2026 – Talks & Poster Sessions

Heparin/Collagen Multilayers Enhance Extracellular Vesicle Production in Human Mesenchymal Stem Cells
Authors: Melanie J. Nelson and Jorge Almodovar (Presenter)
Themes: Bioengineering, Biomimetics, Biosecurity

Leveraging cell-free synthetic biology and protein language models to produce biological polymers
Author: David Garcia (Presenter)
Themes: Bioengineering, Biomimetics, Biosecurity

New strategies for generating hybrid anion-exchange resins for selective treatment of (ultra)short-chain PFAS
Authors: Marylia Duarte Batista, Trevor Gibson, Emily Piazza, Ke He, and Lee Blaney (Presenter)
Themes: Environment and Sustainability

In vitro assessment of Bicuspid Aortic Valve (BAV) hydrodynamics with varying heart rate
Authors: Nadeem Shah (Presenter), Charles D. Eggleton, and Sayantan Bhattacharya
Themes: Bioengineering, Biomimetics, Biosecurity

Effects of Anisotropy and Defects on Mechanical Performance of Brittle Materials

Authors: Keith Bowman (Presenter), Gizaw Melese, Rokia Elgharably, Daniel Combs, Isaac Poole, Cenia Sims, and Ye Lu

Themes: Manufacturing

Acoustic Intelligence for Everyday Healthcare
Author: Dong Li (Presenter)
Themes: Healthcare

Sensorization, with Care: Community-Driven Smart Technology for Assisted Living
Authors: Tera Reynolds and Roberto Yus (Co-presenters)
Themes: Foundations and Applications of Artificial Intelligence, Human-centered Technology and Accessibility, Software Engineering, Environment and Sustainability, Energy, Physical Systems, Engineering and Technology for Social Good, Community Partnerships and Impact with Engineering and Technology

Closed-loop Embodied AI Under Physical Constraints for Health and Agriculture
Authors:  Anuradha Ravi (Presenter) and Nirmalya Roy
Themes: Foundations and Applications of Artificial Intelligence, Healthcare, Human-centered Technology and Accessibility, Communications Networks, Energy, Community Partnerships and Impact with Engineering and Technology

Evaluating the Impact of Pathogen-Mediated Infective Endocarditis on Bioprosthetic Aortic Valve Function
Authors: Corine Jackman Burden and Sayantan Bhattacharya (Co-presenters)
Themes: Healthcare, Bioengineering, Biomimetics, Biosecurity

COHERE: Collaborative Optimization of Human Engagement and Robot Effectiveness
Authors: Sruthi Sundharram, Jake Whitt, Golnaz Moharrer, Andrea Kleinsmith (Co-presenter), Charissa Cheah (Co-presenter), Christine Mallinson (Co-presenter), and Ramana Vinjamuri (Co-presenter)
Themes: Foundations and Applications of Artificial Intelligence, Healthcare, Human-centered Technology and Accessibility, Bioengineering, Biomimetics, Biosecurity

A Unified Multiphysics Modeling Framework for Photodetectors
Authors: Ishraq Md Anjum (Presenter), Seyed Ehsan Jamali Mahabadi, Thomas F. Carruthers, Ergun Simsek, and Curtis R. Menyuk
Themes: Physical Systems

Variational Gibbs State Preparation on Trapped-Ion Devices
Authors: Reece Robertson (Presenter), Mirko Consiglio, Josey Stevens, Emery Doucet, Tony J. G. Apollaro, Sebastian Deffner
Themes: Foundations and Applications of Artificial Intelligence, Quantum Information Science, Quantum Computing, Physical Systems

AI Agents in Offensive Security
Authors: Sairam Bokka (Presenter) and Keke Chen
Themes: Security and Privacy, Software Engineering

Exploring Peer-to-Peer Evaluation with an AI-Supported Learning Tool in an Introductory Programming Course
Authors: Ben Cohen (Presenter), Omobolanle Niyi-Owoeye, Kevin Lemus, Srushti Dharmale, Carine Marette, Patricia Ordóñez, and Edward Dillon
Themes: Foundations and Applications of Artificial Intelligence, Education

How to write bug-free scientific computing software
Author: Tyler Josephson (Presenter)
Themes: Foundations and Applications of Artificial Intelligence, Software Engineering

Sequentially Acquiring Concept Knowledge to Guide Continual Learning
Authors: Shivanand Kundargi (Presenter), Kowshik Thopalli, Tejas Gokhale
Themes: Foundations and Applications of Artificial Intelligence

Abstract: With the growing number of entrepreneurs and talented students at the University of Maryland, Baltimore County there is an urgent need for a centralized solution that allows students to advertise and engage in the services/products offered by students within the campus community. Currently, many student entrepreneurs and campus organizations use disconnected communication channels such as social media, word of mouth, or group chats, allowing for limited visibility and inconsistent outreach. As a result, many students on campus aren’t aware of these opportunities and believe they must leave campus to receive the services they need. The Campus Connect is a digital platform which offers a user-friendly website where on-campus student-run businesses can register, create profiles, and promote their products/services directly to the students of UMBC through campus marketplaces and service listings. The Campus Connect addresses the limited visibility and centralized support for student entrepreneurs while reducing barriers for students seeking affordable, convenient, and campus-accessible services. Initial usability testing with nine student participants indicated that users were able to quickly locate services, understand business offerings, and navigate the platform with minimal instruction. These findings suggest that centralized digital marketplaces can significantly improve awareness of campus-based services and reduce accessibility barriers.

Abstract: Created in a Team-Based Game Development class, Mouse House is a game that explores both three-dimensional environments in collaboration with 2D character style developed in the Unity game engine. The game utilizes and explores Unity gravity mechanics and a first-mouse point of view as you maneuver around enemies that can affect the player’s health using mechanics like radial enemies and fright systems through Unity C# Scripts. Playing as a mouse, Mouse House reimagines everyday household objects at a massive scale, transforming ordinary spaces into immersive and dynamic environments. A 3D environment was constructed with Autodesk Maya to provide an intricate world full of ample hiding spaces, and player interface and 2D assets were created in Krita and edited in Adobe Animate to create a hand-drawn sense of whimsy to contrast the large and harsh environment the player must maneuver. Development of this game provided for a uniquely engaging collaborative experience that allowed for cross-discipline exploration of programming, visual design, and overcoming challenges in a team-based setting.

• 1
  Authors: Mohammad Saeid Anwar, Anuradha Ravi, Emon Dey, Gaurav Shinde, Indrajeet Ghosh, Jade Freeman, Carl Busart, André Harrison, Nirmalya Roy
  Title: CoOpTex: Multimodal Cooperative Perception and Task Execution in Time-critical Distributed Autonomous system

  Abstract:

  Integrating multimodal data such as RGB and LiDAR from multiple views significantly increases computational and communication demands, which can be challenging for resource-constrained autonomous agents while meeting the time-critical deadlines required for various mission-critical applications. To address this challenge, we propose CoOpTex, a collaborative task execution framework designed for cooperative perception in distributed autonomous systems (DAS). CoOpTex contribution is twofold: (a) CoOpTex fuses multiview RGB images to create a panoramic camera view for 2D object detection and utilizes 360° LiDAR for 3D object detection, improving accuracy with a lightweight Graph Neural Network (GNN) that integrates object coordinates from both perspectives, (b) To optimize task execution andmeet the deadline, CoOpTex dynamically offloads computationally intensive image stitching tasks to auxiliary devices when available and adjusts frame capture rates for RGB frames based on device mobility and processing capabilities.We implement CoOpTex in real-time on static and mobile heterogeneous autonomous agents, which helps to significantly reduce deadline violations by 100% while improving frame rates for 2D detection by 2.2× in stationary and 2× in mobile conditions, demonstrating its effectiveness in enabling real-time cooperative perception.

• 19
  Authors: Molly E. Balkan, Shanmukhi Gundu, Sudip Chakraborty, Maloy K. Devnath, Vandana P. Janeja
  Title: Graph-based detection of spatial linkages between sea ice retreat and ice shelf melting over the polar regions.

  Abstract:

  If completely melted, the Antarctic (Arctic) region has the potential to raise the sea level by ~200 (~25) feet. Both the ice sheets over the land and the sea ice extent, which provides a protective buffer to the ice sheets, are melting fast over the polar regions. A recent study shows that interconnections between sea ice retreat and ice shelf melting events exist by identifying several pathways (spatial links) between them using graph theory. However, their graph-based spatial link detection theory has limitations in terms of the spatial distance between the SIE retreat and ice shelf melting and have analyzed only one month of data to identify the spatial linkages over the western Antarctic region only. While sea ice and land ice communities have conducted extensive research to determine the causes of sea ice retreat and ice sheet melt, there remains a knowledge gap on how sea ice retreat gradually propagates towards the ice shelf. This study extends the unique approach from 2000 to 2020 using satellite-based sea ice concentration data and glacier ice depth reanalysis product from ERA-5 to discover the spatial linkages between sea ice retreat events and ice shelf melting events, covering both the Arctic region (boreal summer) and the Antarctic region (austral summer). The algorithm searches for the linkages using Delaunay triangularization and only significant graphs that pass the monte carlo simulation tests are used. Our analysis based on more than 900 graphs generated during the austral summer (November – February, next year) between 2000-2020 shows a strong geospatial variation in the frequency of the spatial linkages. The sea ice retreat and ice shelf melting are strongly related over the Antarctic Peninsula, Weddell region, and Amundsen–Bellingshausen Seas over the western hemisphere and eastern Antarctica under the Indian Ocean in the east. However, the spatial linkages are absent over the Ross Sea region and over the West Antarctic under the Pacific Ocean. As Sea ice retreat and ice sheet melt intensifies from low (Median – Q3) to high (Q3 – upper bound) and anomalous (>upper bound), the number of linkages decreases but the spatial extent of the linkages increases. We also detect the similar spatial linkages over the Greenland region, indicating the robustness of our study. Our future analysis will delve into finding the reasons behind the occurrences of those spatial linkages.

• **Doctorate Student Award**
  2
  Authors: Mostafa Cham, Bayu Adhi Tama, Cheng Gong, Mathieu Morlighem, Jianwu Wang
  Title: LLM-SR-UQ: Uncertainty-Aware LLM-Guided Symbolic Regression

  Abstract:

  Symbolic regression is a promising route to scientific modeling because it can recover human-readable equations from data. However, most symbolic regression pipelines ultimately return a single “best” equation and provide limited or ad hoc uncertainty estimates—an issue that becomes critical in scientific settings where data are noisy, regimes shift, and extrapolation risk is high. Meanwhile, large language models (LLMs) can propose diverse, domain-plausible functional forms, but integrating these hypotheses into a statistically grounded learning-and-uncertainty framework remains underdeveloped. We introduce LLM-SR-UQ, an uncertainty-aware symbolic regression approach that transforms LLM-proposed equation candidates into calibrated predictive statements. Our method combines (i) LLM-guided hypothesis generation to expand the search over structurally meaningful expressions, (ii) Bayesian evidence-based weighting to quantify relative support for competing equations, and (iii) Bayesian model averaging to propagate model uncertainty into predictive mean/variance rather than committing to a single formula. To ensure reliability under model misspecification and unknown noise structure, we further apply conformal calibration, producing distribution-free prediction intervals with finite-sample coverage guarantees under exchangeability. LLM-SR-UQ outputs a ranked set of interpretable equations alongside point predictions and calibrated uncertainty intervals, enabling transparent scientific reporting and risk-aware deployment. We evaluate the approach on controlled physics-style synthetic systems and ongoing geoscience case studies, using accuracy, calibration/coverage, and interval efficiency to characterize the trade-off between interpretability and reliable uncertainty.

• 3
  Authors: Yash Diggikar (presenter), Milton Halem
  Title: A Machine Learning OSSE Framework for Satellite Observations A Multi-Model Study: FourCastNet, Aurora, and GraphCast

  Abstract:

  Observing System Simulation Experiments (OSSEs) are a powerful technique for assessing the impact of hypothetical new observing systems on weather forecast skill. We present one of the first applications of the OSSE methodology to multiple deep learning-based global weather inference models. The forecasting models used are NVIDIA’s Spherical Fourier Neural Operator (SFNO) – FCN V2, FCN V3, Google’s GraphCast, and Microsoft’s Aurora, all state-of-the-art AI models trained on ERA5 (0.25°) reanalysis data. For generating observations from nature, we employ ECMWF ERA5. We conducted a series of OSSE-style experiments with multiple AI models to investigate the impact of periodic assimilation of simulated observations with errors on forecast performance. A 60-day free-running autoregressive forecast served as the baseline. Several OSSE scenarios were explored in which we emulated the influence of prospective satellite observing systems by periodically replacing key atmospheric fields with ERA5 data with assumed Gaussian error distributions. These included simulation of infrared sounder observations through temperature field updates, Doppler wind lidar measurements through wind vector insertions, and geopotential height updates mimicking satellite altimetry or other vertical structure sensors. Data insertions were applied at regular 6 or 12-hour intervals. Additionally, we investigated a multi-variable insertion scenario representing a fully integrated observing network Results from one of the AI OSSE model experiments show that the data assimilation markedly improves the forecast skill relative to the baseline. Forecast error growth was significantly reduced in all insertion scenarios, as quantified by lower RMSE across variables (temperature, winds, and geopotential height) and extended lead times. We will test these findings with multiple AI models to validate the sensitivity of OSSE model performance and its effectiveness in increasing accuracy over time. We will compare our results with physics-based OSSE studies, such as those led by Robert Atlas, which evaluated the forecast impact of hypothetical wind lidar satellite missions and demonstrated substantial gains in forecast skill. In addition, we will test the impact of ensemble forecasts with multiple models and initial datasets. Our work aligns with the broader objectives of frameworks to reinforce the potential of AI-driven OSSEs to complement traditional numerical approaches. Here, machine learning models are tested and evaluated through the lens of classical OSSE methodology. This study highlights the novelty and potential of applying OSSE methodology to machine learning weather models. It represents an innovative application of OSSEs in line with recent community recommendations to broaden OSSE techniques into new modeling frameworks. By bridging data-driven forecasting with classical data assimilation concepts, our experiments point toward a hybrid modeling paradigm that leverages the strengths of both AI and physics-based approaches. In the broader context of OSSE efforts, these results illustrate how next-generation AI forecast systems might be integrated with future observing networks to enhance Earth system prediction skill.

• **Doctorate Student Award**
  4
  Authors: Muhammad Hasan Ferdous (presenter), Md Osman Gani (faculty mentor)
  Title: G-DCD: Generalized Decomposition-based Causal Discovery for Multivariate Multi-Seasonal Temporal Data

  Abstract:

  Multivariate time series data in domains such as finance, meteorology, and epidemiology often feature complex causal structures obscured by composite seasonality and synchronized periodic drivers. Conventional causal discovery algorithms typically rely on fixed lag assumptions and consequently fail to detect long-term seasonal dependencies or identify simultaneous seasonal causal edges arising from multiple overlapping cycles. This limitation frequently leads to the misinterpretation of synchronized seasonality as direct causation, resulting in the detection of spurious relationships. We propose G-DCD, a novel framework for causal discovery based on decomposition that systematically stratifies time series data into trend, multiple seasonal components, and residuals to address these challenges. By applying Multiple Seasonal and Trend decomposition using LOESS (MSTL) to isolate distinct periodic frequencies, our approach employs a period alignment mechanism to detect shared seasonal drivers across varying time scales. We subsequently perform conditional independence testing on these aligned components to identify multiple causal edges associated with different lagged periods and effectively block backdoor paths induced by latent environmental cycles. This framework successfully disentangles actual structural mechanisms from correlations across multiple frequencies, reducing spurious edge detection while preserving essential dependencies. Comprehensive evaluations on extensive synthetic benchmarks and real-world climate data demonstrate that our framework recovers true causal structures with higher accuracy than leading baselines.

• **Doctorate Student Award**
  5
  Authors: Md Badrul Hasan (Presenter – Ph.D. Candidate), Meilin Yu, Tim Oates
  Title: Machine Learning-Enhanced Turbulence Modeling for Hurricane Boundary Layer Simulations

  Abstract:

  Accurate prediction of hurricane intensity is highly sensitive to the representation of turbulent mixing in numerical weather models. At kilometer-scale grid resolutions, unresolved turbulent motions are typically modeled using subgrid-scale (SGS) closures. The commonly employed Smagorinsky model applies a constant coefficient to represent turbulent diffusion; however, this approach can result in excessive dissipation and suppression of organized vortex structures within the hurricane boundary layer. This study investigates machine learning-enhanced SGS closures that predict spatially varying Smagorinsky coefficients based on physically informed tensor invariants derived from strain-rate and rotation fields. The machine learning model is trained on high-resolution large-eddy simulation (LES) data and subsequently evaluated in idealized hurricane simulations using the Weather Research and Forecasting (WRF) model at 2 km horizontal resolution. To assess feasibility and numerical robustness, controlled six-hour simulations are conducted to compare three approaches: (1) a static Smagorinsky configuration with a fixed coefficient, (2) a dynamic Smagorinsky formulation that adaptively estimates local coefficients, and (3) quasi-static machine learning-predicted eddy-viscosity fields constructed offline and applied within WRF. All simulations remain numerically stable under the tested configurations. Results demonstrate that machine learning-based closures reduce outer-core dissipation compared to the static baseline, while preserving coherent vortex structures and realistic energy spectra. Differences relative to the dynamic Smagorinsky configuration reveal variations in spatial mixing patterns and intensity evolution, while maintaining comparable large-scale stability. This study establishes a controlled evaluation framework that connects offline machine learning model development with future fully coupled implementations, thereby advancing physics-informed machine learning approaches for atmospheric turbulence modeling.

• 8
  Authors: Uzma Hasan (presenter), Dr. Md Osman Gani
  Title: Uncertainty Quantification for Learned Causal Graphs through Heterogeneous Evidence Fusion

  Abstract:

  Causal discovery algorithms are AI algorithms which aim to uncover the cause–effect relationships among variables in a scientific domain, typically represented as a causal graph. However, discovering the causal graph from observational data alone involves substantial uncertainty. The outcomes (causal graphs) of the existing causal discovery algorithms are often unstable due to limited data, statistical noise, and unrealistic modeling assumptions. This limitation impacts confidence and interpretability, and restricts adoption in high-stakes domains such as healthcare. Domain experts, relevant literature, and prior empirical evidence can also guide the construction of causal graphs. While we may identify potential causal edges using any of these approaches, there is a lack of a systematic way to quantify confidence in those edges. Without uncertainty estimates, causal graphs risk being misleading, undermining their value for research and decision-making. To address this, our study focuses on estimating how confident different sources of evidence such as data, expert knowledge and literature are about the causal edges in a causal graph, thereby quantifying the uncertainty associated with such graphs. Particularly, we designed a unified framework that leverages the Dempster–Shafer Theory (DST) to integrate the judgement from multiple heterogeneous sources of evidence and thereby quantify the uncertainty in causal graphs. We evaluate this framework on a real-world case study related to the oxygen therapy treatment for ICU patients. By explicitly modeling uncertainty for causal graphs at the edge level, our method aims to improve transparency, robustness, and practical usability of data-driven causal discovery in critical domains such as clinical decision-support.

• **Doctorate Student Award**
  20
  Authors: Muhammad Behroze Hassan, Bayu Adhi Tama, Sanjay Purushotham, Vandana P. Janeja
  Title: XCheck: A Consistency-Based Validation Framework for Englacial Layers Annotations

  Abstract:

  Glaciers and ice sheets slowly flow to lower elevation where they melt. The history of how they have moved over time is captured in their inner (englacial) layers. By studying these layers with ice-penetrating radar across large areas, it validates the behavior of ice sheet models and improves our projections of melting and how it will affect sea levels and climate change in the future. Yet, both manual and automated annotations of englacial layers often suffer from inconsistencies, i.e., same layer across two radargram images don’t match up or don’t align with each other and in the absence of reference data, it becomes extremely difficult to validate them. We introduce XCheck (CrossCheck), a validation framework that leverages natural crossover points in radargram surveys to assess annotation consistency and provides a reference-free validation. Our approach computes dot-product based similarity scores between binary layer masks extracted from radargram columns at GPS-overlapping or geographically intersecting locations. Applied to the publicly available Center for Remote Sensing of Ice Sheets (CReSIS) Greenland radar dataset, XCheck demonstrates scalable and reproducible validation of englacial layer annotations. This contribution addresses three critical gaps. First, it provides a pathway toward constructing AI-ready datasets by enabling robust, large-scale quality control of radargram annotations. Second, it offers a benchmarking capability for algorithmic development, allowing different automatic annotation methods to be compared consistently across large collections of radargrams. Finally, it highlights broader relevance beyond glaciology: the principle of consistency-based validation applies to other remote sensing and Earth observation domains where reference is scarce. By advancing reproducible validation of radar data, XCheck mobilizes radargram as a reliable, scalable, and benchmarkable data source, enabling transdisciplinary research at the intersection of climate science, spatial AI, and data mining.

• 9
  Authors: Jumman Hossain and Nirmalya Roy
  Title: QPRL: Learning Optimal Policies with Quasi-Potential Functions for Asymmetric Traversal

  Abstract:

  Reinforcement learning (RL) in real-world tasks such as robotic navigation often encounters environments with asymmetric traversal costs, where actions like climbing uphill versus moving downhill incur distinctly different penalties, or transitions may become irreversible. While recent quasimetric RL methods relax symmetry assumptions, they typically do not explicitly account for path-dependent costs or provide rigorous safety guarantees. We introduce Quasi-Potential Reinforcement Learning (QPRL), a novel framework that explicitly decomposes asymmetric traversal costs into a path-independent potential function Phi and a path-dependent residual. This decomposition allows efficient learning and stable policy optimization via a Lyapunov-based safety mechanism. Theoretically, we prove that QPRL achieves convergence with improved sample complexity, surpassing prior quasimetric RL bounds. Empirically, our experiments demonstrate that QPRL attains state-of-the-art performance across various navigation and control tasks, significantly reducing irreversible constraint violations by approximately 4 times compared to baselines.

