AI Research Showcase Poster Session

AUTHOR(S)	TITLE	ABSTRACT	POSTER LINK
Ace Rounds Jacob Hartwick Addam Jalal Ben-Abdallah	Automation Methods in Gamma Ray Spectroscopy and Data Analysis	Gamma-ray spectroscopy is used by nuclear physicists to better understand the structure of the nucleus in excited states. By observing the decays from an excited nucleus, we can create level schemes that map allowed energy states and transitions. With this knowledge, we can identify specific energy states relevant to nuclear isomer battery applications. Despite their importance, level schemes are currently only constructed through manual analysis, limiting scalability and increased risk for systematic bias, which in turn limits ability for discovery. We present a fully automated framework that reframes level scheme construction as an image-to-graph translation problem. An end-to-end machine learning (ML) framework converts two-dimensional gamma-gamma coincidence matrices into structured nuclear level schemes with transitions, implemented using an autoregressive graph neural network (GRAN). The framework is supported by Monte Carlo simulation that inputs level transition data, benchmarking the ML solution. Hence, enabling generation of toy sample of level schemes to overcome limited availability of data for training and validation. The resulting level schemes serve as a benchmark for model performance prior to deployment on experimental data. Leveraging ML/AI allows increases in speed, accuracy and reach are bolstering the ability to discover nuclear isomer states of interest in transitions data.
Ashley Conroy Beckett Thomas Eleanor Conway	Leveraging AI for Modern Finance: An Integrated Approach to Valuation and Investment Management	This project presents the development of an AI-augmented financial analysis platform built as part of MGMT 490: AI Finance Application. The system integrates corporate finance and investment management methodologies into a unified, interactive tool designed to support real-time financial decision-making. In the first phase, we developed a Discounted Cash Flow (DCF) valuation application that dynamically retrieves financial data and estimates intrinsic value for publicly traded companies. The model incorporates adjustable assumptions including discount rate, growth rate, and terminal value that performs sensitivity analysis to evaluate how changes in inputs impact valuation outcomes. Building on this foundation, the second phase expands into investment management through portfolio optimization or sentiment analysis. The system processes multi-asset inputs, evaluates risk-return tradeoffs, and generates optimized allocations under user-defined constraints, or alternatively analyzes market sentiment using real-time textual data. Across both phases, the project leverages the DRIVER framework to structure AI-assisted problem solving, emphasizing validation, transparency, and financial reasoning. This platform demonstrates how AI can enhance traditional financial workflows while maintaining human oversight and improving analytical efficiency.
Hien Vu	Wireless Health Monitoring Powered By Machine Learning	Autonomous and continuous health monitoring is essential for applications ranging from elderly care to precision agriculture. While wearable and remote sensors can capture rich bio-signals, large-scale deployment faces three critical barriers: severe energy constraints, a lack of multimodal datasets, and the failure of existing algorithms to extract signals from dynamic environments. This work presents a comprehensive solution bridging the gap between theoretical models and physical deployment. To address energy constraints, we introduce an energy-neutral sensing platform designed for maintenance-free, long-term operation. To address data scarcity, we present MmCows, a large-scale multimodal dataset that serves as a benchmark for sensor fusion and model validation. Finally, we propose a novel machine learning model for processing mmWave radar data that successfully extracts respiration rates from moving subjects,a task where traditional signal processing fails. The integration of these three layers enables a new approach to health monitoring system design. For human health, this facilitates privacy-preserving, non-invasive monitoring of vital signs and falls without frequent battery replacement. For animal health, it enables scalable precision agriculture and the early detection of respiratory disease and distress. By solving the interdependent challenges of energy, data, and models, our work provides a scalable pathway for the next generation of wireless health sensing.
Hojin Kim	Generalizable Data-driven Turbulence Closure Modeling on Unstructured Grids with Differentiable Physics	In this study, we present a differentiable physics framework for learning generalizable sub-grid scale (SGS) closures in three-dimensional incompressible Large Eddy Simulations (LES). In particular, our framework trains a graph neural network (GNN)-based SGS model designed for mesh-agnostic applicability. The trained GNN model is then embedded into a finite element (FEM) solver, FEniCSx. Differentiability of the FEM solver is achieved through the discrete adjoint method, which ultimately enhances the a-posteriori performance and stability of the learned closure.
Hwangyu Cho Tong Chen	CO-MAC: A Center-Out Ordered Stochastic MAC for Low-Latency Inference	Stochastic computing offers efficient approximate arithmetic that aligns well with error-tolerant machine learning workloads, but its deployment is limited by long bitstream latency in stochastic multiply-accumulate (MAC) units. Prior work reduces MAC latency through deterministic bitstream generation and differential accumulation, but these methods do not fully exploit the statistical properties of convolution weights. This work presents a novel stochastic MAC architecture named CO-MAC, which employs center-out weight ordering and an enhanced convolution engine design to reduce effective computation cycles while maintaining high accuracy. The method sorts weights by magnitude, reuses the incremental differences in magnitudes, and applies sign handling after accumulation. This shortens the counter activity, maintains accuracy with long effective bitstreams, and simplifies the MAC hardware by avoiding bidirectional counters. Across convolutional neural network workloads, CO-MAC decreases MAC latency by up to 54.8% compared to prior stochastic MAC architectures, while preserving accuracy and hardware simplicity.
Jiancong Chen	A novel energy-efficient driving strategy for autonomous electric vehicle operations using vision language model and reinforcement learning	Previous studies have shown that driver behavior directly affects vehicle energy consumption and automated driving will drastically shift driving behaviors compared to manual driving. As such, the role of autonomous driving technologies in reducing energy consumption is gaining attention. Autonomous electric vehicles (AEVs) can potentially enhance driving safety and energy efficiency through high-precision real-time maneuvering. However, compared to gasoline autonomous vehicles (GAVs), research on energy-efficient strategies for AEV operations remain underdeveloped. n this context, this paper proposes a novel AEV driving strategy which integrates Vision-Language Models (VLMs) and Reinforcement Learning (RL) to maximize the energy efficiency while preserving driving safety. The methodology involves pretraining VLMs using offline expert driving datasets, thereby enabling them to evaluate safety outcomes as guidance for diverse traffic scenarios. Then, leveraging the safety guidance, a reward framework is developed to generate reward functions that incorporate elements of the surrounding environment: safety, traffic efficiency, and energy efficiency. To enhance the reliability of proposed driving strategy, an artificial potential field (APF) model was integrated downstream of the RL module. This enabled real-time evaluation of safety based on the state information of surrounding vehicles. The proposed strategy was trained and evaluated through extensive experiments on CARLA.
