AIIM Award Winners are denoted in bold.
Biological, Biomedical & Health Sciences
Day 1 – March 30
James Tavernor, Graduate Student Research Assistant, Computer Science and Engineering, College of Engineering
Artificial Intelligence (AI) is an essential tool for improving the accessibility and precision of mental health care, helping to bridge the gap created by a shortage of clinicians and prohibitive costs to see a provider. Many mental health conditions are characterized by shifts in emotional patterns as severity increases (e.g., Bipolar Disorder, Borderline Personality Disorder, Mood Disorders, etc.). Tracking these emotional changes is essential for monitoring these conditions. However, traditional monitoring often relies on periodic self-reporting, which can be infrequent and subject to recall bias. By automating the tracking of emotion, AI provides a quantitative, scalable way to monitor condition severity and identify when clinical intervention is required. Yet, there are risks. In a clinical setting a ‘one-size-fits-all’ model may be misaligned with the speaker’s emotional experience, which risks mischaracterizing their clinical status and highlights the importance of model personalization to the speaker’s perception.
Speech Emotion Recognition (SER) systems face two primary challenges that prevents their ubiquitous deployment. First, models often struggle to generalize across datasets; while research typically focuses on acoustic variations (such as recording devices and acoustic environments), it frequently overlooks annotator shift (the differing opinions between the separate groups of annotators across datasets). Second, standard model training relies on the perception of outside, disconnected annotators, rather than the speaker’s internal state, potentially leading to biased output. Adapting SER models to predict a speaker’s internal feelings rather than external observations represents a fundamental change in the model’s goal that current approaches cannot handle.
Our approach expands SER modeling by shifting from modeling the consensus of a group of outside annotators to individually modeling every annotator. Modeling individual annotators addresses both primary challenges. First we can address the annotator shift. This contrasts with the traditional approach of flattening individual opinions into a single label – our approach retains these perspectives. We can use this information to identify similar perspectives on new datasets or even identify individual-specific patterns (e.g., certain people may perceive loud emotions differently than others). Second, we can further leverage our approach as a novel step towards modeling internal emotional states. We know that labels from an outside group of annotators diverges from the emotion that is self-reported by the speaker. By retaining individual perceptions we can better approximate the perception of the self-report speaker. This offers a significant advantage over training a model on self-reported data, which is often hindered by the scarcity of self-labeled training samples.
We developed a neural network optimized to learn over 12,000 individual annotators (as opposed
to a single label). While modeling at this scale is typically computationally prohibitive, we created a novel prediction method that executes all predictions in a single matrix multiplication. This bypasses the traditional requirement of looping through thousands of linear layers. We further introduce an annotator similarity framework that allows us to extend personalized models to new people. We analyze the correlations between our model’s predictions and the labels in a new dataset to identify similar perception patterns between known and new annotators, which can also be used to adapt to self-report prediction. This allows the model to dynamically address annotator shift without requiring exhaustive re-labeling, providing a scalable solution for both cross-corpus deployment and self-report adaptation. We have shown that this results in significantly improved cross-corpus performance and self-reported emotion prediction performance.
Douglas Craig, Statistician Staff Specialist, Medical School
Emergency Medicine (EM) clinicians make high-stakes decisions that shape the initial trajectory of acute care. However, a structural “feedback sanction” exists in the current healthcare environment: once a patient is admitted or discharged, the treating clinician rarely receives feedback on the downstream outcome unless a negative event triggers a review. This “open loop” limits clinician calibration, hinders practice-based learning, and prevents the identification of latent safety threats. While clinicians desire outcome data to improve decision-making, manual chart review is prohibitively time-consuming and labor-intensive.
To address this, our team developed “Tell Me What Happens Next,” a provider-directed, AI-enabled feedback pipeline. Integrated directly into the Electronic Health Record (EHR) workflow, this tool allows clinicians to flag patients during their shift for educational follow-up. The system then utilizes generative AI to synthesize complex downstream clinical data into concise, educational summaries delivered via email at provider-selected intervals (3, 7, or 14 days). This AI pipeline is a new capability embedded within our ongoing ED clinical education/learning-health-system efforts, enabling outcome feedback at a scale not previously feasible.
Devin Brown, Professor of Neurology, Medical School
Multicenter clinical trials are a pivotal tool in medical research, but demand extensive support from trial leadership for clinical trial sites. Generative artificial intelligence may be used to support various aspects of clinical trials. Our aim was, the first time, to test the ability of a customized GPT to support clinical trial sites by providing rapid responses to protocol and procedure inquiries within a large, complex, multicenter randomized controlled trial. Use of AI in this context could revolutionize clinical trial site support.
Jiaying Zhao, Graduate Student Research Assistant, Biomedical Engineering, Medical School
The current clinical practice of diagnostic magnetic resonance imaging (MRI) exams requires the acquisition of T1-weighted images before and after the intravenous administration of extracellular or tissue-specific contrast agents, such as the liver specific gadoxetic-acid contrast agent. However, these contrast agents are typically expensive and may have side effects on patients with compromised renal function, which limit accessibility and feasibility. Moreover, this approach requires intravenous contrast administration and a 20-minute post-injection delay, increasing scan time, cost, and patient burden.
The goal of the project is to develop AI-driven, contrast-free MRI techniques that preserve the diagnostic value of hepatobiliary phase (HBP) imaging while reducing reliance on intravenous contrast. Specifically, our work focuses on using deep learning models to synthesize HBP-like liver MRI directly from non-contrast acquisitions, enabling time-efficient and safer imaging with comparable quality to images generated by patients who received contrast agents. This project lies at the intersection of biomedical engineering, radiology, and artificial intelligence, addressing a clinically significant problem through AI-enabled image reconstruction and synthesis.
Haowei Tai, Research Fellow, Radiology, Medical School
Urethral closure pressure is a central functional determinant of continence, and reductions in closure pressure are a primary cause of stress urinary incontinence. Age-related atrophy of urethral musculature is thought to play a major role, consistent with an approximate 15% decline in urethral closure pressure per decade. Pelvic floor muscle injury, particularly involving the levator ani muscle, further compromises continence by impairing pelvic support and plays a critical role in postpartum recovery and long-term pelvic floor function. Although both urethral tissue integrity and pelvic floor muscle structure are essential for maintaining continence, existing clinical assessment tools remain limited. Catheter-based pressure measurements can alternative tissue mechanics and provide only global functional readouts without revealing the underlying structural changes. Physical examination and voluntary contraction tests are inherently subjective and insensitive to microstructural remodeling. Conventional ultrasound, while noninvasive and widely available, typically lacks the sensitivity needed to link image appearance to tissue composition, degeneration, or healing when interpreted qualitatively.
Urethral closure pressure is a central functional determinant of continence, and reductions in closure pressure are a primary cause of stress urinary incontinence. Age-related atrophy of urethral musculature is thought to play a major role, consistent with an approximate 15% decline in urethral closure pressure per decade. Pelvic floor muscle injury, particularly involving the levator ani muscle, further compromises continence by impairing pelvic support and plays a critical role in postpartum recovery and long-term pelvic floor function. Although both urethral tissue integrity and pelvic floor muscle structure are essential for maintaining continence, existing clinical assessment tools remain limited. Catheter-based pressure measurements can alternative tissue mechanics and provide only global functional readouts without revealing the underlying structural changes. Physical examination and voluntary contraction tests are inherently subjective and insensitive to microstructural remodeling. Conventional ultrasound, while noninvasive and widely available, typically lacks the sensitivity needed to link image appearance to tissue composition, degeneration, or healing when interpreted qualitatively.
To address these limitations, we developed a noninvasive, machine learning–enabled multiparametric ultrasound (mpUS) framework designed to quantify continence-related tissue microstructural changes and relate them to functional decline or recovery. This approach leverages quantitative ultrasound features that extend beyond visual interpretation and integrates them through machine learning to produce objective and reproducible imaging biomarkers. By jointly examining urethral tissue microenvironmental changes associated with aging and LA muscle remodeling following childbirth, this unified strategy targets the primary anatomical contributors to continence failure and restoration.
Day 2 – March 31
Venkatesh Murthy, Melvyn Rubenfire Professor of Preventive Cardiology, Medical School
Coronary microvascular dysfunction (CMVD) is a critical yet underdiagnosed condition that affects the heart’s smallest blood vessels, leading to ischemia and increased mortality. Currently, diagnosing CMVD requires expensive, specialized imaging such as Positron Emission Tomography (PET), which is rarely available outside of major academic centers. Our project democratizes this diagnostic capability by developing a foundation AI model capable of detecting CMVD using standard, widely available electrocardiograms (ECGs). By bridging the gap between complex physiological data and simple clinical tools, this work aims to transform cardiovascular care for underserved patient populations.
Mohamed Abdelalim, Visiting Research Investigator, Medicinal Chemistry, College of Pharmacy
My research addresses the challenge of discovering potent and selective Src kinase inhibitors, a key protein implicated in cancer and other proliferative diseases. Traditional drug discovery approaches are hindered by expensive, labor-intensive screening and optimization pipelines. As a solution, I developed an integrated app that leverages advanced artificial intelligence (AI) methods including generative models, predictive analytics, and explainable AI to automate and accelerate the identification of novel drug candidates. This project sits at the intersection of computational chemistry, machine learning, and structural biology, embodying a paradigm shift in drug discovery by embedding AI as the engine for molecular innovation.
Renee Liu, Medical Scientist Training Program Fellow, Medical School
Clinical trial enrollment remains one of the most persistent bottlenecks in translational medicine. More than 80% of clinical trials fail to meet their enrollment targets or are delayed due to slow recruitment. [1] Despite widespread adoption of electronic health records (EHRs), identifying eligible patients for clinical research continues to rely heavily on manual chart review and labor-intensive screening workflows. Recruitment processes alone require thousands of hours of human effort, multiple research assistants, and one third of clinical trial costs, yet still fail to capture a large proportion of eligible participants. [2] The core challenge is not a lack of patients, but that clinically relevant information is buried in messy, unstructured clinical notes, imaging notes, and longitudinal care histories that are not easily searchable or interpretable at scale.
