Collaborative Research: CAIG: Artificial Intelligence-Enabled Discovery of Fluid-Rock Interactions in Dense Seismic Arrays

Seismic ways enable monitoring of the Earth's movement and help communities be better prepared againts natural hazards. Temporary dense seismic arrays can record the ground in extraordinary detail, revealing small earthquakes, human-made signals, volcanic unrest, and other subtle vibrations that are often missed by permanent seismic networks. These observations are especially important in places where fluids move through the subsurface, such as geothermal fields, wastewater-disposal regions, and active volcanoes. However, dense arrays are usually deployed for only weeks to months, and the resulting data are too large and complex to analyze fully by hand. This project will develop new artificial intelligence tools that can learn from these massive seismic datasets with limited manual labels, identify meaningful signal patterns, and connect short-term dense-array observations to long-term monitoring records. The work will help scientists better distinguish natural and human-related seismic activity, recognize rare or emerging event types, and improve the long-term value of temporary field experiments. The project will also produce reusable workflows and curated training resources for the research community, while training students and early-career researchers at the intersection of geoscience and machine learning. The project will develop a geometry-aware self-supervised learning framework for temporary dense-array seismic recordings from Oklahoma/Texas and Hawai‘i. The framework will learn representations from paired time-domain and frequency-domain views of array data while accounting for differences in station layout, noise conditions, and deployment geometry. These learned representations will then be linked to interpretable source properties through sequential downstream tasks: source-physics mapping, rare or under-recognized event discovery, and retrieval of similar signals from permanent-network archives. The project will test whether pooled multi-array training can produce a shared representation space that generalizes across regions, separates natural and anthropogenic signals, reduces regional-event location bias, and organizes events into physically meaningful families. For volcanic data from Hawai‘i, the project will examine changes in representation space through time to identify rare event families and possible regime shifts associated with evolving subsurface conditions. Together, these activities will provide a scalable workflow for transforming short-duration dense-array recordings into interpretable, longer-term records of seismic source processes. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.