Mars has a lot more diversity because of its history and geology
For decades, Mars has been observed in extraordinary detail by orbiting instruments that speak in incompatible tongues — different sensors, wavelengths, and formats describing the same world without a common language. A doctoral student at Arizona State University, Mirali Purohit, has answered that fragmentation with MOMO, the first foundation AI model built specifically for the Red Planet, trained on 12 million Martian images assembled from missions spanning generations. Where previous models carried the assumptions of Earth — its forests, its cities, its familiar geometries — MOMO was taught to see Mars as Mars, and in doing so, opens a new chapter in how humanity reads the geological memory of another world.
- Decades of orbital data from Mars have accumulated into an archive so vast and incompatible that scientists could not effectively use it — existing AI models, trained on Earth or everyday objects, simply could not translate what they saw.
- The absence of mature data infrastructure for Mars meant Purohit had to build pipelines largely from scratch, filtering 40 million raw samples down to 12 million validated images through painstaking manual and automated cleaning.
- Rather than forcing all Martian imagery into a single format, the team trained separate models on different sensor types and merged them — creating a system flexible enough to move from pixel-scale boulders to continent-spanning landslides.
- MOMO now outperforms previous approaches across benchmarks, detecting craters, frost, landslides, and boulders, and helping scientists reconstruct Mars' geological history at a scale previously impossible.
- The Kerner Lab plans to release both the model and its 12 million training images publicly, inviting researchers worldwide to study Mars without rebuilding the tools that made it possible.
Mars has been watched for decades by orbiting instruments — high resolution, infrared, thermal — each capturing the planet in a different language. The result is a staggering archive of incompatible data: same world, no shared vocabulary. Into this problem stepped Mirali Purohit, a doctoral student in computer science at Arizona State University who arrived in 2022 with a clear ambition: to make sense of Mars.
Working in the Kerner Lab under AI researcher Hannah Kerner, Purohit developed MOMO — the Mars Orbital Model — the first foundation AI model ever built specifically for the Red Planet. The challenge was not just scale but infrastructure. Unlike Earth observation, which benefits from mature pipelines and established tools, Mars research runs on scattered archives and improvised systems. Purohit built much of the data pipeline herself, starting with 40 million samples and refining them down to 12 million high-quality images through extensive validation.
Rather than forcing all data into one format, the team trained separate models on different imagery types and merged them into a single unified system. The result is something like a general-purpose mind for Mars — capable of moving fluidly between scales, detecting craters, mapping landslides, identifying frost, and spotting boulders. It learns from the planet as a whole, not just one region, which matters because Mars is far more geologically diverse than its barren reputation suggests.
Across benchmarks, MOMO consistently outperforms earlier approaches, particularly on detailed surface mapping. Beyond detection, it helps scientists piece together Mars' geological history — potentially revealing traces of past water or life. The Kerner Lab plans to release both the model and its training images publicly, lowering barriers for researchers everywhere.
For Purohit, this is a beginning. She wants to connect orbital data with rover imagery, stitching large-scale planetary views to the small patches explored on the ground. She is preparing to defend her dissertation this summer and hopes to continue the work as a postdoctoral researcher — and perhaps, one day, much closer to the planet itself.
Mars sits in our telescopes drowning in its own data. For decades, orbiters have circled the planet, their cameras clicking away—high resolution, low resolution, infrared, thermal. They've captured everything from the texture of individual rocks to continent-spanning landscapes. The result is a planetary archive of staggering size and complete incompatibility. Different sensors. Different wavelengths. Different formats. All of it describing the same world in languages that don't speak to each other.
Mirali Purohit arrived at Arizona State University in the fall of 2022 knowing exactly what she wanted to do: make sense of Mars. The doctoral student in computer science joined the Kerner Lab, where she works under Hannah Kerner, a Schmidt Sciences AI2050 Early Career Fellow focused on artificial intelligence built to serve the public good. Purohit wanted to work in planetary science, to look beyond Earth. "If we can explore the moon and really see Mars, we can determine what is actually happening there," she says. That ambition would lead to something unprecedented: MOMO, the Mars Orbital Model, the first foundation model ever built specifically for the Red Planet.
The problem MOMO solves is concrete and urgent. Until now, scientists faced an impossible choice. They could adapt AI models trained on everyday objects—cats, dogs, chairs, tables—or they could use models built for Earth imagery, where forests and oceans and cities dominate the training data. Both approaches failed because Mars data is fundamentally alien to these systems. The models couldn't transfer what they'd learned. Custom-built models, meanwhile, were slow and expensive. What was needed was a single system that could handle it all.