To address this gap, we developed ATLAS (AI for Trial Linking and Automated Screening), an AI-powered system that automatically identifies eligible patients for clinical trials and matches them to appropriate registries and studies using EHR data. ATLAS transforms what was previously a manual, slow, and expensive process into an automated pipeline that operates in minutes at minimal cost.
ATLAS is initially being developed and piloted in ophthalmology through the Kellogg Eye Center, where there is a strong alignment between clinical need, active trial pipelines, and existing interdisciplinary collaborations. Once piloted, validated, and established at Kellogg Eye Center, ATLAS is designed to be directly generalizable to all specialties and hospital systems across the University of Michigan and beyond.
Michael Sweeney, Graduate Student Research Assistant, Biostatistics, School of Public Health
This project focuses on investigating how deep learning- and artificial intelligence (AI)-based genomic models that predict the functional impact of genetic variants can improve the reproducibility of genetic fine-mapping results, a core challenge in statistical genetics. Genetic fine-mapping is the statistical inference procedure of pinpointing which specific genetic variants within a broader associated region are causally influencing molecular traits, such as gene expression levels or disease-related traits, e.g., LDL cholesterol levels. Current standard practices employ purely statistical approaches, such as Bayesian fine-mapping methods, and their results often fail to replicate putative causal genetic variants in independent studies. This lack of reproducibility undermines many downstream applications including causal inference, gene prioritization, target discovery, and translational research.
Recent advances in AI have led to the development of sequence-to-omics (S2O) models that predict molecular phenotypes directly from DNA sequence by leveraging large-scale epigenomic and transcriptomic datasets. Published models such as AlphaGenome, Borzoi, Enformer, and Sei integrate information from thousands of RNA-seq and epigenomic profiles to estimate the functional impact of genetic variants, providing complementary, biologically-informed evidence that can help prioritize likely causal variants beyond what is captured by statistical association alone.
Previous studies utilizing S2O models have shown less accurate prediction of RNA-seq expression levels when directly compared with conventional, purely statistical approaches; however, we hypothesized that, in the context of genetic fine-mapping, S2O models can provide complementary biological information to distinguish highly correlated variants that cannot be disentangled with purely statistical approaches.
This project has three aims that we address. First, we systematically evaluate whether modern AI models, specifically S2O models, can prioritize putative causal variants (using replication in an independent dataset as a proxy for causality). Our second aim integrates AI-derived functional predictions with traditional statistical fine-mapping in one statistical framework to generate functionally-informed posterior inclusion probabilities (fiPIPs) that improve genetic fine-mapping. Our third aim is to apply our framework to large-scale multi-omics datasets from consortia like Genotype-Tissue Expression (GTEx), Trans-Omics for Precision Medicine (TOPMed), and Multi-Ancestry Analysis of Gene Expression (MAGE) to improve causal variant inference from these datasets.
John Oyer, Assistant Researcher, Radiology, Medical School
Mucus plugging is an underrecognized but clinically important imaging phenotype in obstructive lung diseases such as asthma and chronic obstructive pulmonary disease (COPD). Although mucus plugs are visible on chest computed tomography (CT), identifying them manually is prohibitively time-consuming, limiting prior studies to small samples, advanced disease, and coarse summary scores. As a result, the role of mucus plugging in early disease, longitudinal progression, and downstream airway dysfunction has remained largely unexplored.
This project leverages artificial intelligence (AI) to transform mucus plug identification from a labor-intensive, subjective task into a scalable, reproducible quantitative imaging biomarker. We developed an AI-based mucus plug detection pipeline that enables high-throughput analysis of large clinical CT cohorts, allowing investigation of mucus plugging across disease stages, populations, and outcomes at a scale that was previously impossible. This work is tied to an active research program with completed analyses, manuscripts, and presentations.
Natalie Jusko, Graduate Student Research Assistant, Pahrmaceutical Sciences, College of Pharmacy
Clinical adverse events (AEs) experienced by patients at therapeutic dose levels remain a critical issue in oncology drug development, contributing to over 30% of drug failures in late-stage trials. Despite extensive preclinical safety testing, predicting which patients will experience toxicities and when remains profoundly challenging. This is particularly true for small molecule kinase inhibitors (SMKIs), a class of 85 FDA-approved targeted therapies that inhibit not only their intended targets but also dozens of off-target kinases across multiple tissues.
Our research addresses this gap by developing a machine learning framework that predicts both the occurrence and time-to-onset of clinical adverse events in cancer patients receiving SMKI therapy. Unlike conventional approaches that rely primarily on plasma drug concentrations and/or drug-like properties alone, our model integrates three physiologically relevant dimensions: on-/off-target kinase binding profiles (442 kinases), patient-specific drug exposure, and drug-specific tissue distribution across 36 organs. This framework was trained and validated on individual-level data from 3,433 patients enrolled in the registrational trials of 16 FDA-approved SMKIs.
Muzammal Shafique, Graduate Student Research Assistant, Computer Science, Engineering, and Physics, U-M Flint
Intracranial aneurysm (IA) rupture is a catastrophic cerebrovascular event associated with high mortality and long-term neurological disability. Despite advances in neuroimaging and computational modeling, accurately predicting aneurysm rupture risk remains an unsolved clinical problem. Current clinical decision-making largely relies on geometric size thresholds, aneurysm location, or qualitative assessments, which fail to capture patient-specific vascular geometry, hemodynamic instability, and individualized clinical context. As a result, high-risk aneurysms may remain untreated while low-risk aneurysms are unnecessarily exposed to invasive intervention.
The proposed research aims to address this critical gap by developing a multimodal Physics-Informed Neuro-Symbolic AI framework for patient-specific intracranial aneurysm rupture prediction. Situated at the intersection of medical imaging, computational hemodynamics, and artificial intelligence, the framework introduces a unified learning paradigm that integrates 3D vascular imaging (MRA), physics-based hemodynamic modeling, and Electronic Health Record (EHR) data. By embedding physical laws and symbolic clinical knowledge directly into the learning process, the model enables accurate, interpretable, and clinically trustworthy rupture risk assessment.
In current clinical practice, however, rupture risk evaluation remains a highly demanding and time-sensitive task, requiring expert interpretation of complex vascular morphology. Such assessments are often performed under significant cognitive constraints, contributing to variability in decision-making and uncertainty in treatment planning. In this context, an automated multimodal physics-informed graph neural network can play a critical supportive role by systematically integrating patient-specific anatomy, hemodynamics, and clinical context to identify aneurysms at elevated risk of rupture.
Within the broader discipline of medical imaging and computational medicine, this work advances the growing movement toward trustworthy, physics-aware, and clinically aligned AI systems, particularly for safety-critical cerebrovascular applications.
Steven Soliman, Clinical Associate Professor of Radiology, Program Director, Musculoskeletal Radiology Fellowship Program, Program Associate, Clinical Research Lead for Musculoskeletal Radiology, Medical School
Our long-term goal is to establish a cost-effective, AI-driven point‐of‐care medical device for rapid, accurate, and noninvasive detection of metabolic disease and personalized risk stratification. To enable broad access, the device will use AI-driven automation to eliminate the need for specialized ultrasound training. With the rise of portable, affordable, and user-friendly handheld ultrasound devices, often called the “stethoscope of the future,” our approach can be integrated into routine vital sign checks, pharmacies, school screenings, and community events. The current objective is to develop and validate a multimodal AI-driven platform that integrates skeletal muscle ultrasound with readily available, basic clinical attributes (age, sex, race/ethnicity, BMI) for generalizable, interpretable, and personalized prediction of metabolic disease. Our central hypothesis is that early metabolic dysfunction causes measurable, subclinical muscle remodeling that is detectable by ultrasound and, when combined with AI-driven image analysis and multimodal integration with patient-known clinical attributes, yields an accurate, scalable, and personalized biomarker of metabolic disease risk.
Yousif Alyousifi, Research Fellow, Internal Medicine, Medical School
Metabolic dysfunction–associated steatotic liver disease (MASLD) and metabolic dysfunction–associated steatohepatitis (MASH) affect hundreds of millions worldwide yet remain widely underdiagnosed because gold-standard diagnostics (MRI-PDFF, cT1, CT imaging, liver biopsy) are costly, invasive, and not scalable. My research develops and validates scalable, non-invasive AI methods that enable accurate liver disease phenotyping and long-term progression risk prediction (cirrhosis and hepatocellular carcinoma, HCC) using large population cohorts and real-world clinical data. The overarching aim is to support early identification of high-risk individuals, improve surveillance strategies, and enhance clinical trial enrichment.
Mohamed Mohyeldin, Research Fellow, Life Sciences Institute
Chirality is a decisive property for drug safety, efficacy, and regulatory approval, yet absolute configuration (AC) assignment remains a rate-limiting, expert-dependent bottleneck in modern discovery pipelines. Conventional Electronic Circular Dichroism (ECD) workflows often demand tens to hundreds of CPU-hours per molecule and still risk misassignment, making high-throughput stereochemical analysis impractical. DeepChiral addresses this gap through an AI-powered platform that predicts and interprets ECD spectra for chiral small molecules in seconds, transforming AC analysis from an expert-driven task into a scalable, data-centric service tightly integrated with experimental feedback.
Yang Li, Research Fellow, Ecology and Evolutionary Biology, College of Literature, Science, and the Arts
Single-cell RNA sequencing (scRNA-seq) has become a transformative technology for dissecting cellular heterogeneity in complex diseases and for identifying disease-relevant therapeutic targets. However, translating scRNA-seq data into robust candidate gene targets remains highly challenging. Target discovery workflows involve multiple decision-dependent stages, including quality control, normalization, clustering, cell type annotation, and differential expression analysis, where small analytical variations can yield substantially different gene prioritization outcomes. This instability contributes to poor reproducibility across studies, inefficiencies in downstream experimental validation, and uncertainty in mechanistic interpretation.
