COEIT Research Day Talks 2024

The College of Engineering and Information Technology (COEIT) Research Day Talks are scheduled 10:30 am to 12:30 pm in the ITE building. The full agenda for the event is available here.

See below for more information about the talks presented by our faculty and students. Topics and abstracts are presented in thematic areas.

AI/ML | Bioengineering | Education | Environment & Sustainability | Healthcare | Manufacturing | Mathematical Foundations


The connection between vision and language (V+L) is now an integral part of AI, with deep impact in not only in vision, but in adjacent fields such as robotics and human-computer interaction. I refer to this paradigm as semantic vision, where meaning (and by proxy, natural language) serves as a critical source of knowledge for modern computer vision algorithms that seek to understand the visual world. While previous decades in vision were dominated by “the three R’s” (reconstruction, recognition, and reorganization), the success of semantic vision has created a fourth “R” of computer vision: reasoning. The ability to reason requires not only prediction, but a combination of active perception, language grounding, world knowledge, and speculation beyond the observable.

Models that learn from data are widely and rapidly being deployed today for real-world use, but they suffer from unforeseen failures due to distribution shift, adversarial attacks, noise and corruption, and data scarcity. But many failures also occur because many modern AI tasks require reasoning beyond pattern matching — and such reasoning abilities are difficult to formulate as data-based input-output function fitting. The reliability problem has become increasingly important under the new paradigm of semantic “multimodal” learning.

My research provides avenues to develop robust and reliable computer vision systems, particularly by leveraging the interactions between vision and language.

In this talk, I will cover three thematic areas of my research, described below.

(1) Knowledge-Guided Adversarial Discovery of Transformations for Robust Image Classifiers.
Our efforts towards improving the robustness of image classifiers have focused on leveraging domain knowledge in order to discover diverse data augmentations that can expose the classifier to a larger distribution during training. This theme has led to improved robustness in several settings such as attribute-level shift, style-shift, and domain-shift including applicability to domain generalization for satellite images and medical images.
I will discuss important dimensions of machine learning reliability and our recent finding of curious trade-offs between them when using data modification.

(2) Open-Domain Reliability for Visual Reasoning.
Multi-modal tasks involving both vision and language (V&L) inputs combined with the quirks, complexity, and ambiguities of human language, open up intriguing domain discrepancies that can affect model performance at test time. Rigorous, exhaustive, holistic, and trustworthy evaluation and benchmarking is possibly the biggest challenge in multimodal systems today. I will demonstrate our findings of V&L models when dealing with logical compositions, semantic and syntactic perturbations of questions and sentences. This will be followed by a description of techniques to mitigate these failures through knowledge-guided regularization and data engineering, resulting in improvements along several dimensions of robustness, in image-based reasoning, video-based reasoning, and visual question answering tasks.

(3) Evaluation of Generative Vision Models: Challenges and Opportunities.
Text-to-image (T2I) models have seen exceptional advances in less than a decade.
But when it comes to evaluation of generated images, metrics of photorealism dominate the discourse. T2I models have much more to offer than photorealism; to do justice to this potential, we need extensive evaluation tools and benchmarks. My lab is actively pursuing the development of evaluation frameworks to quantify the reliability and abilities of T2I — this includes benchmarking spatial reasoning abilities via a challenging dataset (SR2D) and automated evaluation metrics (VISOR) and evaluation of the ability to learn, reproduce, and compose visual concepts, offered by ConceptBed.

My lab continues to work on addressing robustness issues in computer vision by developing methods that can adapt and generalize to domain shift, algorithms that can detect novelty and anomalies, and algorithms that can learn continually and update their knowledge using multimodal data. My lab is exploring applications of vision-language techniques in mission-critical domains where large-scale image data is unavailable, expensive, or unlabeled, but expert knowledge about images is available in language form. Integrating this natural language knowledge will help guide important decisions while also improving their reliability.

This study investigates the challenge of accurately simulating turbulent flows under mesoscale conditions resembling hurricanes. The work centers on utilizing large-eddy simulations (LES), a technique commonly employed in modeling atmospheric flow. That can be achieved either through an implicit approach or an explicit sub-grid scale (SGS) model. Current SGS models generally consider the transfer of energy from larger to smaller scales, leading to a model that dissipates energy alone. Nevertheless, this paper acknowledges the crucial significance of energy backscatter, which refers to the transfer of energy from smaller to larger scales, in enhancing the precision of dynamic models for complex systems such as hurricanes. This research explores the utilization of machine learning and deep learning as novel approaches to address the constraints in current modeling methodologies for SGS stresses. The methodology entails employing a direct feed-forward neural network to replicate energy transmission at the mesoscale. This encompasses both the conventional forward energy cascade and the relatively unexplored energy backscatter. The neural network is trained using data extracted from meticulous Large Eddy Simulations (LES) of a storm system that bears a resemblance to a hurricane. The objective is to enhance energy dynamics’ comprehension and simulation capabilities in mesoscale turbulent atmospheric conditions, which is crucial for forecasting and comprehending natural calamities such as hurricanes.

Atmospheric gravity waves occur in the Earth’s atmosphere caused by an interplay between gravity and buoyancy forces. These waves have profound impacts on various aspects of the atmosphere, including the patterns of precipitation, cloud formation, ozone distribution, aerosols, and pollutant dispersion. Therefore, understanding gravity waves is essential to comprehend and monitor changes in a wide range of atmospheric behaviors. Limited studies have been conducted to identify gravity waves from satellite data using machine learning techniques. Particularly, without applying noise removal techniques, it remains an underexplored area of research. This study presents a novel kernel design aimed at identifying gravity waves within satellite images. The proposed kernel is seamlessly integrated into a deep convolutional neural network, denoted as gWaveNet. Our proposed model exhibits impressive proficiency in detecting images containing gravity waves from noisy satellite data without any feature engineering. The empirical results show our model outperforms related approaches by achieving over 98% training accuracy and over 94% test accuracy.

Most research on machine learning has focused on learning from massive amounts of data resulting in large advancements in machine learning capabilities and applications. However, many domains lack access to the large, high-quality, supervised data that is required and therefore are unable to fully take advantage of these data-intense learning techniques. This necessitates new data-efficient learning techniques that can learn in complex domains without the need for large quantities of supervised data. On October 30, 2023, President Biden signed an Executive Order on the Safe, Secure, and Trustworthy Development and Use of AI. However, there are three major challenges associated with state-of-the-art (SOTA) AI algorithms: they lack generalizability (i.e., AI models are only as good as the data they are trained on), transparency and interpretability (i.e., AI models are “black box” models: opaque, non-intuitive, and difficult for people to understand), and robustness (i.e., imperceptible perturbations to AI inputs could altering its output). AI systems of the future will need to be strengthened so that they enable humans to understand and trust their behaviors, generalize to new situations, and deliver robust inferences. Neuro-symbolic AI, which integrates neural networks with symbolic representations, has emerged as a promising approach to address the challenges of generalizability, interpretability, and robustness. In this talk, I will present my journey from data-efficient machine learning to neuro-symbolic AI and our recent research on neuro-symbolic approaches in strengthening AI towards safe, secure and trustworthy AI systems, mainly neuro-symbolic reinforcement learning and the verification, validation, testing, and evaluations of neuro-symbolic AI systems.

Developing a computational understanding of vast amounts of data is crucial in the current era of information explosion. Distilling and correlating seemingly unrelated events into ”complex” events and extracting implied outcome knowledge from these events helps us identify, generalize and adapt to unforeseen events. However, complexities encountered in language, such as conditional and counterfactual statements, actual events (ones that happened) versus possible events, expressions of varying levels of certainty and so on, make this challenging. Multiple participants, their deliberate actions and accidental happenings, mixing of descriptive non-functional events and functional events in discourse, etc. add additional layers of complexity. To handle these, we borrow ideas from cross-disciplinary research in cognitive science and psychology on how humans mentally structure events along single participants and we extend the concept of thematic roles for verbal arguments from linguistics research. Using stories from well known NLP datasets, we examine the complex event depicted in a story through the lenses of both an agent-ative and patient-ative role assigned to an identified participant. We develop a multi-step crowdsourcing interface with careful quality controls on Amazon Mechanical Turk, where crowdworkers are guided to (1) infer the collective impact of salient actions that make up a complex event (2) annotate the volitional engagement of participants in these actions and (3) ground the outcome in state changes of the participants. We collect annotations for 8K short newswire narratives and ROCStories of high quality (an average inter-annotator agreement of 0.85 using weighted Fleiss-Kappa) and our dataset, POQue (Participant Outcome Questions), enables the exploration and development of models that address multiple aspects of semantic understanding. We show that current state-of-the-art language models, BART and T5, lag behind human performance in subtle ways through our task formulations that target abstract and specific comprehension of a complex event, its outcome, and a participant’s influence over the event culmination.

