Using Geo-Embeddings to Search for Geothermal Reservoirs
Before Google Maps, finding a good restaurant meant one of two things: visiting every place on your list until you found one that was open and worth staying for, or rolling the dice and hoping the first door you tried was the right one. Today, Google Maps does that work for you (hours, directions, menu, reviews) before you ever leave the house. I wanted to know if I could do something similar for the ground beneath our feet. Geo-embeddings already help detect changes on Earth's surface from imagery; that's the "Google Maps" layer. The question is whether embeddings can point us toward undiscovered geothermal reservoirs, the "overview" Google Maps hands you before you leave the house.
Why This Is Useful
Finding a new geothermal site today usually means combining imagery data with a lot of observational work: site visits, sensor deployments, sample collection. That's a slow process to iterate across hundreds of candidate locations, and too expensive to target every possible one. Think of this as having to walk into every restaurant on your list just to find out which ones are open. This is where geo-embeddings are useful. They let us run a fast similarity search across hundreds of locations at once, instead of sending geologists to assess each site manually.
How It Works
Geothermal prospecting tool built with LEGEND AI platform
Step 1: Look for geophysical anomalies
Geothermal reservoirs sit near hot rock, fluid, or magma, and that heat leaves a fingerprint near Earth's surface: small but measurable shifts in the local gravity and magnetic field. When enough of that hot material has concentrated in one place, those fingerprints cluster together into a detectable anomaly. Looking for these clusters is a well-established first step in geothermal exploration (Alqahtani et al., 2023), and it's how I built my first list of candidate locations.
Step 2: Check whether the surface has changed
A geophysical anomaly on its own doesn't prove there's an active reservoir underneath it, so the second step looks for another piece of evidence: has the surface changed over time? Radar satellites like Sentinel-1 are well suited to this, and InSAR studies have shown that geothermal activity can leave measurable ground deformation (Mellors et al., 2018). I took Sentinel-1 snapshots of the same patch of ground at two points in time, turned each into an embedding using Clay v1.5 model, and measured the distance between the "before" and "after" embeddings. The larger the distance, the more that patch changed.
One caveat: embedding distance captures any surface change, not just ground movement. Vegetation, soil moisture, snow, and construction can all show up too, so this score flags places worth a closer look rather than confirming deformation. It's also worth noting that much of the ground movement at known geothermal sites comes from existing power production, so an undiscovered reservoir may show only a subtle signal.
Putting It Together
Every candidate cluster gets ranked twice: once for how strong its geophysical signature is, and once for how much surface change its embeddings picked up. Those two rankings are then averaged together, providing a score that separates a shortlist of potential candidates from the many geophysical false positives.
None of this replaces a geologist's expertise and fieldwork. What it does mean is that the first, most expensive question in geothermal exploration, "Where should we look?", can be answered with embeddings in days instead of months. That reduces the fieldwork and resources that would otherwise be required. It's the same help Google Maps gives us when finding a restaurant: we still have to walk in and sit down, but we're no longer guessing which restaurant to try first.