Exoplanet Explorer: an open two-stage screening pipeline for transiting exoplanets, and what changes when a light-curve CNN meets real Kepler/K2 data

Abstract Transit surveys record the brightness of many stars, but planets are rare and their dips are small, and a light curve alone cannot reveal what a planet’s atmosphere is made of. We present Exoplanet Explorer, a small, open and fully tested web application that organises the problem as a two-stage pipeline: detection (a 1-D convolutional neural network for light curves and a rule-based video dip detector) and composition (a transmission-spectrum molecule classifier). We retrain the light-curve CNN on real Kepler/K2 photometry (Kaggle “Kepler Labelled Time Series Data”; 5,087 training stars with 37 planets) using exactly the preprocessing the application uses, and evaluate it once on a held-out set of 570 stars with 5 planets. The model ranks planets well (ROC AUC 0.933; average precision 0.412 against a random baseline of 0.009) but at a fixed threshold of 0.5 recovers only 1 of 5 planets. Stratified 5-fold cross-validation over all 5,657 stars (42 planets) confirms this: mean recall 0.12 ± 0.11, average precision 0.21 ± 0.09, ROC AUC 0.84 ± 0.04, with 6 false alarms among 5,615 non-planet stars. A model trained on simulated light curves had recovered 6 of 6 simulated planets; we show that the simulated transits were about three times deeper than the median real transit, so simulated performance was an optimistic upper bound. The video tool classifies all four synthetic test videos correctly. We report these results without threshold tuning on test data, discuss why the problem is only partly solved, and release all code and tests.