Upcoming dates
- -
You are invited to the I-DISC Faculty Forum + Poster Session & Lunch!
This semester, we will be combining the I-DISC Faculty Forum with the External Advisory Council (EAC) meeting. EAC members will offer insights into their research and trending topics. I-DISC faculty are encouraged to have their students (UG/Grads/PostDocs) showcase their research at a poster session and luncheon. Members of the EAC will join us in person, & some by Zoom.
Summary of Events - more detailed information will be posted soon.
9:15 AM -12:00 noon (BC220)
- I-DISC '24-'25 updates
- Research Presentations by I-DISC faculty & EAC members
- Trending Topics - small group discussions
12:00-1:30 PM (C3 Lounge area)
Lunch & Student Poster Session
Schedule of Events
8:45am | Light breakfast available (BC220) |
9:15am | I-DISC Welcome, Overview I-DISC Leadership Team: Kate Arrington (Psychology), Mooi Choo Chuah (Computer Science & Engineering) and Parv Venkitasubramaniam (Electrical & Computer Engineering) |
9:30am | Faculty Presentation 1: "Interdisciplinary Collaboration: Machine Learning Approaches to Clinical Behavioral Health”, Nancy Carlisle, Psychology |
10:00am | Short break |
10:05am | External Advisory Council Member Research Presentation “Navy Center for Applied Research in AI: Interdisciplinary Efforts”. David Aha, Director, Navy Center for Applied Research in AI (NCARAI), Naval Research Laboratory |
10:35am | Faculty Research Presentation 2 "Applications of Algebraic Graph Theory and Algebraic Topology to Optimization, Planning and Control of Robotic and Networked Systems". Subhrajit Bhattacharya, Mechanical Engineering and Applied Mechanic |
11:05am | Short break |
11:15am | “Round Table” Discussions Small group discussions with I-DISC faculty and EAC members Topic: Barriers to Research Researchers across academia, government, and industry are feeling under threat as the current administration shifts funding priorities and academic freedom norms. Join in conversations across disciplines and sectors about the greatest barriers that you are currently experiencing in research. |
12:00 noon | Student Poster Session & Lunch (BC C3 lounge area) |
1:30pm | End |
I-DISC STUDENT POSTER SESSION
(Building C Lounge area C3) 12:00-1:30 PM
(Research authors: Only the students presenting (bold) and the I-DISC Faculty Member Adviser are listed below).
Observer Reasoning about an Agent’s Method for Deceptive Path Planning
Xubin Fang; Adviser: Rick S. Blum
Abstract: Deceptive path planning (DPP) is a well-studied problem in robotics and autonomous systems, where an agent aims to mislead an observer by generating paths that obscure its true intentions or goals. While extensive research has focused on developing strategies for agents to plan deceptive trajectories, the inverse problem—if an observer can infer the deceptive strategy the agent employs and the true objectives from partial path information—remains relatively unexplored. This paper investigates DPP from the perspective of the observer, shifting the focus from path generation to strategy classification and its exploitation. After strategy classification, the observer specifically addresses the challenge of inferring the true goal of an agent in scenarios where multiple potential goals exist, some of which may serve as decoys. To achieve this, we propose a data-driven approach where the observer trains a classifier using previously observed path data. The trained classifier takes partial path information as input and outputs a classification of the strategy the agent employs along with belief probabilities (odds) for each possible goal at a given time step. Preliminary results demonstrate the effectiveness of our approach in distinguishing between strategies and the true goals and decoys, providing a robust framework for adversarial reasoning in autonomous systems.
GIDEA: Generative AI-Powered Interactive Design and Evaluation Platform for Assistant Agent Research
Ziyi Xuan, Yiwen Wu; Adviser: Yu Yang
Abstract: Conducting human-computer interaction (HCI) experiments often requires extensive manual effort, including configuring environments, recruiting participants, and recording interactions. We introduce GIDEA, a generative AI-powered interactive design and evaluation platform for assistant agents to streamline and accelerate HCI research. Our platform employs a three-role interaction pipeline, where researchers define experiments, large language model-driven avatars simulated participants, and a smart assistant agent moderates interactions. This pipeline dynamically generates interaction scenarios, avatar profiles, and adaptive responses based on researcher input. By integrating with Unity, GIDEA enables real-time monitoring and control over simulated experiments, providing researchers with an interactive and adaptable evaluation environment. Through the replication of real-world case studies, we demonstrate that GIDEA reduces the time and effort required for HCI experiments while producing results that align with real studies. This capability has the potential to revolutionize HCI research by transforming traditionally lengthy and labor-intensive processes into a highly efficient, scalable, and adaptive methodology, accelerating innovation and broadening experimental possibilities.
