EXOVEIL: The AI System Changing Transit Detection

EXOVEIL is disrupting the way we detect exoplanets, and it's about time. This system doesn't need the standard phase-folded input. Instead, it works directly with raw flux time series. In simpler terms, it predicts what a star's brightness should look like and flags discrepancies. This approach is nothing short of revolutionary.

Breaking New Ground with Transformers

EXOVEIL leverages a Transformer world model trained on a whopping 16,499 Kepler light curves. It employs transit-masked self-supervised learning to predict expected stellar flux. This isn't just another tool in the kit. it's a leap forward in how we understand and detect celestial phenomena.

The system's matched-filter detector, with variance weighting, is adept at extracting transit signals from prediction residuals. That's where the magic happens. A learned classifier, specifically XGBoost, separates the planets from the noise, achieving an Area Under the Curve (AUC) of 0.938 on the Kepler DR25 dataset. That's precision you can't ignore.

Single-Transit Detection: A Game Changer

Here's what's truly remarkable: EXOVEIL recovers 32% of single transits at a 1000 parts per million (ppm) depth. That's where other classification-based systems fall flat at a big fat zero. In a blind search of 3,737 Kepler stars, it identified 179 new transit-like signals. Among these, 46 are monotransit candidates, a treasure trove previously unexplored.

If the AI can hold a wallet, who writes the risk model? When applied to 47 confirmed TESS planets in the PLATO LOPS2 field, EXOVEIL showed a 100% recovery rate. Zero-shot cross-mission transfer isn't just a buzzword here, it's happening. At PLATO's 25-second cadence, detection reaches a stunning 100 ppm. We're inching closer to identifying Earth analogs.

Pioneering Conformal Prediction

EXOVEIL isn't just resting on its laurels. It marks the first application of conformal prediction to transit detection, boasting a 95.9% empirical coverage. And yes, you can pip install exoveil with pretrained weights and access a candidate catalog. So, why should we care? Because this isn't just about finding new planets. It's about redefining how we search and the tools we trust.

Show me the inference costs. Then we'll talk. But in the case of EXOVEIL, it's clear we've got something special. The intersection is real. Ninety percent of the projects aren't. But this one? It's a keeper.