2025 UCF Research Computing Symposium

For more information about the Student Poster Session, click here: https://rci.research.ucf.edu/2025-annual-ucf-research-computing-symposium-call-for-posters/
Below are the posters and their presenters:

Revealing the Active Site Local Atomic Environment of Oxide-supported Ag Single Atom Catalyst

Syeda Faiza Rubab Sherazi

Single atom catalyst (SAC) supported on a metal oxide surface is a promising candidate for various reactions. Determining the relationship between the active site’s local atomic coordination and its catalytic performance is important for designing SAC. Here, we apply the ab initio thermodynamics approach to investigate the local coordination of Ag atoms that stabilize on CeO2(ııo), ZrO2(īıı), and Al2O3(ııı) through calculated phase diagrams and examine NH3 dissociation on so determined Ag SAC. We find that for the CeO2(ııo)-supported Ag SAC structure, one surface oxygen vacancy nearby is the most thermodynamically favorable, while with the ZrO2(īıı) and Al2O3(ııı)supported ones, no oxygen vacancy nearby is the most Abstract Body: thermodynamically favorable. Our results also show that oxygen vacancy formation is spontaneous near the Ag atom when supported by CeO2(ııo), while non-spontaneous when supported by ZrO2(īıı) and Al2O3(ııı). We compare the energetics of NH3 adsorption and dissociation on Ag/CeO2(ııo) with those on Ag/ZrO2(īıı) and Ag/Al2O3(ııı), finding that the first hydrogen abstraction is the most facile on Ag/CeO2(ııo). We trace the catalytic behavior of these oxidesupported Ag SACs to the differences in coordination, charge states and the availability of unoccupied density of states of Ag atoms, as well as the distance from a H atom of the adsorbed NH3 to nearby surface O atoms. We will discuss the good qualitative agreement of the above results with experimental data that we have obtained.
Authors: S. Faiza Sherazi, Duy Le, Kailong Ye, Shaohua Xie, Fudong Liu and Talat S. Rahman

A Mid-Res Grid of Contribution Functions Characterizing Brown Dwarf Atmospheres

Myrla Phillippe

Using a state of the art radiative transfer code, cloud modeling code, and UCF’s Stokes high-performance computing (HPC) cluster, we created a grid of contribution functions for model brown dwarfs atmospheres from 500K to 2000K. Our grid covers a wide range of cloud properties, temperatures, and gravities, explicitly considering both cloudy and cloud-free models. This enables us to examine how clouds alter the pressures probed across molecular and atomic features. We will present the implications of our grid for the first systematic pressure-mapping effort of brown dwarf atmospheres across a wide range of brown dwarf properties. We will discuss optimal observation strategies to study the 3D structure of brown dwarf atmospheres. We will demonstrate how our grid can contribute to the characterization of 3 brown dwarfs using archival HST ( Hubble Space Telescope ) observations, and discuss how we can use it for planning JWST (James Webb Space Telescope ) observations. The results of this work have been used to produce a user-friendly tool hosted on the cloud-based Jetstream2 (NSF ACCESS) and will be made available to the community through a UCF domain, ensuring secure access.
Authors: Myrla Phillippe, Theodora Karalidi, Elena Manjavacas, Kieran M Manjrawala, Natalia Oliveros Gomez, Jonathan Fernandez

Dynamic Circuit Compilation for Sparse Qubit Connectivity

Sumeet Shirgure

In this work, we show how dynamic circuit compilation methods can transform a densely connected circuit into an ancilla-mediated, sparsely connected dynamic circuit that includes mid-circuit measurement and feedforward operations.
This sparse connectivity is better suited to current quantum hardware, which often has limited qubit connectivity. Compared to compilation methods focused on unitary circuits, our method can reduce both circuit depth and the additional CNOTs required to execute non-adjacent qubit interactions.
Authors: Sumeet Shirgure, Siyuan Niu

