Astrobiology in the Time of Artificial Intelligence

Abstract

The Viking missions showcased multiple spaceflight technologies that represented state-of-the-art capabilities: From digital line-scan imaging to the operation of complex onboard laboratories and software-controlled process autonomy. Since Viking, there have been extraordinary, and still accelerating, advancements in computing technology that impact science, society, and exploration. These developments have occurred in both hardware and software and have resulted in increasingly capable devices, advanced programming tools, and algorithmic innovations. The subset of artificial intelligence known as machine learning has emerged as one of the most transformative of these developments; it has major implications for space exploration and for improvements to the search for evidence of life beyond Earth. Those improvements include the integration of data across different scales and increased sensitivity to complex features in data, as well as the generation of adaptive strategies for sampling environments. In this article, the present and future nature of space exploration and astrobiological research is examined through the contextual lens of Viking and through the history and possible future of artificial intelligence.

Keywords

Life detection—Machine learning—AI—Viking

1. Introduction

1.1. Computing and space exploration

Spacecraft and rocketry design have seen tremendous advances over the past several decades, with developments in computing capabilities perhaps outpacing all other technologies. When the Viking landers arrived on Mars in 1976, they carried a complement of computing hardware that included two redundant Honeywell HDC 402 processors on each lander with 18 kilobytes of plated-wire memory (a type of magnetic-core memory) and an operation speed of about 230,000 instructions per second using 24-bit words (Holmberg et al., 1980a; see also Clark, 2026 in this issue). That electronic memory stored instructions that could, if necessary, operate the landers for their first 22 days on Mars without any contact with Earth. The Viking orbiters had more modest computing needs, but each carried a redundant pair of general electric serial processors with 4 kilobytes of memory and could perform some 25,000 instructions per second (Holmberg et al., 1980b).

By comparison, the Mars 2020 Perseverance rover has two redundant British Aerospace (BAE) RAD750 single-core Central Processing Unit (CPUs), each with 256 MB Dynamic Random Access Memory (DRAM), 256 KB Electrically Erasable Programmable Read-Only Memory (EEPROM), and 2 GB of flash memory—and each CPU can process at least 400 million instructions per second (Farley et al., 2020). That represents an increase in capability over the Viking landers by a factor of more than 1700 in computing speed and 14,000 in memory. Perseverance also carries specialized processors, such as Advanced Micro Devices (AMD) field-programmable gate arrays that support functions such as computer vision (from the rover’s 23 cameras), navigation, and the Scanning Habitable Environments with Raman and Luminescence for Organics and Chemicals (SHERLOC) instrument package (e.g., Wogsland et al., 2023). Today, flight-ready, radiation-hardened, multi-core processors are already capable of 10 times more operations per second than the hardware on Perseverance, along with much faster data flow and better error correction (e.g., Berger et al., 2015).

In concert with these advances in computing power, the autonomy of space exploration has seen dramatic evolution over the past decades. At a basic level, autonomy has, of course, always played a critical role in spaceflight. Early solar-powered interplanetary spacecraft used rudimentary voltage feedback loops from differently placed photoelectric sensors on the spacecraft bus to adjust orientations to ensure consistent power generation (e.g., Mariner 2, Wheelock, 1963), and other navigational and event-based autonomy has been key for most missions (e.g., star sensors; see Goss, 1970). Viking landers deployed a completely autonomous entry, descent, and landing system to handle the approximately 7 min between atmospheric entry and touchdown, since round-trip communications with Earth took some 40 min at that point. This included bootstrapping initial inertial velocity and position data and updating with live radar measurements and controlling the throttleable hydrazine thrusters to target the desired landing region. Autonomy has continued to be especially important for providing support following major mission events, such as orbital insertion maneuvers or planetary landings. More recently, autonomy has been taking on much more sophisticated roles, where sensor data (including imagery) is being analyzed in situ to enable quite complex decision-making. This includes selecting and measuring geological targets based on parameters set by mission scientists ahead of time (e.g., Autonomous Exploration for Gathering Increased Science (AEGIS) on Curiosity and Perseverance (Francis et al., 2017)).

