Mason Journals (George Mason Univ.)
Not a member yet
3256 research outputs found
Sort by
A Temporal Fusion Transformer Framework for 24-Hour Ozone Forecasting Across the Continental United States
No full textGround-level ozone (O3), a harmful air pollutant, poses a significant public health threat, making accurate 24-hour forecasting a critical priority. Predicting ozone concentrations across the continental United States is challenged by complex atmospheric dynamics and the failure of conventional models to capture long-range temporal dependencies. This research introduces a novel forecasting framework that applies a Temporal Fusion Transformer (TFT) to predict hourly ozone concentrations 24 hours in advance. The model is trained on a fused dataset combining ground-truth measurements from approximately 8,000 U.S. AirNOW stations with meteorological and chemical data from the Environmental Protection Agency (EPA) organization. The TFT architecture is uniquely suited for this task, integrating static (e.g., station location) and dynamic (e.g., meteorology, precursor gases) inputs. The model's performance was benchmarked against a Community Multiscale Air Quality (CMAQ) model, achieving a Root Mean Square Error (RMSE) of 6.9 ppb and a Mean Absolute Error (MAE) of 5.3 ppb. This represents a 24.2% and 32.1% improvement over the baseline, respectively. The model’s interpretable attention mechanism identified ozone in parts per billion (O₃), nitrogen dioxide in parts per billion (NO₂), carbon monoxide in parts per million (CO), and organic carbon in micrograms per cubic meter (OC) as the most influential predictors. The TFT model establishes a new state-of-the-art for 24-hour ozone forecasting in the U.S., validating the application of advanced transformer architectures to complex environmental science problems and providing a powerful tool to support more timely public health warnings
AI-Powered Air Quality Monitoring Using ESP32 and BME280 Sensors for Incense Smoke Classification
No full textAir quality monitoring is crucial for understanding environmental and health impacts of pollutants. Accurate, low-cost sensors combined with machine learning models offer new possibilities for real-time detection of air quality variations caused by common sources such as incense smoke. However, developing reliable sensor systems that integrate environmental measurements with predictive algorithms remains a challenge, especially when working with limited hardware.
This project started with designing a sensor platform using a BME280 environmental sensor wired to an ESP32 microcontroller to measure temperature, humidity, pressure, and altitude. Initial setup errors caused sensor damage by exposing it to 5 V instead of 3.3 V power, which was subsequently fixed. After successful recovery was ensured, collecting sensor data from multiple types of incense sticks under controlled conditions was the next step. Using this dataset, I trained a TensorFlow Lite machine learning model to classify incense types based on environmental factors, achieving a validation accuracy of approximately 63%. The model was deployed for offline analysis of recorded sensor data in Google Colab, demonstrating real-time prediction potential, showcasing its ability to classify incense smoke.
This approach highlights the practicality of integrating low-cost environmental sensors with AI models to identify pollution sources dynamically. Future work includes improving model accuracy with more diverse data, implementing live prediction directly on the ESP32 using TinyML, and expanding the system to detect a wider range of air quality factors. These advancements could contribute to accessible, portable air quality monitoring tools for personal and community health application
Spatially Explicit Mapping and Assessment of Urban Heat Islands Using AI
No full textThe Urban Heat Island effect is a threat to urban livability. This phenomenon is most noticeable in dense urban areas, where impervious surfaces absorb and store solar energy as heat, while the cooling benefits of vegetation—shade and evapotranspiration—are diminished. In this study, we integrate satellite-derived datasets with the InVEST Urban Cooling model to quantify heat mitigation potential across the Washington D.C. area. The model computes a Heat Mitigation Index (HMI) by combining land use maps with biophysical drivers: shade, evapotranspiration, albedo, and proximity to “cooling islands.” Generating the high-resolution inputs required (land cover, land surface temperature, albedo, canopy cover) has traditionally been time and labor intensive. Leveraging cloud computing and AI-powered machine learning algorithms, this study develops a model to generate the land use map using Sentinel satellite data. Time-series composites of daytime surface temperature and albedo were computed using the Google Earth Engine cloud computing platform. Results show the average Heat Mitigation Index across the region is 0.30, with tree-covered area achieving values up to 0.54, indicating stronger cooling potential. Simulations reveal that without the influence of vegetation and bodies of water, average urban temperatures rise to 39.17°C, with high-density zones reaching 40.64°C. This approach demonstrates the power of AI-driven analytics to deliver a scalable, data-informed framework that is adaptable to all areas. It enables efficient scenario analysis, supports climate-resilient urban planning, and informs the design of targeted interventions (e.g. green corridors) by identifying priority areas for urban heat risk mitigation and adaptation
