More than 10,000 exoplanet candidates identified from TESS data using machine learning

A team led from Princeton has reduced 80 million light curves to 2.5 million using machine learning, cutting the analysis time from three and a half years to just a few weeks.

The TESS satellite, launched in 2018, produces far more measurements than astronomers can remove by hand. A team led by Joshua Roth, PhD student at Princeton University, with collaborators from MIT, UCLA and the Las Campanas Observatory in Chile, applied an automated processing chain to the catalog of the first observation cycle.

The sorting took place in two stages: the Cambridge Exoplanet Transit Recovery algorithm, and then a random forest classifier that reduced 80 million light curves to about 2.5 million worthy of examination. At the end of the chain, more than 10,000 planetary candidates, including some 9,000 gas giants — especially "hot Jupiters" — and about 100 Neptune-type objects and eleven potential super-Terres. What would have required three and a half years of analysis took place in a few weeks; Convolutive neuron networks must take over from the classifier.

The method is timely. The Roman space telescope, scheduled for launch in August 2026, alone could deliver about 100,000 candidates in transit, and the European Plato mission must follow early 2027 with higher photometric accuracy. Without automatic sorting, these catalogues would remain largely untapped.