Without autonomous capabilities such as these, space systems run considerable risks, both in terms of maintaining operations and in terms of returning the data and discoveries that justify a mission. This is especially true when signal travel times become large compared with the reaction times needed for actions to be carried out.

However, to date most autonomous spaceflight systems have relied on “traditional” forms of software and machine learning (ML), in which data analyses and decision-making are handled with nonadaptive algorithms and predetermined metrics. These often use rigid, rules-based processing and comprehensive statistical measures (e.g., as for AEGIS, Francis et al., 2017). In that sense, autonomy up to this time has largely mirrored the rest of science, where data analyses call on a relatively narrow range of mathematical and statistical principles and tools, from “best fit” functions to Bayesian likelihoods. In contrast, modern artificial intelligence (AI) and data-driven ML can, in principle, enable vastly more flexible and capable methods of wrangling information and discoveries from our explorations of the universe.

1.2. The rise of modern AI

The term “artificial intelligence” was brought into common use in the mid-1950s to help describe the anticipated emergence of a field of research on machines capable of human-like thought and actions (McCarthy et al., 2006). But the idea of such entities already had a very long history, going back to ancient Eastern philosophies and mythological, allegorical stories in ancient Greece. AI is, in truth, a very broad concept. It can include mechanical automata, and it can encompass everything that we currently label as “machine learning,” even if that is just least squares fitting of a straight line to data points.

From the 1940s and 1950s, where remarkable foundational work was done in computer science by people such as Alan Turing (Turing, 1945,1950) and John von Neumann (von Neumann, 1945), through to the early 1970s, the idea of AI boomed. Approaches such as symbolic AI, where logic and rules are used to try to mimic human thought and behavior, were embraced with the hope of achieving thinking machines (Minsky, 1975). However, progress was vastly slower than expected, and by the time of the Viking missions in the mid-1970s, the field of AI research was experiencing a “winter of discontent”—with widespread doubts about the effectiveness of machine intelligence of any kind.

In retrospect, though, some key pieces for future progress were falling into place, even if they seemed obscure at the time. For example, already in 1959, the concept of a “perceptron,” or artificial neuron, had been introduced for classification or recognition problems (Rosenblatt, 1960) and implemented as a computer program. The idea of many interlinked layers of perceptrons, or “deep learning,” was also considered in the late 1960s (Ivakhnenko, 1968).

By 1970, the mathematician Seppo Linnainmaa had described the mathematical basis of differential error backpropagation, a method that would be applied to the iterative, convergent training of networks of perceptrons to arrive at accurate responses to subsequent inputs (Linnainmaa, 1970). This technique would be rediscovered and advanced in the 1980s (Rumelhart et al., 1986), and by the 1990s and 2000s, it would be joined by a growing library of many different ML approaches to handling both very large sets of data and data with extremely complex features (multidimensional correlations, clustering, etc.) that were otherwise extremely challenging for humans to come to grips with (e.g., multivariate dimensionality reduction with Principal Component Analysis (PCA) or the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm) (Hinton and van der Maaten, 2008). These inventions were significant because they, in effect, provided sophisticated, mathematically characterized ways for machines to self-program, rather than be constrained by fixed instructions.

By the 2010s, computing hardware and software developments made greater adoption of these forms of ML feasible and opened the door to implementing so-called deep learning methods, with many layers of artificial neurons interconnected to each other and in more complicated configurations (e.g., Hinton, 2007). Deep learning could exhibit exceptional capabilities for training to discriminate between different classes and modalities in input data (from tabular data to images, audio, and video) with different types of models for different cases, such as artificial neural networks (ANNs) for discriminating tabular data or convolutional neural networks for imaging data. This, along with innovative ideas to capture contextual information in data (e.g., Transformer architectures, Vaswani et al., 2023), has led to the explosive growth in the 2020s of so-called generative AI. In these models, the learned representation of vast datasets (in millions or even trillions of model parameters) can be used to account for context in complex data and therefore predict sequential data accurately (in language and images) and generate derived data outputs (e.g., summarizing text or mapping scientific data into specific modes). Rather remarkably, these same AI model architectures turn out to be adaptable to many different data problems and can be trained generally on one or more types of associated data and then quickly “fine-tuned” for more demanding, specific tasks (e.g., Howard and Ruder, 2018).