The advent of the Internet of Things (IoT) has ushered in an era of unprecedented connectivity, promising to revolutionize how we interact with the digital world. Yet, this technological boon also brings with it challenges, particularly in the realms of accessibility, privacy, and data management. Our research aims to democratize IoT usage by enabling individuals of diverse backgrounds and varying levels of technical expertise to engage with IoT technologies effectively and securely. Through the innovative integration of data management, data privacy, and artificial intelligence, our work at the DAMS Lab (DAta Management and Semantics) seeks to transparently bridge the gap between raw sensor data and user requirements expressed at a higher, semantically meaningful, level.

At the core of our endeavor is the commitment to privacy-preserving IoT data management. By understanding the privacy implications of IoT technology through, among others, analysis of privacy policies and network traffic communication, we strive to illuminate the path towards responsible and informed sensor selection. This process is crucial for fulfilling individual information requests without compromising their privacy or the privacy of others. The talk will highlight our recent papers in this topic including our recent work accepted at the Symposium on Privacy Enhancing Technologies 2024 that involves genAI technology in the process.

We will also delve into the research challenges that our group is currently working on, and plans to work in the near future, in the domains of data management and knowledge representation and reasoning to make this vision a reality. As we continue to explore these frontiers, our ultimate goal remains to foster an environment where IoT technology can be leveraged responsibly and efficiently, thereby truly democratizing its usage for the benefit of all.

Back to Top


In this talk, I will be presenting several current projects in the polymeric biomaterials lab. The first part of the talk focuses on collagen (COL) and heparin (HEP) layered coatings prepared via the layer-by-layer assembly as a robust coating to enhance cell behavior. We performed an exhaustive characterization of the COL/HEP coatings confirming their construction, chemistry, and their stability at body temperature. We evaluated the response of both human mesenchymal stromal cells (hMSCs) and Schwann cells (SCs) cultured on these coatings as well as in the presence of modulatory cytokines. Our results show that hMSCs cultured on COL/HEP coatings have a better response to soluble interferon-gamma regarding proliferation, protein expression, and immunomodulatory properties as compared to the uncoated culture plates. SCs cultured on COL/HEP coatings also demonstrated an enhancement in growth, migration, protein expression, and response to the nerve growth factor as compared to the uncoated culture plates. These COL/HEP coatings have strong potential to enhance the manufacturing of hMSCs or to serve as coatings for nerve implants. The final portion of this talk focuses on the development of a microneedle patch to deliver pain medication to cattle. We developed and characterized a chitosan microneedle patch that is capable of penetrating cattle skin and can sustain a controlled drug release. Overall, we demonstrate how we can leverage polymeric biomaterials for applications in biomanufacturing, tissue engineering, and drug delivery.

Per- and polyfluoroalkyl substances (PFAS) are a class of almost 15,000 synthetic chemicals that are incorporated into nonstick pans, firefighting foams, stain-resistant fabrics, and other industrial and household products. PFAS are used to convey water-, stain-, and temperature-resistant properties. For the same reasons, PFAS do not readily degrade in the environment and, therefore, pollute our water, soil, and air. Based on their persistence in the environment, PFAS are often called “forever chemicals”. The Environmental Protection Agency is poised to implement new drinking water limits on six particular PFAS that will be among the strictest regulations for any chemical contaminant in the United States. These regulations will also spur major investments in PFAS analysis and treatment throughout the country. Our lab is addressing diverse issues related to PFAS pollution in water and soil. The objective of this presentation is to highlight the environmental and human health concerns of PFAS and describe our ongoing projects to identify new opportunities to collaborate with other researchers in COEIT. Examples of some ongoing projects are described below. In cooperation with researchers at San Diego State University, we are creating adsorptive photocatalysts that provide selective adsorption of PFAS from contaminated waters, followed by regeneration processes that employ UV light to degrade PFAS to benign chemicals. Our group is leading an effort to develop new adsorbents specifically designed to remove short- and ultrashort-chain PFAS, which are difficult to treat with current technologies, from contaminated water. Finally, I will be working with NSF later this year to organize a symposium about PFAS use in the semiconductor industry. This presentation will focus on identifying touch points and topics for potential collaboration across the four departments of COEIT, as well as the associated Research Centers, to address the ongoing and future concerns associated with PFAS pollution.

