150813 research outputs found
Sort by
Deep Flows Transmitted by Forced Surface Gravity Waves
No full textWe examine a two-dimensional deep-water surface gravity wave packet generated by a pressure disturbance in the Lagrangian reference frame. The pressure disturbance has the form of a narrow-banded weakly nonlinear deep-water wave packet. During forcing, the vorticity equation implies that the momentum resides entirely in the near-surface Lagrangian-mean flow, which in this context is often called the “Stokes drift”. After the forcing turns off, the wave packet propagates away from the forcing region, carrying with it most of the energy imparted by the forcing. These waves together with their induced long wave response have no momentum in a depth integrated sense, in agreement with the classical results of Longuet-Higgins and Stewart (Deep Sea Research and Oceanographic Abstracts 11, 592−562) and McIntyre (Journal of Fluid Mechanics 106, 331−347). The total flow associated with the propagating packet has no net momentum. In contrast with the finite-depth scenario discussed by McIntyre (Journal of Fluid Mechanics 106, 331−347), however, momentum imparted to the fluid during forcing resides in a dipolar structure that persists in the forcing region—rather than being carried away by shallow-water waves. We conclude by examining waves propagating from deep to shallow water and show that wave packets, which initially have no momentum, may have non-zero momentum in finite-depth water through reflected and trapped long waves. This explains how deep water waves acquire momentum as they approach shore. The artificial form of the parameterized forcing from the wind facilitates the thought experiments considered in this paper, as opposed to striving to model more realistic wind forcing scenarios
From Dialogue to Decision: An LLM-Powered Framework for Analyzing Collective Idea Evolution and Voting Dynamics in Deliberative Assemblies
No full textDeliberative assemblies—representative samples of citizens engaged in collective decision-making through facilitated learning and deliberation—are increasingly recognized as powerful tools for revitalizing democratic governance. Yet, core aspects of how deliberation shapes which ideas advance, how perspectives evolve, and why certain recommendations succeed remain opaque and underexamined. This thesis addresses these gaps by investigating: (1) How might we trace the evolution and distillation of ideas into concrete recommendations within deliberative assemblies? and (2) How does the deliberative process shape delegate perspectives and influence voting dynamics over the course of the assembly?
To answer these questions, I develop LLM-based methodologies for empirically analyzing transcripts from a tech-enhanced student deliberative assembly. The first framework identifies and visualizes the space of expressed suggestions, revealing that seemingly large gaps between ideas and final recommendations often reflect productive deliberative filtering—while also surfacing overlooked viable ideas.
A second analysis integrates post-assembly survey data with transcript-grounded voting patterns to uncover the primary drivers of vote change: edits to recommendations, evolving opinions, and strategic shifts in response to updated priorities. Building on this, I introduce a framework for reconstructing each delegate’s evolving stance across the assembly, linking shifts in perspective to specific deliberative moments and justifications.
Together, these methods contribute novel empirical insight into deliberative processes and demonstrate how LLMs can surface high-resolution dynamics otherwise invisible in traditional assembly outputs. The findings lay groundwork for new tools that support facilitators and delegates during live assemblies, improve transparency for decision-makers, and elevate ideas that may otherwise be missed.
Looking ahead, this work opens pathways for comparative research across assemblies and highlights the potential for human-centered AI to meaningfully enhance deliberative democratic practice. As societies seek new modes of participatory governance amid growing polarization and institutional mistrust, tools that strengthen deliberation without compromising its core human character are urgently needed.S.M
Enabling End-to-End Sensitivity Analysis of Integrated Models
No full textAs space-based precision-pointed telescopes continue to grow in scale and complexity, integrated models are increasingly relied upon to inform early design decisions and support system-level verification. When ground testing of full-system configurations is infeasible, integrated models, including structural-thermal-optical performance models, are essential for predicting performance and validating requirements across multidisciplinary, coupled domains. In early design phases, when uncertainty is high and design decisions have long-term implications for cost and schedule, it is especially important to understand which uncertain parameters most influence system performance. Global sensitivity analysis can help identify dominant uncertainty sources and inform decisions about model reduction, testing priorities, and resource allocation. However, the computational cost of applying global sensitivity analysis to integrated models often exceeds available resources. The presence of cross-disciplinary coupling between subsystem models further complicates analysis efforts. Coupled and dependent variables obscure how specific inputs influence system-level performance, limiting the ability to reduce model dimensionality or focus testing efforts on individual subsystems. There is a need for integrated modeling methodologies that enable tractable global sensitivity analysis of large, feedforward-coupled systems while preserving the accuracy needed to support early-phase design.
