Read more about the awarded projects for the 2026 Propelling Original Data Science (PODS) Grants.
Persona-Based AI Respondents for Factorial Survey Experiments: Benchmarking Synthetic Data for Statistical Inference
Henning Silber (Institute for Social Research)
Yajuan Si (Institute for Social Research)
This project tests whether AI-generated respondents, created from researcher-specified profiles, can help researchers analyze and design factorial survey experiments. Using a publicly available factorial survey experiment, we compare AI-generated responses with human responses to assess when AI respondents recover structural patterns, when they fail, and whether they can support estimation in sparse designs. The project will also examine whether AI respondents can help pilot new factors and levels before future human data collection, and will produce validation standards, practical guidance, and reusable code for responsible use in survey research.
Constructing Interpretable Computer Vision Models from Natural Language
Lauren Gillespie (School for Environment and Sustainability)
Gabriel Poesia (College of Engineering)
This project creates a new approach to computer vision for research problems where images are scarce but expert knowledge is abundant in books, field guides, and technical descriptions. Our method POPLAR translates written human expertise into interpretable visual reasoning systems, allowing vision models to make predictions from the kinds of clues experts use rather than from large image datasets alone. Demonstrating with rare plant identification, a high-impact use case for conservation, the approach is designed as a general framework for building data-efficient, interpretable computer vision models across scientific domains.
A Foundation Model for Mode-Agnostic Discovery of Movement Patterns from Molecules to Metazoans
Nils Walter (College of Literature, Science, and the Arts)
Living systems—from molecules inside cells to multicellular animals—show complex movement patterns that current analysis methods often oversimplify or miss, especially when behaviors are rare or unexpected. We will develop a general-purpose artificial intelligence model that can analyze movement across biological scales from the molecular to the macroscopic, and make it accessible to researchers across disciplines through a user-friendly web platform. This project aims to enable the systematic discovery of nontrivial movement patterns in biology, generating new insights for both basic research and translational medicine.
Counterfactual Causal Knowledge Graphs (C-CKGs): Neurosymbolic Framework for Root Cause Analysis and Actionable Intervention in Smart Manufacturing, Molecular Pathway and Copper Nanoparticle Discovery
Utkarshani Jaimini (UM-Dearborn)
Krisanu Bandyopadhyay (UM-Dearborn)
Kalyan Kondapalli (UM-Dearborn)
Suvranta Tripathy (UM-Dearborn)
When a manufacturing machine produces a defective part, or a stroke patient’s neurons fail to recover after blood flow is restored, or a batch of copper nanoparticle catalysts underperforms due to uncontrolled synthesis variability, the AI systems today can predict that something is wrong but cannot identify the specific upstream cause or recommend the minimal intervention that would have prevented it. This proposal develops Counterfactual Causal Knowledge Graphs (C-CKGs), a new Neurosymbolic AI framework that combines domain knowledge with machine learning to identify root causes and produce precise, actionable interventions across three distinct scientific domains.
INDEX: An Evidence-Linked, Human-in-the-Loop AI Methodology for Harmonizing Clinical Research Data Across Institutions
Florian Schmitzberger (Michigan Medicine)
Alex Janke (Michigan Medicine)
Cong Ma (Michigan Medicine)
Clinical research teams spend extensive time reading medical records to complete case report forms, which slows studies and creates inconsistency across sites. INDEX will develop and test a human-supervised AI workflow that combines structured EHR data and clinical notes to automatically generate answers for the form alongside verified raw evidence from the unstructured notes to help researchers complete forms faster and more consistently. The project will produce reusable software, benchmarks, and guidance that U-M and partner institutions can use to make clinical research data more scalable, auditable, and comparable.
Interpretable AI for Discovering Hidden Collective States in Programmable Active Matter
Steven Ceron (College of Engineering)
Y Z (College of Engineering)
Natural and engineered collective systems can spontaneously reorganize into different states, but predicting when and how these transitions happen remains an open problem because the relevant hidden variables are unknown beforehand and impossible to list out by hand. This project pairs a unique large-scale electromagnet array that can drive and observe distinct collective behaviors in magnetic particulate matter with a topological data analysis framework that automatically discovers the hidden organizing variables, maps out metastable states, and predicts rare transition events that conventional analysis would miss. By closing the loop between AI-generated predictions and real physical experiments, our team will establish a reusable, open-source methodology for AI-driven state discovery in complex collective systems such that it can be adopted across active matter, materials science, and multi-robot systems across all length scales.
AI-enabled Expansion of the Research Frontier
Zhongming Liu (College of Engineering, Michigan Medicine)
Scott Peltier (Michigan Medicine)
This project leverages cutting-edge artificial intelligence and machine learning to map and decode how the brain interacts with the body’s major organs, such as the heart, lungs, and gut, and to reveal how this brain-body interaction shapes emotion and cognition. The scientific findings and research tools delivered from this project will help scientists and clinicians better understand and ultimately treat mental illnesses arising from disrupted brain-body interactions.
Beyond Simulation Echo Chambers: Training Reality-Grounded AI Flow Models from Synthetic X-Ray Radiographs
Jeong Joon Park (College of Engineering)
Ricardo Vinuesa (College of Engineering)
AI models for fluid dynamics are typically trained on computer simulations, meaning they inherit the same errors and blind spots as the simulations themselves. This project develops RADIAN, a new AI framework that learns to reconstruct detailed, physically accurate fluid flow from X-ray images of tracer particles moving through metal structures that conventional cameras cannot see inside. By grounding AI training in measurement-based reconstructions rather than simulations alone, RADIAN aims to accelerate the design of next-generation heat exchangers and energy systems.