• 15
  Authors: Oluwatobiloba Odunsı, Aravind Mohan, Seraj Al Mahmud Mostafa, Jianwu Wang
  Title: CloudBot: Autonomous End-to-End Cloud Deployment from Code to Infrastructure

  Abstract:

  While cloud platforms offer extensive services for running and scaling applications, automatically deploying code from GitHub repositories to cloud infrastructure remains a manual, error-prone process requiring specialized expertise. Current DevOps tools and Infrastructure-as-Code (IaC) frameworks rely on fixed templates and cannot adapt to diverse application requirements automatically. We propose CloudBot, a toolkit that automates the complete deployment workflow by integrating code analysis, IaC generation, and infrastructure provisioning into a unified pipeline. CloudBot employs a pipeline where specialized components, GitHub Analyst, Cloud Architect, and Cloud Engineer, work sequentially to extract requirements, design infrastructure, and generate validated Terraform templates. The toolkit uses large language models enhanced with retrieval augmented generation to map application needs to infrastructure specifications.We evaluate CloudBot with AWS cloud across three use cases: text processing, image analysis, and video processing. CloudBot achieves consistent deployment success, with all configurations deploying and executing correctly. Comparing generated infrastructure against expert-written baselines reveals similarity scores of 18-43% across syntax, semantic, and functional dimensions. While generated configurations successfully deploy, they tend toward over-provisioning compared to minimal expert specifications. As the first end-to-end automated deployment system, CloudBot demonstrates LLM-based infrastructure automation feasibility while establishing quantitative baselines for measuring future improvements toward expert-level generation quality. Index Terms—Cloud Infrastructure Automation, Infrastructure

• 10
  Authors: Sourajit Saha (presenter), Tejas Gokhale
  Title: Zero-Shot Multimodal Retrieval with Multi-Scale Contextual Representations

  Abstract:

  In multimodal information retrieval (MMIR), candidates relevant to an input query need to be retrieved from a database, where the query and database items span different modalities. As real-world databases evolve, repeatedly annotating and indexing data and re-optimizing models across modalities is impractical. We present MULTI-SCORE, a fine-tuning-free, two-stage MMIR approach that couples efficient coarse ranking with fine-grained multimodal re-ranking. Stage-1, guided by Matryoshka representations, filters out non-promising candidates without computing similarities on full-scale representations for the entire database, yielding strong efficiency improvements. Stage-2 re-ranks the filtered candidates by computing their fine-grained multimodal contextual representations with two scoring functions for semantic alignment using Chain-of-Thought (CoT) prompting and question-answering. Our experiments demonstrate state-of-the-art zero-shot retrieval on 12 MMIR tasks across 32 datasets, while outperforming previous state-of-the-art supervised methods on 23 datasets.

• **Doctorate Student Award**
  16
  Authors: Md Sakib Ul Rahman Sourove (presenter), Lujie K. Chen and Shimei Pan
  Title: An LLM-Based Agentic AI System for Automated Construction of Knowledge Taxonomies in Data Science Problem Solving

  Abstract:

  The widespread integration of artificial intelligence into data science practice is fundamentally reconfiguring the roles of data science practitioners. As procedural and execution-oriented tasks such as coding, data preprocessing, and routine analysis become increasingly automated, human contribution is shifting toward complex cognitive task of problem solving including higher-order reasoning and decision making. This shift challenges prevailing models of data science education, which has been focusing on tools and techniques while offering limited support to develop those critical problem solving competencies. A prerequisite for addressing this gap is formalizing the Data Science Problem Solving (DSPS) competency as structured knowledge that can be systematically represented and evaluated, for example, by adopting principled approach of constructing knowledge taxonomies for DSPS. However, constructing a coherent and verifiable taxonomy of such knowledge remains a nontrivial knowledge engineering task. In this paper, we present a multi-agent LLM framework for the automated generation and evaluation of a DSPS knowledge taxonomy, grounded in established taxonomy construction and evaluation methods. The system comprises an \textit{Author} agent that proposes and revises taxonomies and a \textit{Critique} agent that evaluates them and provide feedback based on explicit assessment criteria. Additionally, an Orchestrator agent manages iterative refinement by routing feedback and enforcing stopping conditions. Our results demonstrate the potential of organizing LLMs into collaborative-adversarial architectures that support reflective, iterative knowledge construction and refinement rather than single-pass content generation. When further developed, the proposed multi-agent LLM systems can function as a scalable framework for principled knowledge engineering, enabling the systematic design of DSPS curriculum, assessments, instructional scaffolds, and AI-assisted learning environments at scale.

• **Doctorate Student Award**
  11
  Authors: Tartela Tabassum (Presenter), Roy Prouty, Elliot Gobbert, Jianwu Wang
  Title: LLM-driven user support for high-performance computing resources

  Abstract:

  Large Language Models (LLMs) are transforming how organizations deliver technical support to the research community; however, their use in high-performance computing (HPC) environments remains limited and is not widely explored for individual user support. The primary challenges are domain specificity, the risk of inaccurate answers, and the need to maintain up-to-date knowledge. This project aims to design an intelligent, LLM-driven support system for UMBC’s HPC users that reduces repetitive workloads, supports HPCF’s user support staff, and improves response time. The proposed system will act as a conversational assistant that understands natural-language questions, provides step-by-step troubleshooting assistance to UMBC chip Users, and continuously learns from real support cases. To ensure accurate, trustworthy responses, it will use a Retrieval-Augmented Generation (RAG) layer that connects the LLM to verified institutional knowledge, including the UMBC HPCF wiki page and archived RT tickets with prior staff solutions. A built-in feedback loop will allow the model to grow smarter as new issues are resolved. Beyond simple Q & A, the system will include multi-agent coordination for query routing, content validation, and safety checks. Its performance will be compared directly with student staff using practical measures such as response time, ticket deflection, and answer quality. Together, these components will demonstrate how domain-focused LLMs can effectively reduce repetitive support tasks and enhance overall efficiency in HPCF operations.

• 12
  Authors: Zahid Hassan Tushar (presenter); Sanjay Purushotham
  Title: HyperFM: An Efficient Hyperspectral Foundation Model with Spectral Grouping

  Abstract:

  The NASA PACE mission provides unprecedented hyperspectral observations of ocean color, aerosols, and clouds, offering new insights into how these components interact and influence Earth’s climate and air quality. Its Ocean Color Instrument measures light across hundreds of finely spaced wavelength bands, enabling detailed characterization of features such as phytoplankton composition, aerosol properties, and cloud microphysics. However, hyperspectral data of this scale is large, complex, and difficult to label, requiring specialized processing and analysis techniques. Existing foundation models, which have transformed computer vision and natural language processing, are generally trained on standard RGB imagery and therefore struggle to interpret the continuous spectral signatures captured by PACE. While recent advances have introduced hyperspectral foundation models, they are typically trained on cloud-free observations and often remain limited to single-sensor datasets due to spectral inconsistencies across instruments. Moreover, existing models tend to be parameter-heavy and computationally expensive, limiting scalability and adoption in operational settings. To address these challenges, we introduce HyperFM, a parameter-efficient hyperspectral foundation model that leverages intra-group and inter-group spectral attention along with hybrid parameter decomposition to better capture spectral spatial relationships while reducing computational cost. HyperFM demonstrates consistent performance improvements over existing hyperspectral foundation models and task-specific state-of-the-art methods across four benchmark downstream atmospheric cloud property retrieval and soil parameter estimation tasks. To support further research, we additionally release HyperFM250K, a large-scale hyperspectral dataset from the PACE mission that includes both clear and cloudy scenes.

• 42
  Authors: Dr. David Garcia, Camille Basden (proposed presenter)
  Title: Paper-Based Cell-Free Systems

  Abstract:

  Biosensors have evolved significantly to improve the overall quality of life. Among emerging platforms, cell-free systems are one of interest and offers promising alternatives to other forms of biosensors – the large majority being electrochemical based. Cell-free systems utilize lysed cells to remove the membrane of a cell but preserve the basic functions of a cell i.e. transcription and translation. Removing the membrane avoids transport constraints that may arise using an intact cell and allow direct access to the components needed. Cell-free biosensors are also flexible in that they different DNA can be added to create a new sensor, which can be used to adapt for different targets. The lack of an intact cell also makes it more compatible for high-throughput screening (HTS) enabling optimization and scalability. When making biosensors intended for general use, their ideal characteristics should be simple to navigate, affordable, portable and accessible – similar to over-the-counter biosensors like pregnancy test or covid test. The current work focuses on integrating cell-free biosensing reactions into paper tickets using acoustic liquid transfer technology to create a cost-effective and scalable biosensor. Each paper ticket contains 384 wells, thereby increasing the testing capacity. This approach aims to reduce production cost and provide a biosensor system capable of point-of-care applications.

• **Doctorate Student Award**
  30
  Authors: Rebecca Boese (Presenter) and Dr. Corine Jackman Burden
  Title: Investigating mechanisms mediating pathogen adhesion in infective endocarditis using aortic bovine valves.

  Abstract:

  Infective endocarditis (IE) is an infection of the heart lining and valves that is typically caused by pathogenic bacteria. Several pathogens have been implicated in IE including Staphylococcus aureus and Streptococcus sanguinis. S. aureus is associated with approximately 40% of IE cases, and originates from hospital exposure and intravenous drug use. It damages aortic tissue via the secretion of the pore-forming cytotoxin hemolysin. Streptococcus sanguinis accounts for 10–15% of IE cases and originates from oral periodontitis, migrates through the bloodstream to the heart, where it induces chronic inflammation after adhering to tissue via biofilm formation. To date, few systematic studies have examined biofilm growth rates and gene expression profiles of S. aureus and S. sanguinis on endocardial tissues. To address this gap, we performed adhesion assays where bacteria attach to bovine pericardium tissue and subsequently Medtronic bioprosthetic aortic heart valves. To quantify the level of adherence, we washed the tissue thrice to detect adherent GFP-expressing S. aureus (via plasmid transformation) and mCardinal-expressing S. sanguinis (generated through homologous recombination) on bovine pericardium using confocal microscopy. We use FIJI image analysis software to enumerate fluorescent cells. Preliminary data using sonication demonstrated partial recovery of S. aureus from 1 cm² bovine pericardium tissue compared to controls. Current work employs our direct detection method to visualize and quantify fluorescent colony-forming units (CFUs) on pericardial tissue. This research will provide a better understanding of how biofilm growth rates on endocardial tissue relate to pathogen adhesion.

• **Master Student Award**
  31
  Authors: Sudarshan Bollapragada (Presenter), Corine Jackman Burden
  Title: Microbiome Regulation of Cervical Cancer Progression

  Abstract:

  Recent evidence indicates that cervical cancer progression is significantly influenced by complex interactions within the vaginal microenvironment, particularly between epithelial cells and the human cervicovaginal microbiome (HCM). Within the HCM, commensal bacteria such as Lactobacillus crispatus dominate in healthy vaginal communities and are associated with reduced inflammation and overall mucosal protection. However, the direct immunomodulatory effects of L. crispatus on cervical cancer cells remain unclear. This in vitro co-culture study investigates the impact of L. crispatus on human cervical carcinoma cells by evaluating bacterial growth, host-microbe interactions, and immune responses, with an initial focus on the anti-inflammatory cytokine interleukin-10 (IL-10). Preliminary results obtained using the LUMINEX MAGPIX (3D) system indicate that L. crispatus alters the tumor-associated inflammatory profile, with early analyses suggesting changes in IL-10 expression in co-culture compared to monoculture conditions. These findings indicate a potential immunoregulatory role for L. crispatus within the cervical tumor microenvironment. Future studies will examine MYC protein expression, encoded by the MYC oncogene, through Western blot analysis and transition the model into a microfluidic cervix-on-a-chip platform. Additional investigation is needed to further evaluate the protective effects of vaginal lactobacilli on oncogenesis and cancer-associated inflammation. This work lays the groundwork for future studies using cervix-on-a-chip systems to explore microbiome-driven therapeutic strategies for cancer prevention and treatment.

• **Undergraduate Student Award**
  32
  Authors: Damilola Fapohunda (Presenter), David Garcia
  Title: Measuring pH and Metabolite Changes in Cell Free Systems Using Fluorescent Biosensors

  Abstract:

  Cell free metabolic engineering is a method by which the extracted cytoplasmic contents of the cell are used for biomolecule production. One of the problems we face with this method is a way to track the changes in the cell-free environment over time. Changes in pH, buildup of intermediate and side products, and other unknown factors can affect the rate at which the biomolecules of interest are produced. Traditional methods of tracking these changes, such as high-performance liquid chromatography (HPLC), are effective, but not efficient when tracking these changes in many systems simultaneously. Instead, we propose the use of fluorescent proteins to track the metabolic outputs and pH changes in these cell-free systems over time. ApplyPy, GreenPy, and Lime, are fluorescent proteins that increase fluorescent in response to raised pH and pyruvate concentration. Using these fluorescent proteins, along with our high throughput methodologies, we can measure the changes in hundreds of cell free reactions in real time with fluorescent assays. The data from these experiments will provide a framework for the use of cell free metabolic engineering in the scalable production of biomolecules.

• 43
  Authors: Neveen Faris ( presenter), Rishika Bandi, Lexi Malenfant, Maya Sanyal, Riya Koshy, Vikash Kumar, Mike Tolosa, Venkatesh Srinivasan (mentor), Govind Rao
  Title: Breathable Shake Flask Platforms for High-Yield Recombinant Protein Manufacturing

  Abstract:

  Shake flasks remain the dominant platform for early-stage recombinant protein expression and process development, yet conventional designs impose mass-transfer limitations that restrict culture performance and downstream recoverable yield. Insufficient oxygen transfer and carbon dioxide accumulation in standard polycarbonate flasks constrain working volumes and reduce productivity in aerobic Escherichia coli systems. This study evaluates a molded breathable shake flask platform for the production and purification of glucose-binding protein (GBP). The flasks incorporate gas-permeable surfaces that enhance oxygen availability and off-gas removal without increasing agitation, footprint, or process complexity. Performance was benchmarked against standard flasks across multiple fill volumes, with cell growth quantified by optical density measurements and culture oxygenation monitored using non-invasive optical sensing. Breathable flasks consistently maintained higher dissolved oxygen levels, supporting increased biomass accumulation across all tested conditions. Importantly, these upstream improvements translated directly into higher recoverable GBP following downstream purification, resulting in increased purified protein yield and improved batch-to-batch consistency. The enhanced gas-exchange capability enabled higher working volumes within the same 500 mL format, increasing protein output per run without modifying expression or purification workflows. Collectively, these results demonstrate that breathable shake flasks offer a scalable, low-disruption approach to improving upstream culture performance and downstream protein recovery, providing clear value for biopharma process development and early-stage manufacturing workflows.

• 44
  Authors: Lucas Gois (Presenter), Hanlu Yang, Trung Vu, Erdem Kumbasar, Denis Fantinato, Aline Neves, Vince D. Calhoun, and Tülay Adali
  Title: Domain-Informed Independent Vector Analysis for multisubject fMRI analysis

  Abstract:

  Independent Vector Analysis (IVA) is a widely used technique for multi-subject fMRI analysis. To guide the decomposition and increase the interpretability of the extracted components, constrained IVA methods incorporate prior information, also known as references, into the model to guide the solution towards those. In this context, threshold-free constrained IVA (tf-cIVA) avoids the difficult issue of threshold selection by leveraging the structure of IVA and using a regularization term to encourage correlation with the corresponding component and penalize all cross-component correlations. Although it is shown to provide desirable performance, the strategy in tf-cIVA can suppress correlations between brain networks within the same functional domain, contrary to the expected neurobiological behavior. In this work, we propose a domain-informed tf-cIVA (ditf-cIVA) that defines a regularization term to selectively preserve the correlation within brain domains. We compare our method with tf-cIVA on a resting-state fMRI dataset of 58 healthy controls and 58 schizophrenia patients. Our findings demonstrate that ditf-cIVA produces a more modular spatial functional network connectivity structure, yields more consistent component estimations as measured by higher one-sample t-values, and shows increased sensitivity in detecting significant group differences between the two cohorts.

• 33
  Authors: Maya Haywood (Presenter), David Garcia
  Title: Leveraging Protein Language Models and Cell-Free Protein Synthesis to Identify Effective Divalent Metal Cofactor–Polyphenol Oxidase Pairings for Optimal Catalytic Activity

  Abstract:

  Biological production systems are an emerging area in industrial biomanufacturing where biological pathways aid reaction catalysis to generate chemical products. The success of these systems relies on the efficiency of biocatalysis, where correct enzyme-cofactor pairings enable greater catalytic activity resulting in cost-effective, structurally stable products, and selective reactions increasing overall yield. However, many well-characterized enzymes are studied with a narrow range of metal cofactors, potentially overlooking more effective cofactors that could enhance catalytic performance. This project employs Protein Language Models (PLMs) to predict and classify effective divalent cofactors including Cu2+, Mn2+, Mg2+, Zn2+, and Fe2+ for various polyphenol oxidases. This iterative process uses the results of experimental data to train the model and predict the next set of effective cofactors. Cell-free protein synthesis (CFPS) generates the enzymes needed to test and rapidly validate these computational predictions, feeding the results back into the PLM fine-tuning its prediction. This expedited approach intends to identify effective cofactors giving insight into structural or mechanistic flexibility for each cofactor-enzyme pairing. Initial testing demonstrates that non-canonical divalent ions can act as superior cofactors; specifically, Fe2+ paired with tyrosinase catalyzed reactions with greater efficiency than literature-established Cu2+ under specific concentration conditions. These findings suggest previously unrecognized flexibility in tyrosinase-cofactor pairings, challenging current understandings of its cofactor-binding requirements. Combining machine learning with rapid testing offers new insights into these enzyme-cofactor interactions providing a new pathway maximizing yield for biological production systems.

• **Doctorate Student Award**
  34
  Authors: Emily H. Kruszon (presenter), Sayantan Bhattacharya, Charles D. Eggleton
  Title: Measurement of shear stresses on the coupon surface as a function of RPM in a CDC Biofilm reactor

  Abstract:

  Biofilms are found as a collection of multispecies organisms acting together in order to survive. The life cycle of biofilms are continuous, allowing the community of microorganisms to mature and spread using its own cell fragments. In 2008, 80% of bacterial infections were due to biofilms being spread through the food processing industry. Recent cases involve an outbreak of Listeria Monocytogenes in pre-prepared pasta meals in late 2025 and ongoing 2026. Past research suggests that shear stress can accelerate the attachment and development of biofilms. Researchers have evaluated this focus by using Computational Fluid Dynamics (CFD) simulations, establishing correlations between the magnitude of shear stress on biofilms through growth and detachment processes. However, no studies have experimentally quantified shear stress levels in a bioreactor under operational conditions, aside from CFD analysis. We used a CDC Bioreactor in the experimentation to provide a controlled environment for studying hydrodynamic forces. To determine wall shear stress, we used Particle Image Velocimetry (PIV) to quantify velocity field information for flow within the reactor. PIV is a non-invasive optical method that seeds the flow with particles, illuminates the particles using a laser, records particle images using a high-speed camera, and uses image correlation-based processing to obtain time-resolved velocity measurements on a specific plane. Current images depict increasing vortex structures at larger RPMs after testing ranges 100-500 RPM in the reactor. The momentum transfer intensified the velocity gradient near the coupon surface. This steeper velocity gradient in the boundary layer directly results in a higher shear stress, increasing the friction drag in the area of interest. We aim to compare our findings with past research referenced in similar experimentation. From our findings, shear stress does in fact have a direct influence on a biofilm’s composition, which sets a baseline for more in-depth experimentation on this subject.

• **Doctorate Student Award**
  35
  Authors: Shayan Manuchehrfar, Molly Y. Mollica
  Title: Intrinsic Cellular Factors Underlie Heterogeneity in Single-Platelet Force Generation

  Abstract:

  Cells generate mechanical forces through cytoskeletal contraction and transmit these forces to their environment via adhesion receptors. The magnitude of cell forces is a key indicator of cellular function and health. Despite identical external conditions, individual cells exhibit substantial heterogeneity in force generation, and the intrinsic drivers of this variability remain unclear. Here, we used black dot (BD) traction force microscopy to quantify single-platelet forces and examine how intrinsic platelet properties—including morphology, cytoskeletal organization, receptor activation, and metabolic indicators—associate with contractile force. A fluorescent surface with a pattern of non-fluorescent circles were microcontact printed onto compliant polydimethylsiloxane (PDMS) substrates and functionalized with fibrinogen. Washed human platelets from unique healthy donors were seeded, allowed to contract, fixed, and stained for cytoskeletal components, adhesion receptors, mitochondria, and RNA content. Fluorescence microscopy combined with custom MATLAB-based analysis enabled simultaneous single-cell measurements of force, marker quantity, and marker organization. Across 2,452 platelets from nine unique human research subjects, forces ranged from 4.7 to 206.7 nN (mean 20.4 ± 16.6 nN). Platelet spread area positively correlated with force and circularity showed a weaker positive association, as expected based on prior work. F-actin intensity exhibited a strong positive association with force, whereas its dispersion (i.e., uniformity across the cell) negatively correlated. Intensity of cellular contents including mitochondria, RNA, myosin, tubulin, and vinculin are also positively correlated with force. Finally, integrin αIIbβ3 expression, glycoprotein Ib (GPIb) expression, and GPIb dispersion positively correlated with force, while αIIbβ3 dispersion negatively correlated with force. These results indicate that platelet force heterogeneity arises from interconnected differences in cellular contents, receptor organization, and cytoskeletal architecture. These results will develop fundamental understanding the intrinsic biological sources of cellular force heterogeneity and have implications in hemostasis, thrombosis, and wound healing.

• **Undergraduate Student Award**
  39
  Authors: Katelyn Prasad (presenter), Nadeem Shah, Sayantan Bhattacharya, Corine Jackman Burden
  Title: Design and Validation of a Submersible Mechanical Testing System for Aortic Valve Leaflet Stiffness Characterization

  Abstract:

  Infective Endocarditis (IE) is a bacterial infection that alters the mechanical integrity of heart valve tissue, leading to increased leaflet stiffness, reduced valve opening and impaired blood flow. Quantifying the leaflet structural changes requires a controlled testing platform capable of evaluating tissue strain for known applied stretching force. However, precise measurement of applied force, preserving tissue hydration while testing, and optimizing imaging parameters for optical strain quantification poses a challenge. This work presents the design, fabrication, and validation of a submersible mechanical testing system developed to evaluate stiffness changes in porcine aortic valve leaflets. Conventional tensile testing systems are not designed for soft, hydrated biological tissues and lack appropriate mounting strategies for delicate samples. To overcome these limitations, custom three-dimensional (3D) components were designed and manufactured to create a sealed testing chamber that maintains specimens submerged in phosphate-buffered saline throughout experimentation. The system integrates 3D-printed fixtures with a precision linear stage and force gauge to enable controlled tensile loading and high-resolution force acquisition. Full-field strain was quantified non-invasively using Digital Image Correlation (DIC) with a high-speed Phantom VEO-340L imaging system. Displacement fields were processed to compute two-dimensional strain distributions across the leaflet surface during deformation. Force–displacement data were analyzed using least-squares regression to determine effective material stiffness. The platform was calibrated using reference materials with known mechanical properties to verify measurement accuracy and repeatability. Additionally, Staphylococcus Aureus infected pericardium tissue samples were also tested to quantify and compare changes in tissue stiffness pre and post infection. This validated experimental system provides a quantitative framework for assessing infection-induced changes in valve leaflet mechanics.

