TESS Astronet V2

The TESS mission produced a large amount of time series data, only a small fraction of which contain detectable exoplanetary transit signals. Deep learning techniques such as neural networks proved effective at differentiating promising astrophysical eclipsing candidates from stellar variabilities, and instrument systematics in an efficient, unbiased and sustainable manner. This talk presents a high quality training data set based on observations from the primary TESS mission full frame images, using a thorough process for manual review and labeling. We used this data set to train a neural network derived from the AstroNet architectures developed by Shallue & Vanderburg (2018) and Yu et al. (2019). We show promising performance results on held-out and extended TESS mission data, as well as the TESS Object of Interest catalog. The new model is currently used for Quick Look Pipeline to filter through its detections. We will also present ongoing work on training a new neural network to distinguish planetary transiting signals from eclipsing astrophysical false positives such as on target, nearby and background eclipsing binaries.