To address this fundamental bottleneck, we developed SCTA (Single-Cell Target Agent), a modular and decision-centric multi-agent AI framework designed specifically for stable and interpretable target gene discovery from scRNA-seq data. SCTA reframes target prioritization as a robustness problem, aiming to identify biologically meaningful candidate genes that remain consistent under realistic analytical perturbations. This work contributes to the broader goal of enabling reliable AI-driven precision medicine and reproducible biomedical discovery.
Anna Kay, Graduate Student Research Assistant, Biomedical Engineering, Medical School
Age-related macular degeneration (AMD) is the leading cause of irreversible blindness in the developed world, affecting over one-third of individuals above age eighty. The few approved therapies are limited to slowing late-stage disease, where vision loss is already significant.
In AMD, multiple lesions develop in the retina and evolve over time. While previous works identified lesion correlations, coarse patterns, and some genetic markers, it remains unclear which patients will experience faster disease progression or what predictive patterns exist in lesion development.
Optical coherence tomography (OCT) revolutionized clinical practice by allowing fast, non-invasive, micron-resolution visualization of the retina in cross-section. Routine OCT use has produced large clinical imaging databases, providing an avenue to study lesion evolution. Information on the clinical significance of lesions would support cellular and molecular-based understanding of AMD pathology, predict patient-specific disease-course, and enable earlier clinical trial endpoints for more rapid and cost-effective treatment development.
Achieving sufficient statistical power requires automated OCT interpretation. However, existing AI models segment only a limited set of AMD lesion types and train from significant manual annotations. Automated segmentation of small, widely distributed lesions is a major gap in the field.
We developed a self-supervised training pipeline for precise segmentation of reticular pseudodrusen (RPD). RPD are especially important as they are associated with decreased visual acuity and define a subset of patients more likely to progress to advanced disease.
Segmenting early RPD is challenging since they are small and blend with noise. Only two prior RPD detection models exist; both exclude small lesions and rely on supervised learning. We evaluated the single publicly available model on our own data (same OCT system and scanning parameters) and found that it fails to identify the majority of even large PRD. The other method publishes only two low-resolution outputs, misses 32% of the RPD volume, and detects extended lesions as fragmented components. Both models are unsuitable for RPD quantification and for studying early RPD, which is where intervention in AMD is most needed.
Harrison Greenbaum, MSTP Fellow, Cellular & Molecular Biology, Medical School
Obtaining reliable, noninvasive methods to diagnose cancer has remained an ever-present obstacle. As an MD-PhD student focused on cancer research, I have seen firsthand the grave outcomes of missed or late cancer diagnoses for patients in the clinic who are often afflicted by cancers for which we do not have reliable screening methods. Conversely, there are also too many patients who undergo invasive, costly, and at times painful procedures to evaluate for the presence of a non-existent cancer. In either case, there is a tremendous need for better, cost-effective, and noninvasive cancer screenings for a whole host of cancers.
Historically, cancer diagnoses have often been based upon identifying one key marker of a given cancer, and evaluating to see if a patient has significant elevation of that marker. However, cancer, like many other diseases, often has widespread effects on the expression of many genes that may be individually small but collectively result in monumental effects. If we knew how these genes were affected, we could offer new diagnostic and screening tools for cancer.
With this project, we have built a computational tool based on machine learning methodologies known as SignatureDx to analyze genetic expression in easily obtainable blood draw as a means to diagnose a variety of cancers. Initial usage of this model has elucidated novel peripheral blood markers of various cancers that in our preliminary tests, can diagnose some cancers with extraordinarily high accuracy. As the project continues, we plan to verify these results with hands-on blood based samples from patients with open-access genetic expression data to create verified, novel markers of cancer to uncover novel, noninvasive ways to diagnose cancer. Moreover, as we progress, we intend to publish the entire code as open-access to allow SignatureDx to be a viable tool for other researchers. Although our initial testing has been on cancer diagnostics, the model could be applied to create screening tests for any disease with detectable effects on genetic expression in blood, such as Alzheimer’s or Parkinson’s Disease.
By merging machine learning-based innovations with hands-on biological-clinical verification, we aim to substantially advance methods for noninvasive diagnostic testing.
Taige Lu, Graduate Student Instructor, Medicinal Chemistry, College of Pharmacy
Tuberculosis (TB), caused by Mycobacterium tuberculosis, remains a leading global cause of death, worsened by drug-resistant strains1. Recently approved nitroimidazole drugs such as CGI-1734122, pretomanid3, and delamanid4 are promising but limited by difficulties in scalable synthesis, especially of the crucial building block, 2-bromo-4-nitro-1H-imidazole. Existing debromination protocols suffer from moderate yields and safety concerns5, 6, impeding drug accessibility in under-resourced regions. Combining automated high-throughput experimentation (HTE) with data-driven algorithms7, we utilized our customized generative AI service developed by the University of Michigan, U-M GPT, where large language models like ChatGPT8, 9 are used to screen the vast chemical space of Cernak lab inventory, suggest reagents, and streamline the screening of reaction conditions. HTE rapidly tests thousands of combinations, but practical limits exist due to the vast search space. Machine learning, particularly Bayesian Optimization (BO), addresses this challenge by efficiently exploring reaction parameters10-12. Integrating U-M GPT-aided HTE and BO creates a powerful workflow for rapid reaction optimization, accelerating process development, and improving access to nitroimidazole drugs globally.
Core AI Methods
Day 1 – March 30
Larnell Moore, Graduate Student Research Assistant, Computer Science and Engineering, College of Engineering
Information is rapidly increasing across many domains, leading to large collections of unstructured textbooks, reports, and reference materials. A key challenge posed by this continuous growth is determining how to narrow down and retrieve relevant information distributed across multiple sources. For instance, a student may read multiple sustainability textbooks to learn about everyday practices that ordinary people can apply in their daily lives to reduce carbon emissions. However, these textbooks may cover a wide range of sustainability concepts that are ultimately unrelated to the student’s goal, such as chapters that focus on how corporations or governments can adopt sustainable practices rather than individuals. Additionally, useful information could also be buried within broader discussions of other topics. Thus, this poses a significant challenge with a broad societal impact, as the rapid growth of information demands systems that can aggregate and extract targeted, goal-relevant information across large collections of text.
Retrieval-Augmented Generation (RAG) is a popular solution to this problem. A user submits a query, the system searches for similar passages to the query, and a large language model (LLM) then uses the retrieved passages as supporting evidence to help formulate its response. GraphRAG extends this paradigm by modeling the system’s knowledge base as a graph. Most existing GraphRAG systems, when given a query, first performs a global similarity search over the graph to gather nodes, then expand them based on their structural importance in the graph. Then, the LLM reads the retrieved entities and relations, along with their corresponding passages, to generate a response. However, a global similarity search might accidentally pull in information that appears topically related on the surface but is irrelevant to the user’s informational goal expressed in their query. For example, when a user asks, “What are some daily practices I can do to help reduce carbon emissions,” the system may still retrieve carbon emission mitigation strategies intended for corporations or governments rather than individuals. Also, making expansion decisions based on static graph topology could overlook semantically distant but query-relevant information, leading to missing information in the retrieved context that the LLM may attempt to fill in with its pre-trained knowledge, potentially leading to hallucinations.
To address these challenges, we propose DotRAG, a new GraphRAG method that rethinks retrieval, reframing the paradigm from a retrieve-then-reason approach to a reason-while-retrieving approach. DotRAG leverages inference to make real-time decisions during retrieval, deciding which regions of the graph to explore conditioned on what it has explored so far and which relational paths would be helpful in answering the query before generating a final response.
Jian Hu, Associate Professor of Industrial and Manufacturing Systems Engineering, U-M Dearborn
Our research project addresses one of the most critical barriers to the widespread adoption of Artificial Intelligence in high-stakes domains: the “black box” problem. As deep learning models—specifically Convolutional Neural Networks (CNNs)—reach superhuman performance in tasks like medical diagnosis and biometric authentication, their decision-making processes remain opaque. This opacity creates significant risks regarding trust, safety, and accountability.
Our project introduces the Counterfactual Visual Explanation Process (CVEP), a novel framework designed to provide rigorous, quantitative interpretability for image classification models. Unlike traditional methods that offer static, qualitative visual comparisons, CVEP generates a continuous “counterfactual trajectory”—a sequence of semantically realistic images that transition a query image across a model’s decision boundary. By isolating the minimal modifications required to flip a prediction, we aim to transform abstract neural network decisions into localized, mathematically grounded, and human-understandable visual explanations.
Taha Draoui, Graduate Student Research Assistant, Computer Science, U-M Flint
Machine learning (ML) models are now widely used across scientific research and industry to support decision-making, automate analysis, and generate empirical insights. Despite this widespread adoption, ML pipelines are highly vulnerable to subtle methodological flaws that compromise the validity of experimental results. One of the most pervasive yet least visible of these flaws is data leakage: the unintended flow of information from evaluation data into model training. Data leakage can silently inflate reported performance, mislead scientific conclusions, and undermine trust in deployed systems.
My research addresses this challenge by developing AI-driven methods to detect and correct data leakage directly in machine learning code, with a particular focus on Jupyter notebooks, the dominant medium for applied ML research and education. Rather than treating leakage as a conceptual or dataset-level issue, my work studies how leakage concretely manifests in source code and how AI systems can be used to reason about these failures with precision and accountability. The broader goal is to strengthen the rigor, reproducibility, and trustworthiness of ML-based research by enabling automated, fine-grained analysis of ML pipelines as they are actually written and executed in practice.
