EXOVEIL: A Single Transit Is Sufficient for Exoplanet Detection via Learned Stellar Patterns

This paper introduces EXOVEIL, a novel framework for identifying exoplanets that identifies deviations from a star’s expected luminosity. By learning the typical behavior of stellar brightness, the system highlights instances where observed data diverges from this norm. A key distinction of EXOVEIL is its ability to function directly on raw flux time series, eliminating the need for the phase-folded inputs required by many contemporary detection methods. This capability allows the system to identify planets that exhibit only a single transit event.

The core of the system is a Transformer-based world model, which was trained using self-supervised learning with transit-masked data on a dataset comprising 16,499 Kepler light curves. This model predicts the expected stellar flux, while a matched-filter detector, incorporating variance weighting, isolates transit signals from the residuals of these predictions. To distinguish genuine planetary signals from false positives, the system employs a learned XGBoost classifier, which achieved an Area Under the Curve (AUC) of 0.938 on Kepler DR25 data.

In tests involving single-transit injection-recovery, EXOVEIL demonstrated superior sensitivity, recovering 32% of transits with a depth of 1000 parts per million (ppm). This performance stands in stark contrast to classification-based systems, which inherently score 0% on this specific task. Furthermore, a blind search conducted on 3,737 Kepler stars identified 179 new transit-like signals absent from the DR25 TCE catalogue, including 46 candidates characterized by monotransit events.

The system’s versatility was further validated by applying it without retraining to 47 confirmed TESS planets located within the PLATO LOPS2 field, resulting in a 100% recovery rate and proving its effectiveness for zero-shot cross-mission transfer. Looking ahead to PLATO’s operational cadence of 25 seconds, the detection threshold is projected to reach 100 ppm, nearing the sensitivity required for Earth-analog detection. Additionally, this work marks the first application of conformal prediction to transit detection, achieving an empirical coverage of 95.9%. The complete system, including pretrained weights and a candidate catalogue, is now available for installation via pip install exoveil.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC