Identifying eclipsing binary stars with TESS data based on a new hybrid deep learning model Facilities: TESS

Ying Shan, Jing Chen, Zichong Zhang, Liang Wang, Zhiqiang Zou, Min Li · 2025

Read Identifying eclipsing binary stars with TESS data based on a new hybrid deep learning model Facilities: TESS by Ying Shan, Jing Chen, Zichong Zhang, Liang Wang, Zhiqiang Zou, Min Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-15875.txt.

Introduction

Ying Shan and co-authors train a neural network to spot eclipsing binary stars in data from the TESS space telescope. In these systems, one star passes in front of the other and the light dips in a repeating pattern. TESS has collected far more light curves than humans can inspect by eye.

Their model, CALNet, reads two views of each star at once: the light curve itself and a period chart that highlights repeating cycles. It blends convolutional layers, memory layers, and an attention block to weigh the telling parts of the signal. Trained on 4,225 known systems, it recovered ninety-nine percent of the eclipsing pairs in testing. The team then ran it over four years of TESS observations and, after checking candidates by eye, confirmed 9,351 new eclipsing binaries. The catalog gives astronomers a much larger sample for studying how paired stars evolve.

Catalog record

CREATORYing Shan, Jing Chen, Zichong Zhang, Liang Wang, Zