Physics-Informed Deep Learning for Exoplanet Transit Detection: Discovery of a Sub-stellar Candidate Orbiting the Unvetted K-dwarf TIC 200668668

The search for transiting exoplanets in vast photometric surveys such as NASA's Transiting Exoplanet Survey Satellite (TESS) requires automated, high-precision detection algorithms capable of disentangling shallow transit signatures from stellar variability, instrumental systematics, and eclipsing binary false positives. Here, we present PINN-TransNet v1, a physics-informed dual-branch deep learning architecture that couples 1D normalized time-series flux with 2D continuous wavelet transform (CWT) Ricker scalograms, regularized by analytical Mandel-Agol transit geometry and Keplerian orbital dynamics penalties. Evaluated on a locked 1,000-target benchmark of TESS Sector 1 light curves with zero star-level data leakage, PINN-TransNet achieves a receiver operating characteristic area under curve (ROC-AUC) of 0.941 and precision-recall area under curve (PR-AUC) of 0.918, compared to 0.782 and 0.714 for standard Astropy BLS. At 95% target recall, our physics constraint reduces the false positive rate from 23.8% to 4.2% (5.6x reduction). Deploying our calibrated architecture on unvetted TESS Sector 1 stars devoid of previous cataloged planet candidates or Threshold Crossing Events (TCEs) in the NASA Exoplanet Archive, we report the discovery of a candidate orbiting the nearby K-dwarf TIC 200668668. All datasets, model checkpoints, and reproduction codes are made openly available under an open-source MIT license.