A 50-year Retrospective: How American Journalists Covered the Vietnam War, and the Lingering Aftermath
Full text linkA Comparative Evaluation of Large Language Models for Integration within Argumentation and Graph-based Open Student Modeling Software (ARGOS)
No full textThis project aims to develop Argumentation and Graph-based Open Student Modeling Software (ARGOS), an Intelligent Tutoring System (ITS) that displays a dynamic Open Student Model (OSM) based on real-time conversation. Researchers have explored various conversational agents, assessment-based knowledge displays, and Bayesian Knowledge Tracing (BKT) to model students’ understanding. However, it lacks a system that uses live dialogue to track and update a student’s cognitive progress and understanding in real-time. To determine which Large Language Model (LLM) best suits ARGOS, this study evaluates the abilities of ChatGPT 4o, Gemini 2.5 Pro, Grok 3, Meta AI, and Claude Sonnet 4 to guide student understanding. First, the LLMs were given a system prompt with detailed instructions on how to act as Socratic tutors by asking guiding questions and staying on task. Each model was prompted, one dialogue at a time, with four standardized scenarios involving an imaginary student attempting to factor 12x² + 17x + 6. The scenarios were designed to simulate: (1) providing a correct solution, (2) making a standard conceptual error, (3) making a simple calculation mistake, and (4) going off-topic. Then, four researchers scored the conversations produced by the LLMs using a detailed 30-point rubric that evaluated each model on educational quality, factual accuracy, and instructional following ability. Gemini 2.5 Pro scored the highest, with an average score and standard deviation of 26.625 and 2.55, respectively. Additionally, the one-way ANOVA test resulted in a p-value of 3.05*10-7, proving the data was statistically significant. Thus, the experiment proved Gemini 2.5 Pro as the most effective LLM, ensuring ARGOS will be built on a capable and reliable tutor for real-time conversation. Further research will focus on testing Gemini 2.5 Pro’s ability to accurately quantify students' mastery scores
Moisture Source and Isotopic Signatures of Extreme Rainfall in the Washington, D.C. Area
No full textThe Washington, D.C. area is located at the intersection of multiple air masses, which fuel the intensity of storms and make this region vulnerable to extreme rainfall events. Previous studies have identified four major storm types in this area: general storms, tropical storms, hybrid storms, and local storms. However, the relationship between moisture sources and these storm types, as well as their isotope signatures in rainfall, remains poorly understood. Here, we use Hybrid Single Particle Lagrangian Integrated Trajectory Model to track the moisture origins of extreme rainfall events from 2023 to 2025 and investigate their relationship to storm types. We identified the top 10% of extreme rainfall events during the summer (June-August) and winter (October-December) months, based on meteorological data. Multiple statistical methods were applied to investigate backward moisture trajectories related to these rainfall events, including backward ensemble trajectories, frequency analysis, and cluster analysis. We found that summer rainfall was primarily sourced from the Gulf of Mexico, the Mid-Atlantic Ocean, and land transport from the Northern Pacific. Main winter moisture sources showed more diverse moisture origins, including the North Atlantic, the Great Lakes region, and the North Pacific. Such seasonal shifts in the dominant sources are generally consistent with the monthly average of sparse existing rainwater ¹⁸O observations, which show more enriched (positive) values during the summer months. Moreover, the rainwater ¹⁸O values from individual events exhibit significant variability, underscoring the need for more rainwater isotopic analysis to better understand the link between storm types, moisture sources, and isotopic composition
Exploring proteomic pathways linked to infertility within endometriosis patients through Mass Spectrometry