Predicting Firn Layer Thickness with BiRFNO: A Novel Fourier Neural Operator for Snow Radar Imaging
Heling Wang, Electrical and Computer Engineering; Adviser: Maryam Rahnemoonfar, Computer Science and Engineering & Civil and Environmental Engineering
Abstract: Accurately predicting annual snow accumulation is essential for understanding climate dynamics and monitoring the Greenland ice sheet. NASA’s Snow Radar, part of Operation IceBridge, provides high-resolution data that captures detailed internal ice layer structures across large areas. However, with the abundance of radar data, traditional machine learning models, such as Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs), often achieve lower accuracy due to challenges in capturing the hierarchical structure and thickness-specific information in snow radar data. To address these limitations, we propose the Bidirectional Recurrent Fourier Neural Operator (BiRFNO), a novel model that combines depth-wise convolution with a bidirectional Fourier recurrent structure. On Snow Radar data, BiRFNO achieves a 58.3\% improvement in RMSE over Long Short-Term Memory (LSTM) networks and outperforms other neural operators. This work advances radar-based ice sheet monitoring by providing a scalable solution for precise firn layer thickness prediction and more accurate climate impact assessments.
Evaluating LL Models in Research Paper Metadata Extraction and Adaptability
Shunjie Zhang; Adviser: Zheng Yao, Kai Landskron
Abstract: We evaluate the ability of state-of-the-art AI models to extract structured metadata—such as titles, authors, and publication dates—from academic research papers. Our study compares five leading models on initial accuracy, adaptability to user feedback, and evaluation speed. The results highlight clear differences in model performance and practical usability in research workflows.
Optimizing Inventories in a Multi-echelon Inventory System
Sharmine Akther Liza; Adviser: Lawrence V. Snyder
Abstract: In a multi-echelon inventory system, upstream stages (nodes) are responsible for supplying items to downstream stages, fulfilling their demands whenever sufficient inventory is available. Each downstream stage manages its own inventory while determining order quantities based on the inventory levels of the upstream stage. This interdependency emphasizes a critical relationship—shortages at the upstream stages can significantly impact the ordering decisions of downstream stages. This study explores both the theoretical and numerical aspects of how order quantities at downstream stages are influenced by upstream inventory levels, and how these decisions evolve over a given time horizon. Additionally, we propose a heuristic for determining order-up-to level at each stage within a general multi-echelon system, accounting for stochastic lead times resulting from random stock-outs throughout the system.
Optimal Pricing and Scheduling of Electric Vehicle Charging Stations Using Deep Reinforcement Learning
Morteza Ghorashi; Advisers Parv Venkitasubramaniam, Shalinee Kishore
Research Question: How can we optimally manage an EV charging station (EVCS) operation in real time, given the high uncertainty of the problem?
Goal: To design a reinforcement learning (RL)-based approach to efficiently optimize pricing and charging/discharging scheduling of EVCSs in real time.
Unifying Explainable Anomaly Detection and Root Cause Analysis in Dynamical Systems
Yue Sun; Advisers: Rick S. Blum, Parv Venkitasubramaniam
Abstract: Dynamical systems, prevalent in various scientific and engineering domains, are susceptible to anomalies that can significantly impact their performance and reliability. This paper addresses the critical challenges of anomaly detection, root cause localization, and anomaly type classification in dynamical systems governed by ordinary differential equations (ODEs). We define two categories of anomalies: cyber anomalies, which propagate through interconnected variables, and measurement anomalies, which remain localized to individual variables. To address these challenges, we propose the Interpretable Causality Ordinary Differential Equation (ICODE) Networks, a model-intrinsic explainable learning framework. ICODE leverages Neural ODEs for anomaly detection while employing causality inference through an explanation channel to perform root cause analysis (RCA), elucidating why specific time periods are flagged as anomalous. ICODE is designed to simultaneously perform anomaly detection, RCA, and anomaly type classification within a single, interpretable framework. Our approach is grounded in the hypothesis that anomalies alter the underlying ODEs of the system, manifesting as changes in causal relationships between variables. We provide a theoretical analysis of how perturbations in learned model parameters can be utilized to identify anomalies and their root causes in time series data. Comprehensive experimental evaluations demonstrate the efficacy of ICODE across various dynamical systems, showcasing its ability to accurately detect anomalies, classify their types, and pinpoint their origins.