This study investigates the complex interaction between the Cell Wall Integrity Signaling (CWIS) and Septation Initiation Network (SIN) pathways in Aspergillus nidulans. We used a microscopy-based assay to measure septation phenotypes, aiming to understand A. nidulans’ response to cell wall stress. Micafungin, a known inhibitor of β-1,3-glucan synthesis, was employed to induce this stress. We observed a strong correlation between micafungin treatment and normalized septation in wild-type strains. This finding suggests that the activation of the CWIS pathway under cell wall stress triggers a crosstalk signal that activates the SIN pathway, increasing septation. Further investigations using recombinant strains with targeted knockouts of CWIS and SIN signaling proteins enabled us to identify crosstalk between MpkA (CWIS) and SepH (SIN) kinases. Following this discovery, we broadened our study to screen the A. nidulans kinase deletion library for other regulatory genes connected to these pathways. The results of this study advance our understanding of fungal biology and provide insight regarding development of possible antifungal strategies.

Cultivating cells in shake flasks is a routine operation that is largely unchanged from its inception. A glass or plastic Erlenmeyer vessel with the primary gas exchange taking place across a variety of porous plugs is used with media volumes typically ranging from 100 ml to 2 liters. Oxygen limitation and carbon dioxide accumulation in the vessel is a major concern for studies involving shake flask cultures. In this study, we enhance mass transfer in a conventional shake flask by replacing the body wall with a permeable membrane. Naturally occurring concentration gradient across the permeable membrane walls facilitates the movement of oxygen and carbon dioxide between the flask and the external environment. The modified flask, called the breathable flask, has shown a 40% improvement in mass transfer coefficient (kLa) determined using the static diffusion method. The chemical oxidation method showed a 100% increment in kLa in the newly developed flask. The prokaryotic cell culture studies performed with E. coli showed an improvement of 32-73% in biomass and 41-56% in recombinant product yield. The eukaryotic cell culture study performed with Pichia pastoris expressing proinsulin exhibited a 40% improvement in biomass and 115% improvement in protein yield. The study demonstrates a novel approach to addressing the mass transfer limitations in conventional shake flask cultures. The proposed flask amplifies its value by providing a membrane-diffusion-based sensing platform for the integration of low-cost, non-invasive sensing capabilities for real-time monitoring of critical cell culture parameters like dissolved oxygen and dissolved carbon dioxide. Additionally, it shows that yields similar to conventional shake flasks can be achieved at significantly reduced power input.

Back to Top


This project addresses the evolving landscape of reflection to support learning and self-discovery by introducing a three-domain, theoretically grounded framework. While reflection has been explored across diverse domains, it often lacks a strong theoretical foundation and a universally accepted definition. In this work, we define reflection as “the intentional process by which individuals look back on experiences and integrate them into their embodied worldviews, allowing them to gain insights that lead to changes in knowledge, perspectives, attitudes, or identities and that guide future actions”. The primary outcome of our work is the development of a non-hierarchical framework for reflection-to-learn, outlining diverse characteristics across three domains: Psycho-Motor Skills, Conceptual, and Socio-Experiential. Unlike existing frameworks, it provides a holistic method for assessing reflection for various learning purposes and it challenges a view of reflective learning as progressing linearly. We consider these domains as not mutually exclusive; any one instance of reflection may cross domains and is likely to involve more than one element within a domain. This framework emerged through a grounded, iterative process of analyzing corpora of written reflections of various types, from a variety of learning experiences, across an array of disciplines. It is crafted to serve as a resource for researchers to describe, characterize, and assess reflection in a nuanced manner, while also acting as a practical tool for practitioners to facilitate and support reflective practices. It is designed to function as a resource for researchers, enabling them to describe, characterize, and assess reflection in a nuanced manner. Simultaneously, it serves as a practical tool for practitioners, aiding them in facilitating and supporting reflective practices.

Human-human interactions are fascinating and complex phenomena. However, it is notoriously difficult to study, given the dynamic and multimodal (e.g., audio, verbal, and visual) nature of the interactions. Modern sensors and the computational capability offered by AI/machine learning allow us to process and analyze fine-grained multimodal interaction data at a large scale. In this talk, I will share two studies that characterize and model parent-child interactions using audio/video data recordings, one in math education and another in early childhood parenting intervention contexts. I will discuss the various analytical methods used to derive insights and the implications for designing AI-supported coaching systems for realizing societal impacts at a large scale. I will also examine the challenges of working with multimodal interaction data.

