Active Grants

As artificial intelligence and machine learning evolve, the number of pre-trained models, or PTMs, is rapidly increasing, leading to significant storage needs. Popular model registries like Hugging Face are hosting hundreds of thousands of PTMs, with storage requirements exceeding 500 terabytes.

Current data reduction tools are ineffective for PTMs. The project proposes ELVES, a new data reduction approach combining PTM-aware compression algorithms with computational storage devices and disaggregated CXL for efficient model storage and access. ELVES uses a near-storage compression method called ELF, leveraging near-storage computation in computational storage devices like Samsung SmartSSD to enhance storage capacity and reduce data movement costs.

The project aims to rethink storage stack design for PTMs, proposing three use cases: high-performance model registry caching, tiered model storage with CXL and SmartSSD, and a scalable inference service for LLMs using CXL-integrated storage.

Screening promotes early detection of cancer to decrease mortality. Unfortunately, significant racial disparities exist in lung cancer screening. The current screening guidelines are based on a combination of age and smoking pack-year criteria derived from a national lung screening trial that was primarily (91%) white.

Importantly, smoking behaviors and baseline risks for lung cancer differ greatly between African Americans and whites. Because of this, a risk-based screening strategy may provide a more equitable assessment of eligibility than current screening guidelines. However, the development of personalized risk prediction models for lung cancer in African Americans has been limited. To address this gap and to improve equity in screening eligibility, we propose building a personalized prediction tool using the combined data from three large-scale population-based prospective cohorts with substantial African American representation.

We will employ innovative machine learning techniques and state-of-the- art statistical methods to build a well-calibrated lung cancer prediction model for African Americans. Careful examination will identify sub-populations (such as women, low socioeconomic status, rural, age groups, etc.) that will benefit from screening. A key innovative aspect of this proposal is its community-engaged approach and partnership with a Community Advisory Board, both of which will help translate our empirical findings into the design of a patient-oriented decision aid.

Infectious diseases cause over 13 million deaths annually worldwide. The rapid growth of the human population and its adaptability to various environmental conditions have led to unprecedented interactions between humans and other species. This, combined with globalization, antimicrobial resistance, urbanization, climate change, and ecological pressures, has heightened the risk of a global pandemic. Computation and data sciences can address these complexities and revolutionize real-time epidemiology, offering new ways to reduce the global burden of infectious diseases.

This project aims to advance computational foundations, engineering principles, theoretical understanding, and novel technologies to implement global infectious disease computational epidemiology. The tools developed will enhance decision-making for epidemic planning and response, improve inter-agency coordination, and facilitate outbreak response. The team will collaborate with public health agencies and universities worldwide to deploy these technologies during epidemic outbreaks. Educational programs will be established to foster interest in computational science and train multidisciplinary scientists in public health.

The project will develop a computational theory of spreading and control processes on dynamic multi-scale, multi-layer networks, using artificial intelligence, machine learning, and social sciences. New techniques will enable large-scale simulations of epidemics and social interactions, providing insights into epidemic control. Pervasive computing technologies will support disease surveillance and real-time response. The computational advances will also be applicable to other fields like cybersecurity, ecology, economics, and social sciences. The project will address concerns such as privacy, model interpretability, effective interventions, strategic behaviors, and fairness. The Center for Computational Research in Epidemiology will be established at UVA to support real-time epidemiology and improve outbreak response.

This 5-year prospective, longitudinal cohort study of Veterans and Service Members following discharge from acute rehabilitation with all severity levels of traumatic brain injury (TBI) has the following aims:

1) Determine incidence and association of chronic health conditions with rehabilitation outcome in Veterans and Service Members with TBI. Research questions include discovering the frequency of health conditions in the first five years post-injury as well as determining what health conditions are associated with function in the first five years post injury/

2) Determine environmental and contextual protective and risk factors impacting ongoing life care needs of Veterans and Service Members with chronic TBI. Research questions include identifying the chronic rehabilitation needs the first five years post injury and what environmental and personal factors are associated with unmet needs.

3) Describe perceptions of healthcare needs the first five years post injury. Research question include discovering what patient and family perceived facilitators and barriers exist for accessing rehabilitation and healthcare services to meet needs, as well as what providers perceive as either facilitators or barriers for addressing long-term rehabilitation needs of patients and families.

