At the forefront of AI-in-science innovation, MIDAS’s Schmidt AI in Science Postdoctoral Fellows are redefining the way artificial intelligence is applied across diverse research fields. Their cutting-edge research includes developing AI-driven techniques to uncover hidden patterns in complex genomic data, offering insights into large-scale biological processes like mass extinction events, developing methodologies for discovery and optimization of engineering processes and materials, and more. By bridging traditional research domains with AI-centric study, they enhance our ability to predict natural disasters, optimize aerospace engineering, and improve ecological monitoring. Their interdisciplinary collaborations and innovative solutions emphasize the potential of AI to drive scientific discovery and problem-solving.
Biological Science Research
Schmidt AI Fellow Jacob Berv was highlighted in a Michigan News article for his research on end-Cretaceous mass extinction events. His research tracks how one such event led to drastic changes in bird genomes, contributing to the diversity of over 10,000 bird species today. Berv’s research, published in Science Advances, utilized AI-driven methodologies to analyze genomic data, identifying “genomic fossils” in bird DNA that indicate evolutionary shifts after the extinction event. The study highlighted how changes in DNA composition correlated with shifts in bird traits, such as reduced body size and altered development patterns in hatchlings. This innovative application of AI underscores the enormous potential AI-in-science research has for innovatively analyzing large-scale datasets.
Accordion content.
James Boyko was the lead author of a paper, published in the New Phytologist, exploring how flowering plants adapt to different climates. He applied hOUwie, which models the joint evolution of continuous traits, like temperature tolerance, and discrete traits, such as whether a plant is annual or perennial. This approach enabled a comprehensive analysis across 32 different groups of flowering plants worldwide. Boyko and his collaborators uncovered significant patterns in plant response to climate factors. They found that annual plants are particularly adapted to hot regions because they can endure harsh seasons in the resilient form of seeds. This sophisticated application of AI provides deeper insights into plant adaptation over time, highlighting broad evolutionary trends and unique behaviors across various plant groups.
Schmidt AI in Science Fellow Nathan Fox leverages computer vision to detect invasive species through social media images, providing a novel tool for ecological monitoring. Traditional biodiversity surveys can be slow and resource-intensive, but social media platforms like Flickr and iNaturalist contain vast amounts of user-generated images that, if properly analyzed, could help scientists track the spread of invasive species in near real-time.
Fox’s research applies foundation vision models, including BioCLIP, an AI model trained on over 450,000 species, to identify invasive species in publicly available images. His work focuses on developing an automated pipeline that retrieves social media photos tagged with species names, validates their identity using AI, and maps their occurrence across different regions. By integrating AI-driven species identification with geographic data, Fox and his collaborators aim to enhance early detection of invasive species such as the green iguana (Iguana iguana) or lionfish (Pterois volitans), which pose significant threats to ecosystems worldwide.
This research was recently published as Identifying invasive species sightings from GeoAI-validated social media posts through I-GUIDE, the NSF-funded Institute for Geospatial Understanding through an Integrative Discovery Environment. The study analyzed 23,000 Flickr images tagged with 72 invasive species names across the U.S. Using BioCLIP, Fox and co-author Derek Van Berkel assessed whether these images genuinely contained the species users identified. While some species, such as the European Starling, showed high validation rates, others, like the Alewife, had fewer confirmed matches. Fox’s work also explores the possibility of passive monitoring, using AI to identify species in images without relying on user-provided tags. By making these methods openly available through I-GUIDE, Fox’s work provides ecologists and land managers with scalable, AI-powered tools for real-time invasive species monitoring.
As a Schmidt AI in Science Fellow, Matthew Andres Moreno develops new methods for large-scale simulation of biological evolution, seeking to open new avenues to explore the origins of adaptation, novelty, and complexity in nature. In collaboration with the Neocortex team at the Pittsburgh Supercomputing Center, Moreno conducts agent-based evolution simulations using the Cerebras Wafer-Scale Engine–a recently-introduced 850,000-core computer chip targeted to AI/ML applications. By designing new decentralized algorithms for lineage tracking and engineering high-performance data analysis workflows, Moreno’s work has opened the door to simulation experiments tracing the ancestry of billions of organisms across evolutionary histories encompassing quadrillions of replications. Moreno is mentored by Dr. Luis Zaman at University of Michigan and by Dr. Emily Dolson at Michigan State University.