Leveraging Cloud Infrastructure and AI APIs for Scalable E-Learning Production

Margaret Zorrilla

This project explores the integration of cloud computing and artificial intelligence (AI) to develop a scalable, reproducible workflow for automated e-learning video production. Conducted through a collaboration between the UCF Hospitality+ Center for Innovation and Training and the UCF Research Cyberinfrastructure team, the initiative leverages cloud-native architecture to streamline instructional content development while preserving academic rigor.
A two-stage Extract–Transform–Load (ETL) pipeline was implemented using Amazon Web Services (AWS), Azure OpenAI, and Synthesia. In Pipeline 1, faculty-generated training script objectives—aligned with Bloom’s taxonomy—along with contextual examples and instructional framing, were uploaded in MS Word format to AWS. These documents were automatically extracted and transformed into structured JSON files. Azure OpenAI expanded and validated the content into human-readable scripts. A human-in-the-loop review followed to ensure instructional clarity, pedagogical integrity, and alignment with learning outcomes.
In Pipeline 2, the validated Word documents were uploaded to AWS, reformatted into JSON, and transmitted via API calls to Synthesia. This enabled fully automated video generation with standardized outputs featuring customizable AI avatars, multilingual narration, and dynamic graphic elements. Faculty-specific avatars and voices were developed to enhance authenticity, while template customization supported diverse instructional goals and branding.
This architecture demonstrates how cloud-enabled AI workflows can extend beyond traditional computational domains to support education and workforce training. The integration of automation with human oversight offers a reproducible model for scalable digital content creation across disciplines.
Authors: Dr. Manuel Rivera, Assistant Dean of Research Rosen College, Director UCF Hospitality+ Center for Innovation and Training, Margaret Zorrilla (MA Interdisciplinary Studies, Project Manager UCF Hospitality+ Center for Innovation and Training), Ezequiel Gioia (UCF Research Cyberinfrastructure), Paola Canales-Bigio (UCF Research Cyberinfrastructure)

Platform-Agnostic Modular Architecture for Quantum Benchmarking: Enabling Custom Execution and Flexible Analysis

Neer Patel

We present a platform-agnostic modular architecture that addresses the increasingly fragmented landscape of quantum computing benchmarking by decoupling problem generation, circuit execution, and results analysis into independent, interoperable components. Supporting over 20 benchmark variants ranging from simple algorithmic tests like Bernstein-Vazirani to complex Hamiltonian simulation with observable calculations, the system integrates with multiple circuit generation APIs (Qiskit, CUDA Quantum, Cirq) and enables diverse workflows. We validate the architecture through successful integration with Unitary Foundation’s metriq-gym for execution and results publishing, Sandia’s pyGSTi for advanced circuit analysis, and CUDA Quantum for multi-GPU HPC simulations. Extensibility of the system is demonstrated by implementing dynamic circuit variants of existing benchmarks and a new quantum reinforcement learning benchmark, which become readily available across multiple execution and analysis modes. Our primary contribution is identifying and formalizing modular interfaces that enable interoperability between incompatible benchmarking frameworks, demonstrating that standardized interfaces reduce ecosystem fragmentation while preserving optimization flexibility. This architecture has been developed as a key enhancement to the continually evolving QED-C Application-Oriented Performance Benchmarks for Quantum Computing suite.
Authors: Neer Patel, Anish Giri, Hrushikesh Pramod Patil, Noah Siekierski, Vincent Russo, Changhao Li, Alessandro Cosentino, Nathan Shammah, Avimita Chatterjee, Sonika Johri, Timothy Proctor, Thomas Lubinski, Siyuan Niu

Prompt Optimization Meets Subspace Representation Learning for Few-shot Out-of-Distribution Detection

Faizul Rakib Sayem

The reliability of artificial intelligence (AI) systems in open-world settings depends heavily on their ability to flag out-of-distribution (OOD) inputs unseen during training. Recent advances in large-scale vision-language models (VLMs) have enabled promising few-shot OOD detection frameworks using only a handful of in-distribution (ID) samples. However, existing prompt learning-based OOD methods rely solely on softmax probabilities, overlooking the rich discriminative potential of the feature embeddings learned by VLMs trained on millions of samples. To address this limitation, we propose a novel context optimization (CoOp)-based framework that integrates subspace representation learning with prompt tuning. Our approach improves ID-OOD separability by projecting the ID features into a subspace spanned by prompt vectors, while projecting ID-irrelevant features into an orthogonal null space. To train such OOD detection framework, we design an easy-to-handle end-to-end learning criterion that ensures strong OOD detection performance as well as high ID classification accuracy. Experiments on real-world datasets showcase the effectiveness of our approach.
Authors: Faizul Rakib Sayem, Shahana Ibrahim

Computational Pathways to AI Literacy: Student Engagement with Generative AI in Hospitality Education

