Poster Only Titles and Abstracts (27 October 2026)

Probability of Antibiotic Resistance During Treatment in Stochastic PK/PD-Based Bacterial Model with Distinct Drug and Mutation Modes
Chimezie Izuazu, University of Louisiana at Lafayette
Advisor: Cameron Browne
Abstract: Mathematical models, e.g. differential equations and stochastic processes, have gained considerable attention for understanding evolution of antibiotic resistance. However, most existing models assume standing genetic variation and do not consider the possibility of random or drug induced mutation of reference bacterial strains. Therefore, we propose a pharmacokinetics/pharmacodynamics (PK/PD)-based continuous-time Markov chain considering the competition and mutation between sensitive and resistant bacterial within an infected host during treatment. The proposed model is approximated as a generalized birth–death process with immigration, allowing for explicit derivation of the probability resistant population establishes during treatment. Besides capturing the stochasticity of de novo emergence of a resistant bacterial strain, we explore the effects of different antibiotic modes of action, horizontal gene transfer, nutrient availability and drug pharmacokinetics on antibiotic resistance. We find that replication-targeting (biostatic) drugs suppress resistance more than death-targeting (biocidal) drugs. Like prior works, we obtain maximized resistance at intermediate drug concentrations, however the consideration of de novo mutation magnifies the superiority of higher doses in preventing resistance emergence.
Bayesian Spatiotemporal Modeling of County-Level Heart-Disease Mortality in the United States, 2000–2020
Kamlesh Sarkar, University of Louisiana at Lafayette
Advisor: Kumer Das
Abstract: Heart-disease mortality varies across U.S. counties and over time. Understanding these differences requires accounting for population size, age composition, socioeconomic conditions, and dependence between neighboring counties. This study aims to examine county-level mortality patterns across the United States from 2000 to 2020. Annual heart-disease death counts and population estimates will be obtained from CDC WONDER. Poverty and median household income will come from the U.S. Census Bureau. Urbanization will follow the National Center for Health Statistics classification. A Bayesian negative-binomial model will relate death counts to these county characteristics, with population included as an exposure offset. Age-specific counts will support estimation of age-standardized mortality rates. The model will include a scaled BYM2 spatial component, a nonlinear national temporal trend, and county-specific temporal deviations. Suppressed counts will contribute interval probabilities rather than being replaced by zeros or omitted. County boundary changes and documented reporting problems will be addressed during data preparation. The analysis will estimate county mortality rates, relative risks, and changes between 2000 and 2020. Posterior credible intervals and exceedance probabilities will describe uncertainty. Model assessment will include convergence diagnostics, posterior predictive checks, sensitivity analyses, and temporal validation. This study will provide a national assessment of geographic and temporal variation in heart-disease mortality. It will examine county-level socioeconomic associations while accounting for spatial dependence and uncertainty.
Time Series Analysis of Trait Evolution on Phylogenetic Trees of Amazonian Fish
Eric Sinegal, University of Louisiana at Lafayette
Advisor: Erez Aghion
Abstract: We analyze time series from traitograms describing the evolution of Amazonian fish, from 140 million years ago to today. The data, collected by Dr. James Albert, contains over 6000 individual species. We use statistical tools especially designed to classify anomalous dynamics in stochastic processes, to rigorously test a broudly held hypothesis on the "normal" — non-anomalous, nature of the evolutionary dynamics. Our initial results indicate the trait evolution on ancestral time series contains both aging and rare-event effects which do not correspond with standard Brownian fluctuations. This implies that the traight distribution can violate the central limit theorem.
Collaboration and Landscaping provide an Edge in Stochastic Search
Ayden Guilbeau, University of Louisiana at Lafayette
Advisor: Erez Aghion
Abstract: Optimizing the search for the deepest well in a multidimensional potential landscape is a long standing challange in stochastic processes. To accelerate the search and increase the confidence in the solution of the problem, we develop an Adaptive Concavity method, whereby an ensemble of Random Walkers collaborate and actively reshape the stochastic potential around them. Through this collective effort, the system avoids long trapping events that might lead to false detections, and the time to reveal the true potential minimum is optimized.
When Help Cannot Reach Everyone: Understanding Disaster Preparedness, Accessibility, and Recovery Needs of People with Motor Disabilities and Caregivers.
Busayo L. Ayodele, University of Louisiana at Lafayette
Advisor: Tamanna Motahar
Abstract: As climate change increases the frequency and severity of disasters such as hurricanes, floods, and extreme weather events, people with motor disabilities and their caregivers face heightened risks before, during, and after emergencies. In Louisiana, approximately one in three adults' lives with a disability, and about 15% report on a mobility disability, highlighting the importance of disability-inclusive disaster planning.  People with motor disabilities often encounter barriers related to evacuation, transportation, shelter accessibility, healthcare access, assistive devices, rehabilitation services, and timely emergency information. While prior research has identified the vulnerability of people with disabilities during disasters, limited work has examined the specific preparedness, information, and recovery needs of people with motor disabilities and caregivers, particularly in relation to accessible and real-time disaster support. Using Louisiana as a pilot context and a broader global data collection approach, this research will investigate preparedness barriers, disaster experiences, accessibility challenges, and recovery needs through surveys, follow-up interviews and research through design. Findings will inform the design of accessible, technology-enabled solutions that support preparedness, response, and recovery for people with motor disabilities and their caregivers.
Analyzing Generative AI Use in Abusive Situations: Patterns of Support, Trust, and Safety
Warda Ruheen Bristi and Huxton J Marcus University of Louisiana at Lafayette
Advisor: Tamanna Motahar
Abstract: People are increasingly using AI not only for everyday tasks but also to seek advice on personal challenges, navigate abusive relationships, and manage complex interpersonal situations. Our preliminary analysis on social media data suggests that 27% of users reported AI helped them with mental health support, while only 12% preferred guidance or a better understanding of AI before using it for mental health support. Despite existing research on making GenAI more trustworthy and ethical, gaps remain in evaluating its performance in sensitive contexts. This study has two aims: (1) identify the support people seek and receive from GenAI in abusive situations, and (2) examine how they evaluate these experiences. This study combines qualitative analysis of social media data with a large-scale quantitative survey. It will characterize patterns of GenAI use in abusive situations and examine associations among user behaviors, trust, and perceived support to inform GenAI evaluation in these contexts. 
Missing Not at Random: How Fear of Disclosure and the Disability Data Gap Distort Identity
Mashiyat Mahjabin Eshita, University of Louisiana at Lafayette
Advisor: Tamanna Motahar
Abstract: People with disabilities are undercounted and misrepresented in both official statistics and unofficial sources such as social media and online platforms, leaving a persistent disability data gap. We examine why this gap persists and how it distorts identity. Three forces drive people with disabilities to withhold accurate information. First, many fear that reporting a disability in a census or survey could put their government benefits at risk. Second, disclosure carries social costs: 30% of college-educated U.S. employees report a disability, yet only 3.2% disclose it to their employer. Third, few agencies, organizations, or digital platforms collect this information in ways that feel safe, private, and accessible. Because those least likely to disclose often have the most at stake, the resulting data is not simply incomplete but systematically skewed. Drawing on data feminism, we argue that this gap reflects power: who is counted, by whom, and on whose terms. Drawing on Orientalism, we show how, without self-representation, people with disabilities are defined from the outside and constructed as an "Other" through assumptions and proxies rather than their own voices. As policies, services, and AI-driven tools are built on these data, this distorted identity is reproduced and amplified. Closing the gap requires both statistical and design solutions: privacy-preserving techniques that protect respondents, and data collection systems that earn their trust. 
Confidence-Aware Early Speech Emotion Recognition Under few-Shot Learning
Brett Keown, University of Louisiana at Lafayette
Advisor: Ashok Kumar
Abstract: Speech emotion recognition (SER) typically assumes that a complete utterance is available before classification. This work investigates whether prediction confidence can determine when a few-shot SER prediction can be accepted before the full utterance is observed. The proposed approach combines progressively longer speech prefixes, prototype-based few-shot classification, and confidence-aware early exiting. Experiments are conducted on RAVDESS and EMO-DB under 1-shot and 5-shot conditions. Results demonstrate a controllable accuracy–earliness tradeoff. At the balanced operating point (θ = 0.80), the mean early-exit rate across the four experimental conditions was 63.61%, with an average early-exit duration of 1.09 seconds and final accuracy of 35.50%, compared with 36.38% using full utterances. These results provide initial evidence that confidence can guide early prediction decisions under few-shot conditions.
The Role of Refuge in a Discrete-Time Competition Model with an Allee Effect
Sajan Bhandari, University of Louisiana at Lafayette
Advisor: Azmy S. Ackleh
Abstract: We present a discrete-time nonlinear competition model with Allee effects in refuge. This model consists of two competing species, in which the weaker competitor has access to a refuge patch that allows it to escape inter-specific competition. In the refuge patch, we consider a scenario where the individuals are subject to mate limitations resulting in an Allee effect. We first study the model without the strong competitor and present local and global stability analyses. We then study the model with the strong competitor and establish conditions under which the system is uniformly persistent, thus demonstrating that a refuge patch may enable coexistence despite the Allee effect. We also present some numerical results that demonstrate sudden ecological shifts, such as competitive exclusion of weaker species to coexistence of all species.
Modelling the spectroscopy and dynamics of atmospherically relevant ROO and QOOH radicals
Tanima Barua and Monotosh K Sarkar, University of Louisiana at Lafayette
Advisor: Tolga Karsili
Abstract: During fuel combustion, hydrocarbons (RH) undergo oxidation, forming alkyl peroxy (ROO) and hydroperoxy alkyl (QOOH) radicals that maintain the combustion process. The analogous chemistry also occurs following oxidation of volatile RH compounds emitted into the atmosphere. While a wealth of studies has focused on the ground state unimolecular decay and bimolecular chemistry of ROO and QOOH intermediates, their solar induced chemistry is less well understood, despite previous studies indicating strong absorption at tropospherically relevant wavelengths. Motivated by this, our present contribution presents state-of-the-art computational study aimed at discovering the role of solar induced photochemistry in the removal of ROO and QOOH intermediates from the atmosphere. Our results show that the primary fate of electronically excited QOOH and ROO mirrors their thermal unimolecular decay, generating the same products but via a barrierless potential energy profile. Therefore, the excited state unimolecular decay dynamics occurs on a femtosecond (10-12 seconds) timescale rather than the nanosecond (10-9 seconds) timescale of the analogous (barriered) ground state process. This excited state photodissociation also shows conformer specific absorption. The findings from this research are significant as they demonstrate that excited state photochemistry at atmospherically relevant wavelength may be a competitive removal of ROO and QOOH intermediates, as well as motivating experimentalists to use UV-VIS spectroscopy to detect ROO and QOOH intermediates in their future studies.
Low-Temperature Dry Reforming of Methane over Ni-Based Composite-Oxide Catalysts: Operando DRIFTS–Guided Engineering of Coke-Resistant Interfaces
Godwin U Edor, University of Louisiana at Lafayette
Advisor: Hui Yan
Abstract: Dry reforming of methane (DRM) converts two greenhouse gases, CH4 and CO2, into syngas, but Ni catalysts deactivate rapidly via sintering and carbon deposition, particularly at low temperature. This work examines whether composite-oxide interfaces can sustain DRM at ≤600 ℃. Four 5 wt% Ni catalysts (Ni-SiO2, Ni-SiO2@CeO2, Ni-SnO2@CeO2, and Ni-TiO2@CeO2) were prepared by incipient-wetness impregnation of hierarchically synthesized oxides and characterized by N2 physisorption, SEM/EDS, and XRD. Surface chemistry and activity were probed by operando DRIFTS coupled with online mass spectrometry, H2-TPR, CO2-TPD, and CH4- and CO2-TPSR. At 600 ℃, CH4 and CO2 conversions were 13.5 and 16.7% for Ni-TiO2@CeO2 and 12.9 and 16.1% for Ni-SiO2@CeO2, respectively compared with 10.0 and 16.0% for Ni-SiO2 and 1.5 and 6.8% for Ni-SnO2@CeO2. The ceria-shelled TiO2 and SiO2 catalysts gave H2/CO ratios of up to 0.97 and 0.93, against 0.49 for Ni-SiO2 and 0.27 for Ni-SnO2@CeO2. CH4-TPSR showed that the CeO2 shell lowers the onset of C–H activation from 423 ℃ on Ni-SiO2 to 337 ℃ on Ni-SiO2@CeO2. DRIFTS detected carbonates but no formates, indicating a carbonate-mediated, bifunctional pathway in which CO2 is activated on the oxide and lattice oxygen gasifies carbon species formed on Ni. Ni-SnO2@CeO2 showed only linear CO, consistent with Ni ensemble dilution, but was nearly inactive for CH4 conversion. Ni-TiO2@CeO2 showed negligible adsorbed CO despite strong CO evolution, consistent with a strong metal–support interaction. Conversions declined during 4 h holds at 600 ℃ and distinguishing coking from sintering remains the crucial next step. These results identify ceria-shelled SiO2 and TiO2 as the most promising interfaces for low-temperature Ni-based DRM.