• 36
  Authors: Matthew S. Quintanilla, Ololade D. Lawrence, Ethan F. Folmer, Jayne M. Zeller, Alexander G. Doan, Kelsey Gray, Athira Anilkumar, Nachammai Napiappan, Deepa Madan, Joseph Washington, Mark R. Marten
  Title: Determining the Impact of Dual Approaches on the Mechanical Strength of Mycelial Materials

  Abstract:

  Rising global demand has made nonrenewable resources increasingly scarce. Mycelial materials (MMs) made from filamentous fungi have emerged as a promising renewable alternative. They have seen use in many commercially successful products, but their widespread adoption remains limited by their relatively low mechanical strength. Our research aims to uncover how various features of the mycelium influence mechanical strength to contribute to the development of MMs with “tunable” mechanical properties. We primarily focus on one feature, cell wall strength, in the model fungus Aspergillus nidulans where weakened cell walls result in greater mycelial fragmentation. Past experiments have suggested that, alongside cell wall strength, mycelial density may significantly influence mechanical strength. To determine how each factor may contribute, MMs will be produced from an mpkA deletion mutant (∆mpkA) and an isogenic (i.e., mpkA+) control strain. We hypothesize that for materials made from non-blended fungal mycelia, ∆mpkA strains will lead to materials with lower mechanical strength. To influence mycelial density, biomass will be blended to shorten filaments and facilitate tighter packing. We further hypothesize that blended samples will display stronger mechanical strength. We will also describe methods we are using to produce artificial pores. To determine mycelial density within each material, mercury intrusion porosimetry is tested (through a new collaboration). Together, these approaches will help to provide insight regarding the combined effect of altered cell wall strength and mycelial density, and guide the development of tunable MMs.

• **Doctorate Student Award**
  40
  Authors: Sina Razaghi (presenter), Mehdi Kiani
  Title: MagSonic: A Hybrid Magnetic–Ultrasonic Wireless Interface for Next-Generation Miniaturized Biomedical Implants

  Abstract:

  The elimination of batteries in implantable medical devices enables their miniaturization to millimeter (mm) scale, which is critical for chronic operation. Wireless power transfer (WPT) techniques such as inductive coupling, ultrasound (US), and magnetoelectric (ME) modalities have emerged as leading approaches for powering deeply implanted devices while operating within regulatory safety limits. US and ME links operate at acoustic wave resonance at sub-MHz to MHz frequencies using mm-scale receivers. Magnetoelectric (ME) transducers, composed of layered magnetostrictive and piezoelectric materials, efficiently convert low-frequency magnetic fields into electric voltage and are particularly attractive for miniaturized biomedical implants. To enable simultaneous wireless power and data transfer for closed-loop implant operation, we introduce a hybrid magnetic–ultrasonic interrogation approach, termed MagSonic, realized through a single mm-sized ME transducer capable of generating and receiving both magnetic field and ultrasound. For the first time, wireless power reception through one modality and simultaneous uplink data transmission using the other is demonstrated. At 40 mm depth, the MagSonic link achieves up to 8 mW delivered power and greater than 100 kbps uplink data rate with high robustness to carrier interference and misalignment.Finally, a fully wireless 180 nm CMOS ASIC operating with the MagSonic modality is demonstrated for neural stimulation and recording, establishing a scalable platform for battery-free millimeter-scale biomedical implants.

• 41
  Authors: Nazanin Saberi, Molly Y. Mollica
  Title: Distinct platelet traction forces and morphology on fibrin scaffolds compared to planar fibrinogen

  Abstract:

  Platelets generate contractile forces to compact fibrin scaffolds, yet most platelet force measurements have been performed on planar fibrinogen-coated substrates rather than within fibrin matrices. Thrombin plays a central role in this process by converting fibrinogen monomers to fibrin fibers and by directly activating platelets; however, its specific contributions to platelet force generation within fibrin remain unclear. Here, we formed fibrin fibers directly on black dot traction force microscopy substrates, enabling quantification of single-platelet forces within fibrin networks rather than on planar fibrinogen surfaces. Platelet forces were determined by tracking patterned dot displacements beneath individual platelets. Fibrin networks were generated across a range of fibrinogen (0.625–2.5 mg/mL) and thrombin (0.1–0.75 U/mL) concentrations to systematically examine the effects of thrombin on fibrin polymerization and platelet force transmission. We found that platelet forces measured on planar fibrinogen surfaces significantly increased (p < 0.01) in the presence of thrombin (0.1 U/mL) in washed platelets. In contrast, within fibrin networks, increasing thrombin concentration produced thicker and stiffer fibrin layers, which restricted platelet access to the underlying substrate and led to reduced platelet-induced deformations. Also, we observed significant platelet morphological changes, with a transition from the normal and circular shape to a stellate morphology. These results establish black dot traction force microscopy as a robust approach for quantifying platelet-generated forces within fibrin matrices and demonstrate the critical role of fibrin layer thickness and stiffness in regulating platelet–substrate interactions.

• 23
  Authors: Md Biplob Hosen {presenter}, Houbing Herbert Song , Shuling Yang and Lujie Karen Chen
  Title: NeuroSymRead: Symbolic Governance of Neural Generation for Adaptive Dialogic Reading

  Abstract:

  The deployment of Large Language Models (LLMs) in early childhood education promises scalable personalization but is hindered by stochastic behaviors, specifically the tendency to lose track of learner proficiency and disrupt long-term story consistency. While LLMs excel at fluent text generation, they often lack the executive logic required for consistent, adaptive reading instruction. To address these reliability and interpretability challenges, we introduce NeuroSymRead, a neuro-symbolic architecture that strictly decouples pedagogical governance from content generation. The system employs a deterministic rule-based controller to provide architectural transparency, enforcing hard constraints on difficulty adaptation and narrative progression, while a separate neural engine handles question generation, scaffolding, and semantic evaluation within these symbolic bounds. We validated the system through an LLM-based synthetic user simulation ($N=60$ sessions, 600 turns total) using a Gemini-2.5-pro LLM-as-a-Judge protocol. Results demonstrate that NeuroSymRead effectively eliminates proficiency tracking errors by ensuring all instructional shifts align with predefined pedagogical logic, maintaining high generative quality with an average age appropriateness of 4.66/5 and scaffolding quality of 4.75/5. By generating narrative session audits grounded in the controller’s logic, NeuroSymRead enhances interpretability for educators, providing a verifiable “glass box” alternative to end-to-end neural tutors for early reading assistance.

• **Doctorate Student Award**
  24
  Authors: Vasundhara Joshi, Vasundhara Joshi (presenter), Surely Akiri, Sanaz Taherzadeh, Gary Williams, Andrea Kleinsmith
  Title: Investigating differences in Paramedic trainees’ multimodal interaction during low and high physiological synchrony

  Abstract:

  Physiological synchrony—the unconscious, dynamic alignment of physiological responses such as heart rate and electrodermal activity (EDA)—is increasingly recognized as a crucial element of effective teamwork and interpersonal dynamics. While synchrony has been studied extensively in romantic partners, friends, and therapeutic contexts, there is limited research on how it operates within high-stress, hands-on environments such as paramedic trainee simulations. In this study, we examine how differences in synchrony relate to multimodal interaction—specifically, verbal and nonverbal— between paramedic trainee dyads during simulation training. Quantitative analysis revealed statistically significant differences in Technical Coordination across synchrony levels during the Consult phase, with higher synchrony associated with more effective coordination. Qualitative analysis further highlighted distinct interactional patterns: high-synchrony teams demonstrated mutual gaze, closer physical proximity, aligned body orientation, and cooperative dialogue, whereas low-synchrony teams often displayed disengagement, spatial misalignment, and minimal interaction. These findings underscore the role of physiological synchrony in shaping the quality and effectiveness of multimodal team interaction, offering practical insights for improving collaboration in emergency medical training environments.

• 112
  Authors: Ravi Kuber, Patti Ordonez, Marjory Pineda, Foad Hamidi, Marilyn Iriarte Santacruz.
  Title: Towards Identifying Best Practices for Accessible Makerspace Design

  Abstract:

  Makerspaces offer considerable promise to individuals with disabilities, as these offer an opportunity to design assistive technologies customized to their needs and abilities. However, there is a paucity of guidance related to making makerspaces accessible, which can create challenges. We discuss the outcomes from a workshop conducted with researchers, practitioners and individuals with disabilities to better understand some of the practices used to support accessible makerspace design. Findings have been applied to the AC3 makerspace in Puerto Rico.

• 113
  Authors: Marilyn P. Iriarte
  Title: Engaging Underserved Communities in Computing

  Abstract:

  This research advances a community-centered approach to co-designing culturally responsive support systems for Latino families navigating identity development, intergenerational communication, and postsecondary pathways. Grounded in participatory design and research-through-design, my dissertation examines how youth and parents collaboratively articulate needs, values, and aspirations, and how these insights inform the creation of relational, identity-affirming tools. Across three iterative studies, I conducted semi-structured interviews with Latino youth and parents, facilitated family co-design workshops, and engaged participants in prototype testing and reflective sessions. Data were analyzed using thematic analysis and design synthesis methods to translate lived experiences into functional design principles. This prior work resulted in three prototype concepts: (1) structured parent–teen conversation prompts to scaffold emotionally grounded dialogue; (2) an identity-centered reflection toolkit to help youth articulate purpose, cultural assets, and future aspirations; and (3) a bilingual navigation support system to demystify college and career pathways for families unfamiliar with U.S. educational systems. While initially developed as workshop-based and hybrid toolkits, these concepts revealed a shared need for adaptive, culturally situated, and language-accessible guidance. Future work extends these foundations through the integration of Large Language Model (LLM) systems as conversational infrastructure. Rather than replacing human relationships, the LLM functions as a scaffold, supporting reflective dialogue, generating culturally responsive prompts, translating institutional knowledge into accessible language, and adapting content to family context. Through participatory prototyping, families will co-design interaction flows, tone, and boundaries to ensure alignment with community values. Evaluation will combine usability testing, longitudinal family sessions, and qualitative assessment of identity clarity, communication patterns, and perceived agency. By grounding LLM-mediated systems in participatory design with Latino families, this research advances a model for AI-enabled support tools that center cultural identity, relational trust, and collective empowerment.

• **Master Student Award**
  25
  Authors: Kevin Lemus, Christian Ruiz, Stephanie J. Lunn, and Edward Dillon
  Title: Cultura Connections: Harnessing Large Language Models to Establish Examples of Data Structures Relevant for Hispanic and Latine Students

  Abstract:

  Learning data structures is a core part of computer science education. However, students often struggle with grasping these concepts as they learn to program. The main objective of this work seeks to establish ways for students to better understand how data structures are used and why they matter. The targeted approach is through using familiar experiences as a catalyst to bridge their understanding with these structures on an application level. This approach is important when exploring students from diverse backgrounds, in particular those from different cultural and linguistic backgrounds. In this study, we examined how six generative AI tools (ChatGPT, Gemini, Claude, MetaAI, Mistral, Deepseek), specifically large language models (LLMs) might support culturally relevant teaching by generating explanations of data structures that connect to the experiences of Hispanic/Latiné students in the United States. We created six prompts, each focused on a specific data structure, and used them across six different LLMs. This process produced 216 responses, which we analyzed through discourse analysis to identify cultural references. Our findings show that LLMs frequently drew on themes related to cuisine, family, and cultural events when explaining abstract concepts, suggesting the models adapt explanations toward culturally resonant examples. For instance, cuisine analogies appeared most often using specific foods from each respective country, while folklore and storytelling/oral traditions provided no examples at all. Most responses were written in English, although some incorporated Spanish words or phrases, particularly in scenarios where role-play (LLM acting as teacher or student) was part of the prompt.

• 115
  Authors: Maria M. Lopez-Delgado [presenter], Dr. Ravi Kuber
  Title: Co-designing with Early Childhood Educators for Inclusive Computational Thinking and Creative Technologies Lessons in Puerto Rico

  Abstract:

  This participatory design study looks to tackle the integration of Computational Thinking (CT) and Creative Technologies in Puerto Rican early childhood classrooms. The rationale for this study stems from preliminary findings where educators expressed strong interest in computational thinking and creative technologies but cited concerns regarding cost, safety, excessive screen time, and lack of access to resources. To bridge this gap, this project establishes a partnership between preschool educators, special education teachers, instructional technologists, and makers to engage in a relationship of mutual learning. The research is guided by the following research questions: (1) How can the co-design process support preschool educators in developing Computational Thinking and Creative Technology lessons? and (2) What specific considerations do educators prioritize when designing for mixed-ability early childhood environments? Data collection includes recorded co-design sessions, reflective journals, and group discussions focused on translating CT concepts, such as algorithmic thinking and pattern recognition, into culturally responsive and inclusive practices.

• **Doctorate Student Award**
  26
  Authors: Omobolanle Niyi-Owoeye, Uzma Hasan, Kevin Lemu, Srushti Dharmale, William Parham, and Edward Dillon
  Title: Fostering Comprehension in Introductory Programming through Code Reviews and Verbal Explanations in the Age of Generative AI

  Abstract:

  The rapid emergence and increasing accessibility of generative artificial intelligence (GenAI) tools introduce new complexities for educators and learners alike. While these tools allow students to use large language model (LLM)-based queries to generate functional programming solutions with minimal effort, they also raise concerns about overreliance, which may hinder students’ ability to independently construct and comprehend computational solutions. As GenAI tools become more pervasive in educational settings, it is essential to ensure that students continue to develop and retain proficient computational understanding rather than circumventing it through automated assistance. This specific research examines an instructional intervention implemented in an introductory programming course at a minority-serving institution (MSI) in the Mid-Atlantic United States during the Fall 2024 semester. The intervention integrated the use of the GenAI tool, ChatGPT, within selected take-home assignments and required students to participate in structured code reviews and verbal think-aloud activities. The objective was to assess students’ ability to apply their programming skills and concepts while having direct access to a GenAI tool. This approach also encouraged students to articulate their understanding/comprehension of programming concepts and explain the reasoning behind their code. Additionally, the study aimed to provide initial insight on whether students could continue to develop and showcase their programming skills without becoming overreliant on GenAI tools. Findings indicated that most students demonstrated proficiency in interpreting code and articulating their solutions, both in writing and verbally, even while using ChatGPT. Students also reported that ChatGPT aided their understanding of programming concepts and supported their development of essential skills in code review and evaluation. This study examines how such interventions can support students’ development of program comprehension skills in a GenAI-enabled learning environment, providing insights for educators aiming to adapt instruction while maintaining rigor and depth in foundational computing education.

• 27
  Authors: William Parham III, Olivier Woodgett, Martita Perez, Patrina Pun, Kevin Lemus, and Edward Dillon
  Title: Using Generative AI to Examine Cultural References for Understanding Computational Data Structures

  Abstract:

  Understanding and applying computational data structures are a critical foundation of computing education and programming pedagogy. However, students unfamiliar with them may struggle to conceptualize these computational building blocks more abstractly. In recognition that the same data structure descriptions may not resonate with all populations of students, we sought to explore how large language models (LLMs) can aid in contextual content creation of data structure explanations. This research examines the capacity using LLMs to generate culturally relevant explanations of data structures for novice programmers. Culturally responsive teaching has been shown to enhance student engagement and deepen conceptual understanding, particularly for learners from historically marginalized backgrounds. Using six generative AI tools (ChatGPT, Gemini, Claude, MetaAI, Mistral, Grok) a series of LLMs were constructed to generate explanations for six data structures framed within Black/African-American, White/Caucasian, and Asian cultural contexts. A qualitative thematic analysis of the responses revealed LLM’s consistently produce nuanced analogies grounded in culturally specific schemata, with particularly high relevance in responses referencing Black/African American and Asian contexts. Analogies included food, family structures, and communal celebrations.

• 116
  Authors: Alan T. Sherman, Bharg Barot, Enis Golaszewski Cyber Defense Lab, Computer Science Department, Maria Sanchez Engineering and Computing Education Program, UMBC, inda Oliva Education Department, UMBC, Peter Peterson Computer Science Department, University of Minnesota Duluth
  Title: Introductory Engineering Education

  Abstract:

  We are developing strategies that enhance engineering and computing student learning by integrating rich authentic cybersecurity concepts alongside foundational computer science or engineering principles. The evolving landscape of cyber threats requires all engineering students to understand essential cybersecurity concepts, but many engineering undergraduate programs introduce cybersecurity late in the curriculum, and only as an elective. Embedding cybersecurity concepts within the curriculum prepares students for complex, cross-disciplinary challenges in the real world. Without increasing student or instructor workload, this approach prepares students to consider cybersecurity throughout the design and implementation of engineering systems. In spring 2026, we are piloting this approach in two introductory courses: CMSC-203 Discrete Math and ENES-101 Introduction to Engineering. The project focuses on integrating adversarial thinking, an approach that includes anticipating potential vulnerabilities, threats, risks, and challenges from the perspective of a malicious opponent. It involves analyzing systems to strengthen the safeguards, confidentiality, authentication, integrity, and availability of a system by identifying potential ways to exploit it. Our approach promises to help engineering students enhance their security awareness and improve system designs. The resources and learning materials we are developing include introducing essentials of adversarial thinking, suggesting probing questions for instructors to pose to students, adding an authentic security focus to examples, exercises, and projects, and providing resources on cybersecurity. For example, for the ENES-101 final project, students design a remote observation vehicle for environmental reconnaissance. We suggest having students discuss possible ways in which a polluter might try to interfere with the vehicle. In a CMSC-203 exercise in which students design a Boolean circuit, we ask students to imagine an alarm circuit and to discuss possible ways in which an adversary might sabotage the circuit. We will analyze the effectiveness of our approach using final exam questions, artifact analysis, and interviews of students and instructors.

• 28
  Authors: Sadia Nasrin Tisha, Md Nazmus Sakib, Rebecca Williams, Karen Chen, Sanorita Dey
  Title: Towards AI-facilitated Collaborative Visual Sensemaking

  Abstract:

  Individuals across various fields increasingly encounter complex data visualizations in classrooms, workplaces, social media, and policy discussions. Making sense of these visualizations involves discussing multiple perspectives, negotiating ambiguity, resolving disagreements, and combining partial insights. These skills are best developed through structured collaborative interaction rather than individual analysis alone. However, in academic contexts, it is difficult for a single facilitator to effectively monitor and support multiple collaborative interactions at once. During collaborative sensemaking tasks, instructors cannot facilitate properly to track every discussion, identify confusion, notice unequal participation, or intervene when misunderstandings occur. As a result, students may disengage, agree passively, or develop incorrect interpretations without timely guidance. To address this limitation, a scalable approach is to use an AI facilitator that can support a collaborative group by monitoring the ongoing discussion, detecting breakdowns such as silence, dominance, confusion, or unchallenged misinterpretations, and providing timely interventions to guide the conversation. However, designing such a system requires understanding how effective human facilitation operates in collaborative visual sensemaking in real life. In this study, we conducted a study with 24 participants (12 dyads) who analyzed real-world visualizations of varying complexity in individual and collaborative settings. The study included three stages: individual interpretation, collaborative discussion, and post-task reflection. During the sensemaking sessions, human facilitators intervened at key conversational moments by identifying facilitation points to balance participation, deepen reasoning, and support understanding of the visualization. Through the analysis of these interactions using inductive coding, we identify effective facilitation strategies, breakdown states in collaborative sensemaking, and cases where facilitation can be counterproductive. These insights provide a foundation for designing AI systems that support collaborative data literacy in educational and professional contexts.

• 114
  Authors: Mei-Lian Vader, Krystal Zhang, Grace Hollen, Emily Wingeart (Presenter), Foad Hamidi, Ravi Kuber
  Title: Investigating the Impact of a 3D Modeling and Printing Workshop Designed for Blind and Low-Vision Students

  Abstract:

  Students with disabilities are currently underrepresented in STEM classes at the high school level due to a number of barriers. Community organizations are partnering with universities to provide exposure to STEM, with a view to helping students develop skills and encouraging further study or careers in STEM-related fields. However, it is unclear how community-university collaborations can inform the design of informal learning programs that support STEM interest in learners with disabilities through inclusive hands-on technology-rich activities. In this poster, we describe our university’s collaboration with a community partner that supports blind and low-vision students to provide exposure to the areas of engineering and making. Specifically, we detail the experiences from a four-day workshop on design thinking, engineering and 3D printing. Eight students who identified either as blind or low-vision, participated in the workshop. The students attend both mainstream and specialized high schools. Observations and interviews were conducted over the four-day period. Findings have highlighted the importance of working with mentors from the community organization as partners, designing and tailoring STEM activities, taking the learner’s abilities and interests into account, and the value associated with flexibility determined by student interest and performance in the activities presented. Feedback provided by the community partner has provided an insight into ways in which the workshop can be strengthened for future offerings.

• 111
  Authors: Chaturya Yarradoddi, Anupam Joshi
  Title: WIP: The Impact of Financial Aid and Academic Pathways on Student Success: Analytics, Predictive Modeling, and Interactive Tools

  Abstract:

  This research creates an integrated predictive modeling system which studies how financial assistance, together with academic readiness and enrollment methods, determines student success in completing their undergraduate studies. The research combines data from 25,926 students through their high school records, placement tests, standardized test results, Advanced Placement coursework, demographic details, and multiple years of financial aid records. The pre-enrollment and first-term variables underwent multiple stages of data preparation, which included cleaning, leakage prevention, feature engineering, and deduplication before machine learning model training. The AUC reached 0.846 on the test set after the XGBoost and gradient boosting models received their optimal performance settings. The analysis of model interpretability revealed that financial support amount, mathematics and English placement results, high school academic performance, and transfer student status determined graduation success. The final model is implemented into a Power BI dashboard through Python-based inference, which enables users to run real-time predictions, adjust classification thresholds, and test different scenarios. The research developed an operational evidence-based system that institutions can use to make decisions about financial aid distribution, academic readiness programs, and student achievement strategies.