Marcelo Viejo Rubio

The rapid integration of generative artificial intelligence (AI) into higher education raises urgent questions about how students perceive, apply, and evaluate these tools in professional training contexts. This project examines computationally enabled learning in HMG 6449 Smart Travel Tourism, a graduate course where 62 students engaged in AI-driven assignments across two semesters (Fall 2023 and Fall 2024). Students generated marketing campaigns, designed sustainable tourism proposals, and critically reflected on AI’s role in shaping creativity, judgment, and career identity.
To analyze these data, we applied a six-step thematic analysis process supported by computational text analysis and qualitative coding, producing a conceptual model of student AI engagement. Results demonstrate how students developed selective prompting strategies, multi-model consensus methods, and bias-detection skills, which function as computational practices that sharpen critical thinking and reveal discipline-specific challenges such as localization, cultural adaptation, and ethical boundary-setting. Students further internalized AI use as part of their professional identity, reporting both increased career confidence and moments of impostor syndrome modulation.
The study contributes to the design of scalable AI-integrated curricula, including a scaffolded AI-literacy sequence, cultural-sensitivity modules, and reflective identity journals. Beyond hospitality education, this work illustrates how computational tools can be systematically embedded into coursework to enhance AI literacy, foster responsible adoption, and align student learning with industry-ready competencies. By bridging pedagogy and computation, the project demonstrates pathways for expanding AI training across disciplines and professions.
Authors: Marcelo Viejo Rubio, Dr. Arthur Huang

Gaming for Justice: A Gamified Storytelling Mobile Prototype App for Hospitality Anti-Human Trafficking Interventions

Aili Wu

Human trafficking remains a critical challenge in hospitality, one of the main venues for this crime. Yet no prior research has tested interactive ways to help hospitality consumers (often key bystanders) recognize and respond to trafficking. This study addresses that gap by designing and testing a gamified storytelling mobile app that builds knowledge and confidence.
The app combines four elements to capture attention and deepen learning. Avatar embodiment places players in the role of a detective, giving them a first-person view of the investigation and creating an immersive, playful experience. A clear, suspenseful storyline moves from noticing suspicious activity to catching the trafficker, keeping the game exciting yet easy to follow. Vivid visuals make scenes feel real and memorable, while interactive choices let players steer the story and control the action. A total of 378 hospitality consumers recruited through Prolific played the game and completed a survey. Analysis using PLS-SEM shows that this gamified storytelling approach sparks both enjoyment and cognitive engagement, creating dual pathways that significantly enhance learning and memory. Participants reported stronger knowledge retention, greater confidence in handling trafficking scenarios, and higher willingness to intervene as bystanders. This research shows that gamified storytelling goes beyond training hospitality consumers. It offers a scalable, engaging approach to raise public awareness and inspire action against human trafficking. Breaking communication barriers around this sensitive issue empowers individuals, supports industry efforts, and informs the wider community and policy initiatives, contributing to meaningful social change well beyond the hospitality sector.
Authors: Aili Wu; Dr. Wei Wei

Machine Learning-Accelerated Discovery of Single-Atom Catalysts for Hydrogen Evolution Reaction

Chidozie Ezeakunne

The hydrogen evolution reaction (HER), half reaction in water splitting, plays a crucial role in the efficient generation of green hydrogen, a key step toward building a sustainable energy economy. However, progress has been impeded by reliance on costly and scarce precious metal-based catalysts, highlighting an urgent need to identify high-performance, cost-effective alternatives. To address this challenge, this work introduces a computational framework that combines first-principles density functional theory (DFT) calculations with machine learning (ML) to explore the vast chemical space of single-atom catalysts (SACs) supported on low-cost transition metal nitride and carbide supports. Metal carbides and nitrides were chosen as hosts given their intrinsic stability and electronic properties as support for atomically dispersed atoms to maximize catalytic efficiency. Our high-throughput DFT screening systematically evaluated ~2,000 unique structures. Gibbs free energy of hydrogen adsorption (ΔGH∗) was calculated as the primary descriptor of HER activity, where a value close to 0 eV indicates an optimal balance of H binding strength. This comprehensive screening identified several highly active HER catalyst candidates with near-ideal ΔGH∗ values. Using the DFT calculated data as a training set, we developed a predictive ML model to accelerate future discovery of single-atom catalysts with accuracy comparable to DFT calculations. Our ML model successfully learns the underlying physical relationships between elemental properties and catalytic activity, enabling the rapid and accurate screening of new SACs.
Authors: Chidozie Ezeakunne, Shyam Kattel

Inverse Scattering Problem for Variable Thin Coating Reconstruction using Generalized Impedance Boundary Conditions