Today’s pluralistic and globalized society requires a global perspective in how we think, view ourselves, and relate to others to collaborate successfully with people with different cultural backgrounds. To support the development of a global perspective, universities cultivate international co-operations to enable international mobility for students, faculty, and staff. However, in-person mobility can be inhibited due to various reasons. To support a wider pool of students gaining international experience, modern digital mediated course formats can be used to offer virtual instead of in-person mobility. This contribution evaluates a digital mediated course for engineering students (Global Engineering, GE) held in co-operation between an US and a Portuguese university, where internationally mixed student groups working on design projects in an Hyflex setting. Asides a literature driven discussion about the need for Global Engineers and a detailed description of the GE course, a longitudinal analysis was performed to assess the impact of the GE course on the students’ development of a global perspective. The longitudinal analysis showed strong statistical support that the digitally mediated GE course improved the participating students’ global perspective, assessed with the Global Perspective Inventory (GPI), tremendously. The observed improvements were in three of the six GPI dimensions comparable to an in-person semester abroad. Limitations of the study and future investigations and development will be discussed.

In this presentation, I will describe two projects in which we partner with governmental agencies and community organizations to create ecosystems that enable equitable innovation. Our NSF-funded Rec-to-Tech project transforms underused Recreation Centers in Baltimore City and Pittsburgh into sites for equitable technology-rich learning for hundreds of youth from underrepresented populations in STEM. In the state-sponsored Maryland DIY assistive technology project, we support a state-wide initiative to provide low-cost 3D printed assistive technologies to people with disabilities. Both projects highlight how public universities can form transdisciplinary partnerships with other institutions and organizations to support innovative community-led programs and initiatives.

The Imaging Research Center of the College of Arts, Humanities, and Social Sciences (CAHSS) supports interdisciplinary research collaborations between COEIT and CAHSS faculty. In this talk we will highlight four past and current projects that were the results of collaboration between the IRC and faculty from both colleges:

1. Virtual Ice Museum to Impact Perceptions of Climate Change: A virtual reality immersive experience will be developed to communicate the results of scientific research on climate change and polar ice melt., Vandana Janeja, Sudip Chakraborty, Ryan Zuber, Anita Komlodi, Lee Boot
2. A, B, See: Using Immersive Virtual Reality to Train the Alphabets of Hand Movements: This project aims to build an immersive virtual reality training environment in which the alphabets of complex hand movements can actually be visualized and learned. Ramana Vinjamuri, CSEE; Charissa Cheah, Psychology;
3. Infinite Transformations: A Multimedia Exploration of Embodied Persian Poetry: Infinite Transformations uses an interdisciplinary approach to bridging the fields of bioart, audio design, and augmented reality visualization through combining cutting edge scientific discoveries with medieval concepts in Persian poetry and architecture. Foad Hamidi, Information Systems; Linda Dusman, Music
4. Validation of a Virtual Reality Buffet environment to assess food selection processes among emerging adults: This study tested the validity and user perceptions of a highly immersive and realistic VR food buffet by: (1) comparing participants’ food selections made in the VR buffet and a real-world (RW) food buffet cafeteria one-week apart, and (2) assessing participants’ rated perceptions of their VR experience (0-100 scale). Jiaqi Gong, Information Systems; Charissa Cheah, Psychology

The IRC offers a Faculty Research Fellowship program for faculty from COEIT and CAHSS to collaborate on interdisciplinary research. Information about the FRF call will be shared.

Back to Top

Environment & Sustainability

This study employs the matrix profile and convolution operation of the Convolution Neural Network (CNN) to detect anomalous events in sea ice loss. These methods show that a higher amount of anomalous melting events was detected over the Weddell and Ross regions. This is important since these regions are experiencing a higher loss of sea ice. Our analyses using CNN and matrix profile show that the anomalous melting events are occurring early. Since 2000, the anomalous melting onset dates have advanced at a rate of 3 days per year-suggesting a longer melting period and enhanced melting. Overall, most of the sea ice regions over the Southern Ocean show early melt onset, but the effect is stronger near the coast region. The results are consistent among different datasets and methods used. This suggests an interaction between sea ice and ice sheet regions. Such interactions have not been investigated before. We are investigating the impacts of sea ice retreat over the Antarctic, which is new compared to the Arctic region, on ice sheets melting over the Antarctic since freshwater equivalent storage there is equivalent to 200 feet of sea level rise. Sea ice, in particular, acts as a buffer to the ice sheets over land as a protective blanket. Our approach will enable us to the future of sea ice and land ice interactions using data-driven machine learning approaches.