Jiaru Zhang Manav Gagvani Juntong Peng	Efficient and Explainable End-to-End Autonomous Driving via Masked Vision-Language-Action Diffusion	Large Language Models (LLMs) and Vision-Language Models (VLMs) have emerged as promising candidates for end-to-end autonomous driving. However, these models typically face challenges in inference latency, action precision, and explainability. Existing autoregressive approaches struggle with slow token-by-token generation, while prior diffusion-based planners often rely on verbose, general-purpose language tokens that lack explicit geometric structure. In this work, we propose Masked Vision-Language-Action Diffusion for Autonomous Driving (MVLAD-AD), a novel framework designed to bridge the gap between efficient planning and semantic explainability via a masked vision-language-action diffusion model. Unlike methods that force actions into the language space, we introduce a discrete action tokenization strategy that constructs a compact codebook of kinematically feasible waypoints from real-world driving distributions. Moreover, we propose geometry-aware embedding learning to ensure that embeddings in the latent space approximate physical geometric metrics. Finally, an action-priority decoding strategy is introduced to prioritize trajectory generation. Extensive experiments on nuScenes and derived benchmarks demonstrate that MVLAD-AD achieves superior efficiency and outperforms state-of-the-art autoregressive and diffusion baselines in planning precision, while providing high-fidelity and explainable reasoning.
Jingying Hu	LLM-derived metrics in L2 writing assessment: an explainable AI approach	As large language models (LLMs) are increasingly used in language education, their potential for second language (L2) writing assessment requires systematic evaluation. While existing approaches often rely on prompt-based grading, these methods lack interpretability and consistency. This study investigates two core LLM mechanisms—next-token prediction and contextualized embeddings—to develop interpretable metrics for L2 writing assessment. Using a corpus of 1,196 Traditional Chinese learner essays across four proficiency levels, we compute surprisal, reflecting linguistic predictability, and embedding-based similarity to capture semantic coherence. These metrics are evaluated across multiple LLMs and compared with 11 traditional linguistic features related to fluency, lexical sophistication, and syntactic complexity. Results show that both surprisal and similarity reliably differentiate proficiency levels, with lower surprisal and higher coherence in more advanced writing. Combining LLM-derived and traditional features further improves machine learning performance in proficiency prediction. Interpretability analyses suggest that surprisal captures multiple dimensions of writing development, supporting its use as a holistic and explainable metric. These findings highlight the potential of LLM-derived measures for scalable and transparent L2 writing assessment, particularly for underrepresented populations such as Traditional Chinese learners.
Junhyeok Kil	AI-Driven Heat Transfer Prediction in Additive Manufacturing with PINNs	Accurate and scalable prediction of thermal history is pivotal for optimizing print quality in extrusion deposition additive manufacturing (EDAM). This paper investigates the applicability of physics-informed neural networks (PINNs) for 3D heat transfer analysis during the additive manufacturing process. As finite element method (FEM) requires increasingly fine meshes and smaller time increments to achieve higher accuracy, its computational cost rises substantially. PINNs offer a more scalable solution by eliminating the need for meshes and time increment schemes. We achieve this by recasting the solution to the differential equation as a stochastic minimization problem. While the current focus is a single forward problem, the paper sets the groundwork for future research into parametric problems, where PINNs demonstrate their full potential by efficiently solving a range of scenarios under varying initial, boundary conditions, and material properties. Our results show that PINNs can be used to solve the 3D heat transfer equation on evolving geometries without the need for spatial discretization and time-stepping schemes. This study showcases the growing relevance of PINNs in manufacturing simulations, with future implications for real-time control and optimization in advanced manufacturing.	Poster
Kamelia Sepanloo Dr. Young-Jun Son Dr. Janine E Hinton	Multimodal and Adaptive AI-Driven Extended Reality Simulations for Personalized Clinical Training	Healthcare professionals often make critical decisions under intense time pressure and stress. Therefore, training systems must address not only clinical procedures but also the cognitive and emotional demands of high-acuity environments. This research introduces a multimodal AI-driven extended reality simulation framework that integrates physiological sensing, immersive technology, and adaptive artificial intelligence. The system operates in both mixed reality (MR) and virtual reality (VR), where learners interact with a digital patient and medical equipment. A conversational AI model embedded in the digital patient enables natural dialogue and realistic communication. To enable adaptive training, a fine-tuned transformer-based large language model analyzes multimodal data streams to infer learners’ cognitive load and stress during the simulation. The developed training system consists of six segments in which the digital patient’s condition evolves based on the learner’s actions, with critical events such as hypotension and hypoxia included to increase realism. During the simulation, physiological signals (heart rate, skin temperature, electrodermal activity) and oculomotor data are collected and combined with conversation transcripts, behavioral logs, and survey responses. These fused data sources enable the system to detect patterns related to cognitive load, stress, and decision-making, supporting adaptive training in high-pressure clinical situations.