The current landscape of artificial intelligence and ML is therefore vastly different than when Viking took place. But the central scientific questions of those missions and the challenges they encountered have never been more relevant and potentially more applicable to these new algorithmic systems.

2. Viking, Astrobiology, and AI

The Viking missions were bold and inventive and, when it came to their goal of life detection, both groundbreaking and sobering. The results of the three life detection experiments on the landers, and the consensus evaluation of that data, caused significant revaluation of not only the possibility of life elsewhere but also the suitability of techniques being considered in the search for life in our solar system (McKay et al., 2025). That outcome may have felt disappointing, but in the longer term it helped set the field of astrobiology on a path to a more thorough and nuanced vision of the complex ways that life—whether extant or extinct—imprints itself on the world and provides the foundational, contextual knowledge to support future search for life efforts. Today that vision is manifested in the ideas of scales of confidence in life detection claims (Green et al., 2021), probabilistic “ladders of life detection” techniques (Neveu et al., 2018), and strategies focused on the patterns of molecular composition and assembly history, elemental sequestration, anomalous phenomena, and other measurable features in the organization of matter that is, or has been, exploited and created by living systems (Des Marais et al., 2008; Marshall et al., 2021).

Those strategies align remarkably well with many of the algorithmic capabilities in modern AI and ML, where the multidimensional features in large, complex datasets can be learned, characterized, and exploited to perform numerous tasks ranging from noise reduction to anomaly detection or classification and prediction, as well as the provision of new forms of distilled information about the data features. Deep learning, in particular, utilizes ANNs to attempt, in the broadest sense, to build itself an accurate internal representation of data (in effect a compressed, optimized rendering of the features in data stored in the many neural net parameters) that is typically beyond the capacity of humans. This enables a machine to perform tasks such as classifying instances of measurements using a “holistic” assessment of the data and to extract meaningful information buried in millions, even billions of data points and their correlative properties.

ML is also not a single technique. Hundreds of algorithmic approaches exist, and a major challenge for data scientists and researchers in general is to identify the most efficacious methods for a given problem and to tune models’ hyperparameters (the characteristics of an ML model, such as learning rates or clustering criteria) to maximize their utility. This is both good and bad news for a field like astrobiology, where theory and data are diverse and datasets are growing but still modest compared with many other fields. The good news is that the right ML for a given science problem or mission requirement is out there somewhere; the bad news is that we must identify it and grapple with data that may be heterogeneous and variously sized. Progress is already being made, though, with ongoing efforts to standardize data from very different types of measurement (multimodal data) and use ML to ingest these data and perform biotic versus abiotic classification (e.g., see Benderoth et al., 2025).

In fact, one of the areas in astrobiology that is seeing robust growth in AI and ML development is biosignature detection and environmental assessment related to habitability and biosignature searches. For example, recent work on ML classification of raw pyrolysis–gas chromatography–mass spectroscopy data on physical samples drawn from a range of abiotic, biotic, and synthetic sources shows great promise, with high (>90%) accuracy in identifying biotic features (Cleaves et al., 2023). A similar approach can even discriminate categories of organisms, such as photosynthetic life in ancient rocks (>2.5 Gya, Wong et al., 2025). Methods for interpretable ML (where the way in which a model is making its determinations can be decoded) for biosignatures have been tested for ocean worlds such as Europa (Clough et al., 2025). Other studies have demonstrated strong capabilities for ingesting remote and in situ environmental data to assess and predict targets for biosignature searches using deep learning (Warren-Rhodes et al., 2023). These efforts also extend to exoplanetary spectral analysis, with ML approaches to detecting biosignature features as an alternative to traditional atmospheric retrieval methods (e.g., Duque-Castaño et al., 2025).

Other areas directly adjacent to the core goals of astrobiology where AI and ML are being rapidly deployed include those in planetary science that are engaged with the evaluation of hyperspectral imaging data. Sophisticated work has been carried out using a class of deep learning model called variational autoencoders (VAEs). Autoencoders are a high-heritage ML approach that lies at the heart of many modern deep learning models, including generative AI, where a corpus of data is used to learn a so-called latent (or hidden) representation of that corpus. That latent representation can then be used to identify features in the data (e.g., in images), reconstruct inputs that suffer from noise or incompleteness (e.g., Scharf, 2025), and generate new data that obey the full statistical properties of the original corpus. That latter use is well supported with VAEs where the autoencoder latent representation is built out of the probability distributions of the training corpus rather than discrete data points. That multidimensional probability model can be randomly sampled, for example, to generate fake but statistically compatible data, yet it also appears highly efficacious for tasks like mineralogical novelty detection in Mars rover imagery (Stefanuk and Skonieczny, 2022).

3. An AI Future for Astrobiology?

It seems clear that many recent advances in AI and ML are remarkably aligned with some of the biggest challenges for astrobiology and the search for life elsewhere (Scharf et al., 2023). These algorithms offer powerful new techniques for ingesting and analyzing complex multi-modal data and for integrating almost all aspects of terrestrial and space science.

One of the most exciting emerging applications of AI and ML is in end-to-end scientific exploration that further evolves the nature of autonomy. For example, recent studies have evaluated ML control systems for space missions that analyze data in real time and, in effect, draw scientific conclusions (e.g., Thompson et al., 2011; Theiling et al., 2022). That might mean gauging how new measurements could further validate or refute a hypothesis or determining new priorities for scientific exploration. If AI and ML are tightly integrated with spacecraft or probes, the system can be empowered to modify actions to pursue data to address follow-up questions. Not only can this increase the chances of achieving specific scientific, mission-critical goals, it allows for open-ended in situ research without waiting for human input, which could prove important for the search for life beyond the inner solar system.

There is also the question of how robotics will be incorporated into potential human exploration on Mars or elsewhere in the solar system. Without signal delays, astronauts could be in direct control of a variety of robotic devices—including those engaged in searching for evidence of life—but what would that interaction look like? Even with current AI (e.g., Large Language Model (LLMs) and Agentic AIs), the possibility exists for interfaces that respond to natural language prompts and that can engage in back-and-forth discussion, drawing on the corpus of specialized information that a model is trained on. With human input, an AI model can then be used to create a detailed plan of research that can be quickly translated into software or control instructions. This opens the door to some very sophisticated and advanced options for humans carrying out surface science.

Other emerging areas involve the development of AI Foundation Models (FMs) for astrobiology (Felton et al., 2025). FMs are typically large, deep learning models trained on very large and broad datasets to learn the fundamental features of the data (Bommasani et al., 2022). Once trained—often at considerable cost—an FM can be subsequently fine-tuned relatively quickly and cheaply using much smaller specialized datasets (usually still related to the original training data) to address narrower questions or analyses’ needs. Multimodal FMs can combine LLMs with other forms of tokenizable data (images, tabular information, etc.) to enable sophisticated capabilities such as combining pure research with mission planning, data analysis, and data products (including report and article writing). The interdisciplinary nature of astrobiology makes it a particularly challenging but potentially rich area for FM deployment, perhaps yielding a complete pipeline for research, from ideas to proposals to instrument and mission design to actual exploration.

Beyond these very practical applications of AI and ML in astrobiology, a persistent, and sometimes controversial, question is whether we understand the fundamental nature of life well enough to be sure of recognizing it elsewhere or decoding its function when we do find it. While there are many features of terrestrial, carbon-based biochemistry that seem unlikely to have a viable replacement in any truly alternate (but not improbable) chemical system or substrate, it does seem plausible that the vast library of biochemically functional molecular species (from small to large) found on Earth might be represented somewhat differently elsewhere and utilize different tricks to persist in the face of Darwinian selection. The notion of “alternate life” and “agnostic” models of life and ways to detect such alternate forms continue to be studied (Bartlett and Wong, 2020; Marshall et al., 2021; Chandru et al., 2024), but at the same time, without an underlying, agnostic theory of how life comes into existence and how it must operate, this remains challenging. Although many valiant efforts have been made to construct just such theories for parts of the puzzle, none have yet crossed the barrier of consensus acceptance (e.g., theories of selection across chemistry and biology, Sharma et al., 2023).

