system lays the foundation for future autonomous material synthesis platforms.
In parallel, a polymer-assisted dry ball-milling method was introduced for scalable exfoliation of hexagonal boron nitride (hBN). Using ML-based feature selection, critical polymer properties responsible for high exfoliation efficiency were identified. The method was successfully extended to exfoliate various layered materials, supporting its potential for large-scale manufacturing. Additionally, atomic-layer structured photovoltaics (ALSPs) incorporating 2D heterostructures were fabricated, demonstrating strong performance, stability, and flexibility, with promising applications in self-powered electronics.
In general, these studies establish a robust, data-driven strategy for controlled synthesis and industrial implementation of 2D materials
Nonlocal and nonlinear optical response and STM studies of quantum materials
No full textQuantum materials are unique in their long-range interactions and competing phases tunable by external stimuli. Due to the incommensuracy of the quantum order or competing phases, the volume of quantum materials is partitioned into multiple domains. Light-matter interaction in quantum materials presents a new paradigm as light can tip the balance between many competing quantum many-body phases and give rise to new phenomena.
In the field of light probing quantum materials, most studies focus on ultrashort high-energy probing; rarely has anyone tried to use low-energy light to probe the material in the linear response regime and still get interesting results. In this dissertation, I will present the results of low-intensity light probing of quantum materials. Firstly, I present a nonlocal model of the dielectric function and show it can accurately describe the angle-resolved spectrum of TaS2 in the visible. The competing stacking configurations of the charge domains in this layered material result in significant optical inhomogeneity that necessitates a nonlocal dielectric function. I performed intensity-sweep characterizations and used our model to predict the domain size dependence on light intensity. The non-local parameter extracted from our measurements sheds light on the competition between the two stacking orders.
Next, seeking direct microscopic evidence of light-induced stacking reconfigurations, I present our experimental results from the Laser-STM system probing the surface charge density under stable laser illumination. Despite the noise at room temperature and laser power instability, which prevent an accurate determination of stacking order configurations, the TaS2 topography images uniquely exhibit a clear low-frequency charge-density oscillation on the order of 0.2 Hz. To investigate the dynamics of this light-matter interaction, an optical chopper is used to modulate the laser illumination. I demonstrate the emergence of a breathing charge density wave modulated by the chopping frequency. Furthermore, I propose our conjectures and hypotheses regarding the physics underlying this novel phenomenon. Finally, I will present simulations results that utilize quantum materials to realize advanced phase control and to design an anomalous diffraction grating
[ ] In Progress - An Incremental Degrowth System
No full textThe prolonged vacancy problem in post industrial cities of the rust belt has been uprooted by numerous factors including deindustrialization, white flight, and suburbanization. Surmounting in communal, societal, and legislative pressure, the city of Detroit saw decline in population since the mid 20th century, which continues to disproportionately affect historically marginalized communities within the city limits. Since the beginning of industrialization, growth has become an expectation. This thesis explores how the community can maintain, retain, and repair some essence of the existing condition by prioritizing a system to manage a shrinking city. By reinforcing the local, contesting urban growth, and redefining ownership, the plethora of city owned parcels are incrementally collected and consolidated across multiple residential blocks to restructure areas of growth and provide a public space that gives back to the community. In the vicinity of the consolidated block, underutilized infrastructure and homes are deconstructed with the help of permanent and temporal built interventions. A series of lightweight modular buildings stretch across the central residential block, promoting communal activity and incentives for degrowth while alleviating maintenance efforts and costs to benefit nearby residents
Low-dimensional signatures of Aplysia neuronal population activity revealed by matrix factorization and spatial mapping
No full textOperant conditioning (OC) is a form of learning in which a specific behavior is reinforced by a reward. The memory spans different temporal domains ranging from minutes (short-term OC, STOC) to 24 h or longer (long-term OC, LTOC). This study investigated OC using Aplysia feeding behavior as a model system. Voltage-sensitive dye imaging was used to record activity of 100s of neurons in the buccal ganglia 24 h after training to examine the extent to which STOC and LTOC share common neural correlates. Non-negative matrix factorization revealed two factors that corresponded to the protraction and retraction phases of feeding behavior. Similar to STOC, LTOC resulted in earlier recruitment of the retraction factor. Notably, the number and contribution of neurons in the retraction factor increased in LTOC, but not STOC. These findings suggest LTOC has a different low-dimensional signature compared to STOC
Ultrastrong Vacuum–Matter Interactions in Semiconductors and Magnets
No full textThere is growing interest in using optical cavities to uncover new phases and phenomena, relying solely on vacuum electromagnetic fields enhanced within the cavity, without any external fields. Ultrastrong coupling (USC) between vacuum and matter is a prerequisite. USC occurs when the vacuum–matter coupling rate becomes a significant fraction of the bare frequencies of the systems. In this dissertation work, we studied the USC of vacuum and matter in semiconductors and magnets. First, we realized multimode phonon-polaritons in lead halide perovskites. The USC of two optical phonon modes with the vacuum induced novel vibrational properties. We also demonstrated via photoluminescence measurements that electron–phonon interactions are modified in this system. Next, in a quantum Hall system, we demonstrated the breakdown of the electric-dipole approximation using nanoscale cavities, revealing forbidden electronic transitions. We further investigated quantum–classical correspondence in the nonlinear regime, which was explained by a quantum model. Finally, we observed a magnonic superradiant phase transition in an ErFeO3 crystal. USC between an Fe3+ magnon mode and an Er3+ electron paramagnetic resonance resembled vacuum–matter interactions, leading to the phase transition. These findings provide quantum optical strategies for creating and controlling novel phases in condensed matter via control of the quantum vacuum surrounding the matter
2.2 The Importance of Redlines in the Life Sciences
No full textDeveloped from discussions hosted by the Pathogens and Bioweapons Theme group at the Spirit of Asilomar.This entreaty was created as part of The Spirit of Asilomar and the Future of Biotechnology summit (February 23-26, 2025) in Pacific Grove, CA.This Entreaty recognizes that there are experiments and experimental goals that should not be undertaken or pursued because the risks, as best they are understood, far outweigh the possible benefits. These experiments and goals should be defined by “redlines” that practitioners of science will not violate
Integrated Modeling of Future Energy Systems to Enhance Grid Reliability and Sustainability
No full textThe shift to a sustainable, low-carbon energy economy requires advancements in technology, system design, and policy reform. Electric power systems face challenges such as rising demand, price volatility, geopolitical tensions, and extreme weather events. Recent major outages have exposed vulnerabilities across the current energy system, revealing risks from extreme weather, fuel supply disruptions, and grid instability. Electrifying sectors such as industrial heating and transportation offers a pathway to decarbonize energy and reduce pollution-related deaths, requiring a comprehensive investigation.
This dissertation employs integrated modeling approaches to explore how technological, system design, and policy innovations can enhance the reliability and sustainability of U.S. energy systems. It tackles the computational complexity of linking energy system models with air quality assessments, forecasting future energy scenarios while factoring in air quality and health impacts.
The study shows the value of integrated modeling in plotting a decarbonized, reliable energy future, addressing both environmental and human health dimensions while adapting to regional and sectoral needs in the United States