🔭 astrophysics

One Transit Is All You Need: Detecting Exoplanets Through Learned Stellar Behaviour with EXOVEIL

The paper introduces EXOVEIL, a novel transit detection system that leverages a Transformer-based world model to learn stellar behavior and identify exoplanets from raw flux time series, enabling the detection of single-transit events and demonstrating robust zero-shot transfer across Kepler, TESS, and PLATO missions.

Original authors: Pratik Priyanshu

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pratik Priyanshu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a tiny, dark moth flying across a giant, flickering streetlamp. The lamp isn't steady; it buzzes, dims, and flares up randomly because of the wind and old wiring. For decades, astronomers have tried to find these "moths" (exoplanets) by waiting for the moth to fly across the lamp multiple times in a perfect rhythm. Once they see the pattern repeat, they can predict exactly when the next dip in light will happen and confirm the moth is there.

The Problem: What if the moth only flies across once? Maybe it's a very distant planet with a long year, or maybe the telescope only watched for a short time. If the moth only crosses once, the old "pattern-matching" systems are blind. They can't find it because they are waiting for a rhythm that doesn't exist yet.

The Solution: ExoVeil
This paper introduces a new system called ExoVeil. Instead of waiting for a pattern, ExoVeil learns what the streetlamp should look like if the moth were never there.

Here is how it works, broken down into simple steps:

1. The "World Model" (The Expert Watcher)

Imagine a super-smart security guard who has watched thousands of streetlamps for years. This guard learns exactly how the light naturally flickers, buzzes, and changes due to wind or aging bulbs.

How ExoVeil does it: It uses a type of AI called a "Transformer" (similar to the technology behind modern chatbots) trained on data from the Kepler space telescope. It learns the "normal" behavior of stars.
The Trick: The AI was trained without ever seeing a planet transit. It was shown stars, but whenever a planet did pass in front, the computer filled in the gap with a guess of what the light should have been. This forced the AI to learn the star's natural behavior perfectly, so it never "expects" a planet to be there.

2. The "Residual" (The Surprise)

Once the AI predicts what the star's light should be, it compares that prediction to the actual light coming from the telescope.

The Analogy: If the AI predicts the light should be steady at 100 units, but the telescope reads 99 units for a few hours, that difference is the "residual."
The Discovery: Because the AI knows the star so well, a tiny, unexpected dip in the light stands out like a sore thumb. This is how it finds planets that only cross once. The system flags the moment reality disagrees with the prediction.

3. The "Filter" and "Judge" (Sorting the Noise)

Sometimes, the star flickers for other reasons, or the telescope glitches. The system has to decide: Is this a planet, or just noise?

The Filter: It uses a mathematical "matched filter" (like a template) to look for the specific shape of a planet crossing (a smooth dip down and back up).
The Judge: It then uses a smart classifier (an XGBoost algorithm) to act as a judge. It looks at the shape of the dip and other clues to say, "This looks like a planet," or "This is just a glitch."

What Did They Find?

The authors tested this system and found some impressive things:

The "One-Shot" Win: They tested the system on fake planets that only crossed once. While all other computer systems scored 0% (because they can't work without a repeating pattern), ExoVeil found 32% of them.
New Discoveries: When they ran it on old data from the Kepler telescope, it found 179 new signals that previous searches missed. This included 46 "monotransit" candidates (planets that only crossed once). One strong candidate, KIC 11706231, showed a planet crossing a cool, small star.
The "Zero-Shot" Magic: They tested the system on data from a different telescope (TESS) without teaching it anything new. It successfully found 100% of the known planets in that field. This proves the AI learned the physics of stars, not just the quirks of one specific telescope.
Future Proofing: They simulated how it would work with the upcoming PLATO mission (which takes pictures much faster). They found it could detect very small dips in light, getting close to the size of an Earth-like planet.

Why Does This Matter?

Currently, if a planet takes 365 days to orbit its star (like Earth), and a telescope only watches for 27 days, that planet might only cross the star once.

Old Systems: "I can't see it. I need to see it three times to be sure."
ExoVeil: "I know exactly how that star behaves. I saw a dip that doesn't fit the pattern. That's a planet."

The paper also introduces a new way to measure confidence called "Conformal Prediction." Think of this as the system giving you a "confidence score" with a guarantee: "I am 95% sure this list of candidates contains the real planets." This helps astronomers decide which signals are worth spending expensive telescope time to investigate further.

In short: ExoVeil is a new kind of detective that doesn't wait for a suspect to show up twice. It knows the suspect's face so well that it can spot them the very first time they walk by, even in a crowded, noisy room. The code is now open for anyone to use.