Computational Design of next-generation materials
Tianhao Lian, Srihari Kastuar, Gour Jana; Adviser: Chinedu Ekuma
Abstract: In the context of materials design, computational exploration of new materials with exotic features have become a necessity to save the cost of experimental resources and fine-tune materials search and design. We use first-principles methods to model and characterize magnetic materials which can be explored for spintronics application, and half-metals which show 100% spin polarization and can be further investigated for spin-filtration applications.
Advancing Safe and Intelligent Autonomy Through Formal Methods
Disha Kamale; Adviser: Prof. Cristian-Ioan Vasile
As cyber-physical systems increasingly operate in safety critical domains, ensuring correct behaviors under all possible scenarios is instrumental to avoid serious repercussions.
While existing formal approaches provide the essential rigorous guarantees, they often rely on the restrictive assumptions of -1) complete knowledge of environment, and 2) complete feasibility of assigned missions. Our work focuses on addressing these limiting assumption through formal methods for planning and decision-making.
Large-scale many-body simulation of strong e-e interaction in 2D GeS with random defects
A. C. Iloanya, S. M. Kastuar, Gour Jana; Adviser: C. E. Ekuma, Physics
Abstract: Strong electron-electron interaction and intrinsic defects in 2D materials play a crucial role for enhancing and tuning novel material properties for potential technological applications. Our large scale simulations using a computationally efficient approach, first-principle based Typical Medium Dynamical Cluster Approximation (TMDCA) of 2D GeS intercalated with organometallics in the presence of many-body interaction and random disorder shows strong enhancement of the excitonic absorption and demonstrate its tunability, which is important for efficient optoelectronic devices such as solar cells.
Scalable Networked Feature Selection with Randomized Algorithm for Robot Navigation
Vivek Pandey (Lehigh University), Arash Amini (University of Texas at Austin), Guangyi Liu (Amazon Robotics), Ufuk Topcu (University of Texas at Austin), Kostas Daniilidis (University of Pennsylvania), Qiyu Sun (University of Central Florida); Adviser: Nader Motee (Lehigh University)
Abstract: We address the problem of sparse selection of visual features for localizing a team of robots navigating an unknown environment, where robots can exchange relative position measurements with neighbors. We select a set of the most informative features by anticipating their importance in robots localization by simulating trajectories of robots over a prediction horizon. Through theoretical proofs, we establish a crucial connection between graph Laplacian and the importance of features. We show that strong network connectivity translates to uniformity in feature importance, which enables uniform random sampling of features and reduces the overall computational complexity. We leverage a scalable randomized algorithm for sparse sums of positive semidefinite matrices to efficiently select the set of the most informative features and significantly improve the probabilistic performance bounds.
Nontrivial Chern Phase in 2D ScV6Sn6 Kagome Metal
Chidiebere Nwaogbo, Sanjib K Das; Adviser: Chinedu Ekuma
Abstract: We explore the topological properties of 2D ScV₆Sn₆, a ferromagnetic kagome metal. Ab initio calculations reveal Weyl-like crossings, a Chern number |C| = 1, and a large anomalous Hall effect (257 Ω⁻¹cm⁻¹). A topological-to-trivial transition occurs at ≈0.40 eV/Å. These findings position ScV₆Sn₆ as a promising candidate for quantum computing and electronic devices.