This thesis develops both exact and approximate methods for performing global sensitivity analysis on integrated models. A set of exact propagation techniques is introduced to compute end-to-end sensitivity indices when specific structural conditions are met, including functional linearity, non-interacting transforms, and monotonic intermediate mappings. These methods are evaluated using a suite of benchmark test cases that isolate when the exact sensitivity analysis method is valid and when structural assumptions begin to break down. A modular modeling framework is developed to compute exact or approximate end-to-end sensitivity indices and to enable automated mapping between disciplinary models in the integrated chain. The approach is also applied to a representative linearized structural-thermal-optical performance model, demonstrating how end-to-end global sensitivity analysis can be performed efficiently across thermal, structural, and optical subsystems.
To extend tractable sensitivity analysis to black-box models, several approximate strategies are introduced, including multifidelity surrogate modeling and statistical regression. These methods support both forward uncertainty propagation and variance-based global sensitivity analysis for structurally complex integrated models, without requiring full-system evaluation at every iteration. Together, the exact and approximate strategies developed in this work provide a foundation for scalable end-to-end global sensitivity analysis in early-phase design, where identifying influential parameters and constraining model complexity are essential for evaluating candidate architectures and informing mission decisions.Ph.D
Toward smart carbon capture with machine learning
No full textMachine learning (ML) is emerging as a powerful approach that has recently shown potential to affect various frontiers of carbon capture, a key interim technology to assist in the mitigation of climate change. In this perspective, we reveal how ML implementations have improved this process in many aspects, for both absorption- and adsorption-based approaches, ranging from the molecular to process level. We discuss the role of ML in predicting the thermodynamic properties of absorbents and in improving the absorption process. For adsorption processes, we discuss the promises of ML techniques for exploring many options to find the most cost-effective process scheme, which involves choosing a solid adsorbent and designing a process configuration. We also highlight the advantages of ML and the associated risks, elaborate on the importance of the features needed to train ML models, and identify promising future opportunities for ML in carbon capture processes
Systematic discovery of subcellular RNA patterns in the gut epithelium
No full textBackground Subcellular RNA localization is crucial for the spatio-temporal control of protein synthesis and underlies key processes during development, homeostasis, and disease. In epithelial cells, RNA can localize asymmetrically along the apico-basal axis. Yet, the localization of most transcripts as well as the diversity of patterns that they adopt remains unexplored. Results Here, we use APEX-seq for proximity labeling and MERFISH for spatial transcriptomics to map subcellular transcript localization in intestinal organoids and tissue from adult mice. Many transcripts present localization bias, often localizing in granular structures. We uncover intrinsic and environmental factors that influence the formation of these patterns. Additionally, we identify translation-dependent and -independent localization patterns and pinpoint the role of 3′ untranslated regions and RNA-binding proteins. Conclusions This subcellular RNA atlas presents a detailed resource for understanding intestinal physiology
Scaling contrastive learning batch size by two orders ofmagnitude
No full textContrastive learning has emerged as a powerful framework for unsupervised representation learning, allowing models to learn by maximizing agreement between related samples and distinguishing dissimilar ones. However, contrastive learning frameworks are fundamentally limited by the number of negative pairs a model can observe, and memory-intensive backbones constrain practical batch sizes. We introduce a three-phase, adapter-augmented training framework that scales contrastive batch sizes by two orders of magnitude – surpassing previous state-of-the-art learners in both accuracy and speed. First, we co-train the backbone and adapter on small batches to establish a strong initialization. Next, we freeze the backbone and train the adapter alone with very large batches, exposing it to an enlarged negative pool. Finally, we transfer large-batch adapter gradients back into the backbone via segmented backpropagation. We evaluate our method on the PlacesAudio dataset and show promising results for boosting retrieval performance at each phase. By exposing the model to substantially more negatives per effective batch, we achieve higher accuracy at a faster speed than optimizer-stepping baselines. Ultimately, this approach that scales batch size by hundreds of times can be integrated into any contrastive learning framework for more robust representation learning and abundant negative sampling.M.Eng
Regressionally-Estimated, CDE-Optimized, Integrated Into Launch (RECOIL) Weaponeering
No full textCollateral Damage is a large concern for military
operations. The use of weaponeering software attempts to mitigate
effects on collateral concerns while maximizing effects on
the target. Unfortunately, this software is not available during
dynamic targeting, which is the majority of operations for the
AC-130 and other Special Operations Forces (SOF) aircraft.
Modeling munitions effects against targets and optimizing employment
parameters for Precision-Guided Munitions (PGMs)
enables real-time alleviation for collateral concerns. It also has
the added effect of reserving surplus munitions for large scale
combat operations. This paper outlines the implementation of
weaponeering models onto the AC-130J gunship using regression
estimation and gradient-boosted decision tree machine learning.