Isabela Viana

We solve the inverse scattering problem of reconstructing the thickness variation of a thin coating on a perfectly electric conducting domain using measurements of the scattered field off of the obstacle generated by impinging incident waves. The problem is nonlinear, ill-posed, and computationally expensive. An important application of this problem is the monitoring of fuel rods, where deposits form thin layers that are difficult to inspect directly. Noninvasive methods to estimate coating thickness are therefore essential for safety assessment. Wave propagation is modeled by the Helmholtz equation with generalized impedance boundary conditions (GIBC), subject to the Sommerfeld radiation, and the data is generated by calculating the scattered field from plane waves. Since the transmission problem with thin layers that models our problem is computationally costly, we use the GIBC model, which provides an efficient approximation. For solving the forward problem numerically, we use boundary integral equations accelerated by fast algorithms. The inverse problem is recast as a nonlinear least-squares problem solved with the Gauss–Newton method and a bandlimited regularization. Numerical experiments on star-shaped geometries with thin coatings of variable width show that the method reconstructs the coating thickness with a small error, demonstrating that the approach provides accurate and efficient recovery of variable thin coatings from scattering data.
Authors: Isabela Viana, Carlos Borges

Data-Driven Insights into the Degradation and Failure of Electronic Materials using Interdigitated Comb Sensors

Janice Yeung

The longevity of electronic systems depends on the reliability of their components. For critical applications such as nuclear systems, conventional maintenance practices rely on manual electrical inspections. While effective, this approach is labor intensive and time-consuming, requiring technicians to perform routine site visits. Interdigitated comb (IDC) sensors provide a cost-effective, data-driven approach for monitoring electronic component health and detecting potential failures. The implementation of IDC sensors alongside electronic components enables in situ monitoring and allows technicians to prioritize essential tasks.
In collaboration with Sandia National Laboratories, a standardized workflow was established to evaluate IDC sensors. Samples were first optically imaged in the pristine state and baseline electrical properties were collected. The experimentation includes submerging each IDC board in deionized water or weak organic acid of varying concentrations while subjected to a voltage bias. Current-time data were collected in-situ until onset of failure or short circuit. After drying, post exposure imaging and electrical measurements were performed.
Quantitative analysis of optical images and electrical data were conducted using custom python packages developed for IDC evaluation. The comparison of pristine and exposed samples revealed significant changes in electrical properties, including increased capacitance after submersion, while RGB and grayscale analysis indicated a shift in color channels. The trends are consistent with the presence of material degradation such as dendritic growth. These data driven correlations between electrical behavior and image features provide important insights into IDC degradation mechanism, establishing a foundation for predictive monitoring of electronics materials.
Authors: Janice Yeung, Jarod Kaltenbaugh, Max Liggett, Matthew A. Kottwitz, J Elliott Fowler, Alp Sehirloglu, Roger H. French, Kristopher O. Davis

Curriculum Learning for Inverse Scattering Problems

Nickolas Arustamyan

We address the inverse acoustic obstacle scattering problem in 2D for sound-soft obstacles, aiming to reconstruct the obstacle’s boundary from scattered field measurements. Building on the neural network approach introduced in [1], which outperformed the classical linear sampling method (LSM) for initial domain estimation, we propose a curriculum learning strategy to enhance training. By progressively incorporating multifrequency data—from low to high frequencies—our method improves accuracy, robustness to noise, and efficiency. Compared to [1], this approach yields faster, more reliable training with reduced computational cost.
Authors: Dr. Carlos Borges, Nickolas Arustamyan, Dr. Jeremy Hoskins

DART: Distributed Assignment of Research Tasks for Heterogeneous Compute Environments

Abdur Rouf

In today’s diverse and dynamic research computing environments, scientists face significant challenges in managing computational tasks across a wide range of systems, from personal machines, cloud virtual machines (VMs), to institutional high-performance computing (HPC) clusters. This paper introduces DART, a lightweight, platform-agnostic job distributor tailored for academic research settings. The distributor is designed for minimal setup, broad compatibility, and seamless job submission, monitoring, and rescheduling across distributed resources, reducing research turnaround time and enhanced fault tolerance. Unlike traditional orchestration tools that rely on uniform infrastructure or require complex configurations, our system focuses on simplicity and portability. We detail the architecture, design rationale, and specific use cases while comparing our tool to more complex frameworks like Dask. While more comprehensive systems like DAGMan exist, our tool stands out for its intuitive interface, quick deployment and minimal scripting (using simple JSON configuration files), making it an ideal solution for research software engineers managing diverse and evolving scientific workflows. In a real-world experiment involving over 37,720 jobs across 68 heterogeneous machines, our dynamic job allocation approach achieved up to a 35x speedup over static allocation strategies, completing workloads in days instead of months.
Authors: Abdur Rouf

Spiking Neural Networks for Classification of Neuromorphically Encoded Tactile Sensing Data

