Identifying Exoplanets with Deep Learning. V. Improved Light Curve Classification for TESS Full Frame Image Observations

The discovery of over 5000 exoplanets has relied heavily on human judgment to differentiate true planetary transit signals from various systematic effects and other astrophysical phenomena that mimic them. This manual process is slow, inconsistent, and faces challenges with the increasing volume of astronomical data. This paper addresses these issues by improving upon a deep learning model, Astronet-Triage, which classifies light curves from the Transiting Exoplanet Survey Satellite (TESS) mission to identify potential exoplanets. TESS, with its wide-field survey, generates massive amounts of data, making automated classification crucial for efficient exoplanet discovery. Previous machine learning approaches, like Robovetter and Autovetter for Kepler, and Astronet for various missions, have shown promise. However, the existing Astronet-Triage for TESS suffers from a high rate of false negatives (missed planet candidates). This research aims to significantly improve Astronet-Triage's performance by developing a new model, Astronet-Triage-v2, which reduces false negatives while increasing the rejection of false positives, leading to a more efficient and accurate exoplanet detection process.

Early work in exoplanet detection relied on automated systems like Robovetter (a decision tree) and Autovetter (a random forest classifier) for the Kepler mission. Robovetter proved more robust and was used in producing Kepler's automated planet candidate catalogs. More recently, convolutional neural networks (CNNs), such as Astronet and its variants, have been applied to various datasets including Kepler, K2, TESS, WASP, and NGTS. These models show promise but have room for improvement. Improvements often involve adding input information or modifying data representation. The current work builds upon Astronet, particularly Astronet-Triage, which was used in the TESS Quick-Look Pipeline (QLP) but was found to have a significant number of false negatives. This study improves upon Astronet-Triage to reduce false negatives while improving false positive rejection.

Astronet-Triage-v2 is trained and tested on approximately 25,000 human-vetted transit signals from TESS Sectors 1-39, collected during the Primary and 1st Extended Missions. The dataset incorporates light curves from various aperture sizes, filtered to remove low-frequency variability. The Box Least Squares (BLS) algorithm is used to detect transit signals. Transit signals are carefully labeled with five categories: E (eclipsing signals), S (single transits or incorrect periods), B (contact eclipsing binaries), J (junk – noise or systematics), and N (not sure). The labeling involved multiple independent vetters and consensus-building to ensure high accuracy. The data is then split into training, validation, and test sets. Astronet-Triage-v2 uses a CNN architecture similar to Astronet, but with several key enhancements. It uses multiple detrending settings and creates seven different views of the light curves (global, local, secondary, local half-period, global double period, sample global segments, sample local segments), providing a more comprehensive representation of the data. Scalar data such as period, transit duration, depth, stellar properties, and light curve characteristics are also included as input features. The model is trained using the Adam optimizer and binary cross-entropy loss. An ensemble of 10 independently trained models is used for prediction, combining their results to improve robustness.

On the validation set, Astronet-Triage-v2 achieves an AUC-PR of 0.977, reaching 100% recall at 41% precision and 96.9% recall at 79.8% precision. The test set results show an AUC-PR of 0.965, with 99.6% recall at 39.7% precision and 97.2% recall at 75.7% precision. The model demonstrates good generalization to the 1st Extended Mission data, with an AUC-PR of 0.961 and improved recall compared to Astronet-Triage at similar precision levels. When applied to the TOI catalog, Astronet-Triage-v2 recovers 3577 out of 4140 TOIs at a comparable precision to Astronet-Triage, which only recovers 3349 TOIs – saving at least 200 planet candidates. Analysis reveals that false negatives often arise from borderline label cases (e.g., close eclipsing binaries resembling contact binaries, or noisy transits) and errors in BLS-estimated period and duration, particularly in datasets with multi-year observations.

Astronet-Triage-v2 significantly improves upon its predecessor, Astronet-Triage, by achieving higher recall with comparable or better precision. The model's ability to generalize to unseen data from the 1st Extended Mission demonstrates its robustness and suitability for ongoing TESS data analysis. The recovery of a significantly larger number of TOIs highlights the practical benefits of this improved model. The identification of common sources of false negatives provides valuable insights for further model refinement. While the model doesn't yet differentiate between planets and eclipsing binaries, its enhanced performance makes it a powerful tool for initial triage in exoplanet searches. The availability of the training dataset to the community facilitates further research and improvements.

Astronet-Triage-v2 represents a substantial advancement in automated exoplanet identification from TESS FFI data. Its superior performance over Astronet-Triage, particularly in recovering potential planet candidates and demonstrating good generalization to new data, makes it a valuable tool for exoplanet research. Future work includes incorporating data augmentation techniques to enhance training data, leveraging human vetting results from the TOI catalog, and improving the model's ability to distinguish between planets and eclipsing binaries to move towards a fully automated planet vetting pipeline.

The current model's precision is limited by borderline cases in label assignments (e.g., distinguishing between close binaries and contact binaries, noisy transits, and highly variable stars). Inaccuracies in BLS-estimated period and duration also impact performance. The model also struggles with single-transit events and signals from the SPOC pipeline. The relatively small size of the training set compared to the vast amount of TESS data could also be a limiting factor.