Earth and Environmental Science Research
Three Schmidt AI in Science Fellows–Christin Salley, Nathan Fox, and Alyssa Schubert–took part in a MIDAS Carpentry focusing on Natural Language Processing. In this regularly recurring AI-based research and work group, the three Fellows researched and applied Natural Language Processing methodologies, leading to a conference publication, “Tweeting Through the Flood: Application of BERT Topic Modeling for a Comparative Flood Communication Analysis,” for the 2024 Information Systems for Crisis Response and Management Global Conference.
Schmidt AI in Science Fellow Xin Wei‘s research centers on assessing natural hazards, with a particular emphasis on geohazards such as landslides. He investigates historical losses and community impacts and develops innovative risk assessment tools to mitigate future disasters. His work integrates geotechnical engineering expertise with advanced AI techniques through interdisciplinary approaches that include remote sensing, geospatial modeling, and both data-driven and physics-based modeling. To evaluate past hazard impacts, Xin leverages state-of-the-art natural language processing to analyze diverse data sources, ranging from traditional sources (for example, government geological survey reports) to non-traditional ones (such as social media). For future risk prediction and mitigation, he utilizes remote sensing data (such as optical imagery) to detect existing hazards and identify areas at risk, while integrating physics-based methods with reliability theory to assess potential consequences.
He collaborates with experts across multiple disciplines. Professor Sabine Loos, from the Civil and Environmental Engineering Department, contributes her expertise in geospatial modeling, risk analysis, and user-centered design techniques to develop tools that inform effective and equitable disaster risk reduction and recovery; Professor F. Estéfan T. Garcia, also from the Civil and Environmental Engineering Department, offers insights in geotechnical engineering and advanced numerical modeling; and Professor Elizabeth Bondi-Kelly, from the Electrical Engineering and Computer Science Department, brings specialized knowledge in computer vision, deep learning, remote sensing, and human-AI collaboration. Together, these collaborations drive research that bridges engineering and artificial intelligence to enhance our understanding and mitigation of natural hazards.
Engineering Research
Schmidt AI in Science fellow Elena Shrestha, along with Derrick Yeo, lecturer in robotics, lead research in coordination with the Center for Autonomous Air Mobility and Sensing (CAAMS), a partnership between academia, industry, and government. Their work aims to improve guidance and control techniques for multi-mode unmanned aircraft systems by combining well-established aerodynamic principles with learning-based algorithms for inflight flowfield estimation.
https://robotics.umich.edu/2023/u-m-partners-with-autonomous-air-mobility-center
Schmidt AI in Science Fellow Zheng Guo leverages program synthesis to accelerate scientific computing, with a particular focus on optimizing high-dimensional tensor approximations and computational kernels used across computational science. By automating the search for optimal network structures and contraction orders under diverse optimization objectives, his research enhances the efficiency and accuracy of numerical simulations—an essential component in fields ranging from fluid dynamics to quantum mechanics. Recently, he developed a method that is 10 times faster than state-of-the-art structure search techniques while achieving 1.5x–3x better compression ratios. In the future, this work will be integrated with partial differential equation (PDE) solvers to enhance their performance.
This ongoing research is conducted in collaboration with Dr. Alex Gorodetsky from aerospace engineering, Dr. Brian Kiedrowski from nuclear engineering, and Dr. Xinyu Wang from computer science.
.
Schmidt AI in Science Fellow Haotian Chen, along with his Science mentor Prof. Venkat Viswanathan and AI mentor Prof. Alexander Rodríguez are building the first differentiable electrochemistry framework to bridge the theory-experimental gap in electrochemistry. Electrochemistry simulations are made end-to-end differentiable to obtain gradients of physical processes for learning and optimization. More importantly, differentiable electrochemistry enables first-principles discovery of essential parameters like rate constants, reorganization energy, or transfer coefficients from experiments, to probe kinetics, understand mechanics, and guide experimental design by optimizing these parameters.