This is an area where AI and ML might, just possibly, achieve what humans have not. Already, in other scientific disciplines, AI and ML are proving to be extremely good at learning predictive powers for otherwise extraordinarily complex phenomena. For example, AI models can provide excellent and fast future-casting of Earth’s climate and weather patterns after training on expensive numerical simulation results (for a review see Waqas et al., 2025). The deep learning AlphaFold 3 model (that draws on diffusion techniques that also work well in image generation) has high accuracy in predicting the structure not only of proteins but also DNA, RNA, and ligands based on sequence data alone (Abramson et al., 2024). And in astrophysics, deep learning techniques can enhance cosmological models of the gravitational evolution of matter distributions to generate accurate maps of evolved galaxies and large-scale structures, without running the intensive simulations traditionally used (e.g., Li et al., 2021).

For astrobiology, it is reasonable to ask whether similar kinds of AI models could, in the future, begin to encode the qualities of living systems in a way that allows us to ask more comprehensive and probing questions about the nature of life and its detection. Such models would learn an internal representation of the fundamental properties of biochemistry (from amino acids and RNA to proteins and cellular molecular complexes), together with the characteristics of energy transduction and utilization, as well as the larger structures and behavior of organisms, from cells to populations. The internal (latent) representations formed by such an AI could then help elucidate what it is exactly that makes life, life. Or at a minimum, they could provide a means to predict the outcome of a set of conditions, whether for the origin and evolution of living systems or the outcomes of a whole-organism DNA sequence. This would revolutionize our ability to consider both extinct and extant terrestrial biochemistries and nonterrestrial biochemistries on worlds like Mars or elsewhere. We might not get an easy answer to “what is life?” But it might not matter, because our AI would recognize life’s supremely complex flavors on our behalf.

In that future, a new version of Viking could be designed from scratch as a physical companion to an AI that can predict the best ways to look for evidence of life based on its rich, comprehensive, predictive model and serve as a real-time, in situ expert—the thinking component of the mission. As measurements are ingested by the AI, it can, in effect, run the scenarios to evaluate billions of years of potential biochemical (and environmental) evolution and assess where the data is leading and what measurements are needed next. It could even return the optimal design for the next mission, the inevitable replacement and upgrade of hardware, and the fine-tuned upgrade for itself, trained on the newly acquired data.

An intriguing, but curious, outcome of this type of model fine-tuning or updating is that the AI model itself could become a subject of study, since the way that deep learning models encode data (often termed embeddings) in their internal, or latent, representation is itself data-dependent (Scharf, 2025). The stored parameters in a model (that may number in the millions or billions) represent a potentially unique landscape that reflects the relationship of features contained in the original measurements (such as of a Martian environment). In other words, the AI model itself could contain information that is instructive for humans to understand and that might yield insights to other phenomena “caught up” in the data ingestion or even provide entirely new discoveries about the properties of living systems.

An example would be in identifying so-called “hidden variables” that indicate relationships in the data that are not directly measurable in the real world (e.g., relationships between complex biological processes and their physical and chemical manifestations). In current AI research, a variety of frameworks exist for mapping or visualizing model latent representations to interpret their structure (e.g., Kopf and Claassen, 2021). Consequently, an astrobiology research subfield of the future might be the study of specialist AI models rather than the raw data or traditional products of a mission or experiment.

A Viking 2.0 project in a future of AI would take all the remarkable innovation of the original landers and orbiters and evolve them into a new, hybridized kind of space mission. The design and development might start with a single prompt to an AI FM that operates alongside human engineers, scientists, and project managers. Instrument packages would be optimized both for science and for the ingestion of their data by mission AI, which would in turn adapt on the fly to discoveries and issues to steer orbiters and landers toward the ultimate science goals (and perhaps even new science goals as they emerge). And when Viking 2.0 is complete, its fine-tuned AI models and internal representations of the corpus of mission data will become a “living record” of the mission, available to be studied and interrogated to look for clues to the next scientific questions in the search for life in the universe.

Footnotes

Acknowledgments

Mary Beth Wilhem and Scott Perl are thanked for their organization of this contribution and their editorial input on this article.

Author Disclosure Statement

No competing financial interests exist.

Funding Information

No funding was received for this article.

Associate Editor: Michael A. Meyer

References

Abramson

, Adler

, Dunger

, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 2024;630(8016):493–500.

