Poster Abstracts

Our labs in-vivo mouse exposure studies on the high prevalence of asthma among the communities residing near California’s Salton Sea (SS) revealed a potent neutrophilic inflammatory response to dust collected from the region. Our data revealed that chronic exposure to aerosolized lipopolysaccharide (LPS)/endotoxin from environmental halophilic bacteria in the dust caused this immune response. These findings revealed a distinct disease affecting this population which resembled the nonallergic subset of asthma. The identification of this unique clinical entity led to the development of the first therapeutic on the market for treatment of nonallergic pulmonary inflammation caused by exposure to aerosolized environmental toxins.

Generative AI image models are becoming increasingly integrated into creative, educational, and commercial workflows. However, limited research evaluates how accurately these systems render images using culturally specific language, focusing on textured black hairstyles. This study investigates how four leading generative image platforms represent diverse black women’s hairstyles when prompted using culturally specific terminology.

Using a controlled prompting protocol grounded in hairstyle taxonomy, over 600 image sets were generated across models under standardized conditions. Outputs are being evaluated through recognition alignment scoring, measuring agreement between hairstyle descriptions and perceived hairstyle characteristics. This framework assesses cultural fidelity, texture accuracy, and stylistic precision. 

This work contributes to the AI fairness research, that shifts the conversation from one off anecdotal critiques to a structured computational analysis. This study advances measurable approaches to auditing representational equity within multimodal AI systems. Findings have implications for model training datasets, bias mitigation strategies, and the development of culturally competent generative technologies.

A recent report from the University of California, San Diego (November 6, 2025; updated January 8, 2026) indicates that approximately one in eight incoming first-year students demonstrates deficiencies in mathematics preparedness. This challenge is not unique to UCSD; institutions nationwide report similar gaps. Because mathematics underlies STEM disciplines and many gateway courses impose substantial cognitive demands, these deficiencies have significant implications for student success.

As STEM fields drive economic growth and workforce innovation, higher education institutions seek to improve equity, increase graduation rates, and strengthen employment outcomes. Achieving these goals requires more than content coverage; it demands instructional design grounded in how students learn. Strengthening STEM attainment also requires attention to both higher education (andragogy) and the K–12 pipeline (pedagogy) that shapes college readiness.

High–cognitive load STEM courses raise a central question: Do we teach in ways that align with how students learn? A shift from teacher-centered delivery to evidence-based, learner-centered design is essential. Student success in STEM depends on aligning learning theory, instructional design, and program-level assessment.

Theoretical frameworks such as constructivism and situated learning—drawing on the work of Lev Vygotsky and Jean Lave—emphasize active engagement, social interaction, and contextualized learning. Instructional models that incorporate backward design, transparent outcomes, formative assessment, and inclusive pedagogy demonstrate positive effects on achievement and retention in gateway STEM courses.

At the program level, systematic assessment practices—including curriculum mapping, longitudinal outcome tracking, and continuous improvement cycles—enable departments to evaluate effectiveness and reduce equity gaps. By integrating theoretically grounded instructional design with robust assessment, institutions can move beyond gatekeeping structures toward sustainable models of student development, strengthening persistence, performance, and participation in STEM fields.

We attempt to answer the question "Is it possible to trustlessly compute a supersingular elliptic curve with unknown endomorphism ring that is distributed, secure and scalable?”. We introduce new protocols for implementing the algorithms using general purpose multiparty computation in a committee delegation protocol. We demonstrate the security assumptions of this protocol and show that a semi-honest multiparty computation protocol is enough to obtain malicious security for an isogeny walk in the presence of malicious adversaries. Our solution shows competitive performance to that of prior work, and additionally supports multiple threat models and provides a form of public verifiability that proof-based approaches do not, by drawing randomness from a large number of clients.

Stress is a pervasive factor influencing mental health and overall well-being, with prolonged exposure linked to adverse physical, psychological, and social outcomes. Recent advances in mobile health technologies and wearable sensing have enabled the continuous collection of physiological and behavioral data, creating new opportunities for early stress detection and intervention. In parallel, large language models (LLMs) have demonstrated strong reasoning capabilities across diverse domains, including health prediction from time-series data. However, most existing approaches rely on cloud-based models, raising concerns around privacy, latency, and real-world deployability. This work investigates the use of on-device language models (ODLMs) for stress prediction, emphasizing their potential to support health and wellbeing through privacy-preserving, low-latency inference directly on personal devices. 