Haotian has developed differentiable simulations for all major modes of mass transport: diffusion, migration and convection and both the macroscopic Butler-Volmer kinetics and the microscopic Marcus-Hush-Chidsey kinetics. In addition, it was compatible with coupled nonlinear electrochemical reactions (EC or CE reactions), and both semi-infinite and thin-layer boundary conditions. Compared with data-driven machine learning, differentiable simulations are significantly more efficient, interpretable and accountable. Initial application shows 100x training speedup compared with a neural network surrogate model. Haotian and his mentors believe that differentiable simulation will be the new paradigm of scientific machine learning.
Haotian is actively revising his manuscript with his mentors and will be released on a preprint server in due course. Working as a team, they are also actively exploiting opportunities to build a more accurate and universal battery management system for electric aviation.
.
Physical Science Research
Kevin Napier and Vital Fernández, both Schmidt AI in Science Fellows, participated in a Carpentry on Reinforcement Learning and are part of a group addressing the challenge of optimizing the scheduling of telescope tasks. Astronomers worldwide depend on a limited number of extremely large telescopes for their research. Due to the vast number of compelling questions in astronomy and the scarcity of available telescopes, telescope time is a highly valuable resource. Many research telescopes attempt to maximize their utility by operating in a queue mode, striving to handle a list of tasks as efficiently as possible. However, determining the most efficient order of operations presents a significant challenge due to real-time obstacles such as adverse weather, atmospheric turbulence, or cloud cover, which can alter the optimal solution.
Napier and Fernández’s team is tackling the problem of optimal queue management for telescopes. They are utilizing reinforcement learning algorithms combined with a telescope simulator to devise an optimal policy that accounts for real-time sky conditions. Their project, called “roboqueue,” aims to enable astronomers to conduct more scientific research with the limited time available on large telescopes.
Schmidt AI in Science Fellow Vital Gutiérrez Fernández is applying machine learning to the chemical and dynamical analysis of the light from star-forming regions.
When stars are born, especially massive and young ones, they emit intense light and energy. This UV light is so powerful that it strips electrons from atoms in the surrounding gas, a process called ionization. The photons produced in these ions as the electrons move back to their orbits have very distinctive wavelengths, which produce the wonderful colors seen in astronomical photometry. Moreover, these emissions are so intense that they can be observed across cosmological ages, up to the very first galaxies over 13 billion years ago.
Due to the fact that each element emits light at unique wavelengths, we can measure the chemical composition of this gas. For example, in the local universe, ionized hydrogen is responsible for red colors, while oxygen ions produce blue and red photons. Moreover, as in the case of sound, the light’s shade changes because of the motion of the object: if the gas is moving toward us, the emission lines shift to shorter wavelengths. In the opposite case, the wavelengths become longer as the object is moving away from us.
Thanks to the new generation of astronomical telescopes and instruments, the field has entered a new age of Big Data. In Vital’s research, he is developing machine learning classifiers to automatically distinguish these shapes in the photon distributions to interpret the observations as if they were done by a human. This work is being used by state-of-the-art surveys with the James Webb Space Telescope to diagnose the quality of the data and provide new scientific insights.
Postdoc Accomplishments
Schmidt AI in Science Postdoc Matthew Moreno was among the recipients of the Ecology and Evolutionary Biology Department Justice, Equity, Diversity and Inclusion (JEDI) Award. Matthew is part of the bI/O program, led by Abrianna Soule and faculty in the Department of Ecology and Evolutionary Biology, which aims to bridge academic and incarcerated communities by offering educational science presentations at the Parnall Correctional Facility. This initiative featured researchers presenting concise, engaging talks followed by Q&A sessions, fostering communication skills and educational opportunities for inmates.
AI in Science fellows Nathan Fox and James Boyko hosted a four-day workshop focused on integrating multimodal and crowd-sourced data for biodiversity monitoring and conservation. In addition to research talks, participants engaged in hand-on data challenges designed to tackle a diverse range of real-world data complexities, from exploring bird species identification, comparing human versus AI labels, to navigating the intricacies of crowd-sourced data from platforms such as iNaturalist, Flickr, and Twitter.
AI in Science fellow Kevin Napier was awarded the University of Michigan ProQuest Distinguished Dissertation Award for his dissertation “Novel Methods of Detecting and Characterizing Solar System Objects.” Read more here.