Eugenio Diaz

Restoring a natural sense of touch in prosthetic devices remains a major challenge in neural engineering, particularly in determining how to encode tactile sensor data to evoke lifelike sensations. We collected a comprehensive tactile dataset using a 9-taxel piezoresistive sensor array interacting with 21 3D-printed textures under 5 sliding speeds and 4 force levels, yielding 42,000 trials. The raw sensor signals are converted into biologically plausible spike trains via the Izhikevich neuron model and used to train a three-layer, leaky integrate-and-fire, rate-coded spiking neural network (SNN). The rate-coded SNN achieves high classification performance, correctly identifying individual textures with 93.68% accuracy and broader texture groups with 99.26%. We further investigate data efficiency by truncating the input signals to the onset of first spikes: a rate-coded SNN trained on these shortened spike trains maintains 90.95% (per-texture) and 97.47% (group-level) accuracy. A temporal SNN trained on the same data achieves slightly worse performance than the rate-coded SNN, but much more efficiently and with massively reduced latency for real-time inference pipelines. These results highlight the promise of SNN-based tactile sensing for energy-efficient, real-time texture discrimination, paving methods for improved sensory feedback in prosthetic limbs and realistic haptic feedback in virtual reality.
Authors: Eugenio Diaz, Mohsen Rakhshan

Bayesian Wind Farm Layout Optimization

Taylore Keesler

Maximizing the annual energy production (AEP) of large-scale wind farms is a challenging black-box optimization problem. Each candidate layout requires an expensive wind-farm simulation, gradient information is unavailable, and strict geometric constraints on turbine spacing and site boundaries must be satisfied. We investigate constrained Bayesian optimization (BO) algorithms for efficiently exploring the design space of the wind farm layout optimization (WFLO) problem.
We first evaluate standard BO with Expected Improvement and Lower Confidence Bound acquisition functions on benchmark problems. For the WFLO task, we compare four constrained BO strategies: (1) a vanilla BO baseline using IPOPT (an interior-point method) for constraint handling, (2) the Hybrid BO–IPOPT algorithm for constrained optimization, (3) SCBO (Scalable Constrained Bayesian Optimization), and FuRBO (Feasibility-Driven Trust Region Bayesian Optimization), a trust-region approach that adapts to the feasible set. We also experimented with custom covariance kernels designed to exploit the geometry of turbine layouts.
Empirical results show that FuRBO and the Hybrid BO–IPOPT method consistently outperform both SCBO and the vanilla BO baseline, identifying feasible layouts with substantially higher AEP.
Authors: Taylore Keesler, Michael Melnikov, Sophia Xiao, Allen Zhou

From Signals to Clinical Insights: Transformer-based Feature Extraction and Classification for Epilepsy

Maryam Rahimimovassagh

Epilepsy is one of the diseases that is widely spread with about 50 millions of people affected by it worldwide. It has many categories which make the symptoms diverse. It is important for the clinicians to understand the signals regarding abnormalities as well as the seizure onset zone (SOZ) using Electroencephalograph (EEG) to categorize it and predict the attack in some cases. EEG helps them learn about the type of the disease, its severity, and the prognostic path to manage the severity and the impact of seizure attacks. EEG signals are inherently highly chaotic and have a low signal to noise ratio. Furthermore, there are different sources of error including the way the signals travel, placement of electrodes, and muscle movements. For a clinician, labelling becomes important when the number of the data grows. An automated approach to categorize the normal and abnormal signals becomes important when we are dealing with large amounts of data. Moreover, feature extraction from these signals is important to draw useful conclusions given the chaotic nature of the signals and their high dimensionality. One of the challenges is to extract useful features given the complexity of the brain networks. The current work spans baseline pretrained large language models integrating with feature extraction and machine learning classifiers to help clinician categorize the patients through an interpretable interface rather than a black box model.
Authors: Maryam Rahimimovassagh, Elias Hossain, Mehrdad Shoeibi

Predicting the formation energy and lattice constants of ternary metal oxides using machine learning regression