In the advancement of efficient computing, Compute-in-Memory (CiM) architectures have been developed to address the limitations of traditional von Neumann systems, which suffer from a memory wall bottleneck due to separate processing and storage. This paper presents two CiM innovations designed to boost computational efficiency and energy savings. The first, an energy-recycling 10T static random-access memory (SRAM) architecture (rCiM), enables in-memory logic operations and demonstrates a significant energy reduction by 55.42% and throughput of up to 106.6 GOPS/s, using a resonant write driver in an 8KB SRAM array with TSMC’s 28nm technology. The second architecture optimizes Multiply-accumulate (MAC) operations essential for neural network accelerators through a resonant time-domain (rTD-CiM) approach, eliminating the need for power-hungry analog-to-digital converters (ADCs) and achieves a 2.36 TOPS throughput and 28.05 TOPS/W energy efficiency. Both designs employ series resonance energy-recycling techniques to minimize dynamic power consumption and heat dissipation, marking a significant advancement in CiM technology.

Our group performs research in environmental engineering and science with a focus on the fate, effects, and remediation of toxic pollutants in the environment. We focus on big problems such as identifying pollutant sources and cleaning up of large, contaminated rivers, and managing large volumes of polluted sediments dredged for navigation at port facilities. We use multidisciplinary tools to investigate exposure and bioavailability of organic and metal pollutants to organisms and apply the new understanding to develop novel approaches for risk assessment and remediation. Our basic research investigates fundamental processes of how toxic chemicals bind to material surfaces and biodegrade in the environment to engineer new technologies to clean up contaminated sites. In this presentation I plan to trigger collaborative ideas by providing two examples of how our work is taking basic science understanding through quantitative descriptions in mathematical models that then enable scaleup of full-scale technologies:

1) Working with collaborators in environmental microbiology we accurately identified the key bottleneck in microbial degradation of a toxic organic compound that is the primary risk driver for many contaminated sites in the US, including in MD. In this work supported by the NIH, we quantitatively described the processes through modeling and worked around the bottleneck to engineer a remediation technology that is now translated to field studies.
2) In recent work supported by the DC Department of Energy and Environment, we used new methods developed by our team to identify the main sources of pollution in the Anacostia River. By coupling a mass balance model with an aquatic food web model for the river, we were able to connect the media concentrations to bioaccumulation in fish and evaluate a range of management options to reduce pollutant uptake in fish. The work is informing major remedial decisions for the cleanup of the Anacostia River.

Aerosols have deleterious effects on human health and play an important role in Earth’s climate system. A new EPA call (EPA-G2024-STAR-D1) seeks proposals that address “community-engaged research to advance the use and communication of air pollution data for addressing community-identified air pollution concerns.” A large part of this proposal is advancing methods for analysis and interpretation of low-cost air pollution data, which is already abundant and is growing substantially each year. Dr. Hennigan’s expertise lies in air pollution, including measurements with low-cost sensors. With this talk, Dr. Hennigan seeks to connect with data scientists who may be interested in discussing this opportunity and potentially collaborating on the proposal. The EPA expects to fund eight awards under this call, with a maximum of $1.25M per award.

Back to Top


Patients and caregivers seek informational and emotional support at all stages of managing health conditions. When not receiving adequate information from their healthcare providers, patients and caregivers may turn to online health communities to consult other users. When receiving lab test results, the informational and emotional needs are particularly dire in order to make informed decisions about treatments. The emergence of Large Language Models (LLMs) has opened up exciting opportunities to assist patients and caregivers get health information at time during their healthcare journeys. We report a study where we empirically evaluated the ability of LLMs to provide readable responses with emotional support and personalization to patients’ online medical questions about their lab tests. We also compared LLMs-generated responses to responses written by other users on the online forum. We provide implications for improving LLM responses and opportunities for integrating them into online health communities for better supporting patients and caregivers’ informational and emotional needs.