We systematically evaluate how different data modalities, prompting strategies, and model scales affect stress prediction performance in ODLMs. Using the PMData life-logging dataset, which includes both objective physiological signals (e.g., steps, heart rate, sleep duration) and subjective self-reports (e.g., mood, fatigue, sleep quality), we design prompts that contextualize health data across multiple temporal intervals. Experiments are conducted using several state-of-the-art quantized ODLMs deployed on iOS devices, allowing us to jointly assess prediction accuracy and device-level performance metrics such as latency, throughput, and memory usage.

Our results show that compact ODLMs can achieve competitive stress prediction accuracy while operating entirely offline. Across multiple experimental settings, objective data prompts consistently yield a comparative prediction error to subjective or combined prompts, suggesting that passive sensing can effectively support stress assessment with minimal user burden. Among the evaluated models, a sub-billion-parameter ODLM demonstrates the best balance between accuracy and efficiency, outperforming larger models in both prediction error and inference speed. Additionally, we find that statistical summaries of longer time windows improve performance relative to raw natural language descriptions, highlighting the importance of prompt structure when working with time-series health data. Beyond model performance, we propose a proof-of-concept trust architecture that situates ODLMs within a closed social and clinical loop involving patients, caregivers, and clinicians. Rather than replacing human judgment, the system is designed to augment care by enabling timely insights into stress while maintaining safeguards against harmful or inappropriate outputs. This framing aligns with a well-being-centered approach to AI deployment, prioritizing user safety, autonomy, and clinical oversight.

Overall, this study demonstrates that on-device language models can serve as a practical and responsible component of mobile health systems for stress prediction. By combining wearable data, thoughtful prompt engineering, and efficient on-device inference, ODLMs offer a promising pathway toward scalable, privacy-preserving tools that support mental health monitoring and proactive wellbeing interventions in everyday life

Modeling crowd behaviors is crucial to navigate shared spaces safely and naturally alongside humans. Current state of the art crowd simulators rely on modeling behaviors as discrete-time Markov processes where individuals display singular behaviors at a time. However, often in real world systems behaviors intermingle and interact with one another. Behavior Inspired Neural Networks (BINNs) have shown the capability of mapping physical observable states to latent behavior categories where categories correlate to one another. This deterministic approach, unfortunately, fails to account for the stochastic nature of real-world systems where multiple future trajectories are possible. To address this, we propose a framework utilizing both Behavior Inspired Neural Networks and Conditional Variational Autoencoders (CVAE) to model behavior categories as a distribution. Conditioning on the behavior prior, we can then refine the generation of predicted trajectories using a Diffusion Model allowing for diverse and accurate multi-modal trajectories. Finally, we demonstrate the efficacy of our approach by modeling crowd behavior patterns across different geographical environments.

Personalized affect recognition in healthcare requires models that perform reliably under sparse data and explicitly account for uncertainty for safe deployment. We propose a Gaussian Process Regression framework for affect detection that incorporates uncertainty-aware active learning to support personalized modeling. The model continuously predicts affective state from physiological signals captured by wearable sensors and selectively queries the user when predictive uncertainty exceeds a predefined threshold, maintaining predictive reliability.

As a proof-of-concept, we evaluate the framework using the publicly available WESAD dataset. Findings demonstrate the feasibility of uncertainty-driven querying and highlight the suitability of Gaussian Process Regression for learning personalized models under limited labeled data. This work establishes a methodological foundation for future validation using National Comprehensive Cancer Network (NCCN) Distress Thermometer–labeled clinical data to support distress monitoring in oncology care.

Heavy elements in the universe and on Earth are predominantly produced in explosive stellar endpoints such as supernovae. Understanding when and how these elements formed requires investigating the rates of such explosions across cosmic time. One important class of these events is the Type Ia supernova, which can occur when two white dwarfs merge in a binary system (i.e. a pair of white dwarfs). Type Ia supernovae play a central role in chemical enrichment of our Universe and are also critical tools for cosmological distance measurements. To constrain the formation, physics, and frequency underlying Type Ia supernovae, this project models the evolution of large populations of binary star systems using population synthesis simulations and computes the resulting merger rates of double white dwarf binaries. Our population scale simulations are unique tools for connecting uncertain stellar and binary evolution physics to observable supernova rates and for assessing the impact of model assumptions on predicted event rates.

Understanding electron behavior in low-dimensional systems underpins next-generation electronics. Traditional top-down approaches to producing nanostructures introduce disorder, whereas bottom-up routes can offer high material quality but provide limited control over geometry. Recently, our group developed a technique for growing single-crystal nanostructures of Bi, Sn, and In with predefined geometries defined by molds lined hexagonal-boron-nitride (hBN). I will discuss our efforts to extend this technique to semiconducting nanostructures, such as Te and InSb, which have higher melting points.

Metal-Organic Frameworks (MOFs) are porous nanomaterials with high surface areas that show strong potential for targeted drug delivery applications due to their tunable structures. Researchers designed a novel drug-delivery system to target specific sites, such as cancerous tissues. In this study, zirconium (Zr)-based MOFs were modified with polyethylene glycol (PEG) to investigate their potential as pH-responsive drug carriers. Drug release from model drug-loaded MOFs was analyzed using UV-Vis spectroscopy under neutral and acidic conditions. Results suggest that PEGylation improves drug loading and enables pH-dependent release, highlighting the potential of PEGylated Zr-based MOFs as candidates for targeted cancer drug delivery. 

Ion exchange membranes (IEMs) are filtration materials that enable diverse water and wastewater treatment applications, including water desalination and resource recovery from waste. These applications involve complex mixtures, such as seawater and industrial wastewater, characterized by the presence of multiple salts at varying concentrations. The ions of interest are often present at low concentrations relative to background environmental ions, making selective separations challenging. For instance, e orts to extract lithium from brine are challenged by the low concentration of lithium ions relative to sodium, magnesium, and calcium. Hence, designing IEMs with high selectivity for specific ions in mixtures is essential for emerging resource recovery applications. IEM selectivity is usually characterized by measuring a partition (i.e., sorption) coefficient in single salt solutions. However, single salt partition coefficients cannot be easily used to predict IEM performance in mixed salt solutions. Therefore, we hypothesized that measuring the kinetic parameters (i.e., rate orders and activation energies) for ion sorption would provide a di erent approach to evaluating selectivity. To test this hypothesis, we measured change in ion concentration over time for Li, Na, and K in single and mixed salt solutions over a range of temperatures (25°C, 35°C, 45°C and 55°C) and obtained the rate constants by fitting multiple rate laws and activation energies using the Arrhenius equation.  Through statistical tests, the results showed that there was no significant di erence between the rate of single and mixed salt conditions, except for the zero order rate constants. Preliminary analysis indicated that the pseudo second order rate law best fit the data. Those rates were then used to evaluate the activation energies.

Plant-parasitic nematodes (PPNs) threaten global food production, with Meloidogyne incognita (root-knot nematodes, RKNs) being a major concern due to their ability to damage plant roots, reducing crop yields. Current control methods, including synthetic chemical fumigants, are harmful to humans and the environment. Chalcones, plant-derived precursors to flavonoids, have shown promising nematicidal effects, but their mechanism of action remains unknown. Our hypothesis suggests that chalcones inhibit a nematode protein by binding to its active site, preventing its function. Mutations in the corresponding gene may alter protein conformation, preventing chalcone binding and leading to resistance. To investigate this, our lab used Caenorhabditis elegans as a model organism. whole-genome sequencing followed by bioinformatics approach on ethyl methanesulfonate (EMS) generated resistant mutant lines, were used to identified candidate genes linked to resistance. The current study aimed to identify protein targets by validating putative genes responsible for C. elegans resistance to Chalcone 17 (the candidate genes are sly-1, Y62H9A.8, Y53F4B.21, C30D11.5, C50E10.13) and 30 (the candidate genes are clec-9, cyp-13A10, gst-5, ifb-2, K06A9.1) using CRISPR-Cas9. Single guide RNAs (sgRNAs) were designed, editing efficiency of each guide was tested in vitro, and visualized via agarose gel electrophoresis to detect shifts in gene sequence size after digestion. Microinjection facilitated in vivo delivery of sgRNAs for gene editing. Our findings identified ifb-2, an intermediate filament (IF) protein gene expressed in the intestine, as responsible for Chalcone 30 resistance. For Chalcone 17 resistance, we identified sly-1, a gene encoding a Sec1 family domain protein involved in vesicle transport from the endoplasmic reticulum to the Golgi. This strongly implied that the protein products of these gene were the primary cellular targets of the Chalcones. In silico molecular docking studies were conducted to characterized the atomic-level interactions and confirmed a strong binding affinity between the chalcones and their respective protein targets. Understanding how chalcones affect protein product of these genes will guide future studies on the structural and functional differences between mutant and wild-type proteins, advancing the development of targeted nematode control strategies.