Nicholas Belden

Metal oxides are commonly used as catalysts or catalyst supports in heterogeneous catalytic systems. Multicomponent oxides, such as ternary oxides, can be formed to tune certain catalytic properties. For these reasons, ternary metal oxides offer an expansive material space to design catalytic materials with desired properties. Experimental techniques or computationally expensive density functional theory (DFT) calculations are typically used to examine the properties and structure of ternary metal oxides. However, machine learning techniques may obtain similar results while being computationally cheaper. In this study, we trained machine learning regressors to predict key bulk properties of ternary metal oxides such as formation energy, crystal structures, and band gaps. Predictions were based on features of the metal elements including atomic radius, cohesion energy, and Lennard-Jones potential (LJP) values. Various commonly used regression models were trained and tested on data for ~2,800 ternary metal oxides collected from the Materials Project. Mixed accuracies were obtained for each target variable, but predictions of formation energy per atom showed promising results with an R2 score of ~0.90 and a test RMSE of 0.23 eV/atom. Additionally, feature importance analysis on the gradient boosting regression (GBR) and random forest regression (RFR) models listed the most important features as the electronegativities, cohesive energies, and work functions of the metals along with the minimum potential for the LJP of the metals. These results provide a step towards an efficient method of predicting properties of ternary metal oxides with comparable error to DFT calculations.
Authors: Nicholas Belden, Shyam Kattel

Data-Driven Compression for Large-Scale Simulations in High-Performance Computing

Arshan Khan

The launch of exascale computing facilities has enabled simulations to capture physical processes at unprecedented fidelity, but the resulting extreme-scale datasets present significant challenges for storage, transfer, and analysis. Existing error-bounded lossy compressors achieve only moderate compression and often rely on data-agnostic predictors with predefined functions, limiting both efficiency and scientific usability. To overcome these limitations, we introduce Data-informed Local Subspaces (DLS), a data-driven compression framework that learns spatially local, generalizable features directly from representative samples and uses them to represent large-scale spatiotemporal solutions. The discontinuous variant, DDLS, partitions the computational domain into patches, encoding each independently under user-specified error tolerances. By relaxing continuity across patch boundaries, DDLS achieves higher compression while still preserving essential physical structures. This patch wise encoding produces sparse representations that reduce storage requirements and support rapid analysis directly on compressed data, eliminating the need for full decompression. Case studies demonstrate the method’s broad applicability. For three-dimensional flow past a cylinder at Reynolds number 100,000, DDLS achieved 8x–100x compression while preserving key turbulent structures and accelerating downstream analysis by 15x–600x. Comparable results were obtained for rotor-body interactions and climate modeling data. These findings establish DLS as a scalable, physics-informed, data-driven compression framework that outperforms traditional data-agnostic compressors such as SZ and MGARD. By coupling strong data reduction with accurate preservation of scientific features, DLS advances the state of the art in data management for high-performance computing.
Authors: Arshan Khan, Rohit Deshmukh

Data-informed Feature Learning for Large Scientific Data Compression

Fahim Sakai

Recent launches of multiple exascale computing facilities have pushed the limits of fidelity in scientific simulations. This unprecedented fidelity comes with increased cost of I/O, data storage, data movement, and analysis. Many existing lossy compression algorithms aim to address this issue by reducing the data while retaining mission-critical information. Our in-house Data-informed Local Subspaces (DLS) algorithm combines feature learning with classical computational mechanics to provide an improvement over the existing lossy algorithms by training the compression model directly on the target data. In this work, we aim to improve the training process by leveraging recent advances in machine learning. In particular, we explore sparse coding using neural networks to learn better spatially local features to enrich our DLS models. Although this effort is still in its early stages, we discuss current limitations, our hypothesis, and possible remedies, accompanied by preliminary demonstrations on 2-D images and scientific data.
Authors: Fahim Sakai, Ande Bhanu Naga Peddi Ganesh, Rohit Deshmukh

Multimodal and Multiscale Characterization of Photovoltaics: Understanding Degradation and Failure from Systems Down to Atoms

Max A. Liggett

The mode and rate of degradation for photovoltaic (PV) modules in the field is caused by a litany of stressors, and as a result, an understanding of root cause of failure often requires more comprehensive investigation with multiple data streams on a system-, module-, device-, and material-level to ensure high confidence in conclusions regarding culpability of performance losses experienced in the field. Intrinsic challenges present for reliability research in which data is collected from multiple levels require careful consideration to meaningfully connect and extract information. In this work, we present a multiscale approach for investigating degradation and failure in fielded modules. Additionally, we cover the ongoing development of tools and methods of analyzing multimodal data, particularly electroluminescence (EL) images and illuminated current-voltage (I-V) curves, collected from modules as a means of assisting the multiscale investigative framework.
Authors: Max A. Liggett, Sameera Venkat, Dylan J. Colvin, Andrew M. Gabor, Philip J. Knodle, Joseph Raby, Brent A. Thompson, Hubert P. Seigneur, Roger H. French, Kristopher O. Davis*