Purohit and her team trained MOMO on roughly 12 million Mars images, assembled painstakingly from multiple instruments and missions spanning decades. The scale alone was daunting. But the real work lay in the assembly. Unlike Earth observation, which has mature pipelines and established software infrastructure, Mars research still runs on scattered archives and ad hoc systems. "We realized that we don't have the infrastructure for Mars that we have for Earth observation," Purohit says. "We were lacking pipelines, libraries and packages." She handled much of the work herself, starting with 40 million samples and filtering them down to 12 million high-quality images through extensive cleaning and validation.
What emerged is something closer to a general-purpose brain for Mars. Rather than forcing all data into a single format, the team trained separate models on different types of imagery, letting each learn its own representation. Then they merged them into one unified system. The result is a model that moves fluidly between scales—from identifying tiny boulders to mapping vast geological features like landslides. That flexibility matters because Mars, despite its reputation as a barren desert, is surprisingly complex. "We tend to think of Mars as blank, but it has a lot more diversity because of its history and geology," Purohit says. In one region, cone-shaped features might signal past water activity. A few kilometers away, those same features look entirely different. Models trained on one region often fail in another. MOMO learns from the planet as a whole.
Feed it an image and it can detect craters, map landslides, identify frost, spot boulders. Some tasks are straightforward—atmospheric dust is easy to find. Others, like picking out tiny, pixel-scale boulders, still push the limits. But across benchmarks, MOMO consistently outperforms previous approaches, especially on detailed surface mapping. It doesn't just see Mars. By capturing features across the entire planet, it helps scientists piece together Mars' geological history, possibly revealing signs of past water or even life.
The ambition extends beyond a single model or even a single planet. By turning massive, fragmented datasets into something scientists can effectively use, MOMO points toward a future where planetary science happens at scale, accelerating discovery across worlds. The Kerner Lab plans to release not only the model but also the 12 million high-quality images it was trained on, lowering the barrier for researchers everywhere to study Mars without building tools from scratch.
For Purohit, this is only the beginning. She wants to connect orbital data with rover imagery, stitching together large-scale views of Mars with the tiny patches explored on the ground. She's preparing to defend her doctoral dissertation this summer and will likely continue the work as a postdoctoral researcher. Long term, she wants to take models like MOMO out of the lab and into the real world, where they can continuously process data, adapt, and improve. And if the opportunity ever arises to see Mars up close? "Oh yeah," she says, laughing. "My answer is yes. I would go. Why would I not?"
Citações Notáveis
If we can explore the moon and really see Mars, we can determine what is actually happening there.— Mirali Purohit, doctoral student
We realized that we don't have the infrastructure for Mars that we have for Earth observation, and we were lacking pipelines, libraries and packages.— Mirali Purohit
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that Mars data is so fragmented? Can't scientists just use the images they have?
Because fragmentation means incompatibility. A thermal image and a high-resolution photograph of the same crater are describing different things in different languages. Scientists can't easily compare them or build understanding across the whole planet. They're stuck working in silos.
So MOMO solves that by training on all of it at once?
Not quite. It trains separate models on different imagery types, then merges them. That way each type of data gets to be itself, but they all speak to each other through a unified system. It's like teaching someone to read multiple languages and then having them translate between them.
Why couldn't they just use AI models that already exist, trained on Earth data?
Mars is too different. Earth models learn from forests, oceans, cities. Mars has none of that. The features are alien—different geology, different atmospheric conditions, different scales of erosion. A model trained on Earth can't recognize what it's seeing on Mars.
What does it mean that they're releasing the 12 million images publicly?
It means other researchers don't have to start from zero. Right now, if you want to study Mars, you have to hunt through scattered archives and build your own tools. Releasing the images and the model means anyone can start analyzing Mars data immediately. It democratizes the work.
Can MOMO actually find evidence of past life on Mars?
Not directly. But by mapping geology at scale—identifying water signatures, frost patterns, geological features that suggest past habitability—it gives scientists the tools to know where to look. It's not finding life. It's finding the conditions that life might have needed.
What's next for Purohit?
She wants to connect what the orbiters see from above with what rovers see on the ground. Right now those are separate datasets. If she can stitch them together, scientists could understand Mars at multiple scales simultaneously. That's the real frontier.