Life, Machine Learning, and the Search for Habitability: Predicting Biosignature Fluxes for the Habitable Worlds Observatory

TL;DR

This work addresses how to efficiently prioritize observations for the Habitable Worlds Observatory by predicting biosignature fluxes from reflected-light spectra under constrained telescope time. It develops two ML approaches: a BCNN that provides calibrated uncertainty, and SQuAT, a query-based Transformer that yields interpretable, species-specific spectral attributions. Using an augmented PyATMOS-derived dataset with broad atmospheric diversity and eight biosignature flux targets, both models achieve high predictive accuracy ($R^2$ around $0.97$–$0.99$), with SQuAT offering clearer spectral interpretability through attention maps. The methods enable risk-aware decision-making and targeted scientific output, and they suggest a complementary ensemble approach for robust, explainable biosignature flux retrieval in preparation for onboard or ground-based planning for HWO.

Abstract

Future direct-imaging flagship missions, such as NASA's Habitable Worlds Observatory (HWO), face critical decisions in prioritizing observations due to extremely stringent time and resource constraints. In this paper, we introduce two advanced machine-learning architectures tailored for predicting biosignature species fluxes from exoplanetary reflected-light spectra: a Bayesian Convolutional Neural Network (BCNN) and our novel model architecture, the Spectral Query Adaptive Transformer (SQuAT). The BCNN robustly quantifies both epistemic and aleatoric uncertainties, offering reliable predictions under diverse observational conditions, whereas SQuAT employs query-driven attention mechanisms to enhance interpretability by explicitly associating spectral features with specific biosignature species. We demonstrate that both models achieve comparably high predictive accuracy on an augmented dataset spanning a wide range of exoplanetary conditions, while highlighting their distinct advantages in uncertainty quantification and spectral interpretability. These capabilities position our methods as promising tools for accelerating target triage, optimizing observation schedules, and maximizing scientific return for upcoming flagship missions such as HWO.

Figures (4)

• Figure 1: SQuAT model architecture.
• Figure 2: SQuAT attention map. Top: reflected-light exoplanet spectrum sample. Bottom: per-species attention across wavelength, highlighting prominent regions for each biosignature’s flux prediction. Red dashed boxes mark canonical absorption bands. Attention is normalized per species.
• Figure 3: Error correlation matrix across predicted biosignature species. Each cell shows the Pearson correlation coefficient between the model prediction errors for a pair of species. High positive values (red) indicate that errors tend to co-occur in the same direction, while strong negative values (blue) suggest opposing error trends.
• Figure 4: (a) Predicted versus true biosignature flux for four molecular species (O2, O3, CH4, H2O) and for all species combined. (b) Corresponding model predictions with 95% credible intervals, with samples sorted by increasing true flux.