One in four deaths worldwide is related to dysfunctional blood clotting. Platelet force generation is an emerging metric for the balance of clotting and bleeding due to recent demonstrations of their powerful abilities to predict bleeding risk in trauma patients and detect bleeding dysfunction more sensitively than all existing clinical tests. While existing methods have indicated the clinical and scientific potential of platelet forces, they have been hampered by low-yield, inability to co-measure immunofluorescent cell markers, and/or arbitrary restriction of cell spreading. To address these limitations, we developed a technique (dubbed “black dots”) that enables high-yield co-measurement of cellular forces and immunofluorescent-labeled cell markers in a single image without constraining cell spreading. Applying black dots to measure single-platelet forces, we identify biophysical factors that associate with force generation, determine the effects of platelet storage conditions on function, identify unique cytoskeletal morphologies induced by different blood proteins, and determine the effects of cytoskeletal crosslinkers. As a result of the high yield of data obtainable with black dots, approaches including multivariate mixed effects modeling, K-means clustering, and machine learning were able to be applied to elucidate complex relationships between platelet activation, structure, and force generation, which have implications in bleeding, clotting, and transfusion medicine.

In support of collaborations with other COEIT faculty, (1) this black dots methodology can be applied to measure single-cell forces of other cell types, such as fibroblasts, and (2) this black dots methodology can be further improved in combination with other methodologies such as machine learning approaches (e.g., convolutional neural network), cytoskeletal modeling, and approaches to calculate forces from microscale material deformation.

The distinctive goal of the proposed IUCRC BRAIN site at UMBC is to introduce the following new research areas to the existing national and international sites at UH, ASU, WVU, Georgia Tech, TEC and UMH—Cyber Human Systems, Artificial Intelligence, Neural Signal Processing, Neural Imaging and Stimulation, and Virtual/Augmented/Mixed Reality. Through these research areas our vision is to develop and validate low-cost technologies that can impact individuals with motor and cognitive disabilities in most immediate and personal manner. By leveraging the range of expertise in the above areas available at UMBC and combining enthusiastic support from regional and national AI, Neuro, Biotech and Pharma companies and easy access to top regional and international medical centers, we aim to bring significant contributions to the existing BRAIN sites. The proposed site has specific technological (new research areas), collaborative (projects with existing BRAIN sites), and resourceful (northeast biomedical industry and top national and international medical centers and close proximity to federal agencies and institutes such as NSF, NIH, FDA and NASA) contributions.

Back to Top


A Bitter electromagnet under the designation Adjustable Long Pulsed High-Field Apparatus (ALPHA) has been designed and is under construction at the Dusty Plasma Laboratory (DPL), at the University of Maryland Baltimore County (UMBC). The Bitter Electromagnet Testing Apparatus (BETA), a proof-of-concept model for ALPHA, has been constructed and tested to validate ALPHA’s design methodology and models. ALPHA’s Bitter plates are stacked to generate a mirror ratio slightly above 1 and will have the capacity to produce fields in excess of 7 T, with a bore diameter of 15 cm. Initial planned experiments with ALPHA are short pulse, high field investigations of Paschen-like plasma breakdown in a cylindrical, magnetized configuration, as well as critical ionization velocity limits (CIV) relevant to E x B centrifugal mirror confinement. Details of ALPHA’s construction, measurements of its electrical properties, and planned experiments are presented.

Acknowledgements: Work supported by the Air Force Office of Scientific Research, Grant No. FA9550-19-1-0071.

The design, manufacturing, and assembly of an ultrahigh vacuum (UHV) compatible magnetic compressor nozzle and support structures were developed for the ECLAIR experiment. The ECLAIR experiment aims to validate the novel fusion propulsion concept, called Helicity Drive, that was proposed in 2020. The Helicity Drive is a hybrid approach between magnetic confinement fusion (MCF) and inertial confinement fusion (ICF) with the goal to create an energy-efficient, compact, and rapid way for practical fusion power and space propulsion. This project is a collaboration between Helicity Space, Caltech, and UMBC research teams. The UMBC team worked on validating the magnetic compression scheme proposed by P.M. Bellan in 1979 that was further adapted for the ECLAIR experiment parameters.