Medical imaging datasets often exhibit class imbalances, particularly for underrepresented racial and minority groups, which can reduce the diagnostic accuracy and fairness of AI-driven Computer-Aided Diagnosis (CAD) systems. This study evaluates the use of Conditional Wasserstein Generative Adversarial Networks with Gradient Penalty (CWGAN-GP) to generate synthetic chest X-ray images for minority groups, creating a racially balanced training dataset. Multi-label classification of six thoracic diseases was performed using DenseNet-121.Experiments comparing racially imbalanced, racially balanced, and racially balanced augmented datasets demonstrated that synthetic image augmentation improved minority group performance, increasing the overall validation AUC from 0.767 to 0.858. These findings indicate that GAN-generated images can enhance diagnostic equity without compromising overall model performance, providing a scalable strategy to mitigate bias in medical AI systems.

Large galaxies often host a supermassive black hole (SMBH) at their center. This central region is known as an active galactic nucleus (AGN). The growth history of SMBHs across cosmic time is crucial to understanding galactic and SMBH evolution. Our ability to study this coevolution is hindered by a large fraction of heavily obscured AGN that reside within reservoirs of obscuring gas and dust. I present a method of searching for heavily obscured AGN in the local universe (>100 Mpc) using data from the Chandra X-ray Observatory. We can study the physical processes within AGN by analyzing their X-ray spectra. Obscured AGN exhibit intense iron emission lines in their X-ray spectra. We created images of galaxies isolating light near the wavelength of these iron lines to reveal hidden AGN. I will highlight some initial results from the survey, including the identification of a highly probable new heavily obscured AGN located a mere 15 Mpc from the Milky Way. In addition, we find that iron-band imaging identifies candidate inactive galaxies when combined with other multi-wavelength indicators. Our initial results thus suggest that narrow-band imaging is an effective way to identify heavily obscured AGN, holding prospects as a complementary strategy for current and future instruments to complete the AGN census.

Cells respond to environmental stress by quickly altering gene expression, with translational control being one of the fastest and most critical mechanisms. In yeast, translation initiation is globally downregulated during stress while specific stress response genes are upregulated. A key regulator of this entire process is Ded1, a DEAD-box RNA helicase that plays important roles in translation initiation by unwinding secondary structures on the 5’ end of mRNA and interacting with the eIF4F complex. DEAD-box RNA helicases, characterized by a conserved D-E-A-D motif, are ATP-dependent RNA remodeling enzymes that enable dynamic regulation of RNA metabolism. Previous work has shown that in addition to its normal role promoting initiation, Ded1 also contributes to translation repression under stress conditions that involve inhibition of the Target-of-Rapamycin (TOR) pathway. Specifically, a ded1 mutant that deletes its C-terminal domain (ded1-ΔCT) results in resistance to rapamycin-induced stress, suggesting that the C-terminus plays an important role in stress-related translation regulation. However, the genetic and cellular pathways that interact with Ded1 to promote this translational control are not widely understood.

The primary objective of this thesis proposal is to investigate how Ded1, and its C-terminal domain regulate translation during stress and to characterize factors that modify this function. Building on a Synthetic Genetic Array (SGA) screen that identified genetic interactions with ded1-ΔCT, this project focuses on three candidate genes that were selected based on interaction strength and previously-identified involvement in stress related pathways: VTS1, a member of the CCR4-NOT complex that promotes deadenylation dependent mRNA decay; VPS34, a phosphatidylinositol 3-kinase required for autophagy and endosomal trafficking; and STP22, a component of the ESCRT-I complex that mediates endosomal protein sorting and vacuolar degradation. The goal is to characterize how these genes enhance or suppress the ded1-ΔCT phenotype under rapamycin-induced stress and to determine how their interaction with Ded1 influences translational control.

Preliminary growth assays show that deletion of VTS1 enhanced the rapamycin-resistant phenotype of ded1-ΔCT, while deletion of VPS34 suppressed this phenotype. These results suggest links between Ded1 and RNA binding, autophagy, and membrane trafficking pathways. They also suggest that Ded1 translational repression is integrated with multiple cell stress response mechanisms rather than acting in isolation. To investigate how Ded1 and its C-terminal domain regulate translation during stress, this study will combine genetic and cell biology approaches. Growth assays under rapamycin-induced stress will be used to fully characterize how deletion of STP22 modifies the ded1-ΔCT phenotype, allowing identification of its role as a potential genetic enhancer or suppressor. To assess how these interactions affect translational machinery, protein lysates will be prepared following rapamycin treatment and analyzed by western blotting to examine Ded1 and eIF4G1 protein behavior under stress. In addition, VTS1, VPS34, and STP22 will be HA-tagged to assess whether their protein levels are altered in the ded1-ΔCT background with and without rapamycin. To determine whether they directly interact with Ded1 through co-immunoprecipitation experiments.