Bartlett

, Wong

. Defining Lyfe in the universe: From three privileged functions to four pillars. Life (Basel) 2020;10(4):42.

Benderoth

, Shahid

, Sheng

, et al. Statistical classification of biosignature information: advancing life detection through machine learning. In 56th Lunar and Planetary Science Conference. Lunar and Planetary Institute: Houston; 2025, abstract 2324.

Berger

, Chadwick

, Chan

, et al. Quad-core radiation-hardened system-on-chip power architecture processor. In: 2015 IEEE Aerospace Conference. Piscataway, NJ: IEEE; 2015; doi: 10.1109/AERO.2015.7119114

Bommasani

, Hudson

, Adeli

, et al. On the opportunities and risks of foundation models. arXiv 2022 2108.07258.

Chandru

, Potiszil

, Jia

. Alternative pathways in astrobiology: Reviewing and synthesizing contingency and non-biomolecular origins of terrestrial and extraterrestrial life. Life (Basel) 2024;14(9):1069.

Clark

. Viking Science–1. Astrobiology 2026;26 in press.

Cleaves

, Hystad

, Prabhu

, et al. A robust, agnostic molecular biosignature based on machine learning. Proc Natl Acad Sci U S A 2023;120(41):e2307149120.

Clough

, Da Poian

, Major

, et al. Interpretable machine learning biosignature detection from ocean worlds analogue CO₂ isotopologue data. Earth and Space Science 2025;12(3):e2024EA003966.

10.

Des Marais

, Nuth

, Allamandola

, et al. The NASA Astrobiology Roadmap. Astrobiology 2008;8(4):715–730.

11.

Duque-Castaño

, Zuluaga

, Flor-Torres

. Machine-assisted classification of potential biosignatures in Earth-like exoplanets using low signal-to-noise ratio transmission spectra. Monthly Notices of the Royal Astronomical Society 2025;539(2):1528–1552.

12.

Farley

, Williford

, Stack

, et al. Mars 2020 mission overview. Space Sci Rev 2020;216(8):142.

13.

Felton

, Scharf

, Bartlett

, et al. Foundation models for astrobiology: Paper I—workshop and overview. Astrobiology 2025;25(12):807–821.

14.

Francis

, Estlin

, Doran

, et al. AEGIS autonomous targeting for ChemCam on Mars Science Laboratory: Deployment and results of initial science team use. Sci Robot 2017;2(7):eaan4582.

15.

Goss

. The Mariner spacecraft star sensors. Appl Opt 1970;9(5):1056–1067.

16.

Green

, Hoehler

, Neveu

, et al. Call for a framework for reporting evidence for life beyond Earth. Nature 2021;598(7882):575–579.

17.

Hinton

. Learning multiple layers of representation. Trends Cogn Sci 2007;11(10):428–434.

18.

Hinton

, van der Maaten

. Visualizing data using t-SNE. Journal of Machine Learning Research 2008;9(2605):2579–2605.

19.

Holmberg

, Faust

, Holt

. NASA Reference Publication 1027: Viking ‘75 Spacecraft Design and Test Summary. Volume 1 – Lander Design. Washington, DC: NASA; 1980a. https://ntrs.nasa.gov/api/citations/19810001592/downloads/19810001592.pdf

20.

Holmberg

, Faust

, Holt

. NASA Reference Publication 1027: Viking ‘75 Spacecraft Design and Test Summary . Volume 2 – Orbiter Design. Washington, DC: NASA; 1980b. https://ntrs.nasa.gov/api/citations/19810001593/downloads/19810001593.pdf

21.

Howard

, Ruder

. Universal language model fine-tuning for text classification. arXiv 2018 1801.06146.

22.

Ivakhnenko

. The group method of data handling—a rival of the method of stochastic approximation. Soviet Automatic Control 1968;13(3):43–55.

23.

Kopf

, Claassen

. Latent representation learning in biology and translational medicine. Patterns (N Y) 2021;2(3):100198.

24.

, Ni

, Croft

RAC

, et al. AI-assisted superresolution cosmological simulations. Proc Natl Acad Sci U S A 2021;118(19):e2022038118.

25.

