On the Moon, shadows are absolute. They don’t diffuse or soften like they do on Earth — they fall like a blade. When sunrise hits the volcanic plains of the Marius Hills, every ridge and crater snaps into razor sharp contrast. But some shadows never lift. They pool inside perfectly round holes in the lunar crust, plunging straight down into darkness. For decades, scientists suspected these were collapsed lava tubes — ancient tunnels wide enough to shelter cities. But it wasn’t until AI found hidden caves on the Moon, combing through orbital imagery pixel by pixel, that the scale of those voids became undeniable. What looked at first like simple pits turned out to be entrances to enormous, cathedral-sized caverns — places where light has not touched in billions of years.
One of those voids, scientists now believe, may lead to an enormous cavern where humans could one day build their first neighborhoods off Earth.
And it was discovered not by an astronaut, or a telescope, or a rover — but by an artificial intelligence model, scanning millions of pixels for something humans had overlooked.
This is the story of how a machine learning system named ESSA (Entrances to Sub-Surface Areas) found new candidate cave entrances on the Moon — and why the discovery may change humanity’s entire relationship with our nearest celestial neighbor.
The Problem No One Could Solve by Hand
For decades, scientists have suspected that the Moon holds something extraordinary under its surface: lava tubes — long, hollow tunnels carved billions of years ago when the Moon was volcanic. These tubes, according to planetary geologists, could be so large that some would make the Roman Colosseum feel like a broom closet. Temperature inside them is believed to remain stable, hovering close to -20°C, far gentler than the deadly 127°C to -173°C swings on the surface. The thick roof of rock would block out solar radiation, cosmic rays, and the constant peppering of micrometeorites that strike the surface like sand in a windstorm.
In other words — natural shelters already built and waiting.
But knowing they might exist and knowing where the entrances are are two very different things.
NASA’s Lunar Reconnaissance Orbiter has taken millions of high-resolution photographs of the Moon’s surface over the past 15 years. Each image is rich with craters, cracks, shadows, and dust patterns. Hidden somewhere among them are the skylights — the holes where the roof of a lava tube has collapsed, revealing the tunnel below.
But no human could scan every photo. Even a team of researchers working constantly for decades would barely scratch the surface.
So for years, the Moon’s hidden underground world remained mostly theoretical — suspected, but largely unmapped.
Enter ESSA — A Machine That Looks for Doors
In 2025, a PhD researcher named Daniel Le Corre from the University of Kent set out to solve the problem in a different way.
Instead of staring at images manually or using classical algorithms (which had produced only modest success — finding 16 known pits total), Le Corre turned to deep learning — specifically a type of image recognition system that could learn to notice fine details the way a human might, but tirelessly, across thousands of gigabytes of data.
ESSA was built using a neural network architecture known as Mask R-CNN, paired with a ResNet50 backbone — a system that excels at identifying objects in complicated visual environments.
But here was the catch: there are very few confirmed lunar pit images to train on. So the team supplemented the model with:
- Known lunar pit images from the Lunar Reconnaissance Orbiter
- Martian pit images from Mars Reconnaissance Orbiter missions
- Synthetic images of potential skylight shapes
- Negative examples (false craters, crater shadows, and lighting artifacts)
The idea was to teach ESSA not only what a skylight is — but what it is not.
After training, ESSA was fed raw, full-resolution imagery from the Lunar Reconnaissance Orbiter’s Narrow-Angle Camera (NAC). This wasn’t the clean, pre-processed mosaic imagery used by casual researchers. It was the hard data — the kind a human would struggle to search in bulk.
The AI began scanning.
And scanning.
And scanning.
Then it highlighted something.
and then something else.
Then, after days of computation, it stopped.
Among tens of thousands of images, ESSA flagged two features that had never been recorded before.
The South Marius Hills Pit (SMHP)
The first location sits in a region called the Marius Hills — one of the most complex volcanic landscapes on the Moon. The area is spotted with ancient lava domes, volcanic cones, and collapsed tunnels, frozen in time like a petrified river system.