Ecology & Environmental Science
Day 1 – March 30
Yuhao Zhao, Research Fellow, Ecology and Evolutionary Biology, College of Literature, Science, and the Arts
Ants are among the most ecologically dominant and evolutionarily successful lineages. They are diverse (~14,000 species) and old (~100 million years). Ants shape biodiversity by interacting with nearly all other organisms and accelerate the cycling of energy and materials in terrestrial ecosystems. We propose to use AI to generate precise morphometric data for nearly all ant species, a goal that will enable research about the ecology and evolution of the little creatures that run the world, the ants.
Morphological traits such as body size, head size, and mandible length mediate how ants forage, compete, and defend resources. Understanding trait variation offers a powerful mechanistic lens for understanding how ants interact with their environment and other species. But morphological trait data are rare or nonexistent for most ant species. What little data we do have come from labor-intensive manual measurements which can take months. Recent advances in AI, however, offer an opportunity to overcome these limitations. Powerful machine learning tools enable the rapid and consistent measurements of morphological traits from readily available images.
Our project is a deep-learning–based pipeline to assemble the most comprehensive database of ant morphological traits anywhere in the world, using tens of thousands of high-quality images of specimens hosted on antweb.org. These data will enable us to quantify how morphological traits vary across environmental and geographic gradients and to address whether highly successful ant species exhibit distinctive or more variable trait combinations, ultimately improving our understanding of how biodiversity evolved and how that biodiversity shapes ecosystems.
William Weaver, Schmidt AI in Science Fellow, 2025 Cohort, MIDAS
Long before satellites and DNA sequencers, scientists documented the living world by assembling libraries of life: pressed plants, pinned insects, and preserved animals in museum collections around the world. This immense historical record now underpins modern research on climate change, conservation, species distributions, and evolution. As technology has advanced, decades of curatorial labor have rendered physical specimens into photographs and spreadsheet-based records, now shared through global data repositories such as the Global Biodiversity Information Facility, which houses hundreds of millions of specimen records. Yet the scale and diversity of these collections come at a cost. As a result of their fragmented origins, these data are often inconsistent and incomplete, limiting their utility for large-scale analyses. Many specimens lack precise geographic coordinates or key biological trait information, and existing information is frequently embedded in free-text descriptions.
This research lies at the intersection of biodiversity informatics, evolutionary biology, and artificial intelligence, transforming fragmented biodiversity records into unified, analysis-ready datasets. By mobilizing the scientific potential of centuries of existing collections, it amplifies the value of museum and herbarium holdings and enables research questions that remain inaccessible when data persist as a mosaic of disconnected observations.
Xiaofeng Liu, Schmidt AI in Science Fellow, 2024 Cohort, MIDAS
Across the physical sciences, researchers increasingly seek predictive models that can operate across scales, variables, and data regimes. However, many physical systems, from climate and hydrology to biogeochem-istry and ecology, share two persistent challenges: (1) observations are sparse and heterogeneous, and (2) the governing processes are complex, nonlinear, and strongly interconnected. These characteristics significantly limit the scalability and generalizability of both traditional process-based modeling approaches and many data-driven methods, including statistical methods and conventional deep learning models.
Large pretrained models, also referred to as foundation models (FMs), have revolutionized natural language processing by learning transferable representations from massive datasets (e.g., GPT, Gemini). Inspired by these successes, researchers are increasingly exploring domain foundation models for scientific applications. However, there is currently no clear, generalizable blueprint for building foundation models that are physically grounded, data-efficient, and suitable for complex scientific systems.
This research develops a general blueprint for building foundation models in the physical sciences, using water resources systems as a motivating and rigorous application domain. The proposed frame-work combines knowledge-guided pretraining with modern foundation model architectures by first pretraining models on abundant but low-fidelity simulations from process-based hydrological models, followed by fine-tuning on sparse but high-fidelity real-world observations. This strategy enables foundation models to learn physically meaningful representations, resulting in more efficient, reliable, and scalable prediction of both water quantity and quality, while providing a transferable paradigm for foundation model development across physical science disciplines.
Engineering
Day 1 – March 30
Chenggong Jiang, Graduate Student Research Assistant, Chemical Engineering, College of Engineering
Industrial heterogeneous catalysts deactivate through nanoparticle sintering, where metal-support interactions govern thermal stability. The multifaceted complexity of these interactions—metal identity, particle size, support properties, operating conditions—has precluded systematic design principles for sinter-resistant materials. This research integrated first-principles neural network molecular dynamics with interpretable machine learning to elucidate metal-support interaction physics and enable high-throughput discovery of thermally stable catalyst supports, addressing a fundamental challenge with direct implications for automotive emission control and chemical manufacturing.
Ali Bahrami, Graduate Student Instructor and Graduate Student Research Assistant
Additive manufacturing is a versatile technology that allows for layer-based production of complex structures with a wide range of materials. At the microscale, it enables the fabrication of microelectronics using functional materials, which allows for high geometric freedom and customization at a reduced rate of material usage. Material jetting processes, including electrohydrodynamic jet (E-jet) and inkjet printing, offer distinct advantages in resolution, material compatibility, and direct-write capability. However, the scalability and robustness of these processes are hindered by the complex multiphysics interactions, process variability, and sensitivity to material and operating conditions.
Achieving repeatable outcomes requires careful coordination of electrical, mechanical, and fluidic effects, which exceeds the capabilities of manual or rule-based control. As a result, AI has emerged as a foundation for robust material jetting by integrating modeling, sensing, and control to support reliable and scalable operation. In this work, AI replaces manual inspection and trial-and-error tuning with data-driven perception, learning, and control.
Anwar Ghammam, Assistant Professor of Computer and Information Science, U-M Dearborn
Software quality is no longer determined by source code alone. Modern systems rely on increasingly complex build systems, continuous integration (CI) pipelines, and DevOps configurations that orchestrate how software is compiled, tested, and deployed. Failures or quality decline in these artifacts affect the software reliability, performance, and long-term maintainability. Despite their central role in modern software delivery, these artifacts remain among the least systematically studied components of software systems. This work represents an AI-centered research initiative that advances a paradigm shift in software quality research—from source code-centric to lifecycle-aware intelligence. We position an AI-framework as a collaborator in analyzing and improving quality and maintenance across modern DevOps artifacts—including build systems and CI pipelines—which are traditionally difficult to inspect using manual or rule-based techniques. To achieve this, we integrate multiple AI methodologies in a complementary manner, combining traditional machine learning (ML) with large language models (LLMs), and modern agentic-AI systems to incrementally build an AI- teammate that supports developers throughout modern DevOps workflows. In this context, AI analyzes build and CI artifacts, detects failures, measures quality, and proposes context-aware fixes, or refactorings, while humans retain validation, and final decision-making authority. The central question driving this work is how AI can enable more rigorous, scalable, and responsible maintenance and optimization of modern DevOps-driven software systems.
Xiaohao Xu, Graduate Student Instructor, Robotics, College of Engineering
Our research introduces a novel paradigm in robotics: using Large Language Models (LLMs) not merely as coding assistants, but as the engine of “natural selection” for the automated design of physical machines.
Soft robots—composed of compliant materials—offer distinct advantages in unstructured environments (e.g., search-and-rescue). However, designing them is a bottleneck; it is a complex, intuition-driven process requiring the simultaneous optimization of Morphology (body structure), Control (brain), and Sensing (perception). Traditional human engineering struggles to navigate the infinite degrees of freedom in soft bodies.
Our project, Natural Selection via Foundation Models, automates this by embedding LLMs into an evolutionary algorithm. The AI acts as the “designer,” generating and mutating robot code based on feedback. This award is critical for the next phase of our work: extending our model from optimizing movement to optimizing sensing, thereby creating robots that can not only move but perceive their environment intelligently.
Xiao Zhang, Assistant Professor of Computer and Information Science, U-M Dearborn
Indoor localization is a key part of smart infrastructure, enabling warehouse robot navigation, asset tracking, emergency response and indoor navigation, with its global market projected to reach $4.3 billion by 2030. Traditional RF-based solutions (Wi-Fi, BLE, UWB) have critical flaws including coarse 2-10 meter accuracy, prohibitive UWB costs, RF congestion and cloud-related privacy leaks, while GPS fails indoors, creating an urgent gap for reliable, privacy-protected high-precision self-localization—especially in warehouses. LiFind fills this gap as a novel framework, using optical camera communication and AI to deliver centimeter-level accuracy via pre-installed ceiling LEDs (no dedicated hardware), leveraging smartphone rolling shutter to capture invisible LED patterns, applying AI for real-time LED identification and 3D ranging, and using LSE-optimized trilateration for precise 2D localization to provide a low-cost, scalable alternative to RF solutions and pioneer an AI-driven OCC paradigm.
Yuxin Lin, Graduate Student Researcher, DART Lab, Taubman College of Architecture and Urban Planning
In the next few decades, billions more people will need housing and infrastructure, and reinforced concrete will remain central to meeting that demand. Yet conventional concrete construction remains expensive, labor intensive, and material intensive. Robotic 3D concrete printing (3DCP) could automate construction and reduce material use, particularly for materially efficient concrete structures (MECS). The key challenge is reliability: MECS often feature complex geometries that are sensitive to time-dependent material behavior and a distributed fabrication chain (mixing, pumping/extrusion, and robot motion). In practice, 3DCP remains largely trial-and-error and relies on manual observation and ad hoc tuning.
Abdallah Kamhawi, Graduate Student Researcher, DART Lab, Taubman College of Architecture and Urban Planning
This project develops a scaffolding-free robotic 3D concrete printing (3DCP) workflow for vaulted structures, including non-self-stable and freeform geometries. Its significance lies in solving a persistent constructability problem in architecture and structural design: although vaulted and shell-like forms are materially efficient—carrying loads primarily through compression—they remain difficult to build at scale because conventional delivery depends on labor-intensive formwork, long setup times, and costly temporary supports. Current 3DCP approaches either require significant scaffolding for assembly of prefabricated components or are restricted to inherently self-supporting forms, leaving a broad class of structurally viable vaults inaccessible.
