Single-Transit Parameter Estimation with Calibrated Uncertainties

Abstract

The growing volume of photometric data from current and upcoming exoplanet surveys demands automated analysis pipelines to detect and characterize exoplanet transit signals. Among these, single-transit events present a challenge, since only one event is observed, and standard transit-fitting techniques for periodic signals do not directly apply because they rely on phase-folding. Beyond detection, characterizing these events is equally critical for follow-up, yet existing approaches are either sensitive to stellar variability and require prior detrending, or depend on stellar priors and per-target fitting. This work presents a convolutional neural network that predicts transit depth, duration, and mid-transit time, together with calibrated uncertainties, directly from a single transit signal. Our network is trained on synthetic light curves and evaluated against a trapezoidal least-squares baseline. On 120 confirmed exoplanets, both methods achieve comparable median errors, but the baseline fails to locate the transit in 10% of cases, producing unreliable estimates that would propagate into follow-up prioritization. In contrast, our network fails in only 1% of cases, yields lower mean absolute errors, and does not require detrending or stellar priors. Detrending would distort or remove the transit along with high stellar variability, precisely where the baseline fails and our network succeeds. Finally, we identify 169 new transit candidates in TESS Sectors 27-60 and apply our approach to obtain their physical parameters, creating a catalog that includes single-transit, multi-transit and multi-planet systems. This approach offers a scalable path toward automated characterization of single transit events, a population that will grow with upcoming surveys.