• **Doctorate Student Award**
  37
  Authors: Sanzida Akter*, Lida Xu, Mahmoud Jalali Mehrabad, Mohammad Hafezi, Curtis R. Menyuk
  Title: Chaos Diagnostics in Topological Super-Ring Kerr Microcombs via Lyapunov Exponents

  Abstract:

  Topological photonics uses band topology to guide light robustly against disorder. In a standard configuration, a two-dimensional square lattice of microrings with engineered hopping phases models the anomalous quantum Hall effect (AQHE). This configuration forms chiral edge bands that circulate along the lattice boundary. When pumped in the nonlinear Kerr regime, these rings collectively operate as a travelling-wave super-ring resonator, enabling topological soliton formation. While soliton formation on the edge band has been demonstrated previously, systematic identification of chaotic regimes is essential to obtain the operating regime for topological soliton formation. Here, we use Lyapunov exponents to investigate the chaotic dynamics in this system. We simulate a coupled Lugiato- Lefever model on a 12 × 12 AQHE ring lattice to compute finite-time Lyapunov exponents that reveal transitions from stable topological edge states to chaos. By evaluating the sepa- ration of perturbed trajectories over time, we map the transition from non-chaotic to chaotic dynamics and back again to non-chaotic dynamics as the detuning parameter is swept. The Lyapunov exponent thus identifies operating points that support coherent topological combs and separates the detuning ranges that generate chaotic fluctuations. Our work demonstrates a framework for broader stability mapping in complex topological edge-state dynamics.

• 60
  Authors: Logan Courtright, Alioune Niang, Peter Hedlesky, Patrick O’Mullan, Ergun Simsek, Gary Carter, James Eakin, Tanvir Mahmood, James P. Cahill, Weimin Zhou, Doug Petkie, Curtis R. Menyuk
  Title: Design and Characterization of Thick Silicon Nitride Microresonators from AIM Photonics

  Abstract:

  Microresonator-based optical frequency combs have emerged over the past twenty-five years as an adaptable technology with a wide variety of applications. Of particular interest is microresonator integration with existing silicon photonics platforms in order to move laboratory capabilities into the field. As a result, commercial foundries such as AIM Photonics and LIGENTEC have developed processes for fabricating integrated microresonators on popular platforms such as Si3N4. However, many microresonator designs require extreme fabrication geometries to meet necessary loss or dispersion metrics. These geometries require new fabrication processes, which may alter the optical properties of the material and create deviations from nominal geometric values. Small deviations in geometric parameters can significantly affect the microresonator dispersion. These deviations can prevent soliton generation if device dispersion varies excessively from dispersion simulations used in design. To develop a simulation model that accurately predicts the dispersion of fabricated microresonators, an iterative process involving fabrication, characterization, and simulation is necessary. In this work, we present design simulations, experimental characterization, and follow-up simulations for SiO2-clad Si3N4 ring microresonators fabricated by AIM Photonics. We designed 100-GHZ free spectral range, 750 nm thick ring microresonators to maximize the second-order dispersion D2 using finite element method (FEM) simulations. We determined the microresonator dimensions that achieved a maximal D2/2pi value of 0.98 MHz and fabricated a range of dimensions around those. We then performed experiments to linearly characterize the loss and D2 of the microresonators. We measured the loss and maximum D2/2pi for the fundamental TE mode to be 0.1526 dB/cm at 1550 nm and 0.474 MHz, respectively. We finally carried out follow-up FEM simulations to match our characterization results. We determined that the difference in D2 is probably caused by a change in the microresonator thickness during fabrication and small thickness variations across the chip.

• **Undergraduate Student Award**
  38
  Authors: Zachary Danielson (presenting author), Pradyoth Shandilya, Kartik Srinivasan, Gregory Moille, and Curtis R. Menyuk
  Title: Calculating the Route to Chaos of Dissipative Kerr Solitons Driven by a Modulated Laser

  Abstract:

  Optical frequency combs (OFCs) are fundamental tools in precision metrology, coherently linking optical and microwave frequency domains. For non-laboratory applications, compact and robust OFCs are realized by coupling a continuous-wave (CW) laser into a chip-scale nonlinear microresonator via an integrated waveguide. When pumped with appropriate power and detuning, the Kerr nonlinearity of the microresonator enables the formation of dissipative Kerr solitons (DKSs). Periodic out-coupling of these solitons into the waveguide produces a uniformly spaced pulse train, corresponding in the frequency domain to a microresonator-based OFC, commonly known as a microcomb. The injection of a second laser (a reference pump) at a frequency close to one of the microcomb teeth can induce phase locking of the DKS to the reference pump. This mechanism, Kerr-induced synchronization (KIS), successfully generates ultra-low-noise and broadband microcombs. Furthermore, when the reference pump is phase modulated, the DKS can synchronize to the resulting modulation sidebands (AC-KIS). If a comb tooth falls between the reference pump and its sidebands, the DKS group velocity can exhibit chaotic fluctuations. This regime has potential applications in photonic random number generation.The transition from stable KIS to chaos, however, remains to be investigated. In this work, we model the DKS phase dynamics using a second-order Adler equation, which accurately captures the soliton dynamics in the AC-KIS regime. We compute the corresponding limit cycle solutions using Newton’s method adapted for periodic systems. The Adler equation is then linearized about the limit cycle, and the Floquet multipliers are numerically evaluated to determine linear stability. By analyzing the evolution of the Floquet multipliers at the reference detuning value where the DKS loses stability, we characterize the route to chaos. Identifying this transition mechanism provides insight into how the reference pump detuning can be tuned to deterministically trigger chaotic dynamics, which is essential for controlled applications.

• 59
  Authors: Alexander Dorsey, Ankit Goel
  Title: Experimental Validation of Dynamic Mode Adaptive Control on Multi Rotor Platform

  Abstract:

  This paper will present the experimental validation of Dynamic Mode Adaptive Control on a multi rotor aerial platform. Dynamic Mode Adaptive Control is a data driven control framework that identifies local input output dynamics online and synthesizes optimal state feedback gains without requiring an any model information of the system. While previous work has demonstrated the method in simulation, its performance on hardware subject to sensor noise, actuator saturation, computational limits, aerodynamic coupling, and unmodeled dynamics has not been validated. The controller will be implemented on an embedded flight control architecture and tested on a quadrotor platform constrained to purely rotational motion (on a 3 axis gimbal) and in free flight conditions. The adaptive algorithm estimates discrete time, linear state space matrices from real time flight data and updates an LQR based feedback law accordingly. Experimental results aim to demonstrate stable regulation and disturbance rejection. The paper aims to show the method will maintain performance under parameter variations and external disturbances without retuning. The results will provide the first hardware level validation of Dynamic Mode Adaptive Control on a multi rotor system and demonstrate its feasibility for real time adaptive flight control applications.

• 45
  Authors: Raonaqul Islam, (Presenter); Pradyoth Shandilya, Curtis Menyuk, Ergun Simsek
  Title: Self-trapping of dissipative Kerr solitons for low-noise frequency combs

  Abstract:

  Chip-integrable dissipative Kerr soliton (DKS) frequency combs, also known as DKS microcombs, are useful for on-field precision metrology applications by providing a link between optical and microwave frequencies. Thermal fluctuations that result in variations in the refractive index of the Kerr cavity may cause temporal variability in the resonance frequencies. This disruption, known as thermorefractive noise (TRN), results in timing jitter in the repetition-rate of the generated pulses, causing frequency instability in the downconverted microwave signal. Locking repetition-rate to a low-noise microwave source to reduce intracavity noise has been a topic of great interest. It has been demonstrated that the DKS inside the Kerr cavity can be trapped by phase-modulating the laser pump using a low-noise external RF reference enabling a more stable repetition-rate of the generated microcomb. However, integrating these external low-noise RF sources on-chip remains a challenge to date. In this work, we propose that the generated microwave signal can be fed back to an on-chip electro-optic modulator to trap the DKS inside the cavity. The resulting system eliminates the bulky RF reference, making it chip-integrable. We have numerically shown and verified with appropriate analytical models that TRN is significantly suppressed in our proposed system. Our analysis indicates that the noise suppression achieved with the proposed system is competitive with a typical system that is externally trapped by a low-noise reference. This work paves the way for intrinsic noise-reduction techniques in chip-scale opto-electronic systems.

• **Doctorate Student Award**
  46
  Authors: Revati Kadolkar (Presenter), Govind Rao, Douglas D. Frey
  Title: Application of Mechanistic Modeling for the Development of a Miniaturized Size Exclusion Chromatography (SEC) Column for on-site Quantitative Analysis

  Abstract:

  Size exclusion chromatography (SEC) is a widely used technique in biotherapeutics for the characterization of product related impurities or critical quality attributes (CQAs) such as aggregates (HMW) and fragments (LMW). However, application of SEC for continuous CQA monitoring and process control is often limited by the lengthy, offline and complex high performance liquid chromatography (HPLC) analysis. Expensive columns of long lengths make small-scale operation of SEC difficult. To address these concerns, we have developed a miniaturized chromatography column (μCol) (5 x 0.46 cm, 0.8 mL) fabricated in-house with a polymethyl methacrylate (PMMA) framework, containing HPLC-grade SEC resin in an engraved channel for a low-cost (~$20), simple and rapid (~5 min) on-site quantification of CQAs, as a definitive assay. A mechanistic model was prepared using COMSOL Multiphysics to understand the internal transport phenomena, and to screen various innovative designs to obtain minimal wall-effects, ease of pressure packing, increased path lengths for good resolution, and compact shapes for integration into microfluidic chips. Preliminary scale-down studies were performed using simulations to select suitable kinetic parameters such (column length, inner diameter, resin particle size, flow rate, injection volume, theoretical plate height); screen discrete geometries; and to validation the predicted elution profiles using van Deemter analysis. The physics involved in this model contains various interconnected partial differential equations (PDE) governing mass transfer kinetics and fluid flow. It accounts for the flow disparity inside the column, axial and radial dispersion, adsorption on the particle surface, diffusion from bulk liquid and diffusion within particle pores. Our objective was to develop a predictive model using computational fluid dynamics (CFD) to overcome the experimental challenges of high resource consumption in industry and limited access to target molecules. The developed SEC μCol prototype will also be assessed conceptually for virus clearance studies, polishing and scale-up for preparative analytical chromatography.

• **Master Student Award**
  47
  Authors: Benjamin Kale, Meilin Yu
  Title: Simulations of turbulent hydrocarbon fuel injection and combustion in a cavity-based scramjet
  Abstract:

  Scramjets, short for supersonic combustion ramjet, are the primary means of propulsion at hypersonic speeds (generally defined to be those in excess of Mach 5). As the name suggests, in these engines, combustion occurs at supersonic speeds, which makes them challenging to study. Experimentally, achieving all aspects of these high speed, high temperature flow regimes is difficult, even in state of the art ground testing facilities, and obtaining experimental flight data is costly and challenging as well. This makes simulation a valuable tool for exploring these regimes. However, simulation of these regimes comes with its own challenges. It involves accurately predicting the interactions between several different multiscale and multiphysics phenomena such as shock-boundary layer interaction, turbulent mixing, and combustion dynamics. To do this, the commercial solver, US3D, will be used to perform 3-D simulations of turbulent combustion using turbulence models of varying degrees of fidelity. These will include the Spalart-Allmaras Reynolds averaged Navier-Stokes turbulence model with Catris compressibility correction, detached eddy simulations, and large eddy simulations, each of which will be validated against experimental data. Combustion reactions will be modeled with a finite rate chemistry approach. Flame structure and turbulent mixing phenomena will be investigated using various post-processing techniques including chemical explosive mode analysis and spectral proper orthogonal decomposition.

• 48
  Authors: Parham Oveissi, (Presenter); Turibius Rozario, Ankit Goel
  Title: A Novel Neural Filter to Improve Accuracy of Neural Network Models of Dynamic Systems

  Abstract:

  The application of neural networks in modeling dynamic systems has become prominent due to their ability to estimate complex nonlinear functions. Despite their effectiveness, neural networks face challenges in long-term predictions, where the prediction error diverges over time, thus degrading their accuracy. This paper presents a neural filter to enhance the accuracy of long-term state predictions of neural network-based models of dynamic systems. Motivated by the extended Kalman filter, the neural filter combines the neural network state predictions with the measurements from the physical system to improve the estimated state’s accuracy. The neural filter’s improvements in prediction accuracy are demonstrated through applications to four nonlinear dynamical systems. Numerical experiments show that the neural filter significantly improves prediction accuracy and bounds the state estimate covariance, outperforming the neural network predictions. Furthermore, it is also shown that the accuracy of a poorly trained neural network model can be improved to the same level as that of an adequately trained neural network model, potentially decreasing the training cost and required data to train a neural network.

• 49
  Authors: Soumik Sarker (Presenter), Alok Ghanekar
  Title: Slow-Light-Inspired Steering of Directional Thermal Radiation with Phase-change Materials

  Abstract:

  Directional control of thermal emission with a tunable angular range is critical for the development of next-generation smart meta-surfaces and adaptive thermal photonic systems. However, most existing approaches rely on static or nonadaptive mechanisms to regulate thermal radiation, limiting their dynamic functionality. Here, we propose a phase-change-material-based grating incorporating vanadium dioxide (VO₂) to enable adaptive steering of directional thermal emission. The proposed structure supports slow-mode resonances with flat dispersion characteristics, which facilitate dynamic control over both the emission angle and emissivity. By exploiting the reversible metal-to-insulator phase transition of vanadium dioxide (VO₂), the angular characteristics of thermal emission can be dynamically tuned. The substantial refractive index contrast between the metallic and insulating phases enables active modulation of the dispersion properties, thereby allowing real-time control over the emission direction. This phase-transition-driven mechanism provides a practical route toward adaptive and reconfigurable thermal emitters for smart photonic applications. In the slow-light regime, resonant modes exhibit near-zero group velocity, resulting in enhanced sensitivity to refractive index perturbations. The directional response of the resonance scales inversely with the group velocity, such that flatter dispersion bands lead to stronger angular tunability. Consequently, even a modest refractive index change in VO₂ at a fixed frequency induces a significant shift in the in-plane wave vector Kx, enabling pronounced modulation of the optical momentum and emission direction. This mechanism provides an efficient route toward actively tunable thermal emitters, with potential applications in infrared beam steering, thermal camouflage, energy harvesting, and reconfigurable photonic devices.

• 50
  Authors: Nahidul Islam Shadin (Presenter), Raymond Yu, Dr. Michelle L. Povinelli, Dr. Alok Ghanekar
  Title: Reversible Symmetry Breaking of Directional Thermal Emission in VO₂–SRO Gratings

  Abstract:

  We numerically investigate asymmetric thermal emission in a grating structure composed of vanadium dioxide (VO2) and strontium ruthenate (SrRuO3, SRO). The proposed system consists of a GaAs spacer with alternating VO2 and SRO ridges. It is designed to examine how the insulator-to-metal transition (IMT) of VO2 affects directional emissivity. Our simulations show that when VO2 is in the insulating state, it interacts only weakly with the optical field. This occurs due to its low refractive index and subwavelength thickness. As a result, the thermal emission remains nearly symmetric. In contrast, when VO2 transitions into the metallic state, it behaves similarly to SRO. However, strong angular asymmetry appears in the emitted radiation. This asymmetric response originates from a resonance-driven imbalance in the coupling of forward and backward propagating channels. It produces enhanced emissivity at positive angles. At the same time, emissivity is suppressed at negative angles. These findings highlight that material contrast, structural geometry, and mode confinement must work together to produce asymmetry. Overall, our study demonstrates that VO2–SRO gratings provide an effective and tunable route to asymmetric thermal emission. This opens promising opportunities for radiative cooling, thermal antennas, energy harvesting, and infrared camouflage.

• 51
  Authors: Pradyoth Shandilya (presenter), Gregory Moille, Giuseppe D’Aguanno, Kartik Srinivasan, and Curtis R. Menyuk
  Title: A study of dual-pumped microresonators using three-wave equations

  Abstract:

  Microresonator-based optical frequency combs (“microcombs”) are a promising source of low SWaP-C (size, weight, power, and cost) frequency references. For precision metrology, both the microcomb repetition rate and the carrier- envelope offset frequency must be detected and phase-locked to low-noise references. Detecting and stabilizing the repetition rate is largely an engineering challenge and does not impose stringent constraints on the comb spectrum. By contrast, the most widely used method to measure the carrier-envelope offset frequency, f -2f self-referencing, requires an octave-spanning spectrum, where the highest detectable comb tooth exceeds twice the frequency of the lowest detectable tooth. Achieving such bandwidths in microresonators remains difficult. A recent and increasingly successful approach employs two pump lasers: one to generate the microcomb and a second to drive spectral broadening. The resulting intracavity field is a composite structure comprising multiple components that have identical group velocities while maintaining distinct phase velocities. We refer to each component as a “color” and to the overall waveform as a multi-color soliton. Previous theoretical studies modeled dual-pumped microresonators using a multi-pumped Lugiato-Lefever equation (MLLE). Although MLLE simulations reproduce experimentally measured spectra well, they often require long integration times due to non-stationary dynamics, obscure the individual color contributions within interleaved combs, and prevent the use of standard dynamical methods to study the multi-color soliton’s linear stability. Here, we use a multi-color ansatz with the MLLE and retain only phase-matched terms, yielding a reduced set of equations which we refer to as three-wave equations. The three-wave equations separate the different colors and support stationary solutions, enabling the study of their linear stability. Using this approach, we analyze the stability of multi-color solitons and numerically demonstrate how to compute the threshold second-pump power that triggers instabilities.

• 52
  Authors: Gaurav Shinde (Presenter), Anuradha Ravi, Jared Lewis, Andre Harrison , Henry Gardiner, Md Saeid Anwar, Shadman Sakib, Jade Freeman, Nirmalya Roy
  Title: CAViAR: Quality-Aware Vision-and-Radio Fusion for Relative Range Estimation among Collaborative Autonomous Agents

  Abstract:

  In mission-critical scenarios, autonomous agents often operate in GPS-denied or GPS-degraded environments, making it challenging to localize agents relative to themselves and objects of interest (e.g., potential cover) or adversarial robots. While vision- and radio based modalities have individually been used to estimate relative ranges between agents and surrounding objects, each modality suffers from inherent limitations. Vision-based range estimation degrades significantly under low image overlap, partial visibility, or when an agent is too close to an object, resulting in effective zoom-in and loss of geometric context. Conversely, radio-based ranging is susceptible to multipath interference, signal fading, and environmental variability, which can substantially reduce range accuracy and reliability. To address these challenges, we introduce CAViAR, a quality-aware multimodal fusion framework for accurate relative range estimation among collaborative autonomous agents. CAViAR assigns modality-specific reliability scores and performs statistical fusion to adaptively weigh vision and radio inputs based on their estimated quality. The framework employs modality-specific quality estimators augmented with temporal features and integrates MBConv blocks to enable efficient feature processing on resource-constrained robotic platforms. We validate CAViAR on ROSbot 2 and ROSbot 2 Pro platforms using an in-house dataset collected across diverse indoor and outdoor environments. Experimental results demonstrate that our approach outperforms single-modality baselines by approximately 21% over vision-only and 36% over radio-only range estimates. Moreover, CAViAR adapts robustly to variations in scene structure, viewpoint overlap, and occlusions without requiring fine-tuning on new environments, highlighting its practicality for real-world deployment.

• 58
  Authors: Paris von Lockette
  Title: The Electro-Magnetically Active Composites and Structures (eMACS) Lab at UMBC
  Abstract:

  The Electro-Magnetically Active Composites and Structures (eMACS) Lab at University of Maryland, Baltimore County advances the science and engineering of field-responsive materials that couple electromagnetic stimuli with mechanical response. Our research focuses on the design, fabrication, modeling, and application of magneto-active and electro-active composites whose microstructural organization enables programmable stiffness, shape morphing, and controlled actuation. By embedding ferromagnetic or dielectric inclusions within compliant polymer matrices, we engineer materials whose mechanical properties and deformation modes can be dynamically tuned through external fields. A central theme of the lab is microstructure–property linkage: we investigate how particle volume fraction, alignment, and spatial organization govern emergent macroscopic behavior. Combining multiscale modeling with experimental characterization, we develop predictive frameworks that connect particle-scale interactions to bulk constitutive response. Advanced fabrication strategies, including multi-field curing and additive manufacturing, are used to encode anisotropy and functional gradients directly into composite architectures. Applications span soft robotics, adaptive structures, biomedical devices, and energy-efficient actuation systems. Ongoing efforts include magnetically programmable elastomeric mechanisms, morphing accordion-type structures, and magnetically actuated assistive technologies. Through integration of theory, experimentation, and translational prototyping, eMACS seeks to move beyond proof-of-concept demonstrations toward robust, scalable smart material systems. By uniting mechanics, materials science, and electromagnetic physics, the eMACS Lab establishes foundational principles for next-generation responsive structures capable of remote, wireless, and programmable control.

• 55
  Authors: Liam Wilson (Presenter) and Meilin Yu
  Title: Wake-type Gust Mitigation with Airfoil Pitch Motion

  Abstract:

  This project investigates the unsteady aerodynamic response of a rigidly pitching airfoil undergoing periodic gusts using numerical simulations in ANSYS Fluent. Understanding the effects of gusts on the lift and drag allows for control of the airfoil angle of attack to affect the lift and drag. This CFD study makes use of a 2D pressure-based incompressible flow solver with second-order schemes in both space and time, and a user‑defined function to prescribe the airfoil motion. A convergence study is performed, along with investigating different turbulence models to balance computational cost and accuracy. The gusts are created through vortex shedding over a semi-cylinder, following established research methodologies to mimic animal aerodynamics. Different Strouhal numbers are simulated, informing the pitch frequency of the airfoil, with the purpose of mimicking the biomechanical thrust generated by aquatic and aerial animals. Along with the biomechanics-inspired motion, the inlet conditions are varied to provide a wide range of data that resembles normal conditions that the airfoil will experience in its use case. The simulations will provide insight on aerodynamic parameters to allow improvements in flight and vertical-axis turbine dynamics. These results will provide a foundation for future work on more complex morphing kinematics, leading to further improvement to small Unmanned Aerial Vehicles (UAVs), as well as improved efficiency for wind-energy technology.