Together, these approaches aim to clarify how Ded1 contributes to translation repression during stress and how genetic interactors modify this process. Understanding how Ded1 regulates translation stress response pathways is important because DEAD box RNA helicases, including DED1, are highly conserved and have been implicated in human diseases such as cancer. Therefore, this work will advance our understanding of translation regulation under stress and how Ded1 function is integrated in the networks of cell stress.

Claims of water permeability in nail polish exist, but certifications provided on websites raise concerns. Some lack specific details about the employed methods, while others are outdated or only cover testing of a limited range of product shades. The UToledo Cosmetic Science research group conducted in vitro testing of three water-permeable nail polishes and did not find substantial evidence supporting their water permeability. This ongoing research aims to replicate the in vitro findings using a non-invasive device, the nano Tewameter, to measure water loss (TOWL) through the human nail plate. To date, only a few studies have utilized a Tewameter to measure TOWL, none of which have specifically focused on water-permeable nail polishes. 

The coastal sage scrub is a unique California ecosystem, and one of its most common plants, Artemisia californica, grows today in areas with some of the worst air pollution. These plants release volatile organic compounds that react with nitrogen oxides to form ozone, but models don’t fully capture which compounds are actually emitted. Last summer, our group discovered unusual, “irregular” monoterpenes from A. californica, Santolina Triene, and Artemisia Triene, compounds not typically found in current climate models. Using SPME-GC-MS, I began tracking VOCs emitted by A. Californica at the Bernard Field Station to observe how emissions change throughout the year. One plant seems to be a completely different chemotype, producing none of the irregular terpenes at all. To understand the biological origin of these compounds, I am now sectioning leaves and analyzing specific tissues to identify where these irregular terpenes are produced and to probe the underlying biosynthetic pathway. These results demonstrate that terpene chemistry in sage scrub ecosystems is more complex than previously thought, with implications for how we represent vegetation in air quality models.  

Students: Chigozirim Ifebi, Brent Kong

Prior work has shown that the “refusal direction” in Large Language Models (LLMs) is cross-lingually universal. However, multilingual jailbreaks can still work because the model fails earlier at triggering that refusal circuitry. Additionally, prior work has shown that refusal can be controlled by a single direction and that this direction transfers across many languages. Yet, multilingual jailbreaks remain practical.  This project reframes the question from “is refusal language-dependent?” to “where does the refusal trigger fail across languages?” in order to localize and patch trigger failures under realistic cross-lingual perturbations. Instead of attributing refusal to a single neuron or monosemantic component, we leverage polysemanticity and conduct a refusal subspace analysis to identify distributed directions that jointly encode harmfulness detection and refusal activation. We build a stress suite of meaning-preserving perturbations (translationese, code-switching, transliteration, script mixing, and low-resource paraphrases), measure jailbreak success, and causally probe which representational subspaces fail to separate harmful from harmless prompts across languages. Inspired by recent findings that a universal refusal direction alone is insufficient for robust multilingual safety, we explicitly disentangle the harmfulness and refusal directions and analyze how weak separation in non-English representation spaces enables cross-lingual jailbreaks. Finally, we evaluate lightweight geometric interventions that sharpen harmfulness–harmlessness separation within the refusal-aligned subspace, aiming to improve multilingual robustness without collapsing behavior into English-centric routing.

A conceptual Stirling-driven Small Modular Reactor (SMR) was designed and evaluated as an alternative to conventional small-scale nuclear systems. Instead of using a steam-based Rankine cycle, the design pairs a compact reactor core with a closed-cycle Stirling engine to convert heat into electricity. A detailed simulation modeled thermal output, heat transfer, temperature distribution, engine thermodynamics, mechanical work, and overall efficiency. The model was used to test steady-state operation and transient conditions such as startup and load changes. Results show the system can operate within safe thermal limits while maintaining stable performance and good efficiency. The engine responds smoothly to changes in heat input and power demand, with manageable thermal stress. Overall, the concept appears promising for modular low-carbon power, though it still requires experimental validation.