Societal understanding of how artificial intelligence will transform education in the years ahead remains in its early stages, but a newly funded project from researchers at the University of Virginia may shed light on one key area: Can generative AI tools be used to develop high-quality test items for K-12 schools?

It is funded through a yearlong grant from the National Science Foundation totaling $50,000. The grant was awarded through the NSF’s I-Corps program, which was launched in 2011 as a way to allow research teams to quickly determine the market potential of their innovations through the customer discovery process.

Developing high-quality test items has long been a laborious, time-consuming, and expensive process. The project team hopes that their automatic item generation and evaluation system could be used by a variety of stakeholders — including K-12 testing companies, language testing agencies, and online education platforms — to reduce these burdens while still producing test questions that align with required specifications and ensure fairness.

The team will also collaborate with UVA’s Licensing & Ventures Group on patent applications.

In this AFOSR-MURI project, Dr. Baek will collaborate with the multi-university MURI collaborators to develop a suite of predictive algorithms for cyclotetramethylene-tetranitramine (HMX) explosives. The overall goal of the project is to produce new meso-informed, microstructurally aware reactive burn models for the ignition, initiation, and detonation of HMX.

Baek and the UVA team will contribute to this project with their machine learning expertise. The UVA team will develop convolutional neural network (CNN) models and other AI models to assimilate the linkage between microstructures and property/performance characteristics of HMX.

This project continues previous work by the PIs on deep learning algorithms to detect and measure the shapes of neurons in living tissue. Dr. Baek will be involved in adjusting the algorithm as more data is gathered, improvements in visualization, and the analysis of the shape and volume of the neurons.

ACTIV-6 is a multicenter, double-blind, placebo-controlled, randomized, platform clinical trial designed to evaluate the effectiveness of repurposed medicines in reducing symptoms and preventing disease progression in non-hospitalized participants with mild-to-moderate COVID-19.

This project will develop a novel machine learning model that uses a dual-path neural network (NN) structure to estimate key glycemic metrics from the Ambulatory Glucose Profile (AGP)—Time in Range (TIR), Time Above Range (TAR), and Time Below Range (TBR)—based on two weeks of sparse self-monitoring blood glucose (SMBG) data. Typically, these AGP metrics are derived from continuous glucose monitoring (CGM) data to provide comprehensive insights into glycemic variability.

However, CGM may not be feasible for all patients due to factors like cost, limited insurance coverage, or lack of access to CGM devices. In contrast, SMBG is more widely applicable, as it is cheaper, simpler to use, and requires minimal equipment, making it a practical choice for many patients. However, since SMBG monitoring in free-living conditions is usually reactive, it creates inherent bias, with measurements clustered around perceived hypo- or hyperglycemic events. To address this limitation, our approach innovatively proposes a data-driven, supervised machine learning model to transfer the knowledge of glycemic variability from large CGM observations to facilitate the prediction of key metrics from sparse SBGM data.

Anxiety disorders (ADs) are the most prevalent psychiatric disorders in childhood and adolescence, affecting about one-third of youth and contributing to significant impairment within children, families, and society at large. In recent years, many believed that neuroimaging methods held answers to key questions about how and why ADs emerge in early life, and that imaging the brains of youth with anxiety would contribute to
significant advancements in treatment.

Today, however, rates of pediatric ADs continue to rise and about one-third of youth fail to respond to current “gold standard” treatments for anxiety. There are many potential reasons why neuroimaging research has fallen short in its impact on the field of youth anxiety, including: 1) most neuroimaging studies have been conducted with small samples of youth, contributing to low power and poor reproducibility, 2) ADs (e.g., generalized AD, social AD, separation AD) are often grouped together in research despite high heterogeneity between (and within) diagnostic categories, and 3) measures of brain structure and function are not well-integrated. Integrating multiple neural measures in large samples can provide a more comprehensive and reliable understanding of the brain.

Moreover, we must differentiate between diagnostic categories of anxiety, as collapsing across all disorders may be masking unique associations between the brain and certain anxiety symptoms or disorders. To address these critical limitations, the present study will use novel data fusion techniques to examine how structure-function brain coupling is associated with the development of distinct anxiety symptoms and disorders in a large sample of youth recruited for the Adolescent Brain and Cognitive Development (ABCD) study.