• **Undergraduate Student Award**
  56
  Authors: Monty Yates (Presenter), Reece Robertson
  Title: Fluids on the Bloch Sphere: Quantum Algorithms for Solving PDEs on NISQ

  Abstract:

  Quantum algorithms have a variety of applications, including solving differential equations. In this work, we analyze the effectiveness of NISQ devices at solving partial differential equations (PDEs). To this end, we combine a classical linearization routine with a hybrid quantum-classical adaptation of the Harrow-Hassidim-Lloyd (HHL) algorithm, a quantum algorithm for solving linear systems of equations. We demonstrate our routine by solving Burger’s equation on both superconducting and trapped-ion quantum hardware, and compare the runtime to classical PDE solvers.

• **Doctorate Student Award**
  57
  Authors: Cindy Zozimo Aranda de Almeida (Presenter), John T. Hrynuk, Meilin Yu
  Title: SPOD reconstruction analysis on transitional flow applied to wing-gust interactions

  Abstract:

  Spectral Proper Orthogonal Decomposition (SPOD) decomposes data in time and space by identifying energy-ranked modes that each oscillate at a single frequency. It is used to extract coherent structures or modes from flow data. The most popular application of SPOD is with turbulent flows. In this work, the SPOD reconstruction is going to be studied at low Reynolds numbers, which is characterized by laminar to turbulent transitional flow, a regime which is not frequently the subject of SPOD analyses. Here, an airfoil is subjected to a long duration transverse gust, where two mitigation techniques were applied: first, in which the whole airfoil oscillates continuously, and second, in which only the trailing edge oscillates. SPOD was performed in the wake region, where frequency reconstruction and mode reconstruction are compared and judged by the ARMI (Average Relevance Magnitude Index) metric. Results show that mode reconstruction is more efficient than frequency reconstruction. ARMI demonstrates how each region is far or close from the true data at each addition for both methods, and finally, the L2 norm of the reconstruction helps to conclude that in space the SPOD window chosen was enough to characterize the flow, but in time, the temporal window was not enough to capture the dominant low-signal frequency due to the slow evolvement of the dominant flow structures in time.

• 53
  Authors: Amir Babaei-Gharehbagh (Presenter), Jemma Przybocki, Marissa Cuevas, Isaiah Smith, Maria A.Zawadowicz, Akua Asa-Awuku, Benjamin A. Nault, Peter F. DeCarlo, Christopher J. Hennigan
  Title: Urban-Rural Gradients in Aerosol Liquid Water and pH during the CoURAGE Campaign

  Abstract:

  Aerosol liquid water content (ALWC) and pH are often linked in atmospheric particles due to factors like meteorology and particle composition that influence both parameters. In this study, we characterize urban-rural gradients in ALWC and aerosol pH near Baltimore, MD during the 2025 Coastal-Urban-Rural Atmospheric Gradient Experiment (CoURAGE) field campaign. Simultaneous measurements at paired urban and rural locations show the near-ubiquitous presence of an urban heat island (UHI), with significantly higher temperature and lower relative humidity at the urban site. The average observed differences range between 1-3°C and 8-15% RH, but the differences can exceed 8°C and 25% RH at certain times. The UHI shows distinct seasonal and diurnal cycles that influence pH through effects on particle composition and ALWC. We characterize the resulting differences in semi-volatile species partitioning (NH3-NH4+ and HNO3-NO3-) and discuss the UHI as a key driver of aerosol processes in urban areas.

• **Doctorate Student Award**
  54
  Authors: Emam Hossain (Presenter), Md Osman Gani
  Title: Refining Partial Causal Graphs Through Interventional Representation Learning

  Abstract:

  Causal representation learning aims to uncover latent variables that correspond to meaningful generative factors while capturing the causal relationships among them. Existing approaches typically assume a fully specified causal graph, which is rarely available in real-world scientific systems, or attempt to learn the entire graph directly from observational data, which can be unreliable in high-dimensional and nonlinear settings due to limited identifiability and the presence of spurious correlations. This work introduces a Partial Causal Graph Variational Autoencoder (PCG-VAE), a framework that integrates partial structural knowledge with deep generative modeling to jointly learn interpretable latent representations and refine unknown causal relationships. In this model, known edges are fixed, forbidden edges are constrained, and unknown edges are parameterized and learned using interventional consistency objectives that simulate latent interventions and evaluate how perturbations propagate through the structural mechanism. This formulation provides a practical middle ground between fully supervised and fully unconstrained structure learning. We evaluate the framework on controlled visual environments, including a synthetic pendulum system with interacting physical factors, and extend it to real-world image data using the CelebA dataset. Results indicate that the model can recover semantically meaningful latent factors while progressively updating unknown portions of the causal graph. This work advances scalable causal representation learning in domains where partial domain knowledge exists but complete causal structure is not available.

• 63
  Authors: Nathalie J. Lombard (presenter), Trevor P. Needham, Hilda Khoei Fadaei, Rebecca Donovan, Joel Baker, Upal Ghosh
  Title: Using historical datasets on fish consumption advisory to assess management action effectiveness

  Abstract:

  Historical dataset collected by the states, territories, and tribes to establish fish consumption advisories represents a valuable source of information on PCB contamination levels in the waterbodies, changes over time, and could potentially inform on the effectiveness of management actions and where more action is needed. In this study, we used 3 decades of PCB concentration measured by the Maryland Department of Environment to perform spatial and temporal analysis. The analysis was conducted using lipid normalized data or wet weight tissue from the top 3 species sampled in Maryland. Land use, toxic release inventory (TRI) basic data files from 1987 to 2021, and Superfund sites location and status were also collected for all watershed boundaries of hydrological unit 8 of Maryland. The approach enabled the identification of 6 watersheds highly impacted by PCBs out of 25, including the Upper Chesapeake Bay. Known areas of contamination were better predictors of PCB levels in fish than generic land use or toxic release information. All impacted watersheds showed slow nationally consistent natural attenuation rates ranging from 3% to 6% yr-1, except one, the Gunpowder Patapsco, where concentration in fish has been without a temporal trend since 1996. The overall study showed that PCB source mitigation implemented across Maryland and surrounding States helped recoveries of moderately to highly impacted watersheds, but more efforts are needed to fully control all PCB sources (legacy or ongoing) that are still impacting the rivers and their fish.

• **Undergraduate Student Award**
  64
  Authors: Robert Schroeder (Presenter), Md Badrul Hasan, Dr. Meilin Yu
  Title: Data-Driven Modeling of Dynamic Stall in Vertical-Axis Wind Turbines

  Abstract:

  Lift-based Vertical-Axis Wind Turbines (VAWTs) have shown much promise over the industry standard Horizontal-Axis Wind Turbines (HAWTS), offering advantages including modularity, scalability, and sustainability. Sandia National Laboratory published an influential report in 1981, using experimental data of static symmetric foils to determine their lift, drag, and moment coefficients as a function of angle of attack (AoA) and Reynolds Number. The Sandia study is often referenced in aerodynamic models of VAWTs to improve performance and understand dynamic stall, which typically occurs when a maneuvering foil/wing experiences a very large AoA far exceeding the stall AoA of its static counterpart. Dynamic stall is defined as a sudden increase and decrease in lift coupled with an increase in drag due to the elongated interaction between the leading edge vortex and foil/wing followed by vortex shedding. The applicability of the Sandia study is limited by its methodology, where each blade was statically tested and does not accurately capture dynamic stall effects seen in real-world turbines. Although higher tip-speed-ratios (TSR) of a VAWT decrease the maximum AoA, the temporal AoA change rate also influences stall occurrence and severity. Using high fidelity CFD data for VAWTs, this research investigates nonlinear regression techniques based on angle of attack and its temporal rate to model aerodynamic force coefficients, with the goal of better capturing dynamic stall events. Preliminary results indicate that linear regression provides limited predictive capability, and although polynomial regression improves the fit, it does not fully represent the nonlinear and time-dependent aerodynamic effects, motivating the use of neural networks. Future works for this project will include various blade pitching schemes, further generalizing the model.

• 61
  Authors: Sirisha, Milton Halem
  Title: Towarda an advanced AI fire Forecast model

  Abstract:

  Regional AI models are scarce due to high computation power, lack of data and difficult to compute boundary conditions. The existing AI models are majorly global models with a very few regional models. All the AI models currently are only atmospheric with no land surface variables which is crucial to predict possibility of a fire or fire spread. This project focuses on developing a Fire AI model that included land surface models and comparing it with traditional physics based model NUWRF.

• 62
  Authors: Dingxiang Zhu (Presenter); Ye Lu
  Title: A Novel Digital Image Correlation Framework for Structural Health Monitoring

  Abstract:

  This work presents a novel digital image correlation (DIC) framework for structural health monitoring (SHM). The idea is to leverage the capability of the phase field (PF) method for tracking complex crack morphologies and provide a natural way in DIC to perform damage and crack measurements from experimental speckle images, in addition to displacement and strain fields. We will use several numerical examples to demonstrate the capability of the proposed PF-DIC in terms of capturing various cracks, including the small ones that are difficult to visually identify from speckle images, while improving the measurement accuracy of displacement fields. Additionally, we will show that the proposed PF-DIC can be easily adapted to selectively extract critical cracks under specific loading circumstances for damage assessment. The proposed DIC framework offers numerous opportunities for real-time SHM and automatic damage detection in science and engineering problems.

• **Doctorate Student Award**
  77
  Authors: Sultan Ahmed, Sanjay Purushotham
  Title: FSA-Bench: Large-Scale Benchmarking of Survival Models in Federated and Heterogeneous Settings

  Abstract:

  Survival analysis comprises statistical and machine learning methods for modeling time-to-event data, jointly considering event time and survival status. Prior comparative studies largely rely on single-center datasets and evaluate classical statistical and neural network models in centralized settings. However, single-center data collection often requires extensive cross-institutional harmonization, which is complex and error-prone. Aggregating sensitive survival data can also violate privacy regulations such as GDPR and HIPAA. In practice, survival data are distributed across multiple centers and domains. This motivates Federated Survival Analysis, where models are trained collaboratively without sharing raw data. Despite recent methodological advances addressing heterogeneity and privacy constraints, the performance of survival models in large-scale, heterogeneous, and decentralized settings remains insufficiently explored. Most existing benchmarks continue to assume centralized evaluation, overlooking challenges related to non-IID data, decentralized ownership, and cross-domain variability. To address these gaps, we introduce FSA-Bench, a comprehensive benchmark for systematic comparison of survival models in federated and multi-domain settings. FSA-Bench evaluates CoxPH-based models, traditional machine learning survival approaches, neural network-based survival methods, and state-of-the-art attention-based Federated Survival models using multiple metrics to assess discrimination and predictive accuracy across diverse survival datasets. Our large-scale systematic study, consisting of 1600 experiments across 11 methods, 50 datasets, and 3 settings, demonstrates that survival model performance varies substantially with dataset characteristics and evaluation criteria. These findings reveal key performance trends and limitations, providing practical guidance and a foundation for future research in privacy-preserving and federated survival analysis.

• **Doctorate Student Award**
  78
  Authors: Wanqing Chen (Presenter), Kai Sun
  Title: Predict-Then-Optimize for Large-Scale Anesthesiologist Scheduling with LLM-Enhanced Availability Prediction

  Abstract:

  Anesthesiologists frequently cover multiple clinical sites and assume diverse roles across administrative leadership, perioperative coordination, subspecialty services, and patient care in various settings. Many hold joint appointments across private practice and academic settings or work part-time. The scheduling problem becomes particularly complex in academic anesthesia departments, where faculty anesthesiologists must be coordinated alongside residents, fellows, and certified registered nurse anesthetists (CRNAs) across clinical sites. The heterogeneity in roles, training/educational requirements, and institutional constraints, combined with the national shortage of anesthesiologists, create a large-scale, multi-layered workforce coordination challenge in hospital operations. Anesthesiologist availability is further complicated by heterogeneous scheduling lead times across clinical sites and frequent post-publication schedule revisions driven by clinical urgency and/or operational disruptions. Although prior research has modeled provider availability using structured historical schedules and demand data, these approaches cannot fully capture the latent preferences embedded in unstructured scheduling comments commonly recorded in clinician scheduling platforms. Large language models (LLMs) provide an opportunity to extract and formalize this underutilized information. To address this gap, we develop a predict-then-optimize (PTO) framework for large-scale anesthesiologist scheduling with LLM-enhanced preference signals.
  In the prediction stage, LLMs are used to extract preference-related signals from unstructured free-text scheduling comments. These extracted signals are integrated with structured historical schedule data to enhance predictive models, including logistic regression, random forest, support vector machines, neural networks, and LSTM. The goal is to improve clinician availability prediction. In the optimization stage, the predicted availability probabilities are incorporated into a two-level prescriptive framework. First, robust optimization with a budget-of-uncertainty formulation is used for optimal shift design under clinical demand uncertainty. Second, the availability probabilities are embedded into a multi-objective mixed-integer program that generates detailed work schedules considering optimal and equitable workload distribution. The proposed framework is evaluated using near-real-world historical schedule templates derived from our clinical partners, supplemented with controlled simulated scheduling comments. Experiments under varying comment-density scenarios are conducted to assess the robustness and marginal value of LLM-derived signals, as well as the operational effectiveness of the resulting schedules. The results demonstrate that the proposed method outperforms baseline approaches in incorporating provider preferences, reducing average daily workload, and decreasing workload variance across clinicians.

• **Doctorate Student Award**
  79
  Authors: Ommo Clark (presenter); Karuna Pande Joshi
  Title: The Online Health Safety Gap: VERITAS — A Neuro-Symbolic Framework for Risk-Aware Credibility Assessment in Online Health Discourse Beyond Misinformation Detection

  Abstract:

  People increasingly share rich personal health experiences online and information seekers act on that information to make consequential self-care decisions like adjusting medications, discontinuing treatments and delaying care. Safe and trustworthy online health information is a matter of patient and public health safety. Yet computational systems designed to detect health misinformation only check whether content agrees with consensus, a binary verification that structurally misses the dominant form of potential harm. Two root causes explain this failure. We term the first Narrative Blindness; the systematic inability of computational systems to extract, represent, and reason over the causal and contextual structure determining whether the narrative is potentially safe to act upon. The second we formalize as the Risk Irrelevance Principle; epistemic divergence from established consensus and risk potential are independent dimensions, factual alignment provides near-zero information about harm potential when acted upon (Pearson r = -0.230; 7/7 independence tests satisfied). We present VERITAS (Verification Engine for Risk-aware Information Trust Assessment in health Stories), a neuro-symbolic framework shifting credibility assessment from binary consensus-checking toward risk-aware evaluation of implied behavior. VERITAS reconstructs health stories into Agent-Action-Outcome narrative graphs, computes two novel independent metrics; Narrative Truth Distance (NTD), measuring epistemic divergence, and Narrative Risk Score (NRS), measuring clinical risk potential, and based on the joint NTD-NRS scores assigns narratives into a Credibility Quadrant that replaces binary verdict with graduated, proportionate assessment. Empirically, 68.8% of narratives in our corpus are epistemically aligned with consensus yet potentially harmful, invisible to binary systems by design. VERITAS achieves 66.9% recall versus 6.3% for binary baselines, a 10.6× improvement (p < 0.001) with 100% recovery of this dominant risk class. VERITAS provides trustworthy credibility assessment for online health information seekers, platforms, clinicians, and AI developers, evaluating not just what health narratives claim, but whether acting on them is potentially safe.

• 89
  Authors: Weiding Fan, Sanjay Purushotham
  Title: Worst-Client-Aware Federated Structure Learning for Multi-Environment Time Series

  Abstract:

  Federated learning for multi-environment time series is often evaluated by average risk across clients. When each client corresponds to a distinct environment with its own data-generating distribution, cross-client non-IID heterogeneity can yield models that look good on average yet degrade sharply on the worst-performing clients.The
  The problem is amplified when the learning target includes recovering an interpretable temporal structure, such as a shared lag dependencies via Granger causality test. In such settings, optimizing mean risk may select lag patterns that improve performance on dominant environments yet hurt the worst-case clients, yielding unstable or spurious structure. To address these challenges, we propose a worst-client-aware federated structure learning framework based on a chi-square Distributionally Robust Optimization (χ²‑DRO) objective with risk-coupled aggregation of representation and structure. At each communication round, the server computes robust client weights from per-client validation risk by upweighting poorly performing clients subject to a chi-square constraint. The same weights are then used to aggregate both representation updates and structure-gate updates, resulting in aligned predictive performance and structural learning under a unified worst-case-aware objective. To prevent gate instability caused by client drift, multiple local steps, and partial participation, we further introduce a drift-stabilized structure update implemented via proximal stabilization or dynamic regularization on the shared gate. On synthetic temporal graphs with known ground-truth edges, our method improves recovery of cross-environment stable lag edges and produces more stable structures across communication rounds. On a real multi-environment COVID-19 hospitalization dataset, it reduces worst-client error and client-level CVaR (average error over the worst-performing clients), while producing more consistent structures during training.

• **Doctorate Student Award**
  80
  Authors: Safayat Bin Hakim (Presenter), Aniqa Afzal, Qi Zhao, Houbing Herbert Song
  Title: CyberCane: Neuro-Symbolic, Ontology-Guided RAG for Trustworthy Phishing Detection in Healthcare and Beyond

  Abstract:

  Phishing attacks drive over 60% of organizational data breaches, yet privacy-critical sectors — healthcare, finance, and government — cannot adopt existing AI detectors without violating regulations like HIPAA. Older adults and non-specialist staff face compounded vulnerability, as misclassified critical communications such as appointment reminders or prescription notifications can directly disrupt patient care and delay life-critical decisions. Commercial LLM-based tools expose sensitive patient information in 53.2% of communications when transmitting raw email content to external APIs in controlled comparisons, while traditional rule-based systems remain brittle against modern AI-generated attacks that mimic legitimate domain-specific language with high fidelity. We present CyberCane, a neuro-symbolic framework that resolves this regulatory-security conflict through a privacy-by-design dual-phase architecture. The first phase applies deterministic symbolic rules across DNS authentication signals, URL patterns, and content heuristics, producing instant verifiable decisions without any external data transmission. Borderline cases escalate to a second phase where sensitive data is automatically redacted before retrieval-augmented generation (RAG) semantically matches emails against a curated phishing corpus, grounding AI explanations in concrete historical evidence rather than opaque model outputs. A structured knowledge framework — PhishOnt — encodes phishing attack knowledge as formal reasoning chains, enabling transparent multi-label attack classification with verifiable explanation trails suitable for regulatory audit compliance. Evaluated on DataPhish 2025 — a contemporary benchmark of 12,300 emails spanning template-based to GPT-generated attacks — CyberCane achieves 99.5% precision with a 0.16% false positive rate, detecting 78.6% more phishing attacks than rule-only baselines. For a mid-sized healthcare organization processing 10,000 daily emails, deployment analysis demonstrates 542× ROI with $816K net daily risk mitigation at $1,506 operational cost, with tunable operating points supporting diverse risk tolerances across clinical, financial, and government workflows. The system is open-source, and we welcome collaborations to extend CyberCane to new regulated domains requiring explainable, privacy-compliant AI security.

• **Doctorate Student Award**
  81
  Authors: Tasnim Nishat Islam, Mohamed Younis, Wassila Lalouani, Lloyd Emokpae, Roland Emokpae Jr
  Title: Breathing Cycle Detection for Respiratory Tele-health Systems

  Abstract:

  Recent advancements in wearable devices have enabled the acquisition of lung sounds in real time. By analyzing these signals, key indicators such as respiratory cycles and heart sound components can be extracted, hence enabling the development of telehealth solutions for remote assessment of pulmonary conditions. Particularly, detecting respiratory cycles within the collected sound data plays a crucial role in both clinical and diagnostic applications. Accurate identification of breathing patterns facilitates the assessment of respiratory function and supports early detection of anomalies, including COPD, pneumonia, asthma, and COVID-19. In this paper, we promote a two-step process that first estimates breathing sound signal envelope (in
  time domain) and then analyzes the envelope peaks/valleys to calculate the respiratory cycle. Three methods that follow such a process are proposed. We examine the practicality, scalability, and efficacy of these methods in both healthy and pathological cases, highlighting their potential for integration into real-world respiratory monitoring and screening systems. We further evaluate their performance using the public dataset ICBHI-2017, which allows comparative analysis of various respiratory conditions.

• **Undergraduate Student Award**
  82
  Authors: Urvi Jain (Presenter), Cynthia Chikomoni (Presenter), Dr. Corine Jackman Burden (Faculty Mentor)
  Title: Investigating the Role of Pneumococcal Phenotypic Heterogeneity in Determining Virulence Severity

  Abstract:

  Streptococcus pneumoniae (Spn) is the leading cause of bacterial pneumonia hospitalizations in the United States and kills over 3,000 children under 5 and 5,000 older adults each year. Spn typically asymptomatically colonizes the nasopharynx asymptomatically, but at times, disseminates to other organs causing invasive diseases like pneumonia, bacteremia, and meningitis. Increasing incidences of antibiotic resistance and the growing severity of pneumococcal disease underscore the urgent need to understand the molecular mechanisms that drive the onset of pneumococcal infections. Previous work has shown inherent heterogeneity in the induction levels of PphrA, a promoter of the gene regulatory network TprA/PhrA that contributes to pneumonia and mortality in murine models. However, the influence of heterogeneity on severity of infection is largely understudied. Our lab will address this gap in two complementary ways. First, we will determine the level and duration of heritability of PphrA induction by calculating the intraclass correlation coefficient (ICC) and phenotypic memory time, thereby assessing how long transcriptional states persist across bacterial generations. Second, we will investigate the effects of a wildtype strain and a phrA deletion mutant on pulmonary colonization dynamics and host inflammatory responses using a transwell plate and then a lung-on-a-chip model system that recapitulates key features of the lung microenvironment. Ultimately, this work will advance our understanding of how phenotypic heterogeneity shapes infection outcomes and aims to guide the development of innovative interventions against pneumococcal disease.