By maintaining stability at every intermediate construction state, this project enables these high-performance geometries to be fabricated without scaffolds—improving feasibility, reducing waste, and expanding practical deployment in contemporary construction. The method deposits a primary arch layer followed by a zigzag lamination layer that stitches filaments into a laminated composite capable of cantilevering without scaffolding. Critically, the project integrates generative AI—specifically, reinforcement learning (RL)—to automatically generate stability-aware print paths, task schedules, and process parameters for unseen vault geometries, a decision problem exceeding the capacity of conventional geometric or simulation-based approaches.
In the broader disciplinary context, this research contributes to a major transition in architecture, engineering, and construction (AEC): from geometry-first digital workflows to performance-driven, AI-enabled fabrication systems that co-optimize design intent, mechanics, and production logic. The work integrates architectural computation, early-age concrete mechanics, and construction robotics in one methodological framework aligned with current priorities in the discipline: lowerembodied material use, improved productivity, and reproducible quality in complex construction. Rather than treating AI as a post-processing tool, the project positions generative AI as a core research method for coordinating structural behavior, material kinetics, and robotic execution, offering a transferable model for rigorous, data-informed design-to-fabrication research.
Mohamed Wiem Mkaouer, Associate Professor of Computer Science, U-M Flint
This project develops a scaffolding-free robotic 3D concrete printing (3DCP) workflow for vaulted structures, including non-self-stable and freeform geometries. Its significance lies in solving a persistent constructability problem in architecture and structural design: although vaulted and shell-like forms are materially efficient—carrying loads primarily through compression—they remain difficult to build at scale because conventional delivery depends on labor-intensive formwork, long setup times, and costly temporary supports. Current 3DCP approaches either require significant scaffolding for assembly of prefabricated components or are restricted to inherently self-supporting forms, leaving a broad class of structurally viable vaults inaccessible.
By maintaining stability at every intermediate construction state, this project enables these high-performance geometries to be fabricated without scaffolds—improving feasibility, reducing waste, and expanding practical deployment in contemporary construction. The method deposits a primary arch layer followed by a zigzag lamination layer that stitches filaments into a laminated composite capable of cantilevering without scaffolding. Critically, the project integrates generative AI—specifically, reinforcement learning (RL)—to automatically generate stability-aware print paths, task schedules, and process parameters for unseen vault geometries, a decision problem exceeding the capacity of conventional geometric or simulation-based approaches.
In the broader disciplinary context, this research contributes to a major transition in architecture, engineering, and construction (AEC): from geometry-first digital workflows to performance-driven, AI-enabled fabrication systems that co-optimize design intent, mechanics, and production logic. The work integrates architectural computation, early-age concrete mechanics, and construction robotics in one methodological framework aligned with current priorities in the discipline: lowerembodied material use, improved productivity, and reproducible quality in complex construction. Rather than treating AI as a post-processing tool, the project positions generative AI as a core research method for coordinating structural behavior, material kinetics, and robotic execution, offering a transferable model for rigorous, data-informed design-to-fabrication research.
Day 2 – March 31
Leo Tunkle, Graduate Student Research Assistant, AIMS Lab, College of Engineering
In the Controls section of the AIMS group (Artificial Intelligence and Multi-physics Simulations) led by Professor Majdi Radaideh, we have been leveraging diverse developments in artificial intelligence to improve the control of nuclear reactors. Effective control requires estimating the state of the system of interest and calculating the appropriate control response to achieve a desired outcome. Due to the complex physics and harsh operating conditions of nuclear reactors, both of these objectives are difficult to achieve because of factors such as uncertain or missing sensor data, control actuator wear, fluctuating grid conditions from increasing use of renewables, and timescale dependent reactor behavior. With several journal papers published over the past year and a half, our work aims to address these challenges using a three pronged digital twin approach with significant machine learning components: a generative loop, leveraging generative adversarial networks (GANs) and diffusion models to aid in system state reconstruction based on sensor data; an inverse loop, which appropriately assimilates known data with the reconstructed state to enable uncertainty quantification with variational inference; a forward loop, which determines likely system evolution and makes appropriate control actions using a combination of traditional control methods and reinforcement learning (RL).
Doohyun Kim, Assistant Professor of Mechanical Engineering, U-M Dearborn
Complex chemical and physical processes within combustion-based energy conversion devices can be modeled at several levels. While three-dimensional CFD model offers detailed spatial and temporal resolution, its high computational cost prohibits its usage as an iterative fuel and engine design optimization tool. At the other end, quasi-dimensional (Q-D) modeling approaches provide superior computational efficiency and are widely used for rapid prototyping and exploration of operating domain.
In the literature, Q-D models are typically developed and calibrated for specific fuels, relying on pre-calibrated empirical correlations to predict various combustion phenomena. These correlations prioritize computational efficiency over physical completeness. While this tradeoff is acceptable for fixed fuels, it limits general applicability, particularly for emerging sustainable fuels, which exhibit significantly greater physicochemical variability than conventional ones. As the transportation sector transitions toward diverse energy carriers, the simultaneous optimization of engine hardware and fuel formulation is critical. Consequently, predictive, fuel-flexible simulation tools are becoming essential assets for navigating this complex design space efficiently.
To address this, we introduce Machine Learning (ML) based models to substitute traditional, fuel-specific correlations. Focusing on critical combustion phenomena inside Spark Ignition (SI) engines, which are laminar flame speed (LFS) and ignition delay time, this approach prioritizes fuel flexibility. Unlike static correlations, these models are designed to explicitly capture the non-linear effects of varying fuel composition and properties, ensuring applicability across diverse energy carriers including emerging fuels.
Omer Faruk Erdem, Graduate Student Research Assistant, Nucear Engineering and Radiological Sciences, College of Engineering
Complex chemical and physical processes within combustion-based energy conversion devices can be modeled at several levels. While three-dimensional CFD model offers detailed spatial and temporal resolution, its high computational cost prohibits its usage as an iterative fuel and engine design optimization tool. At the other end, quasi-dimensional (Q-D) modeling approaches provide superior computational efficiency and are widely used for rapid prototyping and exploration of operating domain.
In the literature, Q-D models are typically developed and calibrated for specific fuels, relying on pre-calibrated empirical correlations to predict various combustion phenomena. These correlations prioritize computational efficiency over physical completeness. While this tradeoff is acceptable for fixed fuels, it limits general applicability, particularly for emerging sustainable fuels, which exhibit significantly greater physicochemical variability than conventional ones. As the transportation sector transitions toward diverse energy carriers, the simultaneous optimization of engine hardware and fuel formulation is critical. Consequently, predictive, fuel-flexible simulation tools are becoming essential assets for navigating this complex design space efficiently.
To address this, we introduce Machine Learning (ML) based models to substitute traditional, fuel-specific correlations. Focusing on critical combustion phenomena inside Spark Ignition (SI) engines, which are laminar flame speed (LFS) and ignition delay time, this approach prioritizes fuel flexibility. Unlike static correlations, these models are designed to explicitly capture the non-linear effects of varying fuel composition and properties, ensuring applicability across diverse energy carriers including emerging fuels.
Van Hai Bui, Assistant Professor of Electrical and Computer Engineering, U-M Dearborn
Modern power systems are undergoing a rapid transformation driven by the large-scale integration of renewable energy resources, such as photovoltaic (PV) systems and wind turbines (WTs), and the widespread adoption of electric vehicles (EVs). While these technologies are essential for decarbonization, they introduce significant operational uncertainty due to their stochastic, weather-dependent, and user-driven nature. Conventional power system operation and control methods, including scenario-based stochastic optimization and Monte Carlo simulations, require extensive computational resources and large numbers of scenarios to adequately capture uncertainty. As a result, these approaches are often too time-consuming and impractical for real-time or near–real-time system operation.
This project addresses these challenges by developing an AI-centered framework for probabilistic prediction, uncertainty-aware optimization, and adaptive control in power and energy systems. The overarching goal is to replace computationally intensive traditional approaches with AI-driven models that can be trained offline and deployed online, enabling fast, accurate, and scalable decision-making under uncertainty. By embedding probabilistic machine learning and deep reinforcement learning into power system operation and design, this research advances both the theoretical foundations and practical deployment of intelligent energy systems.
Shuai Che, Research Fellow, Mechanical Engineering, College of Engineering
The Printed Circuit Heat Exchanger (PCHE) is one of the promising heat exchanger candidates for advanced nuclear reactors and high-temperature applications. Compared to traditional shell-and-tube heat exchangers, chemically etched millimeter-scale channels significantly increase heat transfer area density. Among mini-channel designs, zigzag configurations are commonly employed to enhance flow turbulence and heat transfer. However, a primary barrier to semi-circular zigzag PCHE development is the lack of accurate thermal-hydraulic models. With the different flow channel geometries, flow regimes, and working fluids, it is almost impossible to use one general correlation to accurately predict the thermal-hydraulic performance of the PCHEs. Motivated by the efficacy of data-driven methods, this study aims to develop machine learning models to predict the friction factor (fD) and Nusselt number (Nu) for semi-circular zigzag channels in PCHEs. A number of data samples were collected from published literature, including the channel geometric characteristics and operating conditions. Various machine learning techniques were employed, involving kernel methods (KRR and SVR), Artificial Neural Networks (ANNs), and tree-based ensemble methods (Random Forest and Extreme Boosting Tree). The polynomial KRR model was able to accurately predict the Nusselt number Nu. While the ANN models and tree-based models yielded high R2 test scores over 0.97, the maximum percentage errors still exceeded 60%, which was not acceptable in industrial applications. Additional expanded datasets and a detailed analysis of the input feature distribution are essential for further accuracy improvements.