No full textEndometriosis, a chronic gynecological disorder affecting over 10% of womenof reproductive age globally, causes endometrial tissues to spread outside the uterus,leaving scarring, adhesions, and cysts. Furthermore, studies have shown that 25-50% ofwomen living with endometriosis also experience infertility. Most studies done on therelation between endometriosis and fertility have focused on anatomical irregularities andhormonal imbalances interrupting the ovum and sperm from uniting. However, this raisesthe question if the cause of infertility within this demographic may also lie within thefollicular fluid (fluid produced by the antral follicles that protects a developing egg). In thisstudy, we performed DIA mass spectrometry to identify potentially dysregulated proteinsin the follicular fluid affecting ovum viability in 9 endometriosis patients and 5 patientsseeking IVF treatment (set as a control). We used nanoparticles to ensure deeperproteome coverage and the ability to quantify low abundance proteins within the follicularfluid. We found 50 unique endometriosis proteins, 353 shared proteins, 437 controlspecific proteins between our two groups. To identify the key biological processes that aredifferent between the upregulated and downregulated proteins, we used WebGestalt over-representation analysis. In the downregulated protein group, the biological processesincluded humoral immune response, cell killing, and protein maturation. The upregulatedprotein group biological processes were acute inflammation, metal ion transport, andtissue migration. Our results suggest that the follicular fluid proteome may be used toidentify biomarkers related to infertility and endometriosis
Discovering the Trend and Patterns of Wildland Fires in USA and the Role of Climate Change
No full textWith rising global temperatures due to climate change, the risk of wildland fires has grown more significantly as a result of dry and warm weather conditions, especially in the western United States. Understanding the trends and patterns of wildland fire occurrences is crucial for effective land management to mitigate the severe impacts of such fires. The objective of this project is to discover the temporal and spatial patterns of wildland fires across the USA and explore how these patterns relate to climate change. In this study, we examined large wildland fire incidents in the USA from 1984 to 2024, alongside satellite remote sensing data on air temperature, precipitation, and soil moisture, all of which are intimately linked to fire ignition. After 2000, while the number of wildland fires has been relatively less than in the previous two decades, there has been a notable rise in the extent of areas burned. The eastern states experience a higher number of fire incidents, and these fires tend to be significantly smaller compared to those in the western states, so we focus on climate change analysis in the western states. The quantitative evaluation of temperature, soil moisture, and precipitation indicates a correlation between dry conditions and the size of burned areas by wildland fires. Further investigation in this field could develop a machine learning-based model to forecast the severity of damage caused by wildfires, utilizing climate data derived from satellite remote sensing observations and simulations, which would support government agencies and related communities in land management and fire preparation strategies
Characterization of Virginia’s Climate
No full textThe study aims to characterize the general temperature climate of Virginia. It uses the National Oceanic and Atmospheric Administration’s (NOAA) nClimGrid-Daily dataset, a gridded dataset (2.5 arcminute resolution) of temperature and precipitation in the contiguous United States. Gridded data facilitates analysis of spatial characteristics by providing evenly-spaced data, compared to data from irregularly-spaced weather observations. The study examines daily high, low, and average temperatures for 1991-2020, a thirty-year time period which is standard in defining an area’s climatology. Preliminary analysis is confined to coldest and hottest months, January and July. In addition to monthly means, the temperatures of extreme hot and cold days are of interest because of the effects of extreme events on health, transportation, and other aspects of society and ecology. Therefore, this study includes 5th and 95th percentile values as measures of extremes, and it examines geographical features of individual extreme heat events in Virginia.
Contours of uniform temperature align southwest to northeast, with the coldest temperatures occurring in the Blue Ridge/Appalachian Mountains, and January having a larger temperature range than July. Across these metrics the daily high (daytime) temperature range in January is slightly smaller than the daily low (nighttime) range. Most notably, coastal moderation is more apparent at night than during the day, indicating that the warming effects of coastal moderation are more influential than its cooling effect on Virginia’s climate. This characteristic is also found in comparing the days and nights of July, further supporting the finding. The effects of topography (lower temperature at higher elevation) are greater in July than in January. By utilizing the nClimGrid-Daily dataset, this study identified general climatic patterns in Virginia and found unique characteristics in Virginia’s day-night cycle as well as the seasons