The software, built by Floyd Nichols and Grant Major of Virginia Tech’s Department of Geosciences and Ryan Felton of NASA Ames, is a web-based graphical user interface that lets educators, students, and researchers train and interpret machine learning models without writing a single line of code. It is hosted on Hugging Face, the popular open-source machine learning platform, and runs on two virtual CPUs with 16 gigabytes of RAM. The team describes it as an entry point for novices and a rapid prototyping tool for experts, a way to standardize the machine learning workflow across a community that has historically approached it in ad hoc ways.

The need for such a tool is not abstract. Within the last five years, machine learning has crept into nearly every corner of astrobiology: analyzing laboratory data on biosignatures, processing remote sensing and telescope observations, running simulations of planetary environments, and even enabling autonomous decision-making on space missions. NASA’s own strategic documents have repeatedly called for broader adoption. The agency’s Science Mission Directorate Strategy for Data Management and Computing recommended incentivizing and educating the community on AI and machine learning approaches, while the Origins, Worlds, and Life decadal survey explicitly endorsed machine learning for supporting autonomy on remote planetary missions. The NASA ROSES program has signaled the same need through dedicated solicitations for Solar System Science and Foundational Artificial Intelligence for the Moon and Mars.

Yet the barrier to entry remains steep. Machine learning algorithms, even those available as off-the-shelf functions in Python or R, demand a working knowledge of both the underlying mathematics and the programming language itself. Many astrobiologists are domain experts in geology, chemistry, or biology but have neither the time nor the institutional support to become competent programmers. This is precisely the gap AI² is designed to fill, offering a guided pipeline that walks users through the three canonical stages of a machine learning project: data preprocessing, exploratory data analysis with unsupervised learning, and supervised learning for prediction.

The preprocessing stage handles the unglamorous but critical work that determines whether any downstream model will be trustworthy. Users upload a CSV file, and the application immediately computes Pearson correlation coefficients to reveal pairwise linear relationships between variables and variance inflation factors to flag multicollinearity, a statistical pathology in which highly correlated predictors can silently degrade a model’s ability to generalize. The tool then offers options for handling missing values through removal or imputation using mean, median, or most-frequent strategies, encoding categorical variables into numerical form, and scaling features via min-max normalization or z-score standardization. These steps matter because many algorithms implicitly assume features are centered around zero with comparable variance; a single feature with orders of magnitude more spread can dominate the model and prevent it from learning anything useful from the others.

Once the data is clean, the unsupervised learning module offers dimensionality reduction and clustering techniques chosen specifically because they have appeared in prior astrobiology research. For dimensionality reduction, users can apply Principal Component Analysis, which captures maximum variance between features through linear transformation; Multidimensional Scaling, which instead preserves pairwise distances between data points; or t-Distributed Stochastic Neighbor Embedding, a nonlinear manifold learning method that preserves local structure and is particularly useful for visualizing high-dimensional data in two dimensions. For clustering, the options include K-means, which partitions data into k groups of equal variance around centroids; Gaussian Mixture Models, which identify clusters based on probabilistic Gaussian densities; and DBSCAN, which can detect clusters of arbitrary shape, size, and density without requiring the user to predefine how many groups exist.

The supervised learning module covers both classification and regression tasks. Classification options include Logistic Regression, k-Nearest Neighbors, Support Vector Classifier, and Random Forest, while regression offerings include Random Forest Regressor, Histogram Gradient Boosting, Support Vector Regressor, and Ridge Regression. After training, the interface produces standard evaluation metrics: accuracy, precision, recall, and F1 score for classification, and mean squared error, root mean squared error, mean absolute error, and mean absolute percentage error for regression. Cross-validation is incorporated to assess generalizability, and the tool generates interactive visualizations including confusion matrices for classification tasks and predicted-versus-actual crossplots for regression. Crucially, it also produces feature importance plots based on Shapley Additive Explanations, a technique that attributes each prediction to individual input features, keeping the entire workflow within the realm of explainable rather than black-box AI.

What distinguishes AI² from similar tools is its built-in statistical guardrails. The developers recognized that even experienced researchers sometimes select models inappropriately because they skip the diagnostic tests that reveal whether a dataset’s assumptions match a given algorithm’s requirements. The application therefore runs Henze-Zirkler’s test for multivariate normality, the Breusch-Pagan test for heteroscedasticity, and isolation forest analysis to quantify outlier rates, all under the hood. If a user selects a model that is statistically inappropriate for their data, the tool issues a warning and suggests a more robust alternative. For example, ridge regression, often chosen for its speed and resistance to multicollinearity, breaks down when data contains high outlier rates, heteroscedasticity, or non-normal distributions; in such cases the interface recommends tree-based models like random forest or histogram gradient boosting instead.

The team benchmarked the web-hosted version across datasets ranging from 100 to 10,000 observations with 10 features, sizes chosen to represent typical astrobiological datasets from biological measurements to stellar spectral catalogs. Most model scenarios completed in seconds to tens of seconds. The outlier was t-SNE with 10,000 observations, which averaged 194 seconds, a reflection of that algorithm’s inherent computational intensity rather than any flaw in the implementation.

The developers are candid about the limitations. The algorithm menu is deliberately restricted to methods most relevant to astrobiology, and neural networks are excluded for now because of their black-box nature, though simple artificial neural networks are planned for future versions. The team also acknowledges that large language models like ChatGPT and Claude can help novices write machine learning code, but argues that this path is fraught: LLMs require well-crafted prompts that themselves demand machine learning understanding, and studies have shown that novices using LLMs frequently make inappropriate model selections, misuse data, and skip systematic hyperparameter tuning, errors that then get baked into the training data of future models. Compared with existing tools like Google’s Teachable Machine, which handles only image and audio recognition, or Orange Data Mining, which requires moderate machine learning knowledge and lacks statistical assumption testing, AI² is purpose-built for tabular astrobiology data with guardrails baked in. Future plans include video tutorials and integration with the AI-Astrobiology portal at NASA Ames, a community resource that hosts seminars, literature reviews, and tutorial guides. The broader hope is that by lowering the technical barrier, the tool will help a generation of astrobiologists treat machine learning not as an arcane specialization but as a standard part of the scientific toolkit, one that complements rather than replaces human domain expertise in the ongoing search for life in the universe.

Subject of Research: An open-source graphical user interface for applying explainable machine learning to astrobiology datasets

Article Title: Artificial intelligence for astrobiology investigations (AI2): an open-source graphical user interface

Article References: Nichols, F., Major, G., & Felton, R. (2026). Artificial intelligence for astrobiology investigations (AI2): an open-source graphical user interface. Frontiers in Astronomy and Space Sciences, 13, Article 1933473. https://doi.org/10.3389/fspas.2026.1933473

Image Credits: AI Generated

DOI: 10.3389/fspas.2026.1933473

Keywords: astrobiology, machine learning, open-source software, graphical user interface, NASA, biosignatures, data preprocessing, unsupervised learning, supervised learning, explainable AI, Hugging Face, scientific education