The magnetic compressor nozzle designed at UMBC consists of 20 Bitter type magnets that are modeled as a transmission line. The first prototype compressor was used to validate the theoretical predictions by measuring the magnetic fields inside the magnetic compressor nozzle. The second prototype compressor, engineered for integration into the ECLAIR UHV chamber at the Helicity Space laboratory, incorporates system considerations for in-vacuum testing with plasma. Details of the compressor and the upcoming plasma tests are presented.

Dr. von Lockette leads the Electro/Magneto-Active Composites and Structures (e-MACS) at UMBC. e-MACS are fabricated by mixing electromagnetically active fillers into polymeric (plastic or rubber) matrices. The driving and manipulable electromagnetic interactions span a range of length scales. Consequently, his recent work covers how composition and coupled electric- and magnetic-field processing can lead to controllable self-assembly, and subsequent control of magnetic and other properties, in additive manufacturing processes; how electromagnetic fields can be used as novel controlling mechanisms for control of micro- and macroscale scale robots; and how electromagnetically interacting regions in flexible structures can be used to actuate “soft robots” for biomedical and other applications. In addition, he has begun a project to translate these funding along with biologically inspired design rules into the built-environment for low- to moderate-load bearing structures fabricated using green polymer matrix composites. Finally, he is interested in starting discussions on an NSF Center for Research Excellent in Science and Technology (CREST) at UMBC involving the intersection of green materials, rules-driven optimal design, fluid/solid mechanics, nonlinear materials, controls, classical electromechanics, and machine learning applied to the built environment and soft-robotics.

A laser vibrometer can measure the surface velocity of a point on a structure. A continuously scanning laser vibrometer (CSLV) was developed to significantly improve efficiency and spatial resolution of vibration measurement of the structure. As a non-contact system, it can avoid the mass-loading problem in vibration measurement using accelerometers. The CSLV was made by adding two orthogonal scan mirrors in front of a single-point laser vibrometer. During CSLV measurement, two scan mirrors can be controlled to continuously rotate about their rotation axes, and the laser spot of the CSLV can continuously move along a pre-designed scan trajectory on the structure, which is a major difference compared to a conventional scanning laser vibrometer (SLV) that has a point-by-point scanning capability. In this talk, we will discuss our recent developments on a novel general-purpose three-dimensional (3D) CSLV for measuring 3D full-field vibration of a structure with arbitrarily curved surfaces. It can also measure vibrations of difficult-to-access areas of structures with the assistance of reflective mirrors to obtain their 3D panoramic modal parameters through a novel vibration stitching method. Our current work on developing a miniaturized robot-assisted SLV will also be discussed. The methodology can be used to measure 3D vibrations of wearable devices.

Back to Top

Mathematical Foundations

It is well known that zeros and poles of a single-input, single-output system in the transfer function form are the roots of the transfer function’s numerator and the denominator polynomial, respectively. However, in the state-space form, where the poles are a subset of the eigenvalue of the dynamics matrix and thus can be computed by solving an eigenvalue problem, the computation of zeros is a non-trivial problem. This paper presents a realization of a linear system that allows the computation of invariant zeros by solving a simple eigenvalue problem. The result is valid for square multi-input, multi-output (MIMO) systems, is unaffected by lack of observability or controllability, and is easily extended to wide MIMO systems. Finally, the paper illuminates the connection between the zero-subspace form and the normal form to conclude that zeros are the poles of the system’s zero dynamics.

The Lugiato-Lefever Equation (LLE) has played a key role in the theoretical and computational study of optical solitons in microresonators. Recent investigations have demonstrated that by employing dual laser pumping in a microresonator, broadband microcombs can be produced with over an octave of bandwidth. The dual laser pumping also unlocks new nonlinear interactions in the phase velocity domain resulting in the presence of a second repetition rate in the carrier envelope offset. These microcombs are generated by microresonator solitons that have distinct components, which propagate at different phase velocities. We refer to these distinct components as “colors” because their central frequencies differ by large fractions of the total optical bandwidth. Although theoretical studies have utilized a modified LLE (MLLE) that incorporates two pumping terms, the resulting multi-color solitons are non-stationary, hindering the study of their properties and, in particular, their stability. In this study, we introduce a set of three coupled wave equations, termed 3-wave equations, to model the three primary colors of the solitons. Our analysis reveals excellent agreement between the solutions derived from the 3-wave equations and those obtained from the MLLE. Moreover, the solutions derived from the 3-wave equations are stationary, thereby facilitating future analyses of the stability and other properties of the multi-color solitons.

Did you find what you were looking for today?