ASU’s MOMO AI Model Trained on 12 Million Mars Images Beats Existing Tools

Context Mars has been photographed by orbiters for decades, producing a flood of high‑resolution, infrared and low‑resolution data. Scientists now face a bottleneck: turning this massive, fragmented archive into coherent scientific insight. Traditional AI models, trained on everyday objects or Earth‑centric imagery, struggle with the planet’s unique visual signatures.

Key Facts Mirali Purohit, a computer‑science doctoral student at Arizona State University’s School of Computing and Augmented Intelligence, led the effort to build a dedicated foundation model for the Red Planet. Working in the Kerner Lab under Hannah Kerner, Purohit assembled a curated set of about 12 million high‑quality Mars images from multiple missions after filtering an initial pool of 40 million samples. The resulting model, named Mars Orbital Model (MOMO), integrates separate sub‑models for different sensor types, allowing it to handle everything from microscopic rock textures to continent‑scale landscapes. Across a suite of benchmarks, MOMO consistently outperforms previous AI approaches, delivering superior performance in detailed surface mapping tasks such as crater detection, landslide delineation, frost identification and boulder spotting. The model’s flexibility lets it move fluidly between scales, a capability lacking in earlier Earth‑adapted or custom‑built tools. Purohit emphasized her motivation: she wanted to work in planetary sciences beyond Earth, believing that better AI could reveal what is truly happening on Mars.

What It Means MOMO represents the first general‑purpose “brain” for Mars, turning a chaotic data landscape into a usable scientific resource. By outperforming legacy methods, it accelerates the extraction of geological history, potentially highlighting past water activity or habitability clues. The team plans to release both the model and the underlying 12 million‑image dataset, lowering entry barriers for researchers worldwide.

The next step will be monitoring how the open release fuels new discoveries and whether similar foundation models can be built for other planetary bodies.