• 90
  Authors: Yiming Liao, Keke Chen (presenter and faculty mentor)
  Title: Med-HEAL: Analyzing and Healing Hallucinations in Medical LLMs through ICL

  Abstract:

  Hallucinations in large language models (LLMs) pose significant risks in clinical settings, especially when models produce plausible yet incorrect medical information. Despite progress in medical LLMs, systematic benchmarks for detecting and reducing hallucinations are still limited. In this study, we introduce a labeled medical dataset, annotated and validated by both LLMs and human experts, derived from EHRNoteQA. We generate the dataset using a medical-specific open-source LLM, BioMistral, and identify correct and incorrect answers through GPT-4-O and human experts. We also utilize the hallucination-labeled examples in in-context learning, demonstrating that, in the optimal retrieval few-shot setting, the accuracy of the general-knowledge model Qwen2.5 increases from 77.1% to 80.1%. Our results underscore the importance of calibrated, domain-specific, labeled datasets to assess and enhance the reliability of medical LLMs. This work lays the foundation for more trustworthy deployment of medical LLM applications.

• **Doctorate Student Award**
  83
  Authors: Alyssa N. Maguina (presenter), Sithumina Weerarathna, Rishav Gupta. Faculty Mentors: Dr. Janelle Clark and Dr. Dong Li
  Title: Integrating Psychophysics, Tissue Mechanics and Smartphone Ultrasound for Personalized Human-Robot Interaction

  Abstract:

  As we progress in robotic development, the human aspects of customization and comfort in human-robot interaction (HRI) remain underdeveloped. This limitation affects both assistive robots and haptic feedback devices intended to safely convey touch information. While current research emphasizes task performance, movement and injury avoidance, individual comfort thresholds remain insufficiently explored. Comfort depends on body location, interface geometry and tissue stiffness, also our previous work demonstrates strong correlations with tissue properties such as elasticity and hysteresis. In parallel, emerging smartphone-based ultrasound platforms offer a low-cost, portable means of tissue characterization, but their use for modeling comfort and safety limits in HRI has not yet been systematically investigated. In previous work, we used psychophysical methods to define the Allowable Stimulus Range (ASR) representing the range of detectable and comfortable cues. Building on this foundation, this project systematically quantifies tissue mechanics using indentation testing and deformation sensing to estimate stress, strain, Young’s modulus and hysteresis. A mobile application enables ultrasound transmission, data acquisition and visualization. Signal processing methods, including pulse-echo, Doppler and spectral analysis, are then applied to extract tissue composition and deformation features. Contact metrics will allow exploration of the relationship of perceptual thresholds and ultrasound data and support validation of future tissue simulation development. We present an integrated framework that combines psychophysical testing, skin biomechanics, indentation-based tissue characterization, smartphone ultrasound sensing and machine learning to (1) investigate relationships between tissue mechanics and perceived comfort and (2) model and predict individual comfort thresholds in physical HRI. We expect to clarify the influence of robot surface area and body location on individual comfort thresholds and identify links between tissue mechanics and perceived comfort. These findings will inform the design of safer and more personalized assistive and haptic systems, enable smartphone-based assessment of tissue properties and comfort limits for individual users.

• **Master Student Award**
  84
  Authors: Milind Rampure (Presenter). Faculty Mentor/Advisor: Dr. Nirmalya Roy. Co-Advisors: Dr. Anuradha Ravi, Dr. Zahid Hasan
  Title: Robust Human Perception for Search and Rescue Robotics: From Underwater Identification to Contactless Vital Assessment

  Abstract:

  Search and rescue (SAR) operations demand robotic systems capable of perceiving humans under conditions where conventional sensing breaks down. Underwater environments render facial recognition ineffective due to dive equipment and turbidity, while terrestrial scenarios present challenges from smoke, darkness, and extreme lighting variation. This work addresses these challenges through two complementary contributions leveraging body-intrinsic human characteristics and multimodal sensing. For underwater SAR team coordination, we introduce an anthropometric-based diver identification framework extracting metric-scale body proportions from 3D skeletal poses. The system transforms twelve body measurements into 66-dimensional ratio vectors encoding unique physical proportions shoulder-to-hip ratios, arm-to-torso proportions, and limb symmetry invariant to viewing distance and environmental degradation. A Siamese network with contrastive learning enables cross-domain recognition by training on on-land captures augmented to mimic underwater occlusion patterns, then validating on a self-sourced dataset of 3,000 underwater samples across four subjects in diverse swimming postures. For terrestrial SAR vital assessment, we present a multimodal contactless respiratory rate monitoring system deployable across heterogeneous mobile robots. The framework fuses four sensing modalities RGB, thermal, night-vision infrared, and low-light cameras with adaptive signal quality filtering employing modality specific thresholds to accommodate varying sensor noise profiles. A pose-based chest region extraction method enables posture-invariant monitoring across standing, sitting, and lying positions with robustness to partial occlusions. This system is evaluated across three platforms (QUGV: SPOT & Vision60, UGV: Husky A300) spanning quadruped and wheeled configurations with different edge computing architectures, the system achieves reliable respiratory estimation at 2-8 meters, with amplitude-independent normalization maintaining validity at extended ranges where signal strength diminishes.

• **Doctorate Student Award**
  92
  Authors: Pavan Raj Ravi, Dr.Kai Sun, Dr.Jianwu Wang
  Title: A Ground-Truth Structural Causal Simulator for Nursing well-being Analytics

  Abstract:

  In 2022, the National Academy of Medicine released the National Plan for Health Workforce Well-Being, highlighting workforce sustainability as central to the stability of the healthcare system. Yet the nationwide nursing shortage continues to threaten patient safety and hospital performance, with annual turnover rates over 15% . Although nurse workload and well-being have been widely studied, 82% of research relies on cross-sectional designs that preclude causal inference, leaving administrators without rigorous tools to evaluate whether scheduling policies reduce burnout and attrition. Electronic health records (EHRs) record digital traces that could integrate objective operational data with workforce well-being measurement. However, linking subjective well-being constructs, e.g., burnout, moral distress, and cognitive workload, to EHR-derived activity patterns to support credible causal inference and resource planning policy making remains challenging, particularly given that real-world experimentation is costly, ethically constrained, and lacks ground truth for validating causal models. To address this gap, we developed a simulator grounded in a structural causal model (SCM) that generates synthetic longitudinal EHR-like data with known causal ground truth for evaluating causal inference methods and workforce policies. Based on a systematic synthesis of nursing workforce literature, we constructed an SCM comprising 37 causal relationships spanning scheduling exposures, psychological mediators, and patient outcomes. The SCM specifies directional relationships and effect magnitudes, with parameters calibrated using effect sizes from widely recognized empirical studies and meta-analyses on shift duration, burnout dynamics, and staffing–mortality relationships. The simulator incorporates operational constraints, assignment rules, and demand patterns derived from MIMIC-IV data to generate interconnected longitudinal datasets, including (i) shift-level assignments capturing nurse-patient matching and task allocation, (ii) weekly nurse well-being trajectories tracking burnout accumulation and turnover events, and (iii) patient-level outcomes linking nursing exposures to clinical safety endpoints. The SCM operates jointly with the scheduling simulator to dynamically update nurse behaviors in response to well-being outcomes, enabling realistic turnover events and changes in workforce composition over time. The proposed framework was validated by demonstrating simulated operational and well-being outcomes aligned with established ICU benchmarks and reproducing literature-consistent causal effects, supporting the credibility and realism of the modeling approach. This simulator provides researchers and hospital staff with a platform for systematically evaluating and comparing causal inference algorithms against known ground truth, with results validated against established literature effects.

• **Undergraduate Student Award**
  86
  Authors: Jason Rojas (Presenter), Jiajie He, Yash Patel, Yuechun Gu, Zeyun Yu, and Keke Chen (Faculty Mentor)
  Title: Secure-by-Disguise: Clinical Validation of DisguisedNets for Confidential Medical Imaging

  Abstract:

  The integration of deep learning (DL) into clinical workflows has revolutionized medical image analysis; however, the intensive computational requirements of these models often necessitate outsourcing training to public cloud environments. This practice introduces significant privacy risks, as sensitive Protected Health Information (PHI) within medical images remains vulnerable to exposure or reconstruction during cloud-based processing. While DisguisedNets—a framework utilizing Random Multidimensional Projection (RMT) and AES-pixel-level encryption—has successfully enabled confidential training on low-resolution benchmarks like MNIST and CIFAR-10, its efficacy in complex, high-resolution medical domains remains unverified. In this paper, we conduct a comprehensive evaluation of the clinical utility and security of DisguisedNets across classification (breast cancer screening and multi-class wound assessment) and semantic segmentation (colonoscopy polyp and wound delineation). Leveraging high-resolution open-source clinical datasets, we demonstrate that models trained on disguised images maintain high diagnostic accuracy for classification, ensuring that raw anatomical structures remain visually and algorithmically unintelligible to the cloud provider. Our findings also reveal a notable “privacy tax” in pixel-level segmentation. We observe that the disruption of spatial continuity caused by block-wise permutations and encryption poses a significant challenge for the spatial inductive biases inherent in segmentation architectures. This work provides the first empirical evidence that DisguisedNets can successfully scale to real-world clinical classification, offering a secure, HIPAA-compliant pathway for resource-limited clinics to leverage high-performance cloud computing. Conversely, our results highlight that further advancements in image-disguising technologies are required to achieve satisfactory performance in spatially-sensitive tasks like medical image segmentation.

• **Doctorate Student Award**
  87
  Authors: Shadman Sakib (Presenter), Gaurav Shinde, Nirmalya Roy
  Title: Decoupling Perception and Reasoning for Contactless Respiratory Rate with Vision Language and Small Language Models

  Abstract:

  Contactless respiratory monitoring from video is a promising alternative to wearable sensors but remains sensitive to motion artifacts and variations in viewpoint and lighting, which degrade the reliability of respiratory rate (RR) estimates. Traditional signal processing methods often struggle to separate physiological motion from visual artifacts, while large multi-modal models (LMMs) can reason about these artifacts but are often too computationally demanding for privacy-sensitive or embedded deployments. This work aims to improve robustness and deployability by explicitly decoupling visual perception from downstream reasoning. We propose VLS-RR (Visual–Language–SLM Respiratory Rate) auditor, a three-stage framework where video segments are converted into fused motion signals and summary plots. A vision–language model (VLM) then produces a textual description of rhythm, breath count, and artifacts. Finally, small language models (SLMs) perform chain-of-thought (CoT) auditing over this evidence and the signal-derived RR candidates. We evaluate VLS-RR on a 50-video belt-synchronized dataset spanning diverse breathing patterns, viewpoints, and lighting conditions. Compared with signal-only baselines, VLS-RR reduces segment-level MAE by ≈ 30% and improves R2 from 0.85 to 0.92. It also outperforms numeric-only SLM baselines with an additional ≈ 25% MAE reduction. Finally, on an embedded edge device, compact SLM auditors run with 0.82–1.71 s latency per 6 s window, indicating that decoupling perception and reasoning enables accurate, resource-efficient RR estimation from video.

• 88
  Authors: Presenter), Khaled Solaima
  Title: Benchmarking Early Deterioration Prediction Across Hospital-Rich and MCI-Like Emergency Triage Under Constrained Sensing

  Abstract:

  What happens when a 7.0 earthquake hits a city and emergency responders are triaging hundreds of casualties with no hospital infrastructure, no labs, and no imaging, just the readings on a portable monitor? This is the problem we designed our research around. We introduce a benchmarking framework that tests machine learning models under the kind of severe information constraints that define disaster triage: inputs limited to basic vital signs captured within the first hour of patient contact. Using MIMIC-IV-ED, we compare this field-like vitals-only setting against a hospital-rich baseline, and find that models retain substantial predictive power even with minimal inputs. Structured ablation experiments reveal that respiratory rate and oxygen saturation are the most clinically informative signals, and models exhibit graceful degradation as sensing is progressively reduced, a property essential for real-world deployment reliability. This work provides a reproducible, clinically grounded benchmark for evaluating triage decision-support systems designed to operate across the full spectrum of resource availability, from disaster zones to emergency departments.

• 91
  Authors: Zafira Wasma (Presenter), Emam Hossain, Md Osman Gani
  Title: Evaluating Causal and Machine Learning Models for Type 2 Diabetes Risk Prediction

  Abstract:

  Motivated by the limitations of purely correlational models, this study investigates causal, actionable, and interpretable pathways in Type 2 Diabetes (T2D) progression by integrating machine learning with causal inference. We evaluate advanced machine learning models, including XGBoost, Random Forest and Deep Neural Network, alongside causal methods to improve early T2D risk prediction and better understand underlying disease mechanisms. While conventional predictive models identify risk patterns, this work addresses an important gap by examining causal relationships within a 14-variable clinical dataset. By incorporating a causal component, we analyze how specific variables influence T2D progression. The results show that integrating causal analysis with machine learning yields more robust and actionable insights, supporting a transition from risk prediction toward personalized and precision medicine.

• 66
  Authors: Brandon Ables, Dr. Andrea Kleinsmith
  Title: Exploring Embodied Personal Knowledge Management as Systemic Interactivity

  Abstract:

  We present a novel personal knowledge management (PKM) system composed of tangible cubes, color-linked towers, and full-body transcription interfaces. Rather than foregrounding efficiency and productivity, the system foregrounds slow labor, spatial memory, and multimodal engagement as design virtues. Knowledge is transcribed, folded into transparent cubes, and placed within relational towers that cultivate serendipity, spatial associations, and reflection. This work contributes a provocation that reflective intelligence emerges through materially grounded, playful interaction across body, environment, and idea. Through design artifacts and situated experimentation, we argue for PKM as an extended intelligence ecosystem that values novelty over efficiency.

• **Doctorate Student Award**
  67
  Authors: Saquib Ahmed (presenter), Tejo Gayathri Busireddy, Sanorita Dey
  Title: Enhancing Image Comprehension: The Impact of AI-Generated Explanations on Perception of Altered and Synthetic Media

  Abstract:

  In the digital era, the exponential growth of images and videos on social platforms has transformed how individuals perceive information and form opinions. However, the escalating prevalence of altered and synthetic visuals poses significant challenges to media trust. These altered visuals often mislead viewers, propagate confusion, and distort public perception. Social media algorithms, optimized for engagement, can inadvertently amplify the dissemination of such content, making simple tagging insufficient to distinguish authentic from altered visuals. Contextual explanations present a promising approach by offering audiences deeper insights and encouraging more informed interpretations. In this study, we developed contextual explanations for 15 altered and synthetic images and conducted a user study to evaluate their effectiveness. Our findings show that contextual explanations consistently outperformed non-contextual ones across all evaluated metrics. We also assessed the capability of large language models (LLMs) to generate these explanations for diverse audiences. While LLM-generated explanations were generally comparable to those created by human experts, the models exhibited limitations in conveying intrinsic motivations in complex scenarios. We conclude with a discussion of the design implications and ethical considerations of this work.

• **Doctorate Student Award**
  68
  Authors: Zainab Balogun, Bella Dongarra, Samon Nguyen, Candace Gyimah, and Tera Reynolds
  Title: A Human-centered Design Approach for Supporting Self-management of Hypertonic Chronic Pelvic Pain Syndrome

  Abstract:

  Chronic pelvic pain syndrome (CPPS) affects the quality of life of millions of women globally– up to 25% of females of reproductive age and approximately 15% of all adult women. CPPS often presents as a pelvic floor dysfunction that could occur due to a muscle weakness (Hypotonic) or tightness (Hypertonic). Symptoms are characterized by persistent pain in the pelvic floor muscles and organs for at least 6 months. Patients often experience embarrassing symptoms that impact their quality of life, including intimate relationships, mental health, work productivity, and self-esteem. CPPS treatment is complex, involving a multidisciplinary team of care providers. Depending on the patient’s symptoms, a patient’s team of providers could include a gynecologist, a physiotherapist, and a cognitive behavioral therapist. A key component for a successful treatment is at-home pelvic floor physical therapy exercises that patients manage on their own. However, research reports that patients struggle to adhere to treatment plans. Despite increased interest in women’s health research, especially pelvic health, there is still a level of disparity in the research attention and pelvic-floor self-management technologies available for different conditions associated with CPPS. Particularly, technology solutions have often focused on pelvic floor weakness associated with aging and childbirth than pelvic floor hypertonicity which is often characterized with muscle tightness and pain. Therefore, patients with hypertonic CPPS have less support along their care journeys. This study is aimed at addressing the challenges and breakdown in the self-management journeys of patients with hypertonic CPPS, which includes vulvovaginal disorders such as vaginismus, vulvodynia, and dyspareunia. Our goal is to (1) articulate the design space for technology-supported self-management tools for pelvic floor hypertonicity using a human-centered design approach, and (2) understand patients’ and providers’ needs for and perception of self-management technologies for chronic pelvic pain syndrome.

• **Computing for Social Good Award**
  69
  Authors: Rishav Gupta, Dr Dong Li
  Title: Enabling Affordable and Accessible Blood Pressure Monitoring on Smartphones

  Abstract:

  Cardiovascular disease is the leading cause of death worldwide, responsible for an estimated 17.9–19.8 million deaths each year, or about one-third of all global mortality. Frequent blood pressure (BP) monitoring can support earlier detection of hypertension, more timely therapeutic adjustments, and improved long-term risk management, thereby helping to avert many of these events. However, current BP assessment options each have important limitations. Traditional upper-arm cuffs, while clinically reliable, are bulky, require cuff inflation, and are inconvenient for frequent use in everyday settings. Newer devices such as wrist cuffs and smartwatch-based solutions offer better portability but still depend on dedicated, often expensive hardware that not all users can afford or are willing to purchase. Many of these wearables rely on photoplethysmography (PPG), whose performance can be affected by skin pigmentation, ambient lighting, and motion, potentially reducing accuracy and equity across diverse users. In contrast, smartphones are already ubiquitous, carried almost constantly, and thus provide a compelling platform for accessible BP monitoring without requiring additional cuffs, watches, or specialized sensors. These limitations motivate techniques that impose minimal user burden and operate without additional dedicated hardware. In this work, we present a method that enables frequent BP measurements using only a commodity smartphone. Our method transmits ultrasonic signals from the smartphone speaker and records the resulting acoustic responses with the built-in microphone to interrogate blood flow dynamics from the carotid artery. From these signals, we first estimate blood flow and heart rate on a beat-to-beat basis for each cardiac cycle. By analyzing the temporal offset between characteristic features of the heart rate and flow waveforms, we then compute pulse transit time, which subsequently serves as the basis for BP estimation. We are currently collecting data from 38 healthy controls and 38 hypertensive patients to empirically characterize the inverse relationship between pulse transit time and blood pressure and to estimate both systolic and diastolic pressures using only a commodity smartphone. This project is supported by the UMB ICTR ATIP grant and Google Research Scholar Award.

• **Computing for Social Good Award**
  70
  Authors: ASM Mobarak Hossain
  Title: A Multi-Modal Agentic Framework for Auditing Wheelchair Accessibility

  Abstract:

  For a wheelchair user, a standard blue line on a map is often a broken promise. While platforms like OpenStreetMap (OSM) successfully capture where a path is, they frequently fail to convey how it physically feels to travel on it. This information barrier is problematic for wheelchair users. To solve this issue, we present OmniPath, a system that moves from passive mapping to proactive environmental auditing. Our framework fuses the network topology of OSM with the submeter precision of high-density aerial LiDAR (USGS 3DEP) to create a high-fidelity 3D model of the pedestrian environment. Rather than simply routing a user, our agent virtually traverses the network, analyzing the surface in 0.5 meter increments. It rigorously quantifies physical friction points specifically running slope, cross slope, and vertical discontinuities against ADA compliance standards, calculating a weighted severity score to categorize hazards from “Mild” to “Critical.” To ensure real world reliability, we validated the system against 200 physical ground truth field surveys across the National Mall using stratified random sampling. The framework demonstrated strong diagnostic reliability for high-severity hazards, achieving F1-scores of 0.60 for Severe and 0.58 for critical categories. By automating this micro-scale inspection, OmniPath identifies the “invisible” barriers that standard maps miss, effectively transforming a static dataset into accessibility data source that anticipates accessibility challenges before the user ever leaves home.

• **Doctorate Student Award**
  71
  Authors: Md Alomgeer Hussein (Presenter); Rushali Sreedhar; Yachi Jitendrabhai Patel; Sanika Bishnoi; Samon Nguyen; Tera Reynolds
  Title: Understanding the Perceptions of People from Underserved Communities around Generative Artificial Intelligence for Health Information Work

  Abstract:

  Patients must perform significant health information work to seek, understand, organize, share, and use health information to manage their health and care. Generative artificial intelligence (genAI) tools, including ChatGPT and Gemini, are widely available and increasingly used for health information work, such as translating complex medical language and explaining laboratory test results. Yet existing human-centered genAI research has tended to focus on convenience samples of participants that are younger, more educated, and White. Understanding diverse perspectives is critical to ensure the digital divide does not widen and that genAI tools meet diverse needs. Towards this end, we examined how people across age groups and living in low-income neighborhoods in Baltimore, Maryland understand, use, and evaluate genAI for health information work. In partnership with two community-based organizations, we conducted five key informant interviews and six focus groups with 37 participants. Focus groups were conducted by age group, with two sessions for ages 18-29, 30-49, and 50 and older. All sessions were audio recorded, transcribed, and analyzed using inductive thematic analysis. Overall, we found that participants’ current use and perceptions were critical to their opinions on the potential of genAI for health information work, but perceptions were malleable even with low-levels of interaction around genAI. We identified four key themes: (1) limited awareness and prior experience with genAI often lead to initial skepticism and concerns related to trust and privacy; (2) low-level AI and social interactions increased interest in and comfort with genAI; (3) recognition of genAI’s potential for translating and simplifying complex health information and intention to use in the future; and (4) limited effect of information on concerns about genAI (e.g., hallucinations). Based on our results, we make a number of recommendations for health educators and navigators such as providing opportunities for guided interactions with genAI.