Accurate 3D Reconstruction of Ice-Bed topography via Graph Transformer
Zesheng Liu; Adviser: Maryam Rahnemoonfar
Abstract: Accurately mapping the bed topography of ice sheets is essential for understanding ice dynamics and monitoring the impact of climate change. Ice penetrating radar has become a powerful tool, as it emits signals that can penetrate thick ice and directly reach to the bed, providing a direct measurement of the bed topography. However, accurately identifying the ice bed boundary remains challenging due to low signal to interference and noise ratio, high variability of subglacial topography, and presence of artifacts in the collected data. To address these limitations, we proposed a novel geometric deep learning approach. Unlike previous deep learning methods that focus on convolutional operations, our proposed approach leverages the irregular structure of the bed by representing it as graphs. We developed a graph transformer that learns from the ice-surface structural information, aiming to better reconstruct the 3D bed topography. Experiment results show that compared with previous method that focus on reconstruct 3D bed topography from radar echograms, our graph transformer utilized the ice-surface elevation that contains less outliers and thereby achieved a lower mean absolute error.
3D Semantic Segmentation for Post-Disaster Scenarios
Nhut Le, Computer Science and Engineering; Adviser: Maryam Rahnemoonfar, Computer Science and Engineering, and Department of Civil and Environmental Engineering
Abstract: Due to the increment of frequency and severity of natural disasters, I am developing a robust 3D semantic segmentation deep learning method to assist rescuers in their missions. A 3D benchmark dataset which is tailored for post disaster scenarios is also developed.
CQA-CLIP: Incorporating concatenated question-answer pairs to enhance vision language models
Ehsan Karimi, CSE; Adviser: Maryam Rahnemoonfar, CSE & CEE
Abstract: Contrastive Language-Image Pre-training (CLIP) is the most used method to train vision-language models. This method brings the visual and textual representation vectors into the same space, and forces the model to repel unpaired image and text representation vectors, while bringing paired image and text representation vectors closer together. This enables the model to better understand the relationships between visual and textual features. However, this method suffers from focusing on global features rather than fine-grained information in given inputs. Existing methods primarily address the issue of finer understanding of input data by either manipulating the input data itself, information fusing through the model layers, or late fusion through the word-region matching strategy. In this study, we introduce Z-CLIP, a framework designed to train a CLIP-based model with a more refined level of visual and textual feature understanding. We generate a dataset consisting of image, caption, and question-answer pairs, and use both captions and question-answer pairs to train the model simultaneously, allowing it to learn from more detailed and contextual interactions between the modalities.
Watermarks vs. Perturbations for Preventing AI-based Style Editing
Qiuyu Tang; Adviser: Aparna Bharati, CSE
Abstract: The remarkable image editing capabilities of generative models have led to growing concerns regarding unauthorized editing of multimedia. To mitigate against such misuse, artists and creators can utilize traditional image watermarking and more recent adversarial perturbation-based protection techniques to protect media assets. Watermarks generally protect the origin by establishing ownership, but can be easily removed. However, perturbation-based protection is aimed at disrupting editing and is harder to remove. In this paper, we evaluate the effectiveness of the two methods against Stable Diffusion in preventing the generation of usable edits.
HurricanVidNet: A High-Resolution UAV Video-Based Semantic Segmentation Benchmark Dataset for Natural Disaster Damage Assessment
Kevin Dotel, Alec Jang, Kerrick Truong, Hardhik Mandadi, Michael DeFulvio, Caleb Mandia, Ben Shain; Adviser: Maryam Rahnemoonfar. All affiliated with Bina Lab, Lehigh University
Abstract: HurricaneVidNet is a high-resolution UAV-based video dataset created to address the lack of pixel-level training data for real-time disaster damage assessment. Unlike static image datasets, HurricaneVidNet captures temporal dynamics important for accurate scene understanding. Collected from Hurricane Ian footage and annotated with over 100,000 instances, this dataset enables the development and evaluation of robust segmentation models. Preliminary results using Detectron2 show promising performance in distinguishing varying damage levels, supporting future real-time, in-field deployment.