Technical Summary: ExoVeil – Detecting Exoplanets Through Learned Stellar Behaviour

1. The Problem
Current machine learning (ML) transit detection systems (e.g., ExoMiner, AstroNet, RAVEN) operate as classifiers that require phase-folded light curves with a known orbital period. This creates a structural blind spot for long-period planets, which may transit only once or twice during an observation window (e.g., a 27-day TESS sector or a two-year PLATO stare). Since these rare events cannot be phase-folded, existing classifiers cannot attempt detection. Furthermore, the most scientifically valuable targets—Earth-analogs orbiting Sun-like stars—fall precisely into this "single-transit" regime. Separating real planetary signals from false positives (eclipsing binaries, instrumental artifacts, stellar variability) remains a critical challenge, particularly when signals are shallow (10–1000 ppm) and buried in noise.

2. Methodology
The paper introduces ExoVeil, a system that reframes transit detection from classification to anomaly detection via prediction. Instead of classifying folded light curves, ExoVeil learns the expected behavior of a star's brightness and flags deviations. The framework consists of four stages:

World Model (Prediction): A causal Transformer encoder (6 layers, 8 heads, ~3.2M parameters) is trained on raw flux time series using transit-masked self-supervised learning. For planet-hosting stars in the training set, known transit regions are masked and replaced with interpolated baselines, ensuring the model never "sees" a transit during training. The model predicts the expected flux ( $\hat{y}_t$ ) and log-variance ( $\log \sigma^2_t$ ) for the next timestep based on history, utilizing continuous time encoding to handle irregular cadence and a learned detrending module to remove low-frequency stellar variability.
Matched-Filter Detection: The system calculates residuals ( $r_t = y_t - \hat{y}_t$ ). A matched-filter detector convolves these residuals with zero-mean box templates of varying durations (3–25 data points). The detector uses the model's predicted variance for inverse weighting, increasing sensitivity in photometrically quiet regions and reducing false triggers in noisy ones.
Learned Classifier: Initial hand-crafted scoring failed (AUC 0.36) because eclipsing binaries produce deeper residuals than planets. Consequently, an XGBoost classifier was trained on 21 features derived from the world model output (e.g., folded residual SNR, variance-normalized depth, epistemic uncertainty ratio) and stellar parameters. This stage separates planets from false positives.
Conformal Prediction & Uncertainty: The system applies split conformal prediction ( $\alpha=0.05$ ) to provide coverage-guaranteed rankings. Uncertainty is decomposed into aleatoric (mean predicted variance) and epistemic (MC Dropout variance across 30 passes) components, categorizing candidates as confident, data-limited, model-uncertain, or ambiguous.

3. Key Contributions

Single-Transit Capability: ExoVeil is the first ML system capable of detecting planets that transit only once, operating directly on raw flux without requiring a known period or phase folding.
Zero-Shot Cross-Mission Transfer: The system demonstrates the ability to generalize across missions. A model trained exclusively on Kepler data successfully detected transits in TESS data (and vice versa in simulation) without retraining or fine-tuning.
Conformal Prediction in Transit Vetting: This work provides the first application of conformal prediction to transit detection, offering formal empirical coverage guarantees (95.9%) and decomposed uncertainty for candidate ranking.
Open-Source Release: The system is released as a Python package (pip install exoveil) with pretrained weights, a candidate catalogue, and a demonstration notebook.

4. Results

Blind Search (Kepler): Applied to 3,737 Kepler stars, ExoVeil identified 179 new transit-like signals not present in the DR25 TCE catalogue. This includes 46 Tier 1 monotransit candidates (single events with SNR > 10). The strongest candidate, KIC 11706231, shows a 231 ppm dip with an SNR of 31.2 on a cool dwarf.
Injection-Recovery: In single-transit injection-recovery tests, ExoVeil recovered 32% of transits at 1000 ppm depth (FP-corrected). By construction, classification-based systems score 0% on this task.
Cross-Mission Validation: When applied to 47 confirmed TESS planets in the PLATO LOPS2 field without retraining, ExoVeil achieved 100% recovery, including long-period planets (e.g., TOI-4562 b, P=225d) and shallow transits (e.g., GJ 238 b, 160 ppm).
PLATO Cadence Simulation: At PLATO's 25-second cadence (simulated with 50 ppm noise), detection sensitivity reached 100 ppm, approaching the Earth-analog regime (84 ppm).
Classification Performance: On the Kepler DR25 test set, the XGBoost classifier achieved an AUC of 0.938, using only flux time series (compared to ExoMiner's 0.98 which utilizes multiple diagnostic inputs).

5. Significance and Claims
The paper claims that ExoVeil addresses a fundamental limitation in current exoplanet detection: the inability to find long-period planets that do not exhibit periodicity within the observation window. By shifting from classification to prediction, the system learns "normal" stellar physics, making planetary transits appear as systematic negative anomalies.

The authors emphasize that while ExoVeil does not replace existing classification systems, it extends detection to regimes they cannot reach. The zero-shot transfer capability suggests the model has learned genuine stellar physics rather than mission-specific artifacts. The integration of conformal prediction offers a new standard for candidate vetting, allowing observers to prioritize targets based on calibrated confidence and decomposed uncertainty, potentially reducing the cost of ground-based follow-up. The paper concludes that with the higher temporal resolution of future missions like PLATO, this approach could make the detection of Earth-analogs from single or few-transit events feasible.