In detection methods involving massive dataset, such as the transit method, AI streamlines the processing of information and helps identify false signals that could be mistaken for real planets, effectively accelerating the validation process and considerably reducing costs.
We used NASA’s Kepler Objects of Interest (KOI) database due to its periodicity and its high-utility targets and labels for predicting exoplanets using GAN.
Start from Kepler long-cadence light curves fetched with Lightkurve using each target’s KIC/kepid. For every star we download and stitch quarters, then flatten to remove slow instrumental/trend systematics and remove outliers so single spikes don’t dominate the learning signal. Next, we run a Box-Least-Squares (BLS) search on periods in [1, 20] days to estimate the best period, transit epoch (t₀), and duration; using these, we produce a folded representation of the series that concentrates the transit shape into phase space. The resulting 1-D sequence is then normalized to [-1, 1] to place all stars on a comparable scale, and finally padded or trimmed to a fixed length CURVE_LENGTH so the network can operate on uniform tensors of shape (L, 1).