Goal
Create an AI-enhanced multi-scale computational model and integrate with dense data networks to significantly enhance the predictive capability for short and long-term coastal hazards.
Urban and Coastal Flood Modeling
The team willl harness machine learning to create detailed land classification from aerial imagery, LiDAR, and other resources for the State of Delaware. This classification will focus on urban land covers, in particular, various types of impervious surfaces (buildings, driveways, streets, sidewalks, parking lots) and tree cover. This effort will provide critical information for urban/coastal flood modeling and facilitate future groundwater and salinization application.
Sub-Model for Settling Velocities of Cohesive Sediment in Region-Scale Modeling of Coastal Processes
Transport of cohesive sediment (mud) in estuaries, wetlands, and saltmarshes is key to land preservation, water quality and ecosystem health. Due to a complex process called flocculation, it is very difficult to determine its settling velocity of cohesive particles. During fall 2024, we will harness machine learning to create an efficient surrogate model to predict time series of floc size distribution subject to environmental forcing using comprehensive training data provided by a detailed numerical model for flocculation dynamics. This effort will provide a highly efficient and reliable tool to be used as a sub-model for settling velocities of cohesive sediment in region-scale modeling of coastal processes.
Goal – Project 1
Develop a human-centered, agentic AI platform that supports efficient and transparent collection, annotation, and analysis of online social network data while preserving human judgment and user control.
AI4OSN : Human-Centered Agentic AI Platform for Online Social Network Research
AI4OSN is a human-centered, no-code, agentic AI platform designed to support online social network (OSN) research. The platform integrates large language models and multi-agent workflows into an interactive system that assists researchers with data collection, annotation, and early-stage analysis while preserving human judgment and control.
The system enables users to collect real-time social media data through an intuitive search interface, collaboratively annotate multimodal content, and generate structured insights such as summaries, sentiment descriptions, and descriptive statistics. AI assistance is intentionally designed as supportive rather than authoritative, allowing researchers to review, validate, and override AI outputs. Through iterative design and usability testing, the platform demonstrates how agentic AI can reduce technical barriers, lower cognitive workload, and accelerate research workflows for both technical and non-technical users.
Goal – Project 2
Develop a transformer-based topic model (BERT) to analyze a large collection of Russian foreign policy speech texts & images. This helps us to predict how and when foreign governments signal (changing) foreign policy priorities.
Understanding Foreign Policy Priorities Through Multi-Model Topic Models of Foreign Policy Speech Texts and Speech Photographs
The team collected two corpora of Russian foreign policy speeches of 10k-12k speeches each, with associated images. Have run and interpreted BERTopic for texts of each speech corpora, with validated results. This offers new resources for policymakers & analysts to monitor the policy priorities of less transparent regimes. For social scientists, our approach speaks to the dynamics & determinants of foreign policy agenda setting.
MATERIAL SCIENCE
Goal
Develop and integrate a lattice Boltzmann fluid dynamics simulation based on AMReX into the LAMMPS package, extending its capabilities to model the phase behavior of multicomponent emulsions under thermal fluctuations, support complex boundary conditions, and enable GPU acceleration.
Phase-behavior of multicomponent emulsions under the influence of thermal fluctuations
Development and integration of a lattice Boltzmann fluid dynamics framework based on AMReX into the LAMMPS molecular dynamics package. The project focuses on modeling the phase behavior of multicomponent emulsions under thermal fluctuations, supporting complex and physically realistic boundary conditions, and implementing GPU acceleration for scalable, high-performance simulations.
GEOGRAPHY AND REMOTE SENSING
Goal
Develop an open, scalable, and reproducible pipeline to map global irrigated croplands at field-scale resolution, enabling improved monitoring of agricultural water use, food security, and climate adaptation.