The IPERT R25 application aims to formalize and expand a series of Biomedical Data Science Innovation Lab workshops, both online and in-person. These workshops will use an immersive, hands-on approach to help early-career biomedical and quantitative investigators form new collaborations on biomedical science projects. Each year, a specific biomedical topic that could benefit from a new data science perspective will be chosen. During the events, professional facilitators and senior scientist mentors will guide participants in developing innovative research project proposals, fostering community and strengthening team bonds. The program directors will provide leadership and logistical support, while other stakeholders will offer vision, expertise, and feedback. Senior faculty mentors will contribute their knowledge and oversee the development of new projects.

Early-career investigators from various quantitative and biomedical fields are encouraged to apply, with a focus on diversity, including women and underrepresented communities. Participants will receive skills training in data science methods, reproducibility, team science, and responsible conduct in research. They will also learn about research funding mechanisms through interactions with program officers from the National Institutes of Health and the National Science Foundation. Multidisciplinary teams will form during in-person workshops to develop projects involving mathematical modeling, statistical analytics, and other quantitative methods. Follow-up virtual activities will help solidify teams and their research proposals for submission to funding agencies. The program will rigorously assess skills training and team effectiveness. Outreach efforts will include a detailed website and professional-grade visual media. This program aims to build a research community aligned with the goals of the IPERT R25 program.

Over the past decade, advances in machine learning have significantly impacted many real-world applications, achieving breakthroughs across various disciplines. A new era of collaborative learning is emerging, where researchers at different sites work together to correlate disparate data and create sophisticated decision-making models. This necessitates a platform for collaborative, multi-party data analysis, allowing participants to share data with varying degrees of privacy control.

The Bridge project establishes a scalable and trusted hardware and software environment to support general collaborative machine learning. It enables scalable multi-party learning and data analysis in centralized and decentralized settings, with security and privacy guarantees. The project integrates hardware and software innovations with security and privacy mechanisms to support various types of multi-party machine learning. It aims to enable collaborative research in computer and information science and engineering, generating social and economic benefits.

The Bridge platform will develop a unified infrastructure for multi-party learning, integrating cryptographic and noise-based methods to ensure privacy. It provides tools for data access, AutoML, team creation, model vulnerability evaluation, and heterogeneous feature embeddings. The platform ensures scalability by improving asynchronous model updates, communication efficiency, fast convergence, and vertical data partitioning. It aims to build a collaborative learning community and accelerate research in areas like advanced machine learning, data privacy, trustworthy AI, and intelligent internet of things.

This project aims to serve the national interest by developing, implementing, and evaluating simulations that will improve evidence-based mathematics teaching practices. The project will create a suite of open-source, freely available simulations that support teachers to foster rich mathematical discourse in their classrooms. While doing so, it will advance knowledge of how to design simulation tools for teacher learning and will articulate design principles for creating technologies for teacher preparation programs. The project will also create open data sets of the mathematical dialogue and professional development materials so that other teacher educators can develop their own customized scenarios.

The goal of this project is to harness recent innovations in physics-informed machine learning (PIML) to enable the design and discovery of energetic material (EM) microstructures with targeted material properties and performance characteristics.

PIML bridges the gap between the strong predictive capabilities of data-driven models (e.g., deep learning) and the rules and constraints of the physical world (e.g., governing equations and boundary conditions defining the reactive multi-material dynamics). PIML allows rapid predictions of physical quantities without the need for expensive numerical calculation.

Building on PIML, this project will complete an automated experiment and design optimization loop guided by PIML and artificial intelligence (AI).

This project is focused on setting UVA students up for success in the trustworthy artificial intelligence workforce. Trustworthy AI is a system that is designed and developed to be reliable, safe, and fair, while also upholding ethical and legal standards.

The grant will provide seed funding for the Careers in Trustworthy Artificial Intelligence Program (CTAP). Through research, outreach, collaboration, and pilot courses, the program aims to provide UVA students with the skills and knowledge they will need to forge successful careers in the increasingly popular and necessary Trustworthy AI workforce.

