Rapid classification of TESS planet candidates with convolutional neural networks

Abstract

Aims. Accurately and rapidly classifying exoplanet candidates from transit surveys is a goal of growing importance as the data rates from space-based survey missions increases. This is especially true for NASA's TESS mission which generates thousands of new candidates each month. Here we created the first deep learning model capable of classifying TESS planet candidates. Methods. We adapted the neural network model of Ansdell et al. (2018) to TESS data. We then trained and tested this updated model on 4 sectors of high-fidelity, pixel-level simulations data created using the Lilith simulator and processed using the full TESS SPOC pipeline. With the caveat that direct transfer of the model to real data will not perform as accurately, we also applied this model to four sectors of TESS candidates. Results. We find our model performs very well on our simulated data, with 97% average precision and 92% accuracy on planets in the 2-class model. This accuracy is also boosted by another ∼ 4% if planets found at the wrong periods are included. We also performed 3-and 4-class classification of planets, blended & target eclipsing binaries, and non-astrophysical false positives, which have slightly lower average precision and planet accuracies, but are useful for follow-up decisions. When applied to real TESS data, 61% of TCEs coincident with currently published TOIs are recovered as planets, 4% more are suggested to be EBs, and we propose a further 200 TCEs as planet candidates.

The L 98-59 System: Three Transiting, Terrestrial-size Planets Orbiting a Nearby M Dwarf

“…Overall, our DAVE results rule out false-positive features for all three planet candidates of L 98-59, are consistent with the analysis of the Data Validation Report, and indicate that the detected events are genuine transits associated with the star in question. We also note that while the automated vetting of Osborn et al (2019) flagged L 98-59 with "a high likelihood of being astrophysical false positives" (their Table 3), their subsequent manual vetting lists the system as a planet candidate. Additionally, to investigate whether one or more of the transits associated with L 98-59 may result from nearby sources (e.g., a background eclipsing binary), we used lightkurve to extract light curves for nearby field stars.…”

Nigraha: Machine-learning-based pipeline to identify and evaluate planet candidates from TESS

“…Hence, we chose to pursue a different approach Ansdell et al (2018) show that providing additional domain features can improve model performance for deep learning-based classifiers. Therefore, similar to Osborn et al (2020), we provide the following additional inputs to our model: (i) stellar parameters that we download from MAST catalogues namely stellar effective temperature (T eff ), stellar log g (log g), star radius (R * ), star mass, luminosity, stellar density, and (ii) transit parameters computed by TLS namely, transit depth, transit duration, R p /R s , mean odd/even transit depth. Lastly, similar to Osborn et al (2020), we normalize each of these additional data columns by subtracting the median value and dividing by the standard deviation.…”

“…Of these systems, the Autovetter (McCauliff et al 2015), which is a random forest classifier that makes decisions based on the metrics generated by the Kepler pipeline, has been successfully used to E-mail: sriramsrao@gmail.com (SR); aam@astro.caltech.edu (AM) make initial dispositions for Kepler/K2 candidates. Convolutional neural networks (CNNs) that are effective at image recognition tasks have been explored as an alternate approach for classifying light curves (Pearson, Palafox & Griffith 2017;Ansdell et al 2018;Shallue & Vanderburg 2018;Zucker & Giryes 2018;Dattilo et al 2019;Schanche et al 2019;Yu et al 2019;Osborn et al 2020). These models use supervised learning with a CNN where the neural network learns the characteristics from light curves.…”

“…However, the systematics with individual TESS sectors are significantly different when compared to Kepler/K2 (viz., 27 d of data versus 4 yr/81 d of data). Hence, prior deep-learning approaches for classifying TESS data such as Yu et al (2019), Osborn et al (2020) adapt and modify Astronet and Exonet, respectively, and identify new candidates. Our work is similar in spirit in that we start with a hybrid of these two models and develop novel extensions that allow us to explore different aspects of the input parameter space to identify candidates.…”

Identifying Exoplanets with Deep Learning. III. Automated Triage and Vetting of TESS Candidates

“…Since then, researchers have either modified the original AstroNet model or created their own CNNs to classify candidates from ground-based surveys (Schanche et al 2019) and K2 (Dattilo et al 2019). Osborn et al (2019) registered the first attempt to adapt AstroNet for TESS candidates, but the model was trained on simulated data, which are likely to have very different systematics from real TESS data. As a result, the model suffers a deterioration in performance when applied to real TESS data, recovering about 61% of the previously identified TESS objects of interest.…”