Print button

Abstract EANA2024-61


DART-Vetter: A Deep LeARning Tool for automatic vetting of candidates detected in transiting surveys

Stefano Fiscale (1,2,3), Laura Inno (2,3), Alessandra Rotundi (1,2), Angelo Ciaramella (2), Alessio Ferone (2), Christian Magliano (4,3), Luca Cacciapuoti (5), Giovanni Covone (4,6,3), Maria Teresa Muscari Tomajoli (1,2), Vito Saggese (4), Luca Tonietti (1,2,3,7), Antonio Vanzanella (2), Vincenzo Della Corte (3)
(1) UNESCO Chair ``Environment, Resources and Sustainable Development'', Department of Science and Technology, Parthenope University of Naples, Italy (2) Department of Science and Technology, Parthenope University of Naples, Italy (3) INAF, Osservatorio Astronomico di Capodimonte, Naples, Italy (4) Department of Physics “Ettore Pancini”, University of Naples Federico II, Naples, Italy (5) European Southern Observatory, Karl-Schwarzschild-Strasse 2 D-85748 Garching bei Munchen, Germany (6) INFN section of Naples, Naples, Italy (7) Department of Biology, University of Naples Federico II, Naples, Italy


The advent of transiting surveys such as NASA's Kepler and Transiting Exoplanet Survey Satellite (TESS), designed to collect huge amounts of photometric observations, along with the forthcoming ESA PLAnetary Transits and Oscillations of stars (PLATO) mission, has highlighted the need to automate the exoplanet detection process. Machine Learning (ML) techniques have proven their robustness and efficiency in analyzing thousands of potential planetary signals.
We will present DART-Vetter (Deep LeARning Tool for automatic vetting), a ML model designed to distinguish planetary candidates from false positives in transiting survey data. DART-Vetter is a Convolutional Neural Network we trained on publicly available Kepler and TESS data and processes only light curves folded on the period of the detected signal. We applied the model to various sets of exoplanets with independent validation to assess its predictive performance. DART-Vetter achieved precision and recall rates of 84% and 88%, respectively. These results indicate the appreciable capabilities of our model to accurately identify real planetary signals in data from current and future space telescopes. 
We then plan to leverage the robust predictive performance achieved by DART-Vetter by extending its application to the identification of Earth-like candidates within a subset of TESS data.