From Monitoring to Exploration: Hunting Lithium Brine Seeps from Geo-Embeddings
In a previous post, Monitoring Lithium Fields with Geo-Embeddings, we built a detector for new evaporation ponds at Salar de Atacama. Its biggest false positive wasn't a pond at all. It was roughly 270 hectares of open salt flat, where moisture and crust had shifted enough between 2020 and 2024 to look like construction.
For a monitoring tool, that's noise. For exploration, it may be the signal. Wet ground and fresh salt crust on a salt flat often mark the places where brine reaches the surface, and brine is where lithium is. So we turned the question around: instead of filtering out natural surface change, what if we went looking for it?
Why brine seeps point to lithium
Lithium comes from two kinds of deposits: hard rock (pegmatites) and brine. This post is about brine. The world's richest lithium brines sit in hyperarid salars, closed desert basins in places like the Chilean and Argentine Andes. Published work at Laguna Verde in the Chilean Andes (Alam & Muñoz, 2021) traces a clear enrichment pathway:
- Hot, acidic geothermal fluids leach lithium from volcanic source rocks far more effectively than ordinary groundwater.
- Those fluids travel more than 25 km underground before discharging as hot springs — thermal waters there carry roughly 2.8–4.5 mg/L of lithium, about five to seven times higher than non-thermal water nearby.
- In a closed basin where evaporation far outpaces rainfall, that lithium-rich brine has nowhere to go. It concentrates, over time.
The practical takeaway: where geothermal brine discharges at the surface — the seeps, springs, and salt-crusted vents — lithium enrichment is more likely to follow. Find the seeps, and you have a shortlist of places worth exploring.
What satellites can see, and what fools them
From orbit you never measure lithium grade. You see a proxy: a surface expression, such as a crust, a damp patch or a warm spring, that often accompanies something you can't see. Satellites don't see ore. They see the company ore keeps.
From orbit, a brine seep looks a lot like last week's rain. Brine discharge, freshwater springs, rainfall, and temporary wet ground all produce similar signals in satellite imagery, so no single sensor or spectral band reliably separates them. And the alternative — sending geologists to screen thousands of square kilometers on foot — doesn't scale to exploration timelines.
The key discriminant is persistence. True brine discharge stays wet when everything around it is dry, and it leaves a calling card: salt crust and mineral efflorescence at the vent. That crust is the most durable clue of all, because the minerals stay behind after the water evaporates. They record brine activity that may not be visible on any single date.
Start from a description, not a label
The pond detector in our last post started from 50 hand-placed labels. That works when you already know what you're looking for. Exploration doesn't have that luxury: the whole point is that nobody has mapped these seeps yet.
So this search starts from a description instead. We write what a seep looks like in plain language, for example "white salt crust or halite efflorescence on dry desert surface near wet patch", and search the embeddings for the image chips that best match. Each chip is a ~330 m square. Several descriptions run side by side, one per signature, and the results are ranked. Labels can come later, once there are candidates worth labeling.
A multi-tier search instead of a single detector
Rather than betting on one indicator, the search is designed as two tiers and a synthesis step that narrow a large salar region into a short, defensible list. Tier 1 finds where seeps look plausible; Tier 2 checks when those signatures persist.
Tier 1: Spatial & Spectral search: The Wide Net
Several searches run in parallel, each targeting a different surface expression that brine seeps leave behind:
Tier 2
Temporal Persistence: The Noise Filter
Some signatures are stable on their own: salt crust marks long-lived brine activity, and brine-pool morphology is a structural clue. Others — moisture, thermal signals, bare ground — vary with season and weather. For those, we compare imagery across multiple dates and keep only locations whose seep-like expression persists while the surrounding terrain is dry. One-off rainfall and seasonal wet patches drop out. This is also where exploration differs from monitoring. The pond detector compared two yearly snapshots, which is enough to spot new construction. A seep has to prove itself through dry seasons, so the persistence test needs multiple dates within and across years.
Synthesis & Ranking
Candidates that pass both tiers (Tier 1 ∩ Tier 2) are merged and ranked. The highest scores go to sites that match multiple indicators, persist over time, and sit in geologically plausible settings — salar margins, discharge zones, hyperarid closed basins.
Proof of concept: Salt Crust as a Brine-seep proxy across the Andes
Where the pond detector worked inside a single 780 km² box at Salar de Atacama, this test ran regionally. We put the first signature, salt crust and efflorescence as a brine seep proxy, to the test across hyperarid terrain in the Puna Plateau of Argentina and the Atacama Desert of Chile. The initial run screened 70-plus locations and surfaced 26 priority candidates, with top matches scoring 0.61–0.64 against the target signature.
Just as important as the number of hits is where they landed. If the search were picking up noise, the candidates would be scattered randomly across the desert. Instead, [X of 26] sit at the edges of salt flats — the same settings where groundwater and brine are known to reach the surface. The results match what the geology predicts, which is a first sign the proxy is detecting something real.
Screening, not proof — and what's next
This search proposes where the surface evidence is strongest, in rank order, and leaves the verdict to chemistry, hydrogeology and the drill. Rain-wetted ground, freshwater springs and barren pans will still slip through. The difference is that a field team rejects them from a shortlist instead of searching a continent for them.
Salt crust is one of six signatures of Brine Seep, and this test ran a single tier. Next come the persistence filter, the thermal and vegetation signatures, and searches aimed at specific target basins. We'll also run the pond detector from Part 1 first, so evaporation ponds are masked out before anything is ranked. The two searches fit together: monitoring shows where lithium is produced today, and exploration suggests where it might come from next.
Geo-embeddings shrink months of searching into an instant answer by handing field teams exact target coordinates before the engine even turns over.