Read more about this award

This Fellowship will provide funding to support Hicks' research, scholarship, or special projects over a three-year period. The program also offers opportunities for fellows from throughout the University community to share their scholarship and build connections with one another.

Read more about this award

The Sports Science and Analytics Collective (SSAC) is housed in the University of Virginia School of Data Science. It is an assembly of researchers, practitioners, students, and industry partners dedicated to sports, athletes, and data. Our projects range from improving athlete performance, health, welfare to understanding game strategy, and optimizing front office tactics. While we use sports as our laboratory, we aim to extend our findings to a variety of populations and environments. 

Crucial data-intensive applications like big data analytics, AI, and scientific computing need fast, elastic, cost-effective, and user-friendly storage systems. Current cloud storage solutions fall short, causing significant performance issues and high costs.

This project proposes a new cloud-native storage system called Infinistore, leveraging serverless computing for elasticity and cost-efficiency. Infinistore will scale dynamically with changing data access patterns, ensuring strong data consistency, fault tolerance, and a scalable metadata service. The project aims to revolutionize cloud storage pricing by enabling pay-per-access models and will support the development of new serverless applications.

Additionally, it includes an educational component, Infinicloud, to facilitate interdisciplinary learning and programming without server management. All project artifacts will be open source for academic and industry use.

This equipment award will support the PI's ongoing work on modeling and understanding of complex interventions on large and complex networks. The key idea is to investigate theoretical foundations of causal inference on networks and develop new computational tools for modeling, monitoring, interpreting, and evaluating complex interventions on large-scale complex networks.

This multidisciplinary project aims to accelerate the innovation of high-performance materials with complex microstructures by leveraging artificial intelligence. The project will develop new methodologies for human-AI collaboration to optimize microstructural designs for targeted properties and performance. The AI-driven design framework will significantly speed up the discovery process, reduce costs, and be scalable across various materials, benefiting industries and contributing to national security and technological advancement.

Focusing on energetic materials like propellants and explosives, the project will create methods to guide the entire design process using advanced AI techniques. This includes constructing machine learning datasets and developing a physics-informed meta-learning framework, which will be validated with experimental data. Collaboration with the Air Force Research Laboratory will enhance the project’s impact on manufacturing and validation processes, as well as student training and workforce development.

Epigenome data are crucial for understanding genetic variation and gene regulation. Produced by protocols like chromatin immunoprecipitation sequencing, known as ChIP-seq, or the assay for transposase-accessible chromatin method, commonly referred to as ATAC-seq, these data are summarized into genomic intervals defined by chromosome coordinates. Databases now hold vast amounts of these region sets, which are essential for studying gene regulation and disease. Various tools have been developed to analyze these intervals, aiding in biomedical research such as annotating genetic variations linked to diseases.

However, as data sources grow, faster algorithms and new methods are needed to handle the increasing volume. Current algorithms often only analyze pure intervals, not signal values, limiting accuracy. This project aims to address these challenges, developing new, scalable algorithms using approximate computations and the concept of interval set universes.

Additionally, innovative approaches using machine learning and natural language processing techniques are proposed to compare intervals more accurately. These advancements aim to enhance the efficiency and accuracy of biomedical research and open new avenues for exploring genomic data.

This fellowship will support Dr. Martin's application of data science on financial systems with an emphasis on safety and security. Martin is recognized for supporting opportunities for students to learn and engage in this kind of research as part of their educational and research experience.

The LaunchPad for Diabetes program is a University translational research fund supporting collaborative research projects proposing innovative solutions for the treatment of Type 1 or Type 2 diabetes.

Dr. Shakeri and his team will use this grant to support a pilot trial of advanced machine learning tools in artificial pancreas technology. DYNAMO Lab, in collaboration with the UVA Center of Diabetes, is pioneering “trainable” artificial pancreas systems that leverage machine learning to personalize diabetes management for individuals with Type 1 diabetes.

By harnessing vast amounts of continuous glucose monitoring and insulin delivery data, the lab is developing smarter, fully-automated closed-loop control algorithms. The project goal is to improve patient outcomes and enhance quality of life by providing more precise and adaptive insulin delivery.

Read more about this award

Cyberinfrastructure systems, such as high-performance computing and data centers, are essential for supporting research in academic institutions. As scientific projects become more complex and collaborative, enhancing the security and privacy of cyberinfrastructure systems is crucial to address evolving threats. This project uses an existing system as a development platform and integrates novel cybersecurity research to create a more secure and resilient cyberinfrastructure environment. It focuses on developing advanced techniques for secure data management and computation within cyberinfrastructure systems. The project has three main areas: protecting data in centralized computing, protecting data in distributed computing, and protecting data computation. By implementing state-of-the-art security protocols and privacy-enhancing technologies, the project aims to enhance the resilience of cyberinfrastructure systems against evolving threats.

The project also facilitates technology transitions in cybersecurity research, empowering cyberinfrastructure systems and enabling highly secure and trustworthy applications in scientific and engineering domains. It advances cyberinfrastructure security, paving the way for cutting-edge discoveries and innovations while ensuring the security of scientific data and computation. Additionally, the project supports further research, experimentation, and advancements in cybersecurity. The project provides comprehensive training opportunities for students, collaborative research prospects for graduate and undergraduate participants, faculty engagement, and involvement of K-12 students, with a focus on recruiting students from underrepresented groups to promote inclusivity and diversity.

The National Security Data & Policy Institute was established at UVA in August 2024 to advance research and engagement with the national security and intelligence communities.

The Institute's initial activities will center on data and intelligence deliverables already identified in partnership with Office of the Director of National Intelligence. Through this work, the Institute will develop as a durable, interdisciplinary center for security research and government engagement. Institute research teams will collaborate across Grounds, create partnerships with peer institutions worldwide, and engage outside experts across industry, government, and the non-profit sector.

In addition to the research, collaboration, and engagement agendas, the Institute will manage a post-doctoral and graduate research program, develop a virtual research environment for government/university collaboration, offer undergraduate research opportunities, and host conferences and symposia.

Read more about this award

In a collaboration with the U.S. Geological Survey, this project focuses on developing new artificial intelligence tools to identify individual brook trout, a project aimed at gathering data through so-called “fish-ial” technology that can be used to monitor trout health and habitat changes. The project received funding through a five-year grant from the U.S. Geological Survey.

The project team is working to build a database of trout images by focusing on fisheries in West Virginia and Massachusetts, employing various methods to capture the data they need, including crowdsourcing by allowing anglers to upload images of fish through an interactive application. The ultimate goal is that this technology can be used to protect brook trout and their natural environment.

While deep learning has shown remarkable capabilities in areas like image and natural language processing, many advancements are based on heuristic principles rather than provable theories. This project aims to develop a rigorous Bayesian mathematical theory for neural networks, focusing on the role of priors in transfer learning and regularization with limited data. As a result, this project will provide a theoretical understanding of neural networks and deep learning using Bayesian methods.

This study will have three key objectives. First, the project will work to develop a general mathematical understanding of neural networks and deep learning. Second, the project will leverage data from LiDAR, hyperspectral, and high-resolution imagery to derive information about natural and human activities from large-scale geospatial sensor data. And finally, the project will combine information from multiple sensor modalities in a generative learning model to advance the analysis from identifying objects to understanding their activities.

The project will enhance the use of artificial neural networks by moving from simple identification to more complex activity analysis, similar to the capabilities demonstrated by generative learning models like ChatGPT.

Urban Planning with Integrated Natural Systems (UPWINS), is an early-stage applied Research and Development project focusing on the development of monitoring and adaptive management techniques for nature-based infrastructure.

Using both lab and field-scale trials, the project is developing accessible methods of plant and soil manipulation to encourage plant migration and growth that can engage the accelerating, critical aspects of riverine and coastal environments such as erosion and saltwater intrusion. Further, these trials pair with extremely frequent spectral and geometric collections from the plants using multiple scales of terrestrial and remote sensing.

This data collection serves as the backbone of our data science approach to nature-based infrastructure monitoring, in which we seek to develop fundamental spectral signatures that can track the growth and change of key indicator species at large scales. Both sets of techniques are applied to sites within the Chesapeake Bay Region to verify the efficacy and impact of the research.

In the United States, approximately 14% of children with end-stage heart failure or inoperable congenital heart defects die while awaiting heart transplants, yet nearly 40% of donated hearts remain unutilized. When a donor heart becomes available, clinicians have minutes to sort through hundreds of donor and candidate variables to determine whether the organ is a good fit for their patient.

However, the nuanced impact of these variables on patient outcomes remains poorly understood, and no data-driven guidelines or evidence- based tools exist to support clinician decision-making. Consequentially, the prevailing uncertainty and associated risk aversion lead to the unnecessary rejection of many suitable hearts. We hypothesize that the integration of evidence-based tools can increase clinicians’ confidence in these critical decisions, thereby optimizing pediatric donor heart utilization and reducing waitlist mortality without negatively affecting post- transplant survival.

To answer this need, we are developing the first predictive models to assess, at the time of offer, the given candidate’s post-transplant survival if they received the offered heart; and if refused, projected time to next offer and the associated waitlist survival. To do this, we will use machine learning technology to analyze all 30,000 US-based pediatric heart offers made between 2010–2020, the most complete representation of national pediatric donor information ever compiled; we will validate the models using external data from 2020-2022.

We will collaborate with the United Network of Organ Sharing (UNOS), the national organ transplant system, to customize their established simulation platform. We will measure the success of our models and improved interface by sending simulated donor offers to a large team of pediatric transplant cardiologists from around the country.

Read more about this award

Fresh water is crucial for global economic, food, water, and energy networks, but freshwater ecosystems are degrading due to increasing demands and climate change. Monitoring water properties can inform policy and management decisions to address challenges like droughts, floods, and water security. Understanding water properties such as temperature and streamflow helps in studying biogeochemical and ecological processes. Investments in large-scale water data repositories offer opportunities to use machine learning for capturing complex water dynamics.

Graph neural networks are promising for modeling interactions in large river basins, but they face limitations without physical knowledge. This project aims to integrate graph network models with physical knowledge to better model complex, non-stationary water dynamics. It will provide research opportunities for students and contribute to curriculum development. The project will develop new physics-guided graph network models, explore ways to leverage physical knowledge, and pursue innovations in model architectures, learning strategies, and initialization methods.

More than 25-35% of children with velopharyngeal (VP) or palatal anomalies, such as those associated with repaired cleft palate and/or craniofacial conditions, will develop velopharyngeal dysfunction (VPD) as evidenced by hypernasal speech. This can negatively impact quality of life and communication abilities across the lifespan if not corrected during childhood.

VPD often requires surgical management to achieve an anatomy capable of producing normal speech and balanced resonance. However, an extremely high failure rate (up to 32%) is reported for these surgical procedures, and numerous revision surgeries are necessary.

Recent work utilizing MRI has emerged as a methodology to facilitate improved understanding of the VP anatomy. However, previous efforts to understand the VP mechanism have been significantly limited by small sample sizes and time-intensive analytic methods.

The main objectives of this R21 project are to deliver a cutting-edge, artificial intelligence/deep learning framework for automated 3D segmentation and 2D measurement of VP structures, to develop a large dataset to test foundational hypotheses on the relationships between velopharyngeal anatomy and speech resonance, and assess feasibility for integration of anatomic analyses in a clinical workflow.

With the rapid advancement of deep learning techniques, the utility of medical images (such as histology images) as digital biomarkers or diagnostic devices is actively being explored by pharmaceutical companies. Recently, several digital pathology products have already received regulatory approval such as MSIntuit for MSI-H prediction, and PDL1 App for PDL1-TPS scoring.

In order for digital image-based biomarkers to be more widely accepted by the scientific community, one key aspect is the interpretability of such biomarkers. For instance, in the realm of digital pathology, it is not sufficient to demonstrate only a high prediction performance of the machine learning model; pathologists often require the AI model to also provide visual evidence of which part of the histology image (or what are the most relevant interpretable histology features) that are used by the AI model to make the prediction.

In this project, we aim to enhance the interpretability of AI models from the perspective of causality, focusing on the analysis of hematoxylin and eosin (H&E) images in biomedical studies. Causality has been extensively studied in both statistics and computer science, which can benefit many real-world applications in biomedical science, healthcare, digital marketing, econometrics, etc.

