NASA and SETI outline foundation models for astrobiology and the search for extraterrestrial life

A team of researchers from NASA, SETI, and several universities argues that astrobiology needs a shared multimodal foundation model rather than a patchwork of disconnected AI tools. Such a model would help with the search for biosignatures, the planning of space missions, and the analysis of a vast body of scientific data.

Why a new approach is needed

Astrobiology operates on several levels at once: chemistry and molecular structures, planetary observations, field studies of terrestrial analogs, and space mission documentation. The authors start from the premise that no single label or instrument can describe life. Life shows up as a complex process with many signatures, so the field needs models that can connect images, spectra, geochemistry, text reports, and environmental context in one system. This is where foundation models look stronger than classic narrow ML.

The paper summarizes the findings of a workshop that NASA's Ames Research Center and the SETI Institute held in February 2025. The preprint itself appeared on arXiv on October 8, 2025. The researchers note that groundwork already exists: NASA is developing its own large language models, including Goddard LLM and INDUS, as well as the Prithvi geospatial model, and ESA has TerraMind. In other words, this isn't a fantasy decades away but the next logical step: assembling a specialized stack built specifically for astrobiology.

Three working scenarios

The authors suggest treating such a system not as a single chatbot but as the foundation for several applied modes. The logic is simple: one multimodal base can serve different tasks if separate interfaces and use cases sit on top of it. The first mode is searching for signs of life in complex data, the second is helping design and operate missions, and the third is a scientific interface for working with literature, reports, and hypotheses.

• Biosignature detection. The model would cross-reference chemical, morphological, spectral, and ecological features and separate possible traces of life from abiotic mimics.
• Astrobiology Mission Model. A dedicated AI layer would help select payloads, assess instrument constraints, prioritize samples, and support more autonomous spacecraft operations.
• AB-Chat. A specialized astrobiology interface would be able to read papers, technical reports, and mission archives, identify knowledge gaps, and propose new hypotheses.

Crucially, the authors don't propose taking humans out of the loop. On the contrary, both AMM and AB-Chat are described as human-in-the-loop tools: they broaden the team's field of view, but critical decisions stay with scientists and engineers. This matters especially for space missions, where the cost of error is high and onboard autonomy has to go through extensive testing, validation, and edge-case checks before launch.

What's holding it back right now

The main barrier isn't a lack of ideas but the state of the data. Astrobiology information is scattered across different archives, formats, and disciplines: spectral databases, geochemistry, imagery, mass spectrometry, mission reports, field observations, and historical printed materials. To train a genuinely useful model, this data first has to be located, brought to common standards, annotated with metadata, and made ready for machine learning. Sensitive mission data is a separate problem: schematics, internal procedures, and engineering documentation will require secure infrastructure.

That's why the first practical step looks fairly down-to-earth: rather than building a "superintelligence" right away, run existing datasets back through a model. The authors specifically mention visible imagery, VNIR reflectance, elemental and isotopic composition, GC-MS, Raman, XRF/XRD, and topography. Combining these sources with data on Earth, Mars, the Moon, and asteroids could let researchers start teaching the system to distinguish biotic, abiotic, and lookalike signatures. The outcome may not be a hard line between "living" and "non-living" but a multidimensional gradient, which is arguably more realistic for astrobiology.

What it means

If this approach reaches working prototypes, astrobiology won't simply get another LLM; it will get a domain-specific AI layer on top of science and space engineering. For researchers, that's a chance to pull knowledge together from scattered sources faster, plan missions more precisely, and improve the odds that a truly important signal of possible life doesn't get lost in the noise. The key question now isn't whether such a system is technically feasible, but who will be first to build a high-quality, interoperable data ecosystem for it.