Abstract
We present ExoMiner++, an enhanced deep learning model that builds on thesuccess of ExoMiner to improve transit signal classification in 2-minute TESSdata. ExoMiner++ incorporates additional diagnostic inputs, includingperiodogram, flux trend, difference image, unfolded flux, and spacecraftattitude control data, all of which are crucial for effectively distinguishingtransit signals from more challenging sources of false positives. To furtherenhance performance, we leverage transfer learning from high-quality labeleddata from the Kepler space telescope, mitigating the impact of TESS's noisierand more ambiguous labels. ExoMiner++ achieves high accuracy across variousclassification and ranking metrics, significantly narrowing the search spacefor follow-up investigations to confirm new planets. To serve the exoplanetcommunity, we introduce new TESS catalogs containing ExoMiner++ classificationsand confidence scores for each transit signal. Among the 147,568 unlabeledTCEs, ExoMiner++ identifies 7,330 as planet candidates, with the remainderclassified as false positives. These 7,330 planet candidates correspond to1,868 existing TESS Objects of Interest (TOIs), 69 Community TESS Objects ofInterest (CTOIs), and 50 newly introduced CTOIs. 1,797 out of the 2,506 TOIspreviously labeled as planet candidates in ExoFOP are classified as planetcandidates by ExoMiner++. This reduction in plausible candidates combined withthe excellent ranking quality of ExoMiner++ allows the follow-up efforts to befocused on the most likely candidates, increasing the overall planet yield.