Linnainmaa

. (1970) The Representation of the Cumulative Rounding Error of an Algorithm as a Taylor Expansion of the Local Rounding Errors . (Master’s thesis, in Finnish) Helsinki, Finland: Department of Computer Science, University of Helsinki.

26.

Marshall

, Mathis

, Carrick

, et al. Identifying molecules as biosignatures with assembly theory and mass spectrometry. Nat Commun 2021;12(1):3033.

27.

McCarthy

, Minsky

, Rochester

, et al. A proposal for the Dartmouth summer research project on artificial intelligence, August 31, 1955. AI Mag 2006;27(4):12.

28.

McKay

, Quinn

, Stoker

. The Viking biology experiments on Mars revisited. Icarus 2025;431:116466.

29.

Minsky

. A framework for representing knowledge. In Winston

.(ed.) The Psychology of Computer Vision. McGraw-Hill Book: New York; 1975.

30.

Neveu

, Hays

, Voytek

, et al. The Ladder of Life Detection. Astrobiology 2018;18(11):1375–1402.

31.

Rosenblatt

. Perceptron simulation experiments. Proc IRE 1960;48(3):301–309.

32.

Rumelhart

, Hinton

, Williams

. Learning representations by back-propagating errors. Nature 1986;323(6088):533–536.

33.

Scharf

. The model is the message: Lightweight convolutional autoencoders applied to noisy imaging data for planetary science and astrobiology. Icarus 2025;442:116740.

34.

Scharf

, Mayer

, Boston

. Using artificial intelligence to transform astrobiology. Nat Astron 2023;8(1):8–9.

35.

Sharma

, Czégel

, Lachmann

, et al. Assembly theory explains and quantifies selection and evolution. Nature 2023;622(7982):321–328.

36.

Stefanuk

, Skonieczny

. Novelty detection in rover-based planetary surface images using autoencoders. Front Robot AI 2022;9:974397.

37.

Theiling

, Chou

, Da Poian

, et al. Science autonomy for ocean worlds astrobiology: A perspective. Astrobiology 2022;22(8):901–913.

38.

Thompson

, Wettergreen

, Peralta

FJC

. Autonomous science during large‐scale robotic survey. J Field Robot 2011;28(4):542–564.

39.

Turing

. Proposals for the Development in the Mathematics Division of an Automatic Computing Engine (ACE) . The National Physical Laboratory: Teddington, UK; 1945. Available from: https://www.npl.co.uk/getmedia/dda98194-32a1-424c-ad09-33081b4afc1d/turing-proposal-Alan-LR.pdf [Last accessed: June 6, 2026].

40.

Turing

. I.—Computing machinery and intelligence. Mind 1950;LIX(236):433–460.

41.

Vaswani

, Shazeer

, Parmar

, et al. Attention is all you need. arXiv 2023 1706.03762.

42.

von Neumann

. First Draft of a Report on the EDVAC . Philadelphia, PA: Moore School of Electrical Engineering, University of Pennsylvania; 1945. Available from: https://web.mit.edu/sts.035/www/PDFs/edvac.pdf

43.

Waqas

, Humphries

, Chueasa

, et al. Artificial intelligence and numerical weather prediction models: A technical survey. Natural Hazards Research 2025;5(2):306–320.

44.

Warren-Rhodes

, Cabrol

, Phillips

, et al. Orbit-to-ground framework to decode and predict biosignature patterns in terrestrial analogues. Nat Astron 2023;7(4):406–422.

45.

Wheelock

. Mariner Mission to Venus, Compiled 1963. (A Project Gutenberg eBook). NASA Jet Propulsion Laboratory: Pasadena, CA; 1963. https://www.gutenberg.org/files/54585/54585-h/54585-h.htm

46.

Wogsland

, Minitti

, Kah

, et al. Science and science‐enabling activities of the SHERLOC and WATSON imaging systems in Jezero crater, Mars. Earth and Space Science 2023;10(11):e2022EA002544.

47.

Wong

, Prabhu

, Alexander

, et al. Organic geochemical evidence for life in Archean rocks identified by pyrolysis–GC–MS and supervised machine learning. Proc Natl Acad Sci U S A 2025;122(47):e2514534122.