• **Undergraduate Student Award**
  72
  Authors: Kodilinye Mkpasi (Presenter); Corey Benjamin; Dong Li; Anuradha Ravi; Nirmalya Roy; Mary Beth Aichelmann-Reidy
  Title: “O-HygieCare”: Predictive Oral Health Monitoring through Longitudinal Toothbrushing Analytics using Smart Wearable Intelligence

  Abstract:

  Maintaining optimal oral hygiene is critical across all age groups, yet individuals often neglect certain areas, apply excessive force, or disproportionately focus on specific regions while brushing. Such suboptimal behaviors can lead to plaque accumulation, enamel wear, gingival inflammation, and progressive periodontal conditions. Without timely clinical evaluation, these conditions may worsen and result in irreversible damage. At the same time, dental professionals face challenges in assessing treatment effectiveness between scheduled visits, as they lack continuous insight into patients’ daily brushing practices and disease progression. This research seeks to augment dental care with AI-driven tools that monitor routine brushing techniques and support a personalized modeling framework for adaptive, real-time oral health assessment. By integrating longitudinal brushing analytics with automated plaque and gingival index prediction, the system aims to provide clinicians with actionable insights to evaluate treatment efficacy, detect early signs of disease progression, and deliver more informed, data-driven patient care. With the widespread adoption of smartwatches, there is significant potential to use these devices for monitoring toothbrushing behavior and technique. This research explores the use of consumer-grade smartwatches to classify brushing motions (horizontal, vertical, circular) and dental regions (left, middle, right, upper, lower) using IMU and audio data. Key challenges include unequal sampling rates, missing data, sensor drift during rapid motion, synchronization of video annotations with multimodal streams, and variability across smartwatch platforms. Linear interpolation was applied to align heterogeneous signals, and built-in orientation filtering supported stable angle estimation. Prior generic models suffered from inter-user variability in brushing speed and biomechanics, achieving only ~60% accuracy. To improve performance, we propose a two-phase personalized framework combining lightweight neural models (MLP, LSTM, CNN) with user-specific adaptation layers. A global model is trained on pooled data and fine-tuned per user while freezing shared feature extractors, enabling accurate and edge-deployable inference. Custom applications were developed for synchronized multimodal data collection and annotation.

• **Doctorate Student Award**
  73
  Authors: Golnaz Moharrer (Presenter) , Krystal Zhang, Andrea Kleinsmith
  Title: Making Without Purpose, Growing With Intention: Embodied Creative Reflection for Emotional Resilience in Graduate Students

  Abstract:

  Graduate students often experience emotional stress and report a lack of supportive outlets for processing these challenges. Prior work indicates that combining mindfulness practices such as reflection with visual art practices can alleviate stress and reduce mental pressure among graduate students. However, current digital self-reflection tools rarely incorporate creative and/or embodied modalities such as gesture, drawing, or tactile making. This study explored how creative “useless” tactile artifact-making impacts graduate  students’ emotional well-being, self-perceived creativity, and personal growth. Qualitative findings revealed that the creative task was a source of calmness, joy, and accomplishment by removing goal pressure. Students in more technical fields generally preferred non-digital, hands-on, and embodied forms of engaging in creative self-reflection. Quantitative results revealed a significant reduction in negative affect after engaging in the task. This work contributes to C&C and HCI by highlighting open-ended, embodied creativity as a  compelling approach to fostering emotional resilience and identity development in academic contexts.

• **Doctorate Student Award**
  74
  Authors: Qi Zhao, Marjory Pineda, Ketul Kishorbhai Chhaya, Yasmine Kotturi
  Title: AI Literacy in Context: Exploring What AI Literacy Means for Entrepreneurs

  Abstract:

  Entrepreneurs are increasingly adopting AI tools for various business functions, although some leverage these tools more effectively than others. Differences in how entrepreneurs engage with and effectively use AI require an understanding of AI literacy and how to apply such knowledge. However, current approaches to AI literacy tend to assume generalizable competencies (e.g. prompt engineering, knowing capacities and limitations of AI) and user groups (e.g. students and expert users), and do not account for unique contexts such as the small business context. Therefore, through the community-centered design, use, and evaluation of our LLM-powered business planning tool, BizChat, we build on prior work on AI literacy by investigating the existing AI competencies that are relevant to entrepreneurs, as well as the additional competencies not previously explored that are specific to effective AI use within the small business context. Through participatory design workshops, log data, and semi-structured interviews, our findings demonstrate that recognizing what systems use AI, assessing AI outputs, understanding the role of data in AI, and understanding the ethical and legal aspects of AI are skills particularly relevant to entrepreneurs. Furthermore, competencies such as knowledge of collaborative support related to AI, communicating AI-generated content to an audience, and having an understanding of one’s temporal engagement with AI are additional AI competencies that show in the small business context. Our work suggests that competency-based AI literacy frameworks must be constantly evaluated in specific contexts to avoid the risk of being quickly deprecated as the concept of AI literacy continues to evolve.

• 75
  Authors: Kavya Rajendran (Presenter), Dr. Andrea Kleinsmith
  Title: The Emotional Side of Reading News through Social Media for International Students in the US

  Abstract:

  The objective of this research is to investigate how international students studying in the United States consume news on social media and how it affects their emotional well-being. With social media reaching millions of users in turbulent times, it has provided a platform for international students to consistently stay in touch with their families, friends, and engage in other social interactions to find representation and support alleviating depression and anxiety. However, studies also show that the relationship between “doomscrolling” and overall well-being is mediated by psychological distress from news-related stressors encountered in social media. Current studies lack a qualitative representation of the perspectives and emotional impact of social media news on international students. With the numerous changes to immigration procedures that directly or indirectly affect such a large population of international students, it is important to understand how their online news habits impact their emotional state which can also affect progression toward their degree. The poster will outline the results along the topics of their online news reading activity, how social media helps or does not help them in everyday life, how they negotiate their identity as an international student in social media and how this affects their emotions and overall well-being. The discussion will include design implications for social media platforms for improving the emotional wellbeing for users from different cultures.

• **Doctorate Student Award**
  65
  Authors: Sruthi Sundharram (Presenter), Jake Whitt, Golnaz Moharrer, Andrea Kleinsmith, Charissa Cheah, Christine Mallinson, Ramana Vinjamuri
  Title: COHERE: Collaborative Optimization of Human Engagement and Robot Effectiveness

  Abstract:

  COHERE: Collaborative Optimization of Human Engagement and Robot Effectiveness advances human-robot collaboration (HRC) by deriving engineering design principles from human-human teaming and validating them through real-time multimodal robotic systems. Despite rapid advances in robotics, real-world HRC remains limited by challenges in safety, trust, communication, and adaptability. COHERE addresses these barriers by integrating behavioral neuroscience, affective computing, brain-computer interfaces (BCIs), speech interfaces, and robotics within an innovative, classroom-based research and training framework. The project (Aug 2025–July 2026) will implement a three-module special topics course. Module 1 examines human-human collaboration through experiential learning activities while noninvasively recording neural and behavioral signals to characterize cognitive load, affect, and coordination dynamics. Module 2 translates these neurobehavioral insights into actionable engineering principles such as modularity, resilience, adaptability, and multimodal communication to guide collaborative system design. Module 3 applies and evaluates these principles in human-robot interaction tasks. As a proof-of-concept platform, we have developed a real-time Simulink-based multimodal BCI-robotics system using the g.tec Unicorn EEG headset. Neural signals are streamed and processed in real time to detect motor-related cues (e.g., stomping feet), triggering a virtual robotic arm to grasp a ball. In parallel, a speech-to-text module processes auditory commands (e.g., “happy”) to control a second virtual robotic arm. This dual neural, verbal control architecture enables simultaneous, real-time multimodal interaction and serves as a testbed for studying trust, workload, and collaboration fluency. COHERE will generate annotated multimodal datasets, adaptive control algorithms, and a scalable educational framework, advancing the science of collaborative intelligence and human-centered robotic systems.

• 76
  Authors: Krystal Zhang (presenter); Marie Sakowicz; Emily Wingeart; Faculty mentor: Foad Hamidi
  Title: Belonging in the Making: Investigating Inclusive Makerspace Design for Youth with Autism

  Abstract:

  Makerspaces offer opportunities for creative and technology-rich learning, yet their design often overlooks the needs of youth with autism. While inclusive education research has emphasized structured teaching and multimodal engagement, less is known about how spatial, sensory, and cultural factors intersect with pedagogy in makerspace contexts. To address this gap, we conducted interviews with special education staff, including teachers, assistants, and therapists, alongside focus groups with educators and architects, and a panel with disability advocates and special education experts on accessibility in makerspaces and informal learning. Our analysis identified six themes highlighting strategies for predictability, multimodal engagement, and cultural resonance to create inclusive makerspaces. These findings position makerspaces not only as sites for technical skill development but also as infrastructures of belonging and empowerment. Based on our findings, we offer guidance for educators, designers, and policymakers seeking to create accessible, community-engaged spaces that support the participation of youth with autism.

• 103
  Authors: Bahirah Adewunmi (Presenter), Sanjay Purushotham
  Title: TIE Grammar: Stratified Generative Pipeline for Training Generalizable Deep Reinforcement Learning Cyber Agents

  Abstract:

  Developing robust Deep Reinforcement Learning (DRL) agents for cyber security is fundamentally hampered by their inability to generalize due to the non-Independent and Identically Distributed (non-I.I.D.) nature of network topologies, a scarcity of diverse training environments, and the challenge of modeling complex attack physics. To address this, we introduce the Parametric Graph Grammar for Procedural Content Generation (PGG-PCG), a novel stochastic generative model designed to reconstruct realistic network structures. Our approach leverages a parametric L-System formalism, $\mathbf{G}=(V,\Sigma,R,S)$, where the device type alphabet $\Sigma$ and probabilistic production rules $R$ are empirically derived from real-world network data (LANL dataset). This architecture is implemented within a Stratified Generative Pipeline (SGP) that ensures topological validity (Phase I) by synthesizing realistic, structurally valid enterprise topologies (e.g., Fat-Tree, Spine-Leaf) using device adjacency probabilities, thereby avoiding the common issue of structurally invalid samples in pure generative models. We further ensure semantic integrity (Phase II) by injecting device capabilities via Stratified Sampling, and critically, guarantee adversarial realism (Phase III) by integrating the Technique Inference Engine (TIE) as the state-transition oracle. This delegates the complex, non-Markovian attack physics—a major limitation for standard DRL approaches—to a CTI-validated probabilistic recommender ($\text{NDCG@20}=0.18\pm 0.02$). The PGG-PCG framework successfully produces inexhaustible, topologically, and semantically valid training environments, which is the necessary foundation for developing generalizable and robust DRL agents capable of autonomous network defense.

• **Doctorate Student Award**
  104
  Authors: Bahirah Adewunmi (Presenter), Ed Raff, Sanjay Purushotham
  Title: SubstratumGraphEnv: Reinforcement Learning Environment (RLE) for Modeling System Attack Path

  Abstract:

  Automating network security analysis, particularly the identification of potential attack paths, presents significant challenges. Due in part to the sequential, interconnected, and evolutionary nature of system events which most artificial intelligence (AI) techniques struggle to model effectively. This paper proposes a Reinforcement Learning (RL) environment generation framework that simulates the sequence of processes executed on a Windows operating system, enabling dynamic modeling of malicious processes on a system. This methodology models operating system state and transitions using a graph representation. This graph is derived from open-source System Monitor (Sysmon) logs. To address the variety in system event types, fields, and log formats, a mechanism was developed to capture and model parent-child processes from Sysmon logs. A Gymnasium environment (\emph{SubstratumGraphEnv}) was constructed to establish the perceptible basis for an RL environment, and a customized PyTorch interface was also built (\emph{SubstratumBridge}) to translate Gymnasium graphs into Deep Reinforcement Learning (DRL) observations and discrete actions. Graph Convolutional Networks (GCNs) concretize the graph’s local and global state, which feed the distinct policy and critic heads of an Advantage Actor-Critic (A2C) model. This work’s central contribution lies in the design of a novel deep graphical RL environment that automates translation of sequential user and system events, furnishing crucial context for cybersecurity analysis. This work provides a foundation for future research into shaping training parameters and advanced reward shaping, while also offering insight into which system events attributes are critical to training autonomous RL agents

• 105
  Authors: Zeliatu Ahmed (Presenter), Faculty Mentor: Dr. Karuna Pande Joshi
  Title: Ontology Driven Agentic System for Automating Security Compliance in Medical Cyber-Physical WBANs

  Abstract:

  Medical Wireless Body Area Networks (WBANs) operate as medical cyber-physical systems (MCPS) that continuously sense and transmit sensitive patient health data, creating significant security, privacy, and regulatory challenges. Traditional WBAN protections focus on lightweight cryptography but lack semantic reasoning and automated compliance of data regulations such as Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR). This paper presents a novel agentic framework that automates the modeling, analysis, and verification of security, privacy, and compliance properties in cyber-physical wireless body area network (WBAN-CPS) medical systems. The framework represents medical devices, data flows, threats, controls, and regulatory obligations in OWL and operationalizes reasoning through SWRL rules and SPARQL-based compliance queries. A synthetic WBAN-CPS dataset instantiates the ontology, enabling autonomous inference across cyber, physical, and regulatory layers. The agentic architecture comprising a Risk Detection Agent, Compliance Monitoring Agent, and Policy Enforcement Agent collaboratively infers SecurityRisk, PrivacyRisk, ComplianceRisk, and CPS-specific SafetyRisk in a representative medical use case. Evaluation results show that the system provides high expressiveness, regulatory adaptability, and explainability, outperforming traditional WBAN security models. This work establishes a semantic and regulation aware foundation for trustworthy and secure medical cyber-physical systems.

• **Undergraduate Student Award**
  18
  Authors: Jalen Brown, Pearce Packman, Roberto Yus
  Title: Privacy-Preserving Indoor Navigation for the UMBC TRC Building

  Abstract:

  Navigating large, multi-floor buildings remains difficult for visitors and new members, yet many indoor navigation systems depend on expensive proprietary infrastructure or GPS, which is impractical and inaccurate inside concrete-and-steel structures. This student-led project designs, implements, and deploys an indoor navigation system that balances cost, usability, and privacy. We developed a mobile wayfinding application for UMBC’s Technology Research Center (TRC) to guide users to offices and labs within the Computer Science and Electrical Engineering (CSEE) department. After evaluating multiple sensing approaches, the team selected a low-cost architecture based on passive Bluetooth Low Energy (BLE) beacons and conducted an empirical study to determine the minimum beacon density needed to maximize coverage and navigation utility. We installed 50 BLE beacons across corridors, stairwells, and key intersections to support location estimation without collecting personal identifiers—providing a privacy advantage over conventional alternatives. The application, built with TypeScript, React Native, and Expo, renders a real-time, color-coded multi-floor map and computes optimal routes using an A* pathfinding algorithm over a predefined navigational skeleton of walkable paths. Building map data is processed in QGIS and ingested into a self-hosted Supabase backend through a custom ETL pipeline. Ongoing work focuses on improving localization robustness, routing behavior, user experience, and backend security.

• 106
  Authors: Deontic Knowledge Graphs for Privacy Compliance in Multimodal Disaster Data Sharing
  Title: Presenter: Kelvin Uzoma Echenim, Faculty Mentor: Karuna Pande Joshi

  Abstract:

  Disaster response requires sharing heterogeneous artifacts, from tabular assistance records to UAS imagery, under overlapping privacy mandates. Operational systems often reduce compliance to binary access control, which is brittle in time-critical workflows. We present a novel deontic knowledge graph-based framework that integrates a Disaster Management Knowledge Graph (DKG) with a Policy Knowledge Graph (PKG) derived from IoT-Reg and FEMA/DHS privacy drivers. Our release decision function supports three outcomes: Allow, Block, and Allow-with-Transform. The latter binds obligations to transforms and verifies post-transform compliance via provenance-linked derived artifacts; blocked requests are logged as semantic privacy incidents. Evaluation on a 5.1M-triple DKG with 316K images shows exact-match decision correctness, sub-second per-decision latency, and interactive query performance across both single-graph and federated workloads.

• **Doctorate Student Award**
  108
  Authors: Yuechun Gu, Jiajie He, Keke Chen
  Title: Auditing Approximate Machine Unlearning for Differentially Private Models

  Abstract:

  Approximate machine unlearning aims to remove the effect of specific data from trained models to ensure individuals’ privacy. Existing methods focus on the removed records and assume the retained ones are unaffected. However, recent studies on the privacy onion effect indicate this assumption might be incorrect. Especially when the model is differentially private, no study has explored whether the retained ones still meet the differential privacy (DP) criterion under existing machine unlearning methods. This paper takes a holistic approach to auditing both unlearned and retained samples’ privacy risks after applying approximate unlearning algorithms. We propose the privacy criteria for unlearned and retained samples, respectively, based on the perspectives of DP and membership inference attacks (MIAs). To make the auditing process more practical, we also develop an efficient MIA, A-LiRA, utilizing data augmentation to reduce the cost of shadow model training. Our experimental findings indicate that existing approximate machine unlearning algorithms may inadvertently compromise the privacy of retained samples for differentially private models, and we need differentially private unlearning algorithms.

• 109
  Authors: Jiajie He, Min-chun Chen, Xintong Chen, Xinyang Fang, Yuechun Gu, Keke Chen
  Title: ICL-RecSys Privacy Risk

  Abstract:

  Large language models (LLMs) based recommender systems (RecSys) can adapt flexibly across different domains. It uses in-context learning (ICL), i.e., prompts, including sensitive historical user-specific item interactions, to customize the recommendation functions. However, no study has examined whether such private information may be exposed by novel privacy attacks. We design several membership inference attacks (MIAs): Similarity, Memorization, Inquiry, and Poisoning attacks, aiming to reveal whether system prompts include victims’ historical interactions. We have carefully evaluated them on the latest open-source LLMs and three well-known RecSys datasets. The results confirm that the MIA threat to LLM RecSys is realistic, and that existing promptbased defense methods may be insufficient to protect against these attacks.

• 22
  Authors: Dharani Nadendla, Renzhi Hao, Shirui Cao, Yizhu Wen, Rishav Gupta, Mehran Kafai, Hanqing Guo, Dong Li
  Title: Ultrasound Watermark: Real-time Acoustic Watermarking for Voice Scam Protection on Smartphones

  Abstract:

  With the rapid rise of AI-generated voice clones and sophisticated phone scams, confirming the authenticity of a live caller has become an urgent and largely unsolved challenge. This project presents Ultrasound watermark, a real-time caller verification framework that runs entirely on commodity smartphones, requiring no specialized hardware and no interruption to the natural flow of a phone call. During a call, the caller’s smartphone continuously emits inaudible ultrasonic signals from its speakers and captures the reflections of the natural movements of the mouth, lips, and jaw during speech using the smartphone’s microphone. These physiological motion signatures are seamlessly fused with the user speech to generate a dynamic watermark that is invisible to the listener yet deeply tied to the caller’s live physical presence. On the receiving end, a lightweight on-device detector verifies that the incoming audio is genuinely synchronized with real articulatory motion, a check that replayed recordings and AI-synthesized voices fundamentally cannot pass. The system achieves an AUC of 0.95 across diverse attack scenarios like replaying and mimicking, maintains near-transparent audio quality with a PESQ score of 3.99, and introduces less than 200 ms of latency. The system is deployed on both iOS and Android, and delivers a practical, user-transparent defense against the growing threat of voice phishing and AI-powered impersonation.

• 107
  Authors: Hadjar Ould Slimane (presenter), Mohamed Younis, Yousef Ebrahimi
  Title: AI-based Traffic Analysis Attack Mechanism for IoT Systems
  Abstract:

  With the massive growth of wireless service comes the concern about location privacy. Even with the use of pseudonyms and anonymous communication techniques, the threat of traffic analysis continues to be worrisome. Particularly in the context of Internet of Things (IoT) applications, data packets often flow towards a certain user node or a Base Station (BS), which makes such a recipient a sink in the network and elevates its susceptibility to cyberattacks. Existing countermeasures assume that the attacker will employ one of the contemporary analysis engines, such as Evidence theory and transmission rate, as the underlying data correlation mechanism. This paper presents a novel traffic analysis model that is based on Convolutional Neural Networks (CNN). We demonstrate that such a new attack mechanism is so powerful that it could overcome conventional countermeasures and hence warrants extensive investigation to devise protective measures.

• **Undergraduate Student Award**
  110
  Authors: Sejal Patil (presenter), Reece Robertson, Sebastian Deffner
  Title: Secure Quantum Handshakes: Simon’s Algorithm for Quantum Network Verification

  Abstract:

  Quantum networking represents the future of secure communication, enabling fundamentally new capabilities through quantum information exchange. However, these networks require every participating node to be a genuine quantum device, and there is currently no reliable method to verify whether a client entering a quantum network truly possesses quantum computational capability. In this work, we introduce a novel client-validation algorithm inspired by Simon’s algorithm. Building on its exponential quantum advantage, we design a verification protocol that functions as a proof-of-quantum-capability handshake. The algorithm requires a client to solve a structured hidden-period problem that is computationally intractable for classical systems — namely solving Simon’s problem in non-exponential time. We benchmark our protocol on NISQ trapped-ion and superconducting hardware under realistic noise conditions. Our results demonstrate that the algorithm successfully distinguishes quantum devices from classical systems attempting to access the network.

• 6
  Authors: Shaswati Saha, Sourajit Saha, Manas Gaur, Tejas Gokhale
  Title: Side Effects of Erasing Concepts from Diffusion Models

  Abstract:

  Concerns about text-to-image (T2I) generative models infringing on privacy, copyright, and safety have led to the development of concept erasure techniques (CETs). The goal of an effective CET is to prohibit the generation of undesired “target” concepts specified by the user, while preserving the ability to synthesize high-quality images of other concepts. In this work, we demonstrate that concept erasure has side effects and CETs can be easily circumvented. For a comprehensive measurement of the robustness of CETs, we present the Side Effect Evaluation (SEE) benchmark that consists of hierarchical and compositional prompts describing objects and their attributes. The dataset and an automated evaluation pipeline quantify side effects of CETs across three aspects: impact on neighboring concepts, evasion of targets, and attribute leakage. Our experiments reveal that CETs can be circumvented by using superclass-subclass hierarchy, semantically similar prompts, and compositional variants of the target. We show that CETs suffer from attribute leakage and a counterintuitive phenomenon of attention concentration or dispersal.