Md Abul Kalam Azad, Graduate Student Research Assistant, Computer & Information Science, U-M Dearborn
My research addresses a critical software-quality bottleneck: performance inefficiencies that silently waste compute, memory, and developer effort. I build practical AI-driven program analysis methods that move from empirical understanding to actionable detection and patching.
In the first phase, our MSR’23 study of HPC software established the empirical foundation: performance bugs differ systematically from functional bugs. They are harder to diagnose, demand more complex fixes, and are more often resolved by expert developers. This gap motivates intelligent and scalable tool assistance. Translating that need into deployable AI systems requires strong data infrastructure. Yet public real-world performance datasets remain limited in scale and diversity, constraining robust model development and evaluation.
In the second phase, we introduced PCMiner (SC’24), an LLM-guided framework for large-scale identification and curation of performance-related historical commits. Using this dataset, we developed a RAG-enabled LLM detector for inefficient API misuse and observed a 28% gain in detection accuracy. Together, this semantically guided mining framework, one of the largest multi-lingual curated datasets of performance-related commits, and the observed empirical improvements establish a strong foundation for data-driven software performance analysis.
In the current phase, we evaluate multiple LLMs as static memory-leak detectors on a high-quality benchmark derived from our PCMiner pipeline. Initial findings show that even state-of-the-art models struggle with difficult leak cases, especially ownership and lifetime reasoning. To address this limitation, we introduced an LLM-augmentation strategy that improves detection performance by 9 percentage points. We are now extending this work through hybrid pipelines that combine LLM reasoning with traditional static analyzers to improve robustness and deployment readiness. This past-present-future arc positions AI as the core engine for scalable, reliable software performance analysis.
Humanities & Arts
Day 2 – March 31
Ting-Yu Pan, Graduate Student Research Assistant, Electrical and Computer Engineering, College of Engineering
A cappella is a rapidly growing form of collegiate music-making that involves thousands of student performers each year and is characterized by small, peer-led vocal ensembles without conductors. Unlike traditional choral settings, collegiate a cappella groups are typically self-organized, with students rotating through roles such as soloist, arranger, and music director. This structure creates rich opportunities for collaborative learning, but also places significant demands on novice singers to practice independently without professional guidance.
While the recent advancement of AI has been transforming the music industry, little progress of AI technology innovations have been made for a cappella despite its popularity and importance. This is partly due to the lack of AI infrastructure, including datasets and foundational models. To alleviate this, this ongoing project develops (1) ACappellaSet, a multilingual multitrack a cappella dataset for robust source separation, and (2) AcaMate, an AI-assisted rehearsal tool that supports novice acappella singers’ sensemaking of musical information and asynchronous practice. Situated at the intersection of audio ML and human-centered AI, AI is foundational to our work: it enables dataset-driven benchmarking, model adaptation for dense same-source separation, and the human-AI interactive workflow in asynchronous rehearsals.
Through dataset curation, foundational model building, and human-AI interface design, we aim to establish the AI infrastructure for broader AI music research, such as source separation, remote collaboration, and AI-powered music education.
Joshua Ashkinaze, Graduate Student Research Assistant, School of Information
If we think of culture as a feedback loop where individuals and societies shape each other through exchanges of ideas (Boyd & Richerson 1988), then what happens when generative AI enters this “culture loop”? Exposure to LLM output is rapidly increasing, and the ideas we see affect the ideas we create (Nijstad & Stroebe 2006). Yet most experiments on human-AI creativity use static designs that measure one-shot interactions, missing the compounding effects of AI ideas building on AI-influenced ideas over time.
We conducted a large-scale, dynamic experiment (Ashkinaze et al. 2025) with over 800 participants from 40+ countries. Participants completed the Alternate Uses Task—a standard creativity measure—after viewing example ideas from ChatGPT or from prior participants. We varied AI exposure (none, low, high) and AI disclosure (labeled or not). Critically, participants’ submitted ideas became examples for future participants in the same condition, creating response chains that simulate how AI ideas compound through a culture. High AI exposure did not affect the creativity of individual ideas but did increase both the average amount and the rate of change of collective idea diversity. AI made ideas different, not better. There were no main effects of disclosure. Self-reported creative people were less influenced by knowing an idea was from AI, and participants adopted AI ideas more when the task was difficult. As described in Achievements, this work received an Honorable Mention (top 5% of submissions) at the 2025 ACM Collective Intelligence Conference.
Hao-Wen Dong, Assistant Professor of Music, School of Music, Theatre & Dance
This project aims to explore how AI can be integrated into professional video editing workflow and advance foundational AI methodology for next-generation AI-assisted video editing interface. Video editing is a crucial yet laborious step for video production that sometimes takes more time than filming the content. Oftentimes, video creators first shoot a pool of raw video materials and later edit them into a cohesive story, sometimes in various formats and durations for different platforms. While current video generation models can generate compelling short videos, they fall short in inserting real, filmed materials into its outputs.
In this ongoing project, we explore novel machine learning models that can cut, select, and rearrange a long video into a short video. Thus far, we have already made significant contributions into this emerging field of AI-assisted video editing, including new datasets, foundational models, evaluation metrics, and integrated systems. Our results have been published at ICLR 2025 and NeurIPS 2025, two top machine learning conferences. In the first research paper published in ICLR 2025, we have compiled a new publicly-available dataset consisting of 1,269 high-quality documentaries paired with their teasers, which addresses the long-standing issue of data accessibility and reproducibility in this field. Moreover, we have developed a narration-centered teaser generation system that can effectively compress 30+ minute documentaries into 3-minute teasers leveraging pretrained large language models (LLMs) and language-vision models. In the second research paper published in NeurIPS 2025, we have developed a new retrieval-embedded generation framework that allows an LLM to quote multimodal resources while maintaining a coherent narrative. We have also proposed a novel long-to-short video editing model for generating shorts that feature a coherent narrative with embedded video insertions extracted from a long input video. Unlike most current multimodal generative AI frameworks that focus only on taking multimodal content as inputs rather than generating multimodal outputs, our proposed methodology enables LLMs to generate structured multimodal outputs. Our project contributes towards next-generation video editing interfaces that can be integrated into the existing creative workflows of video creators using multimodal LLMs and retrieval augmented generation (RAG).
The rapid proliferation of generative AI has precipitated a moment of profound uncertainty for the creative community. While AI offers powerful tools for ideation and production, it simultaneously generates great anxiety regarding the future of human creativity. It is our goal through this research project to transform AI from a threat into an enduring creative partner. It is our utmost goal to develop AI that augments, not replaces, human creativity.
Math & Physical Sciences
Day 1 – March 30
Haotian Chen, Schmidt AI in Science Fellow, 2024 Cohort, MIDAS
Scientific modeling has advanced across four paradigms: empirical, analytical, numerical, and the state-of-the-art, data-driven paradigms. Data-driven machine learning, however, are “black-box” that does not explicitly correlated underlying physics and data, and thus are non-interchangeable, non-extrapolative, and vulnerable to curse of dimensionality. Differentiable Electrochemistry (DiffEC) is thus proposed as a new framework that integrates thermodynamic, kinetic, and mass transport equations with differentiable programming enabled with Automatic Differentiation (AD). We introduce the first end-to-end differentiable electrochemistry simulation framework. By making electrochemical transport and kinetics fully differentiable, our contribution spans two complementary perspectives:
- We developed five general-purpose Differentiable Electrochemistry simulators that span fundamental to advanced electrochemical mechanisms.
- We leveraged Differentiable Electrochemistry to overcome long-standing bottlenecks in system identification, and thereby advance mechanistic understanding of electrochemical response.
Subsequently, Differentiable Electrochemistry has been applied to fuel cells and Li-ion batteries for essential kinetic analysis, establishing its broad applicability and transformative potential for next-generation electrochemical system design. Looking ahead, Differentiable Electrochemistry represents a paradigm shift in how electrochemical systems are modeled and interpreted, illustrating how physical laws and data can be unified through differentiable programming to enable trust-worthy and extrapolative discovery across chemistry, materials, and energy disciplines.
Ricardo Vinuesa, Associate Professor of Aerospace Engineering, College of Engineering
Turbulence is one of the most challenging open problems in classical physics. Despite decades of theoretical modeling, high-fidelity simulations and experiments, the causal mechanisms that govern turbulent flows remain only partially understood, and effective fluid-flow control strategies in realistic regimes are still elusive. This limitation has profound consequences for aerospace, energy and environmental systems, where turbulence directly impacts efficiency, emissions and sustainability.
The overarching aim of my research is to fundamentally transform how turbulence is understood, modeled and controlled by placing artificial intelligence (AI) at the center of the scientific process. Instead of using AI as a post-processing or acceleration tool, my research develops AI methods for discovery, control and simulation of turbulent flows. My work focuses on explainable artificial intelligence (XAI), deep reinforcement learning (DRL) and AI-based surrogate modeling, forming a unified research program that has already produced multiple high-impact results.
Chuqi Chen, Assistant Professor of Mathematics, College of Literature, Science, and the Arts
Partial Differential Equations (PDEs) are the bedrock of scientific modeling in physics, engineering, and applied sciences. However, in modern applications involving high-dimensional systems, complex geometries, and inverse problems, classical numerical methods often become computationally prohibitive. This bottleneck has motivated the rapid adoption of Artificial Intelligence (AI) solvers. Yet, a critical gap remains: while Neural PDE solvers demonstrate strong empirical potential, they frequently suffer from slow convergence, instability, and a lack of interpretability.
My research addresses this reliability crisis by moving beyond ”black-box” experimentation. I focus on establishing a rigorous mathematical framework to improve the reliability, efficiency, and interpretability of AI methods for scientific computing. Specifically, my project answers two fundamental questions:
- Quantification: How can we mathematically quantify the training difficulty in neural PDE solvers?
- Methodology: How can we overcome these difficulties by designing specialized neural architectures and training strategies?