Koushiki Basu	Machine-Learning Based Identification of Causal Attributes of Corona Formation	Nanoparticles (NPs) entering biological environments rapidly acquire a biomolecular coating, called “biocorona”, composed of lipids and proteins that governs its pharmacokinetics, biodistribution, and cellular uptake. However, identifying the causal molecular attributes governing corona formation remains a major challenge due to biomolecular heterogeneity and complex environmental conditions. In this study, we present a machine learning framework integrating lipidomic and proteomic perspectives to uncover these determinants. For lipids, experimental lipidomic data were used to train a Sparse Gaussian Process classifier to predict adsorption across varying nanoparticle sizes and serum conditions, achieving strong performance (AUROC up to 0.83). SHAP analysis revealed that lipids with extended hydrocarbon chains and charged or hydrogen-bonding headgroups are strong adsorbents, while bulky, nonpolar structures exhibit weak affinity. For proteins, a surface-based representation combining aggregation propensity (AGGRESCAN3D) and Manifold Kernelization of Molecular Surfaces (MKMS) encodes spatial and electronic features governing interactions. These representations are further modeled using SPDNet to learn on SPD matrices, preserving geometric structure critical for protein adsorption prediction. This unified framework enables mechanistic interpretation and identification of causal attributes across biomolecular classes, providing a foundation for rational NP design.
Luca McAllister	Computer Vision for Large Assembly Defect Detection	The ability to detect defects in large assemblies is an important part of quality control in any manufacturing environment. Surfaces are visually inspected, typically by a qualified worker. However, these quality control workers can make mistakes and miss small scratches, foreign objects, and missing or incorrectly installed parts. Research is being done using computer vision to increase the effectiveness of visual inspection. A current promising machine learning approach has been Patch Distribution Modeling. Performing well on industry defect datasets like Mv-tech, PaDiM has been at the forefront of computer vision for quality control. There has, however, been a lack of testing and research done at the macro level. Currently tested datasets are single-camera, sub-foot-squared defects. This research works to use a multicamera approach to detect defects in large, multipart assemblies in the 10 feet3 and above range.
Mahir Rahman Samuel Joseph Nihar Kodkani	Harvest: Adaptive Photonic Switching Schedules for Collective Communication in Scale-up Domains	As chip-to-chip silicon photonics gain traction for their bandwidth and energy efficiency, their circuit-switched nature raises a fundamental question for collective communication:when and how should the interconnect be reconfigured to realize these benefits? Establishing direct optical paths can reduce congestion and propagation delay, but each reconfiguration incurs non-negligible overhead, making naive per-step reconfiguration impractical. We present Harvest, a systematic approach for synthesizing topology reconfiguration schedules that minimize collective completion time in photonic interconnects. Given a collective communication algorithm and its fixed communication schedule, Harvest determines how the interconnect should evolve over the course of the collective, explicitly balancing reconfiguration delay against congestion and propagation delay. We show that the schedules synthesized by Harvest significantly reduce collective completion time across multiple collective algorithms compared to static interconnects and reconfigure-every-step baselines.
Manav Gagvani Sivamurugan Velmurugan Hayden Chang Nysa Pragya Kumar Benjamin Namikas Thomas Joseph Peterson Shrey Sharma Pratyush Mathur	Toward Robust Trajectory Scoring for Parameter-Efficient End-to-End Driving	Recent end-to-end driving systems increasingly utilize large backbones incorporating large vision-language models for planning, limiting their ability to inference in real-time. Recent works model planning as selection among multiple plausible futures rather than regression of a single trajectory. We explore this idea in a parameter-efficient setting for autonomous driving given only camera inputs. After conducting ablations on model inputs, we found that several intuitively useful additions provided limited benefit, motivating a stronger focus on proposal generation and scoring. Our baseline method uses a lightweight visual backbone, generates multiple candidate trajectories, and predicts a scalar error proxy for each proposal. It demonstrates high performance on the Waymo Vision-based End-to-End Driving dataset, achieving a rater feedback score of 7.683 and ranking 12th/50 on the 2025 leaderboard. To improve the robustness and generalizability of trajectory scoring, we investigate a data augmentation strategy that constructs synthetic suboptimal proposals, shaping the scoring manifold to be smoother and more suitable as a direct optimization target.	Poster
Mashiat Mustaq	Globscope; Toward a Global View of the Loss Landscape	Understanding the structure of neural network loss landscapes is important for gaining insight into model merging, hyperparameter selection, generalization, and the relationships between distinct solutions. Visualizing the structure of loss landscapes is very challenging because of the high dimensionality of the parameter space of neural networks. Prior work has primarily focused on visualizing the loss landscape around one single solution or basin, missing how different minima or basins relate to each other. We introduce Globscope, a framework for providing a global view of the loss landscape across multiple solutions or basins. Globscope learns a low-dimensional non-linear manifold of model parameters using an autoencoder framework, enabling both latent-space visualization and reconstruction of full model weights. Then it summarizes the relations among minima and connecting regions on this manifold through topological data analysis. Our framework produces continuous, interpretable visualizations that reveal global connectivity patterns in the landscape. We compare Globscope with kernel-based methods and demonstrate how it performs in preserving the global structure across diverse solutions. We further show how Globscope can be used to analyze two applications: revealing global low-loss solution pathways between distinct solutions using mode connectivity algorithms, and visualizing permutation symmetries of different solutions using re-basin approaches.	Poster
Megha Das	Adaptive Cleaning Optimization in Pharmaceutical Processes via Reinforcement Learning	Small-molecule pharmaceutical manufacturing frequently experiences fouling, progressively degrading efficiency in continuous stirred-tank reactors (CSTRs). Traditional fixed-schedule cleaning ignores the nonlinear, time-varying nature of fouling, leading to inefficient downtime and increased operational costs. Cleaning is fundamentally a long-horizon, sequential decision-making problem under uncertainty. While Model Predictive Control (MPC) is the industry standard, its deterministic nature and high online computational demands limit its effectiveness against stochastic fouling. Formulating this as a Markov Decision Process enables reinforcement learning (RL) to offer a pre-trained, real-time decision-making alternative. Deep RL agents are trained offline using a physics-informed digital twin of a CSTR subjected to stochastic fouling. By balancing exploration and exploitation, the agent optimizes the trade-off between production yield, efficiency, and downtime. For benchmarking, the RL strategy is evaluated against fixed schedules and MPC-MIP approaches to assess relative, comparable performance. Using a multi-objective reward formulation, the agent navigates the Pareto front to maximize yield while minimizing solvent consumption, offering a viable pathway for optimal industrial cleaning policies.