VHS 1256 b, HIP 99770 b, AF Lep b: Expected Thermal Light Polarization

Maxwell Galloway

Direct imaging of planetary mass companions and brown dwarfs has revealed similar spectra to L/T transition brown dwarfs, including hints of rotational variability. Polarization, being sensitive to macro- and micro-physical properties of clouds, can break degeneracy in potential cloud structures from flux-only observations. The dramatic increase in signal-to-noise ratio in flux that can be achieved for exoplanets due to JWST has furthered the need to be able to properly characterize these objects’ signals. Modeling of these exo-atmospheres, however, can be difficult without adequate computational power when utilizing state of the art radiative transfer models across several wavelengths. We used a climate model code (PICASO), a 3D radiative transfer code (ARTES), and UCF’s Stokes high-performance computing (HPC) cluster to model potential light curves and select spectra for VHS 1256 b, HIP 99770 b, and AF Leporis b in the NIR. We explored a range of potential cloud formations, sizes, and cloud sedimentation parameter values. In this presentation, we will discuss the effect of temperature, gravity, cloud parameters, and inclination on the observed signals and potentially observable trends. The models presented here are part of a larger grid computed over hundreds of thousands of CPU hours through Stokes that will be given open-access to the community and can be used to aid in the characterization of directly imaged exoplanets and brown dwarfs.
Authors: Maxwell Galloway, Theodora Karalidi, Sagnick Mukherjee, Max Millar-Blanchaer, Connor Vancil, Joseph Harrington

Decoding Michelin-Starred Restaurant Experience: A Theory-Driven, Topic-Sentiment Weighted Analysis of Online Reviews

Ngoc Tran Nguyen

This study examines how high-end restaurant experiences are constructed and evaluated by consumers through computational methods to bridge theoretical and empirical gaps in hospitality research. Drawing from 28,882 online reviews of 96 Michelin-starred restaurants in California and New York, the research follows a two-stage analytical procedure. In Stage 1, bigram-based keyword extraction was used to develop a domain-specific dictionary encompassing both predefined constructs (e.g., atmosphere, food, and service quality) and emergent constructs (e.g., wine quality and holistic experience). In Stage 2, constructs were quantified using a novel integration of Guided Latent Dirichlet Allocation (GuidedLDA) and VADER sentiment analysis, producing topic–sentiment weighted scores that reflect both thematic relevance and emotional valence.
These quantified measures were then used to empirically test a theory-driven quality–experience–satisfaction–intention framework through Generalized Structural Equation Modeling (GSEM). Results show that all four quality dimensions (i.e., atmosphere, food, service, and wine) significantly predict holistic experience, which in turn drives satisfaction and revisit intention. The emergence of wine quality and experience as distinct constructs enriches theoretical understanding of symbolic consumption in luxury services.
Practically, this approach provides a scalable, cost-effective method for extracting actionable insights from unstructured review data, which eliminates reliance on biased or costly surveys. Fine-dining managers can use these insights to optimize service design, enhance wine programs, and personalize customer engagement strategies. More broadly, this study illustrates how advanced computational methods, specifically, text mining, topic–sentiment weighted analysis, and structural modeling, can transform unstructured consumer narratives into strategic intelligence for experience-centric industries.
Authors: Ngoc Tran Nguyen, Jeong-Yeol Park, Hyoung Ju Song, Ji Eun Lee

N2 splitting via a bare bimetallic catalyst in an L-shaped geometry

Shiv Patel

Nitrogen (N₂) splitting is a crucial step in producing nitrogenous derivatives, such as ammonia, which is essential for fertilizers and the global food supply. Haber–Bosch process, currently used commercially, requires extreme pressures and temperatures and is therefore highly energy-intensive and unsustainable. The discovery of alternative methods for N₂ activation under milder conditions is thus an urgent challenge. Bimetallic transition metals offer a different and viable approach for reducing the activation barrier for N2 bond cleavage.
In this paper, we used density functional theory (DFT) to screen more than one thousand Group 6–10 transition metal bimetallic complexes, modelled in various spin multiplicities in an L-shape conformation. After initial geometry optimizations, complexes were sorted based on binding free energies as an indicator of stability and reactivity. Complexes with metal–metal–nitrogen angles between 70–110° emerged as the most promising for N₂ activation, narrowing the dataset to approximately 150 possible candidates.
This study illustrates the usefulness of computational screening for the rapid identification of promising systems from large chemical space. By checking certain geometric designs and binding energy relevant to N₂ splitting, we can create a rational catalyst design for N₂ activation. This conclusion helps emphasize the significance of new ways to activate N₂. This research can help set future studies with ligands and solvents. Furthermore, this study is an advancement in the continued pursuit of efficient and sustainable catalytic pathways to ammonia synthesis and other nitrogen conversions.
Authors: Shiv Patel, Shengli Zou, Ph.D

Unified Multimodal RAG for Cross-Brand Marketing Analytics Across Modalities

Sabab Ishraq

Contemporary marketing strategies orchestrate intricate narratives across heterogeneous media channels, yet existing analytical paradigms remain constrained by modality-specific limitations. This creates blind spots where important patterns across different media types go unnoticed. This research addresses the fundamental challenge of unified multimodal understanding in marketing analytics by proposing a cross-modal retrieval-augmented generation framework that bridges semantic gaps between visual, textual, and auditory brand communications. We explored how temporal video understanding, paired with cross-modal attention mechanisms can identify previously undetectable patterns in brand messaging consistency across different modalities. The hybrid retrieval approach introduces marketing-specific knowledge recognition as a semantic layer, enabling fine-grained analysis of pricing strategies, promotional messaging, and call-to-action effectiveness. The study contributes three key innovations: (1) a unified embedding space that preserves both semantic and temporal relationships across text, image, video, and audio, (2) marketing entity-aware retrieval that captures domain-specific commercial intent, and (3) cross-modal consistency metrics for brand messaging analysis. We evaluated 71 major brands across diverse industries. Our framework demonstrates significant improvements in capturing nuanced marketing strategies that traditional unimodal approaches miss, revealing novel insights into audio-visual alignment patterns and cross-platform brand positioning inconsistencies. This means better competitive intelligence and more coherent brand strategies for marketing teams. It opens new directions for researchers in commercial content analysis and cross-modal AI applications.
Authors: Sabab Ishraq, Juncai Jiang, Chen Chen

Scientific Machine Learning for Digital Twins

Nam T. Nguyen

Digital twins are dynamic, virtual representations of physical objects, systems, or processes that remain connected to their real-world counterparts through live sensor data. In control systems, constructing accurate digital twins is particularly important, as they provide a virtual environment for evaluating the performance of controllers before deployment in practical settings. Scientific machine learning (SciML) algorithms, which integrate scientific principles into machine learning frameworks, have shown strong potential for precisely approximating complex dynamical systems. This poster presents our results on implementing SciML models that incorporate stability, boundness, and monotonicity properties into the learning technique of real systems using experimental data. Furthermore, we apply an active learning strategy to collect informative data for training the digital twin, while explicitly accounting for the cost of data collection. The results demonstrate that our SciML models outperform standard machine learning models by approximately 30 percent in identifying system dynamics. In addition, the active learning algorithm achieves comparable accuracy with only 10 experiments, in contrast to 30 experiments required by the passive learning approach, according to simulation results.
Authors: Nam T. Nguyen, Binh Nguyen, Truong X. Nghiem

Quantum Algorithms to Improve Density Functional Theory

Hafiz Arslan Hashim

Quantum computers offer new approaches to enhance Density Functional Theory (DFT), particularly for strongly correlated systems where conventional functionals are ineffective. In this work, we develop a hybrid quantum–classical approach to construct spin–dependent exchange–correlation (XC) potentials from ground-state data obtained using the Variational Quantum Eigensolver (VQE). We demonstrated the method on a Fermi–Hubbard model, which is paradigmatic for strongly correlated fermionic physics. We describe calculations and results for the spin-resolved XC potential and the exchange–correlation (Exc) energy as the system size reaches six sites. The XC potentials derived from VQE already outperform Hartree–Fock DFT and direct VQE energy estimates.
Authors: Hafiz Arslan Hashim, Volodymyr M Turkowski, Eduardo R Mucciolo

Barren Plateaus, Complex Entanglement, and Structured Circuits in Variational Quantum Algorithms

Suman Mandal

Variational Quantum Algorithms (VQAs) are useful tools for solving many-body quantum problems on near-term quantum hardware. A major challenge, however, is the onset of barren plateaus, where gradients vanish and optimization becomes very difficult. In this work, we argue that barren plateaus are unavoidable in deep, unstructured circuits, as they develop what we call complex entanglement. To study this phenomenon, we look at both the Cluster–Ising model and the Toric Code Hamiltonian, tracking the entanglement spectrum through level-spacing statistics. At the same time, we compare structured circuit families—including Finite Local Depth Circuits (FLDC), Global Local Depth Circuits (GLDC), and Brickwall layouts with Cartan blocks—that remain trainable at modest depths. Our results show that while deep circuits exhibit Wigner-Dyson statistics and loss of trainability, FLDCs and GLDCs with Cartan blocks achieve low energies at shallow depths, making them promising candidates for practical VQAs.
Authors: Suman Mandal, Maximillian Daughtry, Eduardo R Mucciolo