Rather than treating neural networks as black-box surrogates, my work systematically investigates their optimization landscapes, spectral structures, and training dynamics in order to address these two central questions. From a theoretical perspective, we analyze convergence and learning dynamics using tools such as Neural Tangent Kernel (NTK) theory and spectral methods. These analyses directly inform the principled design of architectures, initialization strategies, and optimization procedures to mitigate training difficulty and improve robustness.
This project is situated within the broader field of scientific machine learning, which seeks to integrate data-driven models with physical laws and numerical analysis. By combining mathematical theory, algorithmic design, and large-scale numerical experiments, my research advances a rigorous and responsible framework for deploying AI in complex scientific applications.
Yunus Zeytuncu, Professor of Mathematics, U-M Dearborn
4OPS started from a simple observation that anyone who teaches mathematics quickly recognizes: two arithmetic problems that look similar on the surface can feel very different to a learner. One may be solved almost immediately, while the other causes hesitation, trial-and-error, or disengagement. The goal of the 4OPS project has been to understand this gap in a principled way and to build a research and deployment platform around it.
4OPS is a publicly available mobile arithmetic puzzle app that has been downloaded hundreds of times and is actively used as both a learning tool and a research environment. At the research level, the project examines how difficulty in arithmetic reasoning can be defined, measured, and explained using structural rather than performance-based approaches. At the systems level, it explores how artificial intelligence can make such a structure usable at scale, in real educational settings. The result is a project in which AI is not an add-on, but part of the conceptual and technical foundation.
Shaghayegh Emami, Graduate Student Research Assistant, Physics, College of Literature, Science, and the Arts
Experiments at CERN’s Large Hadron Collider (LHC), collide protons to probe the physics of fundamental particles at the smallest accessible scales. There, our goal is to identify rare events within the large flux of particles produced in collisions and to perform measurements across a wide range of phenomena to understand fundamental forces and interactions in nature. The billion-channel particle detectors collect millions of images each second, producing a massive amount of data, requiring algorithms that filter data for storage and analysis in real-time, commonly referred to as trigger algorithms. Trigger algorithms must be highly selective, often filtering out 1-in-1000 events operating under severe bandwidth, computational, and storage constraints, with collision rates of 40 million events per second and storage volumes of hundreds of Petabytes per year.
Traditionally, data filtering strategies rely heavily on detailed prior knowledge of the physics processes being probed and precise simulations of the detector and instrumentation. As a result, data selection rules are designed around predefined physics objectives and expected detector performance. While robust under stable conditions, such rules are effectively static and inflexible: they do not automatically adapt to time-dependent variations in the detector environment, or the emergence of new features in the data that were not previously modeled or statistically significant. Consequently, trigger algorithms design and operation require substantial manual effort and expert oversight, and can take weeks or months to evaluate and deploy.
This potentially leads to both known and unknown human biases, raising concerns that novel phenomena may be missed or that uninteresting data may be preferentially selected. These limitations highlight the need for autonomous trigger systems that respond dynamically to real-time changes, reduce bias, and maximize the possibility of rare signal collection.
Our research team brings together expertise in detector instrumentation, experimental particle physics, and computer science, with a shared focus on extending the discovery reach of LHC experiments through optimized data selection. We are motivated by the dual challenge of maximizing physics sensitivity while operating within strict resource constraints. Our team is developing a flexible framework to deploy adaptive AI data-filtering systems, to maximize discovery potential in the face of evolving experimental conditions. This focus is motivated by the fact that the trigger system, the real-time filtering layer in the collision environment, serves as the gatekeeper of recorded data and, therefore, of experimental discovery.
Day 2 – March 31
Elad Zelingher, Donald J. Lewis Postdoctoral Research Assistant Professor, Mathematics, College of Literature, Science, and the Arts
Together with Nate Harman (University of Georgia) and Andrew Snowden (University of Michigan), we resolved an over 30-year-old open problem and found an explicit formula for units of rank ideals in nite matrix monoid algebras. We discovered our formulas by running computational experiments in SageMath, written partially with the help of large language models, and by asking LLMs for assistance with pattern identication. Our project demonstrates that with AI assistance, one can unravel complicated patterns and make new discoveries in mathematics.
Matthew Lynch, Graduate Student Research Assistant, Nuclear Engineering and Radiological Sciences, College of Engineering
At the intersection of materials science and nuclear engineering lies the field of nuclear materials, where research investigates how the radiation from a nuclear reactor slowly degrades the materials that comprise it. By understanding these mechanisms of degradation, we can drive the development of advanced reactor designs (including next-generation fusion concepts) that deliver higher efficiency, greater safety, and reduced costs. However, a significant bottleneck in this research is the quantification of radiation-induced damage on a material’s microstructure. Traditionally, this quantification involves the tedious manual labeling of dozens to hundreds of individual features across as many electron microscopy images. These features are difficult to identify, not only because they represent less than 1% of the image, but also due to substantial image noise caused by the radiation. This labor-intensive process slows progress and limits both the scale and accuracy of materials analysis.
Melody Zhang, Graduate Student Research Assistant, Chemical Engineering, College of Engineering
The rational design of nanomaterials remains a central obstacle to translating nanoscale science into real-world technologies. Despite extensive experimental and computational advances, materials discovery is still dominated by costly trial-and-error approaches to navigate the vast combinatorial design space of nanoscale systems. My doctoral research addresses this challenge by developing foundational AI-driven frameworks for inverse design, enabling predictive, scalable, and physics-informed discovery of nanomaterials. By integrating machine learning with statistical mechanics and molecular simulation, my work replaces empirical search with data-driven design principles.
My research focuses on machine-learning method development for colloidal nanomaterials, which exhibit emergent self-assembly behavior underpinning applications in catalysis, drug delivery, consumer products, and beyond. Meanwhile, their tunability in surface chemistry and geometry makes rational design intractable using traditional forward modeling, which directly simulates these systems. However, conventional molecular mechanics–based interaction models used in these simulations are typically rigid and require manual reparameterization for each new system, limiting their ability to explore large design spaces efficiently. In contrast, machine-learned interaction models (ML-IAMs) provide a flexible, data-driven framework for representing nanoparticle interactions, making inverse design of chemically detailed nanomaterials computationally feasible. In my work, I actively integrate AI to define, optimize, and interpret representations of particle interactions.
Valentin Fondement, Research Fellow, Nuclear Engineering and Radiological Sciences, College of Engineering
Mixed-field radiation environments are challenging to measure reliably in demanding conditions, such as in-core reactor locations (<400°C), where conventional detector materials and photodetectors face severe mechanical and thermal constraints. Our project addresses this gap by pairing an unusually rugged inorganic scintillator, a YAP:Ce, with an interpretable machine learning approach to enable Particle IDentification (PID) in mixed fields even when scintillation pulse-shape differences are subtle and light collection is degraded.
YAP:Ce is attractive for extreme deployments because of its high melting point (~2000 °C), hardness, and fast scintillation. However, its intrinsic Pulse-Shape Discrimination (PSD) capability is typically considered insufficient for practical PID using standard Charge Comparison Methods (CCM). This work demonstrated experimentally that AI can unlock useful PSD performance from a material without acknowledged PSD capability, thereby broadening the set of commercially available scintillators that can be used for harsh-environment measurements.
Changwen Xu, Graduate Student Research Assistant, Mechanical Engineering, College of Engineering
Predicting the properties of crystalline materials is essential for understanding structure–property relationships and accelerating the discovery of functional materials [Wang and Pan, 2008, Pilania et al., 2013, Ahmad et al., 2018, Tabor et al., 2018, Pyzer-Knapp et al., 2022]. Conventional approaches based on experimental characterization or density functional theory (DFT) calculations provide reliable estimates but are often computationally expensive and difficult to scale for high-throughput screening [McCullough et al., 2020, Chen et al., 2024]. As a result, only a small fraction of the vast chemical and structural design space of crystalline materials can be explored in practice [Hellenbrandt, 2004, Davies et al., 2016]. Machine learning models offer a promising alternative by learning structure–property relationships from data; however, many existing approaches rely heavily on
labeled datasets, adopt representations that insufficiently capture crystallographic structure, or lack integration with physical principles, limiting their generalizability and interpretability [Fujinuma et al., 2022, Phuthi et al., 2024].
To address these challenges, I have been working on developing CLOUD (Crystal Language mOdel for Unified and Differentiable materials modeling) [Xu et al., 2025a], a transformer-based foundation model designed to provide a scalable and physics-informed framework for crystal property prediction. Rather than treating artificial intelligence as a purely data-driven surrogate for expensive simulations, CLOUD is developed as a foundational framework that encodes crystallographic symmetry, structural constraints, and physical laws directly into the learning process. The project integrates concepts from machine learning, crystallography, and thermodynamics, reflecting a fundamentally interdisciplinary approach to materials research.
Mengqi Lin, Graduate Student Research Assistant, Statistics, College of Literature, Science, and the Arts
Sensitivity analysis quantifies how strong an unmeasured confounder must be to overturn causal conclusions from observational studies. Analysts often pair sensitivity analysis with matching to approximate randomized treatment assignment within matched sets. However, matching is rarely exact: residual imbalance in measured covariates remains within sets. This inexact matching can introduce systematic bias in sensitivity-analysis test statistics, meaning sensitivity conclusions may partly reflect remaining imbalance in measured covariates rather than robustness to unmeasured confounding.
This project addresses that problem using regression adjustment. We first adjust outcomes for observed covariates via regression and then conduct sensitivity analysis using the regression residuals as the test statistic. The theoretical goal is to show that regression residualization removes the leading dependence of the statistic on measured covariates and that the remaining bias is asymptotically negligible as match quality improves with increasing sample size. In particular, our analysis targets a quantitative rate for the bias control under inexact matching, on the order of I^(−2/p), where I is the number of matched sets and p is the covariate dimension. Establishing and empirically validating such rates is challenging because sensitivity analysis involves worst-case optimization that can be non-smooth (the worst-case configuration may change under small perturbations induced by inexact matching).
