Presentation #402.02 in the session “Mining TESS Data with Machine Learning and Other Advanced Methods”.
Photometric data archives are overflowing with data that no human eye has ever or will ever look at. Automated methods that can effectively filter data are needed so that researchers can focus on the science rather than the search for relevant information. This work uses convolutional neural networks to identify planet candidates from transit signals in the Transiting Exoplanet Survey Satellite (TESS) 30-minute cadence data. We demonstrate the effectiveness of the network to distinguish between light curves which include and exclude transiting events. Perhaps of more interest, we analyze the effectiveness of the network to distinguish between eclipsing binary signals and planetary transit signals. Notably, the features the network learns to look for are entirely self-learned, allow our pipeline to be easily generalized to other types of phenomena. As the inference process requires only a few tens of milliseconds per light curve, the entire TESS 30-minute cadence dataset can be processed in a short period of time, and new data can be processed immediately after being produced. This work additionally compares various network architectures and mechanisms to analyze their impact on the network’s ability to discriminate between TESS light curves.