Muskan Goel	Protein Stability, Abundance, and Oligomerization Constrain the Architecture of Protein Interaction Networks	The structural configuration(s) that proteins adopt ultimately dictate their propensity to interact with other proteins. Proteins that persist in a discrete conformational state may form a limited number of specific interactions while those that sample a broader ensemble of disordered structures may instead associate with numerous other clients. These intrinsic tendencies potentially constrain the manner in which proteins navigate wider interaction networks. In this work, we aggregated and surveyed a wide variety of biophysical, biochemical, and cellular descriptors of the S. cerevisiae proteome in order to identify biases in the connectivity of the protein interaction network. Using mass spectrometry-based interactome measurements and various protein stability estimates, we find that a disproportionate number of abundant, and unstable proteins act as network hubs. Moreover, an analysis of supervised machine learning models trained to discriminate between hubs and non-hub proteins identify these two most predictive features. Interestingly, we find that half-lives of hub proteins depend on whether they tend to form static complexes or instead participate in a wider array of dynamic interactions. Moreover, dynamic hubs have a significantly higher proportion of chaperone interaction compared to dynamic hubs. Finally, we note that the observed connectivity biases associated with abundant, unstable proteins only pertain to network hubs, but not to the bottlenecks that connect them
Omkar Mamidpalliwar	Unpacking introductory students’ epistemic engagement with simulation and AI- integrated learning environment in physics	As artificial intelligence (AI) becomes increasingly integrated into students’ academic practices, studies have raised concerns regarding its potential influence on learning. While contemporary research highlights risks associated with AI-supported learning, less is known about the learner characteristics that shape productive and personalized use of these tools. We investigate how students’ epistemological beliefs, i.e., beliefs about the nature of knowledge and learning, mediate their engagement with AI. As part of an ongoing study, we developed an online learning environment that combines interactive simulations with a custom-built AI platform to support students’ sensemaking of introductory physics concepts. We analyze multiple data from around 1000 students, including their responses to the Epistemological Beliefs Assessment for Physical Sciences (EBAPS) survey, the number of questions posed to the chatbot, the number of simulation interactions, and time spent engaging with simulations. This work aims to explore how epistemological beliefs shape AI-mediated learning behaviors and to inform the design of more effective, learner-centered AI-supported physics environments.
Paridhi Agarwal	Advancing Domain-Centric Tutors in LLM Frameworks	Large Language Models (LLMs) have demonstrated significant potential in advancing educational technology through personalized learning and intelligent tutoring systems. This project focuses on enhancing existing LLM-based AI tutors and refining them to serve specific academic domains in order to improve pedagogical effectiveness. The research conducted to augment the tutor includes performing focused literature reviews on LLM-driven tutoring models to identify desirable features that were lacking in the current system. These insights inform strategies to optimize system performance and improve instructional quality. Enhancements include using a fine-tuned training dataset built on a curated knowledge base, incorporating structured teaching dialogues and sample prompts, extracting essential information from domain-specific sources, and enabling the system to interpret visual and symbolic data such as charts, diagrams, graphs, and mathematical formulas. Based on these findings, we design and implement research-informed feature improvements configured to meet system and deployment requirements. These enhancements provide structured problem-solving guidance and contextualized explanations tailored to students’ queries with improved accuracy and content relevance. The resulting system offers a cost-effective, scalable and adaptable framework for customized tutoring across multiple fields of study.	Poster
Pinaki Mohanty	Slithering through Gaps: Capturing Discrete Isolated Modes via Logistic Bridging	High-dimensional and complex discrete distributions often exhibit multimodal behavior due to inherent discontinuities, posing significant challenges for sampling. Gradient-based discrete samplers, while effective, frequently become trapped in local modes when confronted with rugged or disconnected energy landscapes. This limits their ability to achieve adequate mixing and convergence in high-dimensional multimodal discrete spaces. To address these challenges, we propose \emph{Hyperbolic Secant-squared Gibbs-Sampling (HiSS)}, a novel family of sampling algorithms that integrates a \emph{Metropolis-within-Gibbs} framework to enhance mixing efficiency. HiSS leverages a logistic convolution kernel to couple the discrete sampling variable with the continuous auxiliary variable in a joint distribution. This design allows the auxiliary variable to encapsulate the true target distribution while facilitating easy transitions between distant and disconnected modes. We provide theoretical guarantees of convergence and demonstrate empirically that HiSS outperforms many popular alternatives on a wide variety of tasks, including Ising models, binary neural networks, and combinatorial optimization.	Poster
Pinaki Mohanty	Entropy-Guided Sampling of Flat Modes in Discrete Spaces	Sampling from flat modes in discrete spaces is a crucial yet virtually untouched direction. Flat modes represent robust solutions and have broad applications in combinatorial optimization and discrete generative modeling. However, existing sampling algorithms often overlook the mode volume and struggle to capture flat modes effectively. To address this limitation, we propose Entropic Discrete Langevin Proposal (EDLP), which incorporates local entropy into the sampling process through a continuous auxiliary variable under a joint distribution. The local entropy term guides the discrete sampler toward flat modes with a small overhead. We provide non-asymptotic convergence guarantees for EDLP in locally log-concave discrete distributions. Empirically, our method consistently outperforms traditional approaches across tasks that require sampling from flat basins, including Bernoulli distributions, restricted Boltzmann machines, combinatorial optimization, and binary neural networks.	Poster
Pretty Mitra Alvaro Miguel	Data-driven Safety Analytics of Lithium-Ion Batteries	Accurate prediction of thermal runaway is crucial for optimizing the reliability and safety of lithium-ion batteries (LIBs). Traditional physics-based models often fall short due to the complex multi-scale, multi-phase interactions during thermal runaway. To address this, we present a data-driven framework using artificial intelligence (AI)-driven machine learning to predict thermal stability of LIBs. Experimental data from accelerating rate calorimetry (ARC) is used to analyze and predict key safety characteristics, including thermal runaway onset and heat release, across diverse cells with varying state-of-charge, chemistries, and aging conditions. We extract four critical thermal safety metrics: self-heating onset temperature (T_1), thermal runaway onset temperature (T_2), peak temperature (T_3), and end-of-window temperature (T_w), providing a comprehensive view of thermal behavior under abuse conditions. Input features include cell metadata and ARC-derived observations, enabling robust predictions despite limited data. Our AI-driven LSTM-based pipeline captures nonlinear relationships and integrates time-series learning with experimental and physics-informed data, offering scalable and accurate prognostics. This approach demonstrates the potential of data fusion as an alternative to traditional models, enabling improved battery safety assessment, diagnostic tools, and design strategies for safer energy systems.