Flood extent prediction using Vision Language models
Oluwanisola Ibikunle; Adviser: Maryam Rahnemoonfar
Abstract: to follow
PASS Project: Air Pollution and Asthma Prevalence Analysis in Pennsylvania
Giovanni Sanchez; Advisers: Hyunok Choi, Parv Venkitasubramaniam; Xiang Gao
Abstract: This project investigates the relationship between asthma hospitalization rates and environmental factors such as air pollution and meteorological variables in Pennsylvania from 2019 to 2022. Using PHC4 hospitalization data and environmental data including NO_2, SO_2, CO, O_3, etc, we analyze seasonal and geographic patterns of asthma prevalence. Data are aggregated at the geographic level using ZCTA codes. Exploratory data analysis, correlation studies, and machine learning models are applied to identify key predictors of asthma incidence and to generate spatial risk maps. The goal is to better understand environmental contributors to asthma and support data-driven public health interventions.
Typical Medium Dynamical Cluster Studies of the Rashba Spin-Orbit Coupling in 2D Disordered Systems
Yongtai Li, Gour Jana; Adviser: C.E. Ekuma, Physics
Abstract: Anderson localization, a disorder-induced metal-insulator transition, has been a subject of study for decades. The single-parameter scaling theory predicts that even minimal disorder in low-dimensional electronic systems will localize all electronic states. However, recent research suggests the possibility of an Anderson localization transition in 2D disordered systems with spin-orbit coupling (SOC). In this work, we extend the Typical Medium Dynamical Cluster Approximation (TMDCA) by incorporating Rashba-type SOC to investigate the interplay between disorder, Rashba SOC, and spatial correlations in the context of the Anderson metal-insulator transition in 2D systems. Our findings provide new insights into the combined effects of disorder and SOC, as well as the associated critical phenomena.
Pennsylvania Asthma-COPD Syndromic Surveillance (PASS)
Giovanni Sanchez, Xiang Gao; Advisers: Hyunok Choi, Parv Venkitasubramaniam
Abstract: The Pennsylvania Asthma-COPD Syndromic Surveillance (PASS) project examines chronic respiratory health risks by integrating hospital admission data with environmental, meteorological, and socioeconomic indicators. Using spatial interpolation and machine learning, we analyze ZCTA-level hospitalization patterns across Pennsylvania. Preliminary findings for asthma suggest that factors such as nitrogen dioxide, wind direction, income level, and age distribution may influence admission rates. This work supports future predictive modeling and more equitable public health strategies.
Acceleration of Approximate Maps for Matrices Arising in Discretized PDEs
Rishad Islam; Adviser: Arielle Carr
Abstract: Generally, discretization of partial differential equations (PDEs) creates a sequence of linear systems A_k x_k = b_k , k = 1, 2, ..., N with well-known and structured sparsity patterns. For solving closely related systems in sequence, we can use preconditioner updates such as Sparse Approximate Map (SAM) instead of computing a preconditioner for each system from scratch. A SAM acts as a map from one matrix in the sequence to another nearby one for which we have an effective preconditioner. We seek to compute an optimal sparsity pattern to efficiently compute an effective SAM update. In this poster, we examine several sparsity patterns for computing the SAM update in an effort to characterize optimal or near-optimal sparsity patterns for linear systems arising from discretized PDEs. The allowable number of nonzeros in the sparsity pattern should strike a balance between the accuracy of the map and the cost to apply it in the iterative solver. We can show that the sparsity pattern of the exact map is a subset of the sparsity pattern of the transitive closure of a graph representation of A_k , G(A_k ). Additionally, we make use of the heterogeneous computing environment to accelerate the computation of the SAM. The inherently parallel nature of the SAM algorithm naturally lends itself towards efficient implementation in GPU and distributed computing systems. We present preliminary results in this area.
Transforming Color Correction for Colorblindness Using Neural Networks and Computational Simulations
Lucas Yang, Parkland High School; (c/o Beibei Wang, Marketing, College of Business)
Abstract: Color blindness affects 300 million people globally but remains an underrecognized disability, with current solutions often costly or ineffective. This project develops a computational color correction method to enhance color perception for colorblind individuals. It uses the LMS color space to simulate colorblindness through RGB-to-LMS conversion, then applies a correction model inspired by fluid dynamics—using pressure to push confusing colors apart and viscosity to keep similar colors together. A convolutional neural network (CNN) trained on Ishihara-style test plates and MNIST digits achieved 96% accuracy. Testing showed a 60% improvement in recognition accuracy after correction (from 34% to 54%), outperforming an existing tool that only improved accuracy by 20%.