GMIA-NEXT: Next-Generation Global Map of Irrigated Areas Using Data Fusion
The DARSE team is supporting GMIA-NEXT, a next-generation global irrigation mapping effort that produces 30-meter maps of irrigated and rainfed cropland for the 2023/2024 growing season and provides a foundation for extending the dataset into a yearly time series since 2000. The workflow fuses multi-sensor satellite observations such as Landsat-derived seasonal vegetation indices with agroecological zones, hydroclimatic and topographic variables, and sub-national irrigation statistics to generate pixel-wise irrigation probability and binary irrigation maps. To scale globally, we compute predictors in Google Earth Engine and run model inference and post-processing on UD’s DARWIN high-performance computing cluster using parallel tiling and job arrays. All code, ground truth, and data products are openly released to support reproducible science and decision support for water and agricultural planning.
EDUCATION
Goal
Leverage machine learning to generate actionable, data-driven insights into high school factors that predict college readiness.
Predicting College Readiness from Longitudinal High School Data
The DARSE team is supporting an education project that applies predictive machine learning models to multi-year high school student data to estimate college readiness using academic and behavioral indicators such as course enrollments, grades, assessment results, and attendance. The workflow includes data integration across multiple sources, feature engineering, model training and evaluation, and the development of interactive dashboards that translate model outputs into interpretable metrics for educators and stakeholders. The ML models will learn trends across cohorts, school types, Delaware counties, and connect high school trajectories to first-year college outcomes using National Student Clearinghouse (NSC) enrollment patterns such as enrollment, attendance, institutions, and attendance. Overall, the project provides a scalable framework for understanding which high school experiences and supports are most strongly associated with postsecondary readiness and success.
ATHLETICS
Goal
Assist baseball pitchers perform better by interpreting, distilling, and communicating pitch data as actionable information to players and coaches
Data-Driven Baseball Analytics: Predictive Models from TrackMan Radar Data
Developing machine learning models to quantify pitch effectiveness by estimating Expected Runs and Expected Pitch Impact from vast amounts of TrackMan data. Together with a “Stuff+” model that predicts the likelihood of a swing and a miss, these models are designed for deployment in a live performance analytics dashboard to assist coaching decisions, pitcher development, and player recruitment.
ART CONSERVATION
Goal
Integrate building operational data with temperature and relative humidity sensor measurements to identify risk conditions and support preventive conservation decisions across museum spaces.
Winterthur Museum: Environmental Monitoring Analysis and Dashboard Development
The team is combining room-level temperature and relative humidity measurements with building operational data (e.g., energy use and HVAC/system logs) and staff survey responses to quantify comfort and preservation risk metrics. We then deliver clear, actionable visual summaries and dashboards that help museum staff and researchers identify at-risk spaces, track trends over time, and prioritize preventive conservation interventions.
PLANT & SOIL SCIENCES
Goal
Automate the daily export of soil greenhouse gas (GHG) data from instruments to a remote server, ensuring seamless data flow for real-time access and analysis as well as ensure the accuracy, reliability and usability of the data collected through automated chambers.
Automated Soil GHG Data Quality Assurance and Quality Control System
Reduces manual effort through automated data validation, increasing the overall efficienty of data collection and processing. The team has identified gaps in data due to missing recordings, sensor faultures, or power outages and are developing an automated Python-based QC/QA framework to validate data. There are implemented checks for missing data, range violations, and identifying outliers.
UD OPERATIONS
Goal
Develop a secure, in-house conversational AI system tailored to University of Delaware needs that supports campus operations, improves access to institutional knowledge, and reduces repetitive workload through governed, auditable AI workflows.
HenChat: An In-House UD Chatbot with Fine-Tuning and Agentic Workflows
In collaboration with First State AI Institute, the DARSE team is helping build HenChat, an internal chatbot for the University of Delaware designed to provide reliable, UD-specific assistance while meeting institutional needs around privacy, governance, and operational control. The project focuses on adapting an open-source large language model for UD use through targeted fine-tuning and retrieval from approved campus knowledge sources, enabling accurate responses for common questions and workflows. On top of the core model, we are developing agentic workflows that can execute multi-step operational tasks—such as routing requests, generating draft communications, summarizing tickets, and assisting staff with standard procedures—while maintaining traceability and human-in-the-loop oversight. This effort creates a scalable, campus-ready platform that improves response quality, reduces repetitive workload, and provides a foundation for expanding AI-enabled services across UD.