ExoVeil: How a Single Transit Shadow Reveals Exoplanets ✦
EXOVEIL is a transit-finding system that operates on raw flux time series and can detect single transits. Based on a Transformer model trained with transit masking on 16,499 Kepler light curves, it predicts the expected stellar flux; a variance-weighted matched filter extracts signals from the prediction residuals. An XGBoost classifier separates planets from false positives (AUC 0.938 on Kepler DR25). When injecting single transits with 1000 ppm depth, the system recovers 32%—a task where all classification methods show 0%. A blind search of 3737 Kepler stars revealed 179 new transit-like signals, including 46 single-transit candidates. Without retraining, the system detected all 47 confirmed TESS planets in the PLATO LOPS2 field, demonstrating cross-mission transferability. At PLATO’s 25-second cadence, transits down to 100 ppm are detectable, approaching the Earth-analog regime. Conformal prediction was applied for the first time, with an empirical coverage of 95.9%. The code, pre-trained weights, and candidate catalog are open-source (pip install exoveil).
Context
Classical transit-detection algorithms require knowing the period and a phase-folded signal. If a planet crosses the disk just once over the whole observation — typical for long-period Earth twins — these methods fail. That’s exactly where the PLATO mission and future habitability research focus. ExoVeil flips the paradigm: instead of classifying known signals, the system learns to predict a star’s brightness and reacts to any systematic deviations, like an anomaly detector.
Methods
ExoVeil’s core is a Transformer neural net (6 layers, 8 attention heads) trained to predict the next brightness data point from the entire past history. To make transits maximally anomalous, during training segments with known planets were masked and replaced with interpolated background — the model never saw actual transits. Trained on photometry curves of 16,499 stars, including sun-like and red dwarfs, the model then computes residuals, which are processed by a matched filter weighted by inverse prediction variance, enhancing sensitivity on quiet stretches and suppressing noise. False alarms like eclipsing binaries with white dwarfs are filtered by an XGBoost classifier on 21 features, and conformal prediction yields formal coverage guarantees for candidate ranking.
Results
A blind search among 3,737 Kepler stars flagged 179 transit-like anomalies not in the DR25 catalog; however, visual checks revealed some are caused by post-flare cooling, tracking errors, and edge effects. In a single-transit injection test, ExoVeil recovered 32% of events at 1000 ppm depth — while traditional classifiers can’t operate without phase-folding. Transferred to TESS data without retraining, the system detected all 47 confirmed planets in the PLATO LOPS2 field, including long-period and small ones. PLATO cadence simulations (25-second exposures) showed sensitivity reaches 100 ppm, entering the Earth-analog regime. All these results, however, need follow-up spectroscopy for confirmation.
Implications
The new approach tears down the barrier between detection and classification. It’s especially valuable for finding long-period planets that traditional methods miss, effectively doubling the usable observation window. Planet hunting in our Galaxy is entering a new era. Moreover, conformal prediction’s formal guarantees help allocate costly telescope time more efficiently for confirmation. Down the road, as atmospheric spectroscopy based on Rydberg laws matures, such candidates could become targets for biomarker searches — water, methane, and carbon dioxide — in their atmospheres.
Future development
The method will evolve along several fronts: boosting sensitivity via lighter architectures (e.g., Mamba instead of Transformer) for processing full light curves; integrating with PLATO simulations to precisely gauge performance in the Earth-analog domain; and building hybrid systems that blend prediction with classification. A key step will be automated recognition and culling of known false-alarm classes.
Impact
This will directly shape the data processing pipelines of future missions (PLATO, ARIEL), ground-based follow-up strategies, and telescope time allocation.
Next steps
Immediate steps include validation on simulated PLATO data (PlatoSim), adding automated false-positive rejection filters, and confirming candidates through radial velocities that exploit the Doppler effect.
Key open problems
The results feed directly into modern astrophysics’ central quest: finding potentially habitable worlds and gauging the prevalence of life in the Universe.
🎯 Fun fact: the first detector version, running on hand-crafted rules, scored an AUC of 0.36 — worse than random (0.5). It turned out that eclipsing binary residuals were deeper than planetary ones, so the system treated them as ‘more reliable’ signals.
Key numbers
- Trained on Kepler stars: 16,499
- Recovery of single transits (1000 ppm): 32%
- New candidates in blind search: 179
- Recovery of TESS planets (LOPS2): 100% (47/47)
- Detection depth at PLATO cadence: 100 ppm