Wavelet-Based Multiscale Analysis of Illumination and Noise for Robust Low-Light Image Enhancement
Azwaad Labiba Mohiuddin
Advisor: Ashok Kumar
Abstract: Low-light image enhancement is challenging because illumination degradation is often accompanied by noise amplification and loss of fine image structures. This work investigates the statistical characteristics of illumination degradation and frequency-dependent image information and develops a wavelet-based multiscale framework for robust low-light enhancement. The proposed approach combines a differentiable illumination-adjustment curve with wavelet-domain processing and exposure feedback. An input image is decomposed into multiple frequency components, allowing illumination correction to be concentrated on the low-frequency component while preserving higher-frequency structures that may be susceptible to noise amplification. The exposure-feedback mechanism adaptively adjusts enhancement according to the estimated illumination state of the image. We quantitatively analyze the behavior of the proposed components through component-wise ablation and robustness experiments under varying noise levels. Experiments on the LOL-v1 low-light image dataset show that combining wavelet processing with exposure feedback improves enhancement quality over the individual components, achieving a PSNR of 22.864 dB and an SSIM of 0.796. The results demonstrate that incorporating frequency-dependent characteristics into illumination modeling provides a robust framework for recovering information from degraded low-light observations.
Integrating Geochemistry and Microbial Activity in the Pecan Island Area with Coastal Change
Diana (Ana) Kreamer, Johnathan Pillsbury, Bibi Morium, Zubair Ahmed, Imran Ullah, Isaac Asiamah, Anasiberuje Sanni, Samuel Mensah, Grace Landry, and Dreux Davis, University of Louisiana at Lafayette, and Paul Helfrich, University of Maryland
Advisor: Alysia Cox
Abstract: Abstract: Imminent and future coastal developments in the Pecan Island area have prompted need for baseline environmental data to evaluate coastal change over time. The Laboratory Enumerating Geobiochemical Exploration and Natural Dynamics (LEGEND) is conducting a full aqueous geochemical (~ 65 parameters) and microbial composition (metagenomics) and activity (metaproteomics) quantification of the area. This analysis will provide a microbial and geochemical baseline to monitor ecosystem health with time.
G-SELC: An R Package for Sequential Optimization of Large Discrete Search Spaces
Charles Wu, University of Georgia
Advisor: Abhyuday Mandal
Abstract: Many scientific and industrial optimization problems require searching through a very large number of possible combinations of factor levels when only a limited number of evaluations are feasible. This project develops an R package implementing G-SELC (Genetic-algorithm-based Sequential Elimination of Level Combinations), a sequential optimization approach that combines genetic algorithms with Gaussian process modeling and the Expected Improvement (EI) criterion to efficiently explore large search spaces. G-SELC adaptively combines the search capabilities of genetic algorithms with model-based information from previously evaluated combinations, using EI to balance exploration of uncertain regions with exploitation of promising regions. The package will provide user-friendly functions for defining the optimization problem, fitting the Gaussian process model, evaluating expected improvement, selecting candidate batches, and carrying out the sequential search. Its performance will be demonstrated through numerical examples and benchmark optimization problems, with the goal of making this flexible optimization strategy accessible to researchers and practitioners working with large discrete and mixed search spaces.