Stellar Orthography: How ExoVeil Reads Missed Planets ✦
Scientists have created EXOVEIL, a system that finds exoplanets from single dips in starlight, without needing repeated signals as before. It uses a ‘predictor’ neural network (a Transformer) trained on thousands of light curves from Kepler, to reconstruct the star’s ideal brightness; then a special filter catches the tiniest deviations. Among the finds are 179 new candidates, including 46 single transits missed by standard catalogs. The system works without retraining even on data from other telescopes, and it’s edging toward detecting Earth-like planets. It’s like teaching a clock to notice a single missing second instead of waiting for minute-long glitches.
A star's light curve is not chaotic noise, but a text written by pulsations, spots, and the steady glow of a thermonuclear furnace. Any planet slipping across the disk leaves a typo in this text: a tiny dip in brightness lasting hours. Classic exoplanet search methods required the typo to repeat—word after word, orbit after orbit. But what if the planet passes only once in years of observations? That's exactly how potential Earth twins on wide orbits behave. Catching them with old tools is like searching for a single missing 'not' in a multi-volume novel without knowing the language.
The ExoVeil system works like a genius proofreader who has mastered the grammar of starlight. At its core is a transformer neural network with six layers and eight attention heads, trained on 16,499 photometry curves from the Kepler telescope. The trick: during training, sections with known planets were masked and replaced with smooth background interpolation—the model never saw real transits. It only learned to predict the next brightness reading, much like a language model predicts the next word based on full context. When a real light curve is fed in, the difference between prediction and observation becomes an anomaly map.
The residuals are processed by a matched filter—a detector that weights each deviation inversely to noise variance. Calm stretches of the curve where the star 'is silent' get more weight, while turbulent ones get less. This mathematical trick is captured by a simple signal-to-noise ratio formula:
\[ \text{SNR} = \frac{\sum_i (r_i \cdot m_i) / \sigma_i^2}{\sqrt{\sum_i m_i^2 / \sigma_i^2}} \]
where \(r_i\) are prediction residuals, \(m_i\) is the transit template, and \(\sigma_i^2\) is local variance. Thanks to this approach, the system recovers 32% of single transits as shallow as 1000 ppm—without the phase folding that traditional methods require. Then an XGBoost-based classifier filters out false alarms, like eclipsing binaries with white dwarfs, and conformal prediction provides formal guarantees for ranking candidates.
A blind search on 3,737 Kepler stars yielded 179 transit-like anomalies not listed in the DR25 catalog. Transferring to TESS data without retraining confirmed the method's power: all 47 known planets in the PLATO LOPS2 field were detected, including long-period and small ones. Simulations of the future PLATO mission cadence with 25-second exposures show sensitivity will improve to 100 ppm—right down to Earth-like scales. Transit depth is given by another elegant formula:
\[ \delta = \left(\frac{R_p}{R_s}\right)^2 \]
showing that an Earth-sized planet blocks only 0.008% of a sun-like star's light. ExoVeil learns to discern such minuscule shadows, effectively doubling the available observing time for analysis and paving the way to search for Earth analogues in our Galaxy. And if one day the PLATO telescope captures a single fleeting dimming, ExoVeil could not only spot it but also suggest: 'This is an Earth-sized planet, and its year lasts 365 days.'
The method's future lies in integration with PLATO simulations. Lighter architectures like Mamba will allow processing full light curves without cutting corners. Automatic filtering of false alarms will make the search routine. Confirming candidates will require radial velocity measurements using the Doppler effect and subsequent spectroscopy of atmospheres, where absorption lines are described by the Rydberg formula. In the long run, such planets will become targets for biomarker searches—water, methane, carbon dioxide. ExoVeil doesn't just hunt worlds; it teaches us to read the language of stars, so that one day we may find a text where, among familiar symbols, the word 'life' emerges.
🎯 The first version of the detector, based on hand-crafted rules, yielded an AUC of 0.36—worse than random guessing (0.5). The residuals from eclipsing binaries turned out deeper than planetary ones, and the algorithm deemed them more reliable signals.