Rahul Prabhu Rishikesh Madhuvairy	Fusing Handcrafted Spatial Descriptors with a Lightweight CNN for Semiconductor Wafer Map Defect Classification	Automated defect classification in wafer maps is critical for semiconductor yield management and quality control, but pure deep learning models often underperform on rare or spatially subtle defect types and lack interpretability. Handcrafted spatial features can capture physical defect characteristics, yet their integration with modern CNNs is underexplored. We systematically evaluate 8 physically motivated spatial descriptors: radial mean and std dev, directional entropy, aspect ratio, fail fraction, and zone‑wise failure densities, by training a lightweight CNN (101k parameters) augmented with each descriptor, both individually and in full combination. To the best of our knowledge, this is the first systematic ablation study to quantify the synergistic effect of fusing physically-informed spatial descriptors with a modern, edge-optimized CNN for this task On the WM‑811K benchmark (8 defect classes, 25k labeled wafers), the vision‑only baseline achieves 60% test accuracy and 0.615 weighted F1. Nearly every single descriptor individually underperforms the baseline, with the best descriptor (fail fraction) reaching only 60.1%. However, the full fusion of all eight descriptors significantly outperforms the baseline, reaching 72.7% accuracy (+12.7 pt) and 0.728 wF1. This synergy demonstrates that descriptors provide complementary information that is only realizable in combination. Per‑class analysis reveals that the combination model substantially improves challenging classes.
Rajiv Ranjan Udaiveer Singh Anjali Aggarwal Shashank Tamaskar Dharmendra Saraswat	SPEAR: Self-Supervised Sample Efficient Pixel-Level Multi-Modal Spectral Fusion for Earth Observation Applications	SPEAR is a pixel-level, multi-modal foundation framework for Earth Observation. It combines optical imagery, radar, and climate data into one system that supports agriculture, environmental monitoring, and engineering planning, even when some data are missing. With about one million parameters, it delivers a highly efficient, lightweight analysis for crops, soil, and infrastructure.	Poster
Taewoong Yoon Soomin Lee Arghadip Das	Exploiting Temporal Similarity for Energy-Efficient Audio Event Classification	Audio event classification (AEC) enables environmental awareness in applications such as sensing, surveillance, and occupancy monitoring, but deploying AEC on battery-powered microcontrollers (MCUs) remains challenging due to tight energy and compute budgets. Although acoustic ML models are relatively compact, continuous feature extraction and fixed-rate inference still consume substantial power, especially on low-cost ultra-low-power MCUs with limited acceleration support. We observe that real-world audio streams exhibit strong temporal similarity: adjacent clips from the same event often produce similar feature tensors, while different events are more separated in representation space. Motivated by this, we investigate a distance-guided early-exit AEC architecture that adaptively selects inference depth based on tensor similarity rather than classifier uncertainty alone. We show that tensor distances correlate with sample difficulty and the depth needed for correct classification. Building on this insight, we design a lightweight gating mechanism for multi-exit networks that reduces average inference cost and energy while preserving accuracy, enabling more practical always-on AEC on commodity ultra-low-power platforms.
Tzu-En Feng Pin-Chen Su	TalentOne: Quantifying Semantic Vector Matching Against Human HR Baselines in Enterprise Recruitment	1. Problem Statement Enterprise recruitment is currently crippled by a dual-sided data failure: hiring managers provide vague job descriptions, while candidates submit undifferentiated resumes. HR professionals are caught in the middle, relying on legacy Applicant Tracking Systems (ATS) that use brittle boolean keyword matching (e.g., equating “watched an AWS video” with “architected AWS at scale”). The objective of this research is to replace keyword matching with a Semantic Vector Engine and quantitatively validate its performance against human recruiter baselines. 2. Methodology & System The TalentOne architecture utilizes multimodal LLMs to extract contextual meaning from resumes, bypassing traditional PDF parsing errors. Crucially, this project includes a quantitative validation phase: an AI-vs-human ranking test. The system’s output will be compared against manual rankings produced by HR professionals at SYSTEX Corporation (a 5,000+ employee enterprise). Success is defined by achieving a correlation of >= 0.6. 3. Real-World Impact By evaluating the correlation between semantic vector space and human intuition, this research bridges the gap between theoretical LLM capabilities and practical enterprise workflows. This creates hiring ecosystem where candidates are viewed based on contextual impact of their experience, rather than their ability to optimize algorithms
Udeme Idem Justin Hess Robert Loweth	Scaffolding or Integrating? How Engineering Faculty Are Designing Generative AI into Learning	Generative AI tools like ChatGPT have thrown engineering faculty into unfamiliar territory. Students can now generate code, draft reports, and brainstorm design concepts in seconds. This fact raises hard questions about what learning looks like when a machine can do so much of the work. We hear a lot about Gen-AI policy at the institutional level, but less about how individual instructors are actually handling these questions in their own classrooms. This poster shares preliminary findings from an ongoing NSF study on “Prompting Socially Engaged Divergent Thinking in Engineering Design by Leveraging Generative AI”. The research asks: What approaches are engineering faculty taking to shape generative AI policy in their classrooms, and what values and concerns guide their decisions? To investigate these questions, we conducted semi-structured interviews with 17 engineering faculty across U.S. universities, all of whom teach design-focused courses. Using thematic analyses to identify, interpret, and characterize patterns in how engineering faculty understand, negotiate, and implement generative AI policies in their classrooms. Some instructors use a Guided Use paradigm: they actively scaffold student learning. Another paradigm is the Pedagogical Alignment paradigm: Gen-AI is intentionally integrated into assignments to achieve specific learning outcomes. This study offer a vocabulary for reflecting on when, why, and how we might integrate or restrict Gen-AI in our future class.