This project seeks to improve the capability of statistical causal inference models for accurately estimating treatment effects and providing informative explanations for biomedical images, especially the H&E images. To achieve this goal, we investigate the paradigm of Explainable Causal Inference, developing new statistical and explainable machine learning models for heterogeneous treatment effect estimation from observational data.

This project will apply innovative machine learning methods to develop clinically-relevant computational models which analyze genetic and morphological signatures of Crohn’s disease to objectively and accurately predict specific outcomes of disease. This work is needed because predicting clinically-relevant outcomes of this debilitating autoimmune inflammatory disease at the time of diagnosis is currently impossible.

Successful completion of this work will establish a bench-to-bedside machine learning platform which uses clinical metadata along with deep features of Crohn’s disease from patient-derived biopsies, thus improving clinical management of one of the most chronic, costly, and consequential diseases in the US.

The REGARDS program at the University of Virginia’s School of Data Science will prepare students for careers in genomics research or additional advanced studies in the field, such as a doctoral program, as they pursue an M.S. in data science. The program will be run in collaboration with the Center for Public Health Genomics, part of UVA’s School of Medicine.

The new coursework will leverage the School’s long-running capstone projects where students work with an external organization to tackle a data science challenge. The REGARDS program will provide opportunities for students to gain valuable research experience in UVA genomics labs.

Additionally, the School of Data Science will develop new courses in genomics data science as well as modify existing offerings to include more genomics examples and lab exercises.

This project comes from the developers of the UVA Automated Insulin Delivery (AID) System, which in 2012 was the first wearable artificial pancreas running on a smart phone, and in 2017 was built in Tandem’s t:slim X2 insulin pump to become Control-IQ, an AID system now used by over 270,000 people with diabetes worldwide.

After this successful clinical translation, our academic mandate now is to open new horizons, design and test a new class of AID algorithms -- Adaptive Motif-Based Control (AMBC) -- which uses innovative adaptation-to-profile process to modulate the UVA-AID individual parameters for a person, by learning from this person’s CGM and insulin delivery patterns, and from the patterns of others stored in population databases.

Thus, unlike any other AID algorithm, the AMBC will use both a person’s own history and the history of others, to forecast and adapt to metabolic disturbances. The AMBC is expected to (i) be superior to current state-of-the-art AID systems in terms of glycemic outcomes and technology acceptance, and (ii) enable the transition from hybrid to full closed loop.

Extracting actionable knowledge from time series data is critical for driving decision-making in finance. But despite a long history of research, nearly all methods consider time series data in isolation, ignoring vast sources of relevant text data.

We instead aim to develop machine learning models that use text descriptions and documents to reason about financial time series, ultimately producing practical models that can infer unobserved events and their causes, intuitively describe complex data, and summarize the results of data analyses.

While large Language Models (LLMs) are a natural starting place, it remains contentious whether they bring advantages. We propose that for time series analysis, LLMs are best used to unlock new time series reasoning tasks. Towards this goal, we will evaluate and improve LLM-based time series models
for two tasks fundamental to inducing such multi-modal reasoning, using recent datasets.

This project uses Multi-Marginal Stochastic Flow Matching (MMSFM) to capture the dynamics of prices and volumes of high frequency financial transactions without dimensionality reduction. This framework enables both accurate modeling of historical patterns and generation of realistic market scenarios, with direct applications to risk management, trading strategy development, and market stability analysis.

Family caregivers provide essential support to nearly half a million veterans and service members with a history of traumatic brain injury, often working more than 60 hours per week. However, this care is not without cost: Many caregivers experience anxiety, depression, even suicidal thoughts, as well as a range of other health problems.

To help address this crisis, this study will adapt, test, and implement on a national level a six-session telehealth caregiver intervention called Resources for Enhancing All Caregivers’ Health, or REACH, which has already been developed. The study will mark the first time in the history of the U.S. Department of Veterans Affairs that a plan will be developed to implement this proven intervention on a national level at the more than 100 clinics in the Department’s Polytrauma System of Care

The input from caregivers who share their lived experience will not only guide the development of the intervention, but they and other subject matter experts will be key to decision-making. Given the goal of delivering an effective and meaningful caregiver intervention to the VA’s entire Polytrauma System of Care, the project’s approach to implementation planning will depend upon the engagement of clinicians who will ultimately champion the uptake of the REACH program.