Social Science
Day 1 – March 30
Kyle McCullers, Graduate Student Research Assistant, Sociology, College of Literature, Science, and the Arts
Romantic desirability is increasingly being articulated in the language of markets. Across contemporary dating cultures, individuals present themselves not only as partners but as “entrepreneurs of themselves” (Vallas & Christin, 2018): emphasizing what they “bring to the table”, their commitment to self-improvement, the degree to which they willingly engage in productivity-generating activities beyond typical 9-5 employment, etc. Dating, then, has become another institution through which the “enterprising subject” (Foucault, 2008) is (re)produced. This project investigates how such neoliberal ideologies are reshaping perceptions of romantic desirability among Black Americans, using the popular YouTube dating series Pop The Balloon or Find Love as an empirical case. On the show, participants publicly decide whether to “pop” a balloon to reject a potential partner or keep it unpopped to signal interest, then explain their decisions. A noticeable trend is how contestants draw on neoliberal ideologies to position themselves as desirable and evaluate others. Contestants frequently highlight entrepreneurial endeavors (many claim to “own multiple businesses”), demonstrate financial literacy through references to investing and building generational wealth, fetishize meritocratic individualism, and frame romantic relationships as quasi-business partnerships requiring mutual “value-add.” These patterns suggest that neoliberal ideologies, including entrepreneurialism, market-based subjectivities, and individual responsibility, have become central to structuring desirability in our contemporary moment.
Siyu Yu, Assistant Professor of Management and Organizations, Stephen M Ross School of Business
This project examines how the integration of generative artificial intelligence (GenAI) into team-based work reshapes the social foundations of status inequality in collaborative settings. As GenAI tools such as large language models become routine infrastructure in organizations, they are increasingly embedded directly into team workflows—supporting idea generation, problem solving, and decision making in real time. While prior research has focused primarily on productivity, efficiency, trust, or ethical risks associated with AI, much less is known about how AI-mediated collaboration alters informal social hierarchies within teams: who is perceived as competent, who gains influence, and who emerges as a leader during interaction.
The central aim of this research is to understand whether generative AI amplifies or disrupts existing status advantages rooted in diffuse social characteristics such as race and gender. Decades of research in status characteristics theory demonstrate that when task-relevant information is ambiguous, group members rely on socially constructed status beliefs to infer competence and allocate influence. These processes historically advantage dominant-group members, particularly White men, who are more readily presumed competent and worthy of deference. The introduction of GenAI presents a theoretically important intervention into this process. By making high-quality task outputs rapidly available and visibly shareable during collaboration, AI has the potential to shift the informational basis of competence judgments away from diffuse status cues and toward observable contributions.
Accordingly, this project asks whether AI-assisted teams exhibit weaker translation of diffuse status characteristics into perceived competence and status attainment than teams that rely on human effort alone. The broader significance of this work lies in advancing theory on status dynamics, inequality, and technology by identifying AI not merely as a productivity tool, but as a social infrastructure that reshapes how advantage is constructed, perceived, and enacted in real time. The findings have implications for organizational design, diversity and inclusion, and the governance of AI-enabled work systems.
Shivani Kumar, Research Fellow, School of Information
As a postdoctoral research fellow at the University of Michigan, I study how AI systems reason about human-centered phenomena, particularly moral values and persuasion, across languages and cultures. I also develop computational frameworks that make these processes more interpretable, structured, and socially grounded. One line of this work examines how AI represents pluralistic moral values [1]. A central contribution is UNIMORAL [5], a novel multilingual, cross-cultural dataset and modeling framework that conceptualizes moral reasoning as a multi-stage cognitive and computational pipeline. Most existing benchmarks assess isolated moral judgments within a single cultural context [4, 8, 7], overlooking the broader reasoning process, such as weighing ethical preferences, emotions, cultural norms, and consequences. UNIMORAL addresses this gap by offering a unified platform to study these stages holistically and pluralistically, enabling a psychologically grounded evaluation of large language models (LLMs), particularly in multilingual and cross-cultural settings. Pluralistic moral values in AI are examined along two complementary strands:
Day 2 – March 31
Parker Howell, Graduate Student Research Assistant, Economics, College of Literature, Science, and the Arts
Jobseekers direct their search toward vacancies they believe are well-matched to their skills and candidacy (Lang et al., 2005), but empirical work reveals that search frictions, discrimination, and imperfect information (especially regarding soft skills), can lead to misguided and inefficient search activity (Conlon et al., 2018; Carranza et al., 2020).
These frictions, combined with a severely biased labor market (Lang and Spitzer, 2020; Hyland et al., 2020), lead to significant misallocation of labor and skills (Moscarini, 2001). Information frictions on skill transferability can further slow the re-matching of workers to available vacancies. Entry-level front-line positions have a high transferability of skills across positions (Gathmann and Sch ̈onberg, 2010), yet this transferability is not salient to jobseekers (Menzio and Shi, 2011). This friction leads workers to search for employment too narrowly and mainly focus on sectors in which they have experience (Belot et al., 2019).
We address these issues via a large-scale randomized controlled trial that recruited 1,117 entry-level jobseekers across seven U.S. metropolitan areas. Participants were randomly assigned to receive job recommendations based on preferences, AI-predicted performance, both, or neither. By embedding AI directly into the job search process and randomizing access to its outputs, the project generates causal evidence on how AI-driven information alters search behavior and labor market outcomes.
Long-Jing Hsu, Schmidt AI in Science Fellow, 2025 Cohort, MIDAS
People living with dementia are resilient, continuing to adapt to adversity through acceptance, lifestyle change, and meaningful engagement after diagnosis. However, this adaptive process often unfolds gradually, can fluctuate from day to day, and typically requires ongoing support from care teams. This project investigates how AI-enabled robots can recognize and support psychological resilience in people living with dementia. Specifically, the project centers on how an AI-enabled robot can listen to and interpret lived experiences, providing supportive, non-directive responses that help individuals remain engaged and purposeful despite cognitive limitations. Conceptually, a user interacts with the robot and may describe a recent challenge, such as forgetting how to prepare a familiar meal. The robot identifies expressions of challenge recognition and adaptive coping and responds with reflective listening and resilience-supporting suggestions rather than task-level correction. To achieve this goal,
the project pursues the following AI-focused aims:
Mainak Sarkar, Assistant Professor of Marketing, U-M Dearborn
Customer lifetime value (CLV) is a cornerstone metric in customer relationship management, yet most models reduce it to a point estimate, implicitly assuming that managers are risk-neutral and indifferent to how that value may materialize (Kumar et al., 2008; Venkatesan and Kumar, 2004). In contrast, we conduct a survey of management consultants from a global strategy firm, managers report many reasons to distinguish between two customers with identical expected CLV but different CLV distributions, citing considerations from firm valuation to sales force efficacy. This paper argues for a distributional view of CLV and develops a flexible dual-LSTM model that jointly predicts customer behavior and firm marketing strategy. The model generates rich CLV distributions that capture self-fulfilling prophecies, diverging paths, and even bimodal CLV distributions (features often missed by traditional approaches). In an empirical application, the dual-LSTM model consistently outperforms seven benchmark models, including Pareto/NBD, regression-based, and state-of-the-art deep learning alternatives. Beyond predictive accuracy, we show that bimodal customers are more responsive to marketing than unimodal customers, and marketing to these bimodal customers can nudge them towards the upper mode of their predicted outcome. Therefore, our model enables firms to prioritize customers not just by expected value but also by the shape of their CLV distribution.
Lu Xian, Graduate Student Research Assistant, School of Information
Privacy policies are often complex. An exception is the two-page standardized notice that U.S. financial institutions must provide under the Gramm-Leach-Bliley Act (GLBA). However, banks now operate websites, mobile apps, and other services that involve complex data sharing practices that require additional privacy notices and do-not-sell opt-outs. We conducted a large-scale analysis of how U.S. banks implement privacy policies and controls in response to GLBA; other federal privacy policy requirements; and the California Consumer Privacy Act (CCPA), a key example for U.S. state privacy laws. We focused on the disclosure and control of a set of especially privacy-invasive practices: third-party data sharing for marketing-related purposes. We collected privacy policies for the 2,067 largest U.S. banks, 45.2% of which provided multiple policies. Across disclosures and controls for the same bank, we identified frequent, concerning inconsistencies—53.8% of banks with multiple privacy policies indicated in GLBA notices that they do not share with third parties but disclosed sharing in other policies. This multiplicity of policies, with the inconsistencies it causes, may create consumer confusion and undermine the transparency goals of the very laws that require them. Our findings call into question whether current policy requirements, such as the GLBA notice, are achieving their intended goals in today’s online banking landscape. We discuss potential avenues for reforming and harmonizing privacy policies and control requirements across federal and state laws.
The project sits at the intersection of privacy, human-centered computing, and public policy. Its significance lies in its empirical findings of prevalent, alarming inconsistencies in consumer-facing legal documents and the implications for privacy law reform. The paper also introduces a rigorously validated AI-assisted research pipeline for large-scale, nuance-sensitive analysis of regulatory texts, grounded in an assessment of current AI limitations and resulting in a high-quality annotated dataset that enables future automated analysis.
Mengqi Lin, Graduate Student Research Assistant, Statistics, College of Literature, Science, and the Arts
My research develops and rigorously evaluates AI methods that automate a central bottleneck in cognitive diagnosis modeling (CDM): estimating the Q-matrix (Tatsuoka 1990), a binary item-by-skill structure specifying which latent attributes each assessment item requires. The Q-matrix is foundational for interpretability and validity in diagnostic assessment (e.g., mastery feedback and targeted instruction), yet it is typically constructed by expert judgment—costly, subjective, and difficult to scale. This project delivers a principled, data-driven pipeline that learns Q-matrices accurately, ensures identifiability by construction, and scales to practical regimes that are out of reach for most existing approaches.