The AGM-176 model achieved an average of 0.81 R2 across all
armored target sets with a MAE of 0.041. The HFR9E model
also achieved an average R2 of 0.81, with a MAE of .040. This
shows each specific probability prediction has an average error
of 4 percent, which is acceptable for in-flight weaponeering.The Department of the Air Force Artificial Intelligence Accelerato
The Situation Awareness Framework for Explainable AI (SAFE-AI) and Human Factors Considerations for XAI Systems
No full textRecent advances in artificial intelligence (AI) have drawn attention to the need for AI systems to beunderstandable to human users. The explainable AI (XAI) literature aims to enhance human under-standing and human-AI team performance by providing users with necessary information about AI sys-tem behavior. Simultaneously, the human factors literature has long addressed importantconsiderations that contribute to human performance, including how to determine human informa-tional needs, human workload, and human trust in autonomous systems. Drawing from the human fac-tors literature, we propose the Situation Awareness Framework for Explainable AI (SAFE-AI), a three-level framework for the development and evaluation of explanations about AI system behavior. Ourproposed levels of XAI are based on the informational needs of human users, which can be deter-mined using the levels of situation awareness (SA) framework from the human factors literature. Basedon our levels of XAI framework, we also suggest a method for assessing the effectiveness of XAI sys-tems. We further detail human workload considerations for determining the content and frequency ofexplanations as well as metrics that can be used to assess human workload. Finally, we discuss theimportance of appropriately calibrating user trust in AI systems through explanations along with othertrust-related considerations for XAI, and we detail metrics that can be used to evaluate user trust inthese systems
Modular Development Platforms and Creative Ecosystems: Design & Deployment for Wide Impact Across Fields
No full textPhysical, digital, and conceptual tools and building blocks are fundamental enablers and accelerators of humanity’s progress in technology, science, medicine, art, and even in abstract fields like mathematics, philosophy, and social sciences. Hardware development platforms present a special class of tools and building blocks, facilitating and accelerating innovation, prototyping, and research. They drastically reduce prototyping time and complexity, improve efficiency for experts, democratize access to innovation, and even inspire entirely new ideas. This research investigates how to design, develop, and deploy development platforms in ways that maximize their real-world impact potential. It focuses not only on the technical and engineering aspects, but also on the complete ecosystem a platform needs in order to have impact, including community building, engagement with users and volunteers, content strategy, online presence, publicity, deployment, feedback loops modularity, financial viability, and symbiotic relationships. A comprehensive Design & Deployment Framework is introduced as a conceptual tool for creating high-impact platforms and creative ecosystems, recognizing and fostering the positive feedback loops that sustain them and that shape their evolution and growth. This framework is applied in the development and deployment of multiple novel platform and ecosystem projects, including FlowIO, SleeveIO, and ModiStrap, as well as the ecosystem SoftRobotics.IO. Those works have benefited thousands of people around the world, providing researchers, designers, and engineers with powerful, reconfigurable, modular enabling artifacts that streamline prototyping, accelerate research, and lower barriers in fields like soft robotics, haptics, assistive technology, shape-changing interfaces, interactive arts, and more. A multitude of research, art, and engineering projects made possible by FlowIO and SoftRobotics.IO are presented, as well as over a dozen case studies showcasing how other users across disciplines have adopted, utilized, and extended these systems to advance their own creative, educational, and technical endeavors. Additionally, this thesis also investigates various deployment models for hardware and introduces a new hardware deployment model for equitable access to expensive hardware that may otherwise be financially out of reach for many users, as well as an “earned open-source” model, which preserves the essence of the traditional open-source model, while eliminating many of its pitfalls.Ph.D
C. elegans as a Platform for Multimodal Neural Data Integration
No full textSystems neuroscience has traditionally been fragmented into investigations at discrete levels of organization, creating methodological and conceptual gaps that hinder unified understanding of neural function. This thesis examines the nematode Caenorhabditis elegans as a platform for integrating diverse neural data modalities, offering a pathway to bridge these gaps. The hermaphrodite C. elegans, with its completely mapped connectome, optical transparency, genetic tractability, and stereotyped nervous system of only 302 neurons, presents an opportunity for comprehensive measurements across multiple dimensions of neural function. The review is organized around three fundamental neural data modalities accessible in C. elegans: (1) molecular genetic profiles, (2) network connectivity, and (3) neural activity dynamics. Historically studied in isolation, these complementary data types are increasingly being bridged through technological and computational innovations. We examine experimental advances enabling whole-nervous-system measurements of these modalities, as well as data standardization efforts and computational frameworks for cross-modal integration. While understanding the relationship between neural activity and behavior remains a fundamental goal of systems neuroscience, this thesis focuses on neural data acquisition and integration rather than behavioral analysis, which has been extensively covered elsewhere.1 We conclude with some original proposals to overcome current limitations in multimodal data acquisition and synthesis, and suggest future directions toward a holistic understanding of how molecular components, network connectivity, and cellular physiology collectively give rise to neural function in C. elegans. These integrative approaches establish a roadmap that may eventually scale to more complex nervous systems and advance our understanding of neural computation across species.S.M