In the images ESSA reviewed, a dark circular opening appeared. Not the wide, shallow bowl of an impact crater, but a sharp-edged, steep-sided hole — the hallmark of a skylight. Measurements showed it was 50 to 80 meters across, large enough for a helicopter to fly through if there were air.
Even more compelling: it lies only 4.3 kilometers from a known linear collapse feature — a mark on the surface where a lava tube is believed to run below.
It was as close to a signpost as the Moon is capable of offering.
A doorway.
The Bel’kovich A Pit (BAP)
The second pit was even more intriguing.
It lies near the Moon’s north polar region, inside a lava-flooded impact crater known as Bel’kovich A. The poles of the Moon are unique — because of the way sunlight strikes them, some areas are in permanent shadow, never warmed by the Sun. These shadowed pockets are believed to contain water ice — one of the most valuable resources in space.
If astronauts could reach ice inside or near this pit, they could:
- Drink it
- Breathe it (after splitting it into oxygen and hydrogen)
- And even use it to make rocket fuel
Water on the Moon isn’t just life support — it’s the key to a transportation system between worlds.
Le Corre’s team didn’t claim to have proven ice is present. But a cave entrance in a polar environment immediately raises the stakes.
This wasn’t just about shelter anymore.
This was about self-sustaining settlement.
“We’ve barely looked yet.” — The Bigger Meaning
Here is the part of the discovery that stunned researchers most:
ESSA found these two candidates while scanning only 0.23% of the Moon’s surface — less than one quarter of one percent.
If two major skylights appear in that tiny fraction, then:
There could be dozens. Maybe hundreds.
And every one of those could lead to a space large enough to build:
- Underground research stations
- Hydroponic farms
- Storage facilities
- Crewed habitats
- Even small cities
We are not talking about bubbles or tents on the surface anymore.
We are talking about real architecture.
On another world.
The Road to Proving What’s Below
Right now, ESSA has identified entrances. What lies beneath them must be confirmed with instruments such as:
- Ground-penetrating radar on lunar rovers
- Miniature descent drones lowered into pits
- High-resolution thermography imaging
- Possibly a “cave hopper” robot capable of jumping across lunar floors
International missions already in development — such as Japan and India’s LUPEX, or NASA’s planned VIPER rover — could be redirected or expanded to survey these sites.
Robots go first.
Humans follow only after the voids are mapped.
What This Means for Our Future on the Moon
If these caves are deep, stable, and interconnected, then this AI discovery does not just reveal two holes in the ground.
It identifies the starting points of the first permanent human settlement beyond Earth.
Changes mission designs.
Landing site priorities move.
It changes how we think about living in space.
Surface bases were always temporary.
Underground is the future.
The Giant Leap That Goes Down Instead of Up
The Apollo era imagined exploration as a step outward — flags planted, boots on regolith, footprints on open plains.
But the next giant leap may be different.
It may be the first human climbing down — into shelter carved by ancient fire, preserved by cold, revealed by algorithms — to switch on a light inside the Moon.
Humanity’s future off Earth may not begin by reaching upward.
It may begin by descending into the dark.
Into a doorway our species couldn’t find on its own.
A doorway a machine found for us.
FAQ: How AI Found Hidden Caves on the Moon
Q1: What did the AI discover on the Moon?
A1: The AI system ESSA identified previously unknown lunar cave entrances, known as skylights, which could lead to lava tubes suitable for human habitats.
Q2: Why are lunar caves important?
A2: These caves offer natural shelter from extreme temperatures, radiation, and micrometeorites, making them ideal for future Moon settlements.
Q3: How did the AI find hidden caves on the Moon?
A3: ESSA analyzed millions of high-resolution images from NASA’s Lunar Reconnaissance Orbiter, learning to detect skylights with deep learning techniques.
Q4: Could humans live in these caves?
A4: Potentially, yes. If the caves are deep and stable, they could support research stations, habitats, and even small cities on the Moon.
Q5: What’s next for exploring these caves?
A5: Robots and instruments like ground-penetrating radar and mini drones will map the interiors before any human missions enter them.