Yansong Qu Jiancong Chen	Found-RL: foundation model-enhanced reinforcement learning for autonomous driving	Reinforcement learning (RL) enables end-to-end autonomous driving (AD) with real-time inference, but often suffers from sample inefficiency and limited semantic interpretability in complex scenarios. Foundation models, especially vision-language models (VLMs), offer rich contextual knowledge, yet their high inference cost makes integration into high-frequency RL training impractical. To address this, we present Found-RL, a platform for enhancing RL for AD with foundation models. Its core innovation is an asynchronous batch inference framework that decouples VLM reasoning from the simulation loop, removing latency bottlenecks and enabling real-time RL learning from VLM feedback. We further introduce multiple supervision mechanisms. For action guidance, we develop Value-Margin Regularization (VMR) and Advantage-Weighted Action Guidance (AWAG) to distill expert-like VLM suggestions into the policy. For dense supervision, we use high-throughput CLIP for reward shaping and address its dynamic blindness and probability dilution through Conditional Contrastive Action Alignment, which conditions prompts on discretized speed and command to produce a normalized margin-based bonus from action-anchor scoring. Found-RL provides an end-to-end modular pipeline for VLM-enhanced RL and shows that a lightweight RL model with millions of parameters can achieve near-VLM performance while sustaining real-time inference (~500 FPS). Code, data, and models: https://github.com/ys-qu/found-rl
Zitao Song Cedar Site Bai Zhe Zhang Brian Bullins David F. Gleich	Decoupling Variance and Scale-Invariant Updates in Adaptive Gradient Descent for Unified Vector and Matrix Optimization	Adaptive methods like Adam have become the \textit{de facto} standard for large-scale vector and Euclidean optimization due to their coordinate-wise adaptation with a second-order nature. More recently, matrix-based spectral optimizers like Muon \citep{jordan2024muon} show the power of treating weight matrices as matrices rather than long vectors. Linking these is hard because many natural generalizations are not feasible to implement, and we also cannot simply move the Adam adaptation to the matrix spectrum. To address this, we reformulate the AdaGrad update and decompose it into a variance adaptation term and a scale-invariant term. This decoupling produces \textbf{DeVA} (\textbf{De}coupled \textbf{V}ariance \textbf{A}daptation), a framework that bridges between vector-based variance adaptation and matrix spectral optimization, enabling a seamless transition from Adam to adaptive spectral descent. Extensive experiments across language modeling and image classification demonstrate that DeVA consistently outperforms state-of-the-art methods such as Muon and SOAP \citep{vyas2024soap}, reducing token usage by around 6.6\%. Theoretically, we show that the variance adaptation term effectively improves the blockwise smoothness, facilitating faster convergence. Our implementation is available at \url{https://github.com/Tsedao/Decoupled-Variance-Adaptation}
Zonghan Lei	Unsupervised Chemical Partitioning: An AI-Driven Approach for Mapping Pathological Metal Patterns in Neurodegenerative Tissue	The spatial distribution of metals is a critical indicator of neurodegenerative diseases like Alzheimer’s, yet manual identification of these complex chemical environments is labor intensive. This research presents an AI-driven framework that utilizes unsupervised machine learning to segment multi-elemental X-ray Fluorescence (XRF) data into biologically significant “chemical phases.” We implemented K-Means clustering on high-dimensional XRF scatterplot data derived from human neurodegenerative tissue, which grouped the pixels (area of tissue measured in microns represented by each row of data) into distinct chemical signatures. These phases were projected onto spatial maps, color-coding pixels to visualize the architectural distribution of the tissue’s chemistry. Our results successfully identify complex elemental co-localizations, pinpoint toxic heavy metal accumulations, and reveal recurring pathological patterns, transforming raw spectral data into a navigable map of disease progression. The next stage involves correlating these identified clusters with direct biomarkers of Alzheimer’s. By integrating our spatial maps with chemical staining, we aim to validate specific elemental “proxies” for Amyloid-beta (Aβ) and Tau proteins. Our goal is to train a supervised model capable of predicting protein aggregate locations based solely on unique metal fingerprints, providing a non-destructive, AI-based diagnostic tool for neurodegenerative research.
Seoljoo Kang	Let’s Build with GenAI: Designing Learning Experiences Through Computational Thinking and Planning	This poster presents three initiatives repositioning Generative AI (GenAI) as a proactive educational collaborator. Grounded in self-directed learning, these projects foster learner autonomy and reduce technical barriers through technology-integrated cycles. The first initiative, “Let’s Code with GenAI,” provided 34 K-12 educators with an asynchronous module that used Gemini for real-time coding and computational thinking (CT) support. Results demonstrated significant gains in teaching efficacy (p < .001) and coding/CT performance (p = .040), suggesting that GenAI can simultaneously build confidence and technical skills. The second initiative, “Coding and CT with Micro:bit and GenAI,” extended this learning through hands-on Python activities with physical hardware, reinforcing educators’ confidence for classroom integration. Finally, the Intelligent Academic Scaffold System (IASS) converts rigid doctoral handbooks into dynamic “AI Plan of Study” roadmaps. By simulating a Graduate Coordinator persona via Chain-of-Thought prompting, the system enforces complex credit rules while providing personalized suggestions. This work reduces hidden-curriculum barriers for graduate students, offering a scalable infrastructure for academic advising. Together, these service-learning experiences demonstrate a commitment to designing inclusive, ethically grounded AI support for complex academic tasks.	Poster
Saeka Rahman Reepa Saha Md Motiur Rahman Ardavan Vakil Hope Teng Muhammad Arafin Khan Benjamin Larimer Rita Basu Andy Basu Amy Warriner Miad Faezipour S. Abdollah Mirbozorgi	Edge AI-Enabled Vagus Nerve Analysis for Adaptive Neuromodulation in Diabetes	Purpose/Aim Comprehending the neural mechanisms underlying Type 2 Diabetes (T2D) is critical to developing next generation bioelectronic therapies. This research work-in-progress presents a novel neural interface that integrates edge artificial intelligence (AI), machine learning (ML), and flexible implantable electronics to record, analyze, and modulate vagus nerve activity in real time for potential diabetes treatment and reversal. Methods The proposed system combines low-power, semi-analog signal processing with embedded AI/ML algorithms capable of performing on-chip feature extraction and classification of neural signals under stringent power and communication constraints. By leveraging edge-based learning, the platform will enable adaptive neuromodulation of vagus nerve activity, providing a closed-loop framework for studying the stomach-brain-pancreas signaling involved in diabetes. The edge AI-driven analysis facilitates localized inference and decision-making directly at the neural interface. Results/Discussion Validation in small animal models aims to correlate vagal signal dynamics with metabolic states, laying the foundation for energy-efficient, autonomous implants that could support diabetes reversal through intelligent, personalized neural control.