Complex networks enable formal and extensible representations of real-world systems in a wide range of domains, such as brain networks, transportation networks, power grids, and infectious disease networks. Answering causal questions on complex networks will effectively inform decision-making.

For instance, in military operations, the commander would be eager to know the potential causal effects after an intervention is applied on the complex network, e.g., what will be the overall performance level (i.e., outcome) of this team if the unmanned aircraft is dispatched to other battlefields (i.e., intervention).

Estimating causal effects requires the proper modeling of entities, interventions, and outcomes, while the highly heterogeneous and dynamic nature of complex networks significantly increases the difficulty of causal inference. To date, the role of interventions in causal inference on complex networks has been largely underexplored.

The proposed research project will investigate theoretical foundations of causal inference on networks and develop new computational tools for monitoring and evaluating complex interventions on networks.

Rapid technological advances are helping to define our future. This project focuses on integrating artificial Intelligence and the Internet of Things to create smart, communicative, and powerful mobile and embedded devices. It aims to develop new systems, algorithms, and tools for deep integration at a large scale, ensuring the scalability of future machine learning systems across many distributed devices. This involves combining advanced machine learning algorithms with co-designed hardware, computer architectures, and distributed edge-cloud systems, while also addressing security and privacy concerns. The project considers the diverse performance and energy constraints of devices and the vast amount of data they produce. Its impacts include laying the foundation for future systems, advancing fields like machine learning, edge computing, Internet of Things, hardware, and software, and contributing to society through education, research dissemination, and outreach. The primary goal is to create a co-designed framework of hardware, software, and algorithms for extreme-scale machine learning systems. The project includes five research thrusts: developing scalable hardware and software, designing frameworks for weak embedded devices, addressing system and data unreliability, ensuring secure and privacy-preserving machine learning systems, and implementing testbeds for evaluation.

Graph Neural Networks (GNNs) have extended the success of Deep Neural Networks to relational data points, such as environmental sensor observations distributed across different spatial locations.

This project aims to develop new cyberinfrastructure for training large-scale GNNs to effectively learn from large-scale, real-time, geo-distributed, and heterogeneous data. The project will overcome computational and communication bottlenecks, benefiting various domain science applications that rely on massive spatiotemporal prediction.

More efficient agricultural practices are necessary to feed a growing and increasingly urban population. Controlled-environment agriculture (CEA) can increase production efficiency, reduce pests and diseases, and reduce resource usage.

However, wider adoption is prevented by the fact that CEA is currently an energy-intensive form of agriculture; current CEA facilities consume substantial energy, used to power supplemental lighting (such as LEDs) as well as heating and cooling systems.

Energy costs (and dependency on the power grid) can be reduced, and crop production optimized, by using networked cyber-physical systems (CPS) together with advanced data analytics and integrated, renewable (solar) energy and energy storage.

This project brings together a team of interdisciplinary researchers from agricultural sciences, control systems, and computer science to a create a networked sensing, control, and renewable energy system for CEA, with the goal of improving crop growth and yield while minimizing the energy cost. Proposed data analytics, embedded in every component of the proposed system, will enable self-adaptation and autonomy of CEA as well as advance CPS research.

MOSAIC intends to incorporate an intelligent tutoring system using natural language processing and ChatGPT to provide feedback to students to meet the needs of diverse learners with disabilities. The project seeks to facilitate improved academic vocabulary knowledge, content area learning, and middle and high school readiness for students with and at risk for disabilities and English learners.

Read more about this award

UVA's Resource and Data Exchange (RDE) will transform the organization and dissemination of research data by expanding the iTHRIV Research Concierge Portal and Research Data Commons to service the needs of a diverse range of research fields across the University. The RDE will provide a central resource for team-based science and collaborative research.

The mission of the RDE is to provide a customizable platform for all researchers, facilitating data sharing and individual and team-based collaboration across disciplines.

This will eliminate redundancy, improve data discovery, and allow researchers to tackle complex problems with a wider range of information. It will eventually encompass data from across the research community, ultimately positioning UVA as a leader in data resource sharing and research infrastructure.