• 21
  Authors: Gaurav Sharma, Milton Halem, Yaacov Yesha
  Title: A Digital Twin for the Safety of Smart Ports

  Abstract:

  Modern container ports are complex systems where throughput depends on vessel arrivals, crane productivity, gate capacity, labor availability, and external disruptions. Small operational constraints can escalate into large scale congestion, making accurate forecasting and stress testing essential for resilience planning. This work presents a data driven deep learning digital twin framework for forecasting container throughput and simulating operational disruptions at the Port of Los Angeles. A long term historical extension of synthetic Daily dataset of vessel records from 1995 through 2025, required for training a Temporal Fusion Transformer (TFT) neural net, is generated from Monthly TEU throughput data available from 1995 to 2025 records trained on limited overlapping Daily data beginning from January 2020– 2025 employing the Gemini 12B large language model to reconstruct operational patterns. Operational variables include vessel arrival times, crane productivity, gate throughput, dwell time and yard utilization indices. TFT model captures multivariate temporal dependencies across the infrastructure capacity and generates weekly forecasts for a 3-month horizon and monthly forecasts for a 12-month horizon. For validation, during 2024 the model predicted an annual throughput range of 8.35 to 8.85 million TEUs, closely capturing the observed 8.63 million TEUs handled by the port and accurately tracking seasonal import peaks between Q4 2023 and Q1 2024. The digital twin is currently being extended with a knowledge graph-based representation to enable scenario conditioned forecasting and quantify TEU throughput degradation under infrastructure failures or demand shocks, supporting data driven analysis in maritime logistics systems.

• 7
  Authors: Javed Tamboli, Karuna Joshi
  Title: Security Compliance for Smart Manufacturing using Knowledgegraph based Digital Twin

  Abstract:

  The combination of Information Technology (IT) and Operational Technology (OT) in smart manufacturing, driven by smart factory innovations and Internet of Things (IoT) devices, generates vast, diverse, and rapidly evolving Big Data, which in turn increases cybersecurity and compliance issues. Adherence to security standards, such as NIST SP 800-171, which requires rigorous access control and audit reporting, is currently obstructed by the resource-intensive and error-prone aspects of manual evaluations. We have developed a semantically rich knowledge graph-based digital twin to automate security compliance of the smart assembly line, specifically focusing on categories specified in NIST SP 800-171. We have used Semantic Web technologies like RDF, OWL, and SPARQL using the Jena Fuseki server to build our system. Our approach improves data integrity and structure identification in IT/OT systems by tackling the Big Data 5Vs. The qualitative assessment of our digital twin shows a scalable approach with reduced compliance violations and enhanced audit effectiveness. In this paper, we describe our design in detail along with the validation results. This study propels future investigations by integrating Knowledge graphs and reasoning with industrial compliance, establishing a basis for automated compliance in smart manufacturing

• **Doctorate Student Award**
  13
  Authors: Wenkai Tan, Jayaprakash B. Shivakumar, Rishikesh S. Govindarajan, Nicholas Reed, Daewon Kim, Eduardo Rojas, Yongxin Liu, Huihui Wang, Houbing Song
  Title: Anomaly Detection for Additive Manufacturing with Antenna-Based Embedding Sensors: An Explainable, Instructive Learning Framework

  Abstract:

  Deep learning models have demonstrated pattern recognition capabilities in additive manufacturing defect detection, yet their reliance on labeled defect data and opaque decision boundaries hinder deployment in quality-critical production environments. To address this issue, we propose a zero-bias neural framework where deep networks provide latent feature extraction while geometric statistical constraints ensure interpretable anomaly detection, advancing trustworthy manufacturing intelligence. The framework formalizes anomaly detection by computing cosine similarity between latent features and learned class templates in a zero-bias layer, then applying Mahalanobis distance thresholds to quantify deviations from normal distributions, thereby establishing verifiable statistical decision boundaries with geometric interpretability anchored to frequency-domain antenna signal characteristics. We demonstrate that the framework enables threshold-based sensitivity adjustment without retraining and supports visual latent space analysis for transparent reasoning, operating exclusively on normal-class sensor data without requiring defect labels.

• **Doctorate Student Award**
  14
  Authors: Zahid Hassan Tushar (presenter); Sanjay Purushotham
  Title: Security of Biomedical Large Language Models: Threats, Defenses, and Open Challenges

  Abstract:

  Large language models (LLMs) are increasingly deployed in biomedical and clinical settings, where failures can compromise patient safety, privacy, and trust. Unlike general purpose applications, biomedical LLMs operate under strict regulatory constraints and interact with complex systems such as electronic health records, retrieval pipelines, and clinical workflows. This paper systematizes the security of biomedical LLMs through a system level analysis of threats and defenses across the model lifecycle. We introduce a unified taxonomy of attacks grounded in four security objectives: privacy, safety, integrity, and reliability, and characterize adversarial capabilities ranging from black box users to insider and deployment layer attackers. We analyze how vulnerabilities arise during training, alignment, inference, and deployment, and map existing defenses to the attack surfaces they mitigate. Our analysis shows that many high impact attacks, including prompt based PHI extraction, adversarial persuasion, and retrieval poisoning, are immediately deployable in real world systems, while effective defenses remain fragmented and costly. We conclude that securing biomedical LLMs is an ongoing systems security challenge that requires defense in depth strategies, inference time verification, and continuous adversarial evaluation rather than one time alignment.

• 17
  Authors: Afia Zuhaira (Presenter) , Mohamed Younis
  Title: PUF-Based Reconfigurable QAM Modulation for Secure Communication

  Abstract:

  Cyberattacks pose persistent threats to wireless communication, especially in low-power IoT devices where conventional cryptography is impractical. This work leverages Physical Unclonable Functions (PUFs) to deterministically configure a session-specific modulation profile. For each session, the PUF response to a challenge dynamically determines both the modulation order, and also permuting the bit-to-symbol mapping to create a unique, obfuscated QAM constellation. This ensures that each communication instance is uniquely encoded and obfuscated at the physical layer. The receiver uses the same challenge-response pair to infer the session specific mapping and to demodulate the signal. This dynamic constellation generation thwarts adversaries by preventing pattern reuse and rendering modulation recognition by machine learning models ineffective. Importantly, the session-specific and device-dependent nature of the scheme ensures that even if a node is compromised, other nodes remain secure. The approach operates entirely at the signal level and introduces minimal computational burden for IoT and embedded communication platforms.

• **Doctorate Student Award**
  99
  Authors: Mst Maksuda Bilkis Baby (Presenter), Khushika Shah, Naiyue Liang Emma, Faculty Mentor: Lei Zhang
  Title: Separating Secrets from Placeholders: A Hybrid CNN-CodeBERT Framework for Three-Class Credential Leakage Detection

  Abstract:

  Credential leakage in public source code repositories poses a critical security threat, with over 23.8 million secrets exposed in 2024 alone. Existing detection tools suffer from high false-positive rates because rigid pattern matching and binary classification schemes fail to distinguish genuine credentials from placeholder or weak credentials. We propose a three-class classification framework that explicitly models placeholder or weak credentials as a distinct class, leveraging CodeBERT-based semantic understanding combined with character-level pattern recognition. We evaluate our approach on a newly constructed dataset of 9,426 samples spanning 10 programming languages. Our model achieves a Matthews Correlation Coefficient of 0.86 and a macro F1-score of 0.90, achieving 93% recall and 89% precision for genuine credential leaks while reducing high-severity alerts by 33.0% (from 373 to 250) without sacrificing security coverage. Compared to prior character-level approaches, our method improves placeholder or weak credential detection from 54% to 81% F1-score while maintaining strong cross-language generalization, with 9 of 10 languages achieving F1 above 0.80 under leave-one-language-out evaluation.

• **Doctorate Student Award**
  98
  Authors: Venkat Ganapathy, Samin Semsar, Evelyn Kempe, Aaron Massey, Sai Nikhil Reddy Pendhyala, Carolyn Seaman, Sreedevi Sampath
  Title: Managing regulatory ambiguities within SDLC

  Abstract:

  Unresolved ambiguities in regulatory requirements can undermine compliance & audit, traceability, and accountability in regulated software domains. This research demonstrates that effective ambiguity management is essential for compliance & audit readiness and persistent regulatory orientation across all phases of the software development life cycle (SDLC). Without organized documentation and traceable resolution of regulatory compliance, organizations risk misinterpretation, unreliable software implementation, and impaired evidence of due diligence. Ambiguity modeling provides a systematic method to document and analyze ambiguities embedded within regulatory texts. The Ambiguity Heuristics Analysis Builder (AHAB) supports this process by enabling developers to build Regulatory Ambiguity Models (RAM) that capture, trace, and administer regulatory uncertainty throughout development. By embedding ambiguities methodically within the RAM, AHAB allows for interpretation of ambiguities to be documented by software development stakeholders. To enhance traceability from regulatory requirements to software development and execution artifacts, we recently introduced new features that expand AHAB’s capabilities. These include: (1) enhanced logging of imported models and integration with ongoing development logs; (2) a dedicated user interface component for recording rationales behind ambiguity resolution decisions, supported by a JSON diff tool to track changes over time; and (3) “Linked Artifacts,” which associate ambiguity resolutions with related software artifacts such as solution design, Git repositories, test cases, issue trackers, and maintenance documentation. These features reinforce coordinated decision-making, preserve proof of compliance, and confirm demonstrable due diligence in regulated environments. The study evaluates these features through focus groups that include students and software professionals with SDLC background as AHAB users, assessing the effectiveness of these new features in improving accountability, enhancing the software development processes, and representing defensible compliance & audit.

• 97
  Authors: Ainaz Jamshidi (Presenter), Dongchan Kim, Lei Zhang
  Title: From Vulnerabilities to Weaknesses: A Retrieval-Based Siamese Transformer Approach

  Abstract:

  Reliable alignment between vulnerability disclosures and standardized weakness taxonomies underpins many core security workflows, from triage and prioritization to automated risk assessment and large-scale vulnerability analytics. Yet, the construction of CVE–CWE mappings remains largely manual, which limits scalability and introduces variability in quality. Prior automated approaches predominantly frame this task as closed-set classification or shallow similarity matching, which struggle to adapt to evolving taxonomies and are often confounded by semantically similar weakness categories and subtle sources of label leakage. In this paper, we reformulate CVE–CWE mapping as a leakage-aware semantic retrieval problem and propose a siamese transformer-based bi-encoder that learns a shared embedding space for vulnerability descriptions and canonical CWE definitions. We introduce a two-phase training strategy that combines in-batch negatives with hard-negative mining to explicitly improve discrimination among closely related CWE classes. At inference time, CWE prediction is performed via cosine-similarity-based retrieval over precomputed CWE embeddings. We evaluate our approach on the 2025 CVE–CWE corpus and obtain strong retrieval performance averaged over five random seeds with the average of 85.43%, 93.73% and 96.20% for Recall@1,5,10, respectively. Compared to the only cosine-similarity description-embedding baseline, our model yields large gains, improving top-1 accuracy by +52.41 points and top-5 by +28.35 points. In the future, we plan to extend this framework by introducing hierarchy-aware evaluation metrics, expanding the dataset to broader vulnerability corpora, and improving performance on rare and underrepresented CWE classes.

• 96
  Authors: Dr. Muhammad Ali Yousuf, Anjali Jha
  Title: A-Eye: Empowering Independence in Individuals with Color Vision Deficiency through Generative AI-Assisted Dietary and Daily Navigation

  Abstract:

  Color Vision Deficiency (CVD) significantly limits visual accessibility in daily tasks, particularly in environments where color indicates safety, such as food preparation, shade identification, pairing clothes and traffic navigation. Existing assistive solutions often rely on expensive specialized optical hardware or human-in-the-loop volunteer applications, which can compromise user independence and privacy. This research introduces “A-Eye,” a software-first accessibility application designed to support individuals across all six primary forms of CVD (protanomaly, deuteranomaly, tritanomaly, protanopia, deuteranopia, and tritanopia). Utilizing advanced Vision Generative AI models, the initial phase of A-Eye features a “Food Buddy” module that processes image inputs from standard smartphones or integrated smart glasses to accurately assess food ripeness, spoilage, and cooking status. To overcome common computer vision challenges associated with variable ambient lighting, the system integrates automated lighting correction and contextual prompt engineering. Analysis results are delivered instantly via customizable audio cues and text, providing a discreet, judgment-free user experience. By eliminating the reliance on human volunteers, A-Eye fosters true autonomy for the user. Future work explores the integration of A-Eye with vehicular dashcams to provide real-time traffic signal identification, further expanding the application’s capacity to seamlessly integrate individuals with CVD into everyday environments.

• 93
  Authors: Peiying Liu (Department of Diagnostic Radiology & Nuclear Medicine, University of Maryland School of Medicine, Baltimore, Maryland, USA); Elham Karimigharighi (Department of Diagnostic Radiology & Nuclear Medicine, University of Maryland School of Medicine, Baltimore, Maryland, USA and Department of Computer Science and Electrical Engineering, University of Maryland, Baltimore County, Baltimore, Maryland, USA); Hanzhang Lu (Department of Radiology, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA); Fariba Badrzadeh (Department of Diagnostic Radiology & Nuclear Medicine, University of Maryland School of Medicine, Baltimore, Maryland, USA)
  Title: Normalized cerebrovascular reactivity mapping using hypercapnia and hyperoxia challenges

  Abstract:

  Cerebrovascular reactivity (CVR), an index of cerebral vessel’s capacity to dilate, is an important biomarker of the brain’s vascular function. CVR mapping is typically performed using BOLD MRI with blood CO2 maneuvers. However, BOLD-based CVR mapping techniques are susceptible to multiple confounding factors such as vessel density (more veins, more BOLD signal) and hematocrit levels. Here a normalized CVR mapping approach was developed using BOLD MRI with hypercapnia and hyperoxia challenges. This technique utilizes the BOLD-response to O2 to normalize the BOLD-response to CO2 and estimate a “CBF-CVR” which is directly related to the brain’s microvascular function. The BOLD fMRI signal is modeled as ΔBOLD/BOLD0=M·(1−(CBF/CBF0)^α·([dHb]/[dHb]0)^β), where M=TE·A·vCBV0·(c_ery·Hct)^β·(1−Yv,0)^β includes imaging parameters (e.g., TE) and baseline physiological parameters (e.g., venous oxygenation (Yv), vCBV and hematocrit (Hct)). This M factor is present in the BOLD responses to CO2 and O2 inhalation. By combining the BOLD-CO2 and BOLD-O2 responses, one can solve for ΔCBF/CBF0 and derive a normalized CVR as CBF-CVR in the units of %CBF changes per unit of EtCO2 change, with minimized effects of BOLD confounding factors while maintaining the sensitivity of BOLD-CVR. A proof-of-principle study was performed in 7 healthy subjects, followed by a test–retest reproducibility study in 11 subjects. BOLD-CVR maps presented ultra-bright large veins and large contrasts between gray and white matter. Compared with BOLD-CVR, CBF-CVR showed smaller inter-voxel coefficient of variance (CoV) and smaller gray matter/white matter ratio, and showed little dependence on vCBV, providing better assessment of microvascular functions. The inter-session CoV was comparable between BOLD-CVR and CBF-CVR, suggesting a minimal effect of potential noise propagation by the normalization. Future work will validate this approach by ASL-MRI. Overall, this approach can provide a more direct index of microvascular function and would be practical and promising in cerebrovascular disease applications.

• **Doctorate Student Award**
  94
  Authors: Jamel Lawson – Presenter, Dr. Sreedevi Sampath
  Title: Vibe Coding and Software Quality Assurance: A Systematic Mapping Study
  Abstract:

  The rapid adoption of large language model (LLM)-based coding assistants has introduced “vibe coding,” a development approach defined by intent-driven, natural-language programming and a reduction in manual code implementation. While these tools offer the potential for increased productivity, their impact on software quality assurance (QA) is not yet fully understood. This systematic mapping study reviews 39 studies published between 2023 and 2025, including 37 peer-reviewed primary studies and 2 grey literature sources, to assess the current evidence linking vibe coding practices to QA outcomes. The study addresses three research questions: (1) the documented impacts of AI hallucinations on QA processes, (2) how developer practices with AI tools influence QA outcomes, and (3) the current evidence base and remaining gaps. Our findings indicate that 32 of 39 studies (82%) explicitly report hallucinations, with some studies identifying security vulnerabilities in up to 40% of generated code. Developer practices are inconsistent, with execution-based validation (“run-and-see”) observed in 15 studies while rigorous manual review is largely confined to regulated environments. The corpus is dominated by empirical evaluations in industry settings, and 20 studies report quantitative QA metrics. No longitudinal maintenance studies were identified. This mapping study offers a structured overview for researchers and practitioners, highlights critical gaps such as the absence of longitudinal studies and standardized metrics, and provides a foundation for future systematic reviews as the field evolves.

• 100
  Authors: Joey Mule (PRESENTER), Shahmir Rizvi, Tejas Gokhale, Riadul Islam
  Title: STOP-ET: Spatio-Temporal Optical Pipeline for Event-based Threats

  Abstract:

  Event cameras offer low-latency perception with high temporal precision, but their asynchronous data streams are uniquely vulnerable to adversarial perturbations. Minor manipulations of spike timing or polarity can induce severe misclassifications in downstream models. We introduce the Spatio-Temporal Optical Pipeline for Event Threats (STOP-ET), a fully rule-based adversarial defense framework that restores reliability in neuromorphic vision without adversarial training or supervision. To comprehensively evaluate robustness, we formulate two new attack paradigms: removal and manipulative injection, that delete or synthetically implant structured event clusters, exposing failure modes overlooked by prior additive or gradient-based attacks. STOP-ET monitors temporal statistics and motion continuity to detect event-stream inconsistencies. Upon detecting abnormal activity, STOP-ET applies motion-aligned denoising, spatial masking, and temporal reconstruction to suppress injected or missing events, while an optical-flow-guided prediction stage preserves coherence under severe corruption. Evaluations on NMNIST (digit classification), DVSGesture (gesture recognition), and 1Mpx (automotive data), alongside real-world physical experiments, demonstrate the efficacy of STOP-ET for improving adversarial robustness of event cameras across diverse motion regimes.

• 95
  Authors: Dr.Carolyn Seaman (Faculty mentor) and Mahsa Radnead (Presenter)
  Title: Explainable AI for Identifying and Managing Test Debt

  Abstract:

  Rapid technological advancements have fundamentally transformed software development practices, enabling organizations to build increasingly complex systems at unprecedented speed. This acceleration has also increased pressure on development teams, often resulting in rushed deadlines and compromised engineering decisions. Such short-term trade-offs frequently introduce technical debt. Technical debt is a metaphor that reflects the implementation of suboptimal solutions to achieve short-term goals at the expense of long-term software maintainability. Organizations worldwide must invest significant effort to manage technical debt in order to maintain software quality. In many cases, the impact of technical debt is difficult to avoid. There are multiple causes of technical debt such as schedule pressure, lack of skilled engineers, insufficient documentation, inadequate processes and standards, poorly maintained test suites, delayed refactoring, and knowledge gaps. These causes contribute to various types of technical debt, including code debt, test debt, requirements debt, and defect debt. Test debt specifically arises from inadequate decisions related to software testing activities, such as insufficient test and code coverage. There is currently limited empirical evidence regarding the accuracy, efficiency, and explainability of machine learning–based techniques for identifying test debt. This gap limits practitioners’ ability to trust, understand, and act upon model predictions, such as performing targeted refactoring or improving test quality. Understanding the causes of test debt enables organizations to make informed decisions about whether to incur test debt and how to address it. This research aims to explore a new approach to identifying test debt and explaining to software developers and testers. Our approach identifies test debt by using machine learning and explains the reasons and causes of the debt found with Explainable Artificial Intelligence (XAI) techniques.

• **Doctorate Student Award**
  101
  Authors: Hasan Masum, Mehreen Rashid (Presenter), Tarannum Shaila Zaman
  Title: Evaluating Large Language Models for End-to-End System-Level Concurrency Bug Reproduction

  Abstract:

  In modern multicore computing environments, concurrency bugs such as race conditions, deadlocks, atomicity violations, ordering faults etc. pose significant challenges to software reliability and maintenance. These bugs are difficult to detect, reproduce, and fix due to their non-deterministic nature. Open source bug tracking platforms, such as Red Hat Bugzilla, Debian Bug Tracking System, and GNU Project issue trackers, contain rich natural-language descriptions, stack traces, execution environments, and developer discussions that implicitly encode failure-triggering conditions. However, bug reports are written in unstructured natural language and often lack critical information needed to reproduce the bugs. As a result, extracting actionable reproduction information from these reports demands substantial manual effort and deep domain expertise. In this study, we investigate whether Large Language Models (LLMs) can automate this process end-to-end, as they can understand and reason over natural language, which makes them a promising approach. We created a curated dataset of real-world system-level concurrency bug reports from major open-source repositories, with manually annotated bug-triggering system-call interleavings and relevant inputs as ground truth. We first evaluate the ability of state-of-the-art LLMs to extract bug-triggering system-call interleavings and relevant inputs from these bug reports. We then use the extracted information to synthesize executable system-level test cases using LLMs and assess their effectiveness in reproducing the reported failures. Experimental results demonstrate that LLM-generated test cases can reproduce a significant portion of historical concurrency failures. Our study provides insights into the capabilities and limitations of LLMs in reproducing system-level concurrency bugs and establishes a benchmark for future automated concurrency debugging tools.

• **Undergraduate Student Award**
  102
  Authors: Christian Wilkins (Presenter), Dong Li , Rishav Gupta
  Title: Smartphone Human Detection using Acoustic Sensing

  Abstract:

  The standard methods of detecting human presence face a variety of issues such as privacy concerns, expensive hardware, and interference vulnerability. A solution that avoids these problems, leveraging physics and existing hardware, is human detection through acoustic sensing done entirely on the user’s smartphone. Although other work has been completed to develop an acoustic sensing solution to human detection, a system implemented entirely on a smart phone is a current research gap. The FSI lab seeks to build a real-time Human Presence Detection system for mobile devices. We plan to demonstrate that such a system is feasible and to develop an adequate physics-informed detection algorithm given the hardware constraints. Thus far, we have developed an offline human detection pipeline using only the speakers and microphone available on a smartphone. Currently, the system is able to detect human presences within a distance of 1m from the phone on pre-recorded experiment .wav files. Our work demonstrates that a smartphone is capable of such a system, provided that computational demands of signal processing are met. We implement an FMCW chirp sensing pipeline by using the smartphone as an I/O device to emit FMCW chirps and receive their time delayed versions. We then abstract the signal processing to our computers which convert the signals to the frequency domain and extract the physics changes across chirps. Lastly, we apply a human detection algorithm by thresholding those physics changes to classify presences and non-presences. Our next research milestone is to make the system real-time and our overarching vision is to implement it entirely on the smartphone.