Garvit Agarwal Yousef Mohammed Y. Alomayri Seunghyun Cho Feng Li Xukai Zou	Temporal Split Learning: A Privacy-Preserving Intrusion Detection Framework for Distributed Medical IoT Networks	The rapid digitization of healthcare has led to an explosion of Internet of Medical Things (IoMT) devices, creating a complex and vulnerable attack surface. While Intrusion Detection Systems (IDS) are critical for defense, traditional centralized approaches require the aggregation of sensitive network telemetry, often violating privacy regulations such as HIPAA and GDPR. Federated Learning (FL) offers a privacy-preserving alternative but imposes significant computational and bandwidth overheads unsuitable for resource-constrained medical edge devices. To address these challenges, we propose Temporal Split Learning (TSL), a novel distributed deep learning framework tailored for IoMT security. TSL partitions the neural network execution between the edge (client) and the cloud (server), transmitting only obfuscated intermediate activations (“smashed data”) rather than raw data or model weights. We enhance standard Split Learning with Recurrent Neural Networks (RNNs) to capture the temporal sequentiality of Advanced Persistent Threats (APTs) and malware behavior. We evaluate TSL using a high-fidelity, 50-node virtualization testbed generating realistic HL7 and DICOM traffic mixed with live malware detonation. Our results demonstrate that TSL achieves a 98.4% detection accuracy, which is comparable to centralized methods, while reducing CPU consumption by over 94% compared to Centralized and Federated Learning. Furthermore, we conduct a rigorous Model Inversion Attack (MIA) analysis to quantify privacy leakage, demonstrating that our temporal smashing technique significantly increases the difficulty of raw feature reconstruction by adversarial servers.
Shiva Shokouhmand Md Motiur Rahman Saeka Rahman Indya Mooney Guoliang Zhang Smriti Bhatt Namasivayam Ambalavanan Colm P. Travers S. Abdollah Mirbozorgi Miad Faezipour	AI-Powered Pacifier with Intraoral Sound Analysis for Neonatal Pulmonary Monitoring	Respiratory monitoring in neonates remains a clinical challenge due to the limitations of existing pulmonary function assessment tools and the infant’s delicate physiology. This work-in-progress presents an artificial intelligence (AI)-driven framework for intraoral respiratory analysis using a smart pacifier platform that integrates acoustic sensing, embedded electronics, and deep learning. The proposed AI model is based upon a deep convolutional neural network that processes intraoral breathing sound reflections passively, or in presence of active sound stimuli. By learning complex temporal-spectral features from the breathing sound signals, the AI model will estimate key cardiorespiratory measures (e.g., respiratory rate, lung capacity) and the pulmonary condition (e.g., existence of underlying obstructive respiratory condition) in real time. Preliminary results highlight the correlation between breathing sound patterns and respiratory mechanics, demonstrating the potential feasibility of continuous, noninvasive pulmonary monitoring in neonatal settings. This AI-enabled intraoral sound sensing approach establishes a foundation for intelligent neonatal respiratory assessment, enhancing early detection and personalized management of pulmonary disorders.
Andres Torres Jorge Neira Taehee Kim Darren Chang	Integrated Design of a 2-DOF Robotic Ankle-Foot Prosthesis with Sensing, Control and Vision	Transtibial amputees need prostheses that support multidirectional locomotion, yet most powered devices leave frontal-plane inversion-eversion (IE) dynamics uncontrolled during turning. This work presents a proportional gait turn controller for the CABLE-Prosthesis, a two-degree-of-freedom cable-driven ankle-foot prosthesis with active dorsiflexion-plantarflexion and IE actuation. A shank-mounted IMU measures stride-level foot heading change to continuously interpolate Fourier-series-encoded IE trajectories across straight walking, spin turns, and step turns. The controller operates within a hierarchical architecture combining a phase portrait estimator, Kalman-filtered phase smoothing, and admittance control. Human-subject trials (N=4) confirmed statistically significant IE kinematic differences between assisted and unassisted turning (p<0.05) without disrupting sagittal-plane dynamics. Future work will focus on implementing computer vision for human motion prediction, including real-time foot placement tracking and obstacle detection to enable proactive adaptation during complex terrain negotiation.
Yoonhyuck Woo	Probing Knowledge Graph Reliability and Semantic Coherence with Language Models	Knowledge graphs (KGs) are widely used as structured representations that support reasoning, inference, and integration across heterogeneous data sources. Yet, despite their central role in modern AI systems, the extent to which KGs preserve consistent and coherent relational structure remains insufficiently examined. This paper evaluates how well KGs maintain semantic coherence and whether they are sufficiently expressive and complete under realistic constraints on representation formats and available resources. We propose a systematic probing framework that leverages language models in two complementary ways: (1) an embeddingbased analysis that measures the stability of relational semantics across alternative verbalizations, and (2) a ranking-based evaluation that tests the consistency of relational interpretations under controlled prompts. Together, these methods provide an empirical assessment of the robustness of KG semantics. Our results highlight both the strengths and the limitations of KGs as practical semantic representations and offer suggestions for future work on KG evaluation.
