Collaborative Research: CAIG: Machine Learning Accelerated Hydrothermal Modeling For Earth Science

Within faults, volcanoes, and geothermal fields, hot water moves through rock, carrying heat and chemicals. These ‘hydrothermal’ systems drive volcanic unrest, concentrate the metals needed for clean energy, sustain geothermal power, and may have facilitated the emergence of life. They are also notoriously hard to predict. They are mostly underground, and small changes in the rock can produce very different behavior. The flow mixes liquid water and steam, and it spans scales from tiny cracks to entire mountains. The computer models used to study these systems are accurate but slow and limited by the uncertainty of what is happening in real systems. They are slow enough that testing more than a few scenarios is out of reach in most situations. This project will use machine learning to speed these fluid flow models up by factors of thousands while keeping the underlying physics intact. With that speed, researchers will be able to ask questions that were out of reach before. How likely is an eruption, given the changes observed on the ground? Where should the search for critical minerals focus? How much energy can a geothermal field supply? The project will also train students in both Earth science and AI, publish its software publicly, and bring geohazard lessons to K-12 classrooms. The project, HydrotherML, will combine physics-based simulations of hydrothermal flow with a new class of machine learning models designed to respect the symmetries and physical laws built into the governing equations. The team will first generate thousands of simulations covering realistic ranges of rock properties, heat sources, and fluid conditions from the meter to the kilometer scale. They will then train models that reproduce the full simulations at a fraction of the computer time and measure how well those surrogate models perform for new geometries and flow regimes. Because surrogates are fast, it becomes practical to run thousands of cases, measure uncertainty, and search for behaviors that individual simulations miss, such as the feedback between flow and the rock properties that concentrate ore deposits or trigger hazards. The work will also advance machine learning itself by demonstrating how geometric constraints improve accuracy when training data are scarce, a situation common across the sciences. All models, datasets, and software will be released as open-source tools that plug into widely used hydrothermal modeling codes. 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.