Yi Zhang	Do Programmers and AI See the Same Problem? Quantifying Cognitive Misalignment in Code Generation	The integration of AI assistants into software development raises fundamental questions about how task complexity is evaluated and the extent to which these evaluations align with human perception. Current evaluations focus primarily on functional correctness, overlooking this cognitive alignment. We introduce and empirically examine cognitive misalignment: the discrepancy between human and AI perceptions of a task’s cognitive demands. Using Bloom’s Taxonomy, we prompt five LLMs to classify 2,520 tasks from three code generation benchmarks, and establish human reference annotations for 150 tasks via expert consensus. Results show systematic misalignment: humans predominantly classify tasks as “Apply” or “Analyze”, whereas several LLMs overestimate the “Create” dimension. This gap varies by model and task type and may contribute to observed interaction frictions and productivity paradoxes. Our findings motivate the development of cognitively aware benchmarks and evaluation methods that better reflect human judgments of task complexity.
Jui-Cheng Chiu Yu-Chao Wang Shengyang Luo Qi Yang Tongyan Wang Nabin Khanal Dr. Yingjie Victor Chen	MIRAGE: A Micro-Interaction Relational Architecture for Grounded Exploration in Multi-Figure Artworks	Appreciating multi-figure paintings requires understanding how characters relate through subtle cues such as gaze, gesture, and spatial arrangement. We present MIRAGE, an evidence-centric framework that scaffolds exploration of these micro-interactions. These cues are often distributed across complex scenes and difficult to systematically identify, while existing vision-language models (VLMs) frequently produce ungrounded interpretations without traceable visual evidence. MIRAGE constructs a structured intermediate representation capturing identities, pose, and gaze hypotheses, enabling explicit coordination of relational evidence. Without such organization, models tend to collapse multiple interaction hypotheses into unstable narratives, even when low-level signals are available. Our representation allows users to verify how high-level interpretations are anchored in visual evidence. By separating spatial grounding from narrative generation, MIRAGE supports inspection and reasoning over figure relationships through a verifiable evidence layer. Evaluation against painting-only VLM baselines shows improved identity consistency, reduced relational hallucinations, and greater coverage of subtle interactions, demonstrating the value of structured grounding for transparent, human-led visual understanding.
Nabin Khanal Jui-Cheng Chiu Tongyan Wang	LLM Augmented Forestry Document Digitization	Digitizing complex, domain-specific documents is challenging, especially when documents contain handwritten text and irregular tables. OCR systems often fail on such data. Vision-based large language models (LLMs) have improved the accuracy, but when these models fail, prompting the LLMs to fix errors is often difficult. Making corrections across scattered locations or targeting a specific table cell is cumbersome, and such localized instructions are often not understood by LLMs. To address this, we present a human-AI collaborative system that combines layout parsing, OCR, and LLMs for interactive refinement. Users can provide instructions for global edits or sample changes for localized fixes, while also being able to propagate changes across the document, reducing repetitive effort. Beyond accurate digitization, we found that users explored creative manipulations, such as extracting data from visualizations or applying enhancements to improve presentation. Our findings show that transparent human-AI collaboration enables efficient, reliable digitization empowering users to re-shape documents.
Hao Wang	Agentic AI for Smart Analytical Visualization of Clinical Studies Using Graph Data Models on Research Data Commons	Analytic visualization of longitudinal clinical data requires understanding study designs, temporal concepts, and analysis strategies. In cBioPortal, predefined pipelines suffice for similar designs. However, data commons like ARDaC and IPO host heterogeneous studies, making cross-study visualization labor-intensive, error-prone, and hard to reproduce. LLMs could automate this process, but existing tools hide intermediate reasoning from users, often misreading temporal designs and lacking protocol awareness. Graph data models in these commons require graph-aware query generation (e.g., GraphQL), limiting current tools. We propose an agentic visualization framework where reasoning agents interpret study protocols, understand user intent, and generate clinically interpretable visualizations. Agent-based decomposition with protocol- and schema-aware retrieval improves temporal and semantic accuracy beyond single-agent or template-driven systems. It incorporates study protocols and data structures, with contextual retrieval, rule-based validation, and secure execution, ensuring transparency. An agile process with rapid iteration and feedback shapes a user-aligned methodology. No existing framework integrates protocol-aware cohort reasoning, multi-agent LLM orchestration, and schema-constrained visualization generation in a data commons. The system will be demonstrated on ARDaC (liver disease) and IPO (oncology) datasets, integrated with the ARDaC Analytical Engine (Anagine).
Professor Nikhilesh Chawla Satyaroop Patnaik Eshan Ganju Hamidreza Torbati Sarraf Hamdan Ashfaq	Harnessing the Power of AI in 3D X-ray Microtomography	X-ray computed tomography (XCT) is a unique technique for three-dimensional microstructural characterization. Post-acquisition workflows are inefficient due to acquisition constraints, noise, and labor-intensive image analysis. This work presents a suite of deep learning strategies developed within our group to address these challenges across multiple material systems. We will describe efforts to achieve high-throughput high-fidelity reconstruction from reduced projection datasets, directly enabling faster 4D acquisitions without sacrificing volumetric image quality. GAN (generative adversarial network)-based models were employed to enhance high-throughput low-exposure absorption and diffraction contrast XCT datasets, suppressing noise and artifacts while preserving microstructural features. For semantic image segmentation of 3D microstructural features, UNet++ and a 2.5D ResNet-18 based architecture demonstrated the critical role of data augmentation in improving segmentation generalizability across thermo-mechanical processing steps. Finally, we will describe the use of image classification models that enable automated classification of microstructural patterns. Together, these approaches demonstrate that AI integration at every stage of the X-ray tomography and 3D image analyses pipeline – reconstruction, denoising, segmentation, and classification – substantially improves throughput, reproducibility, and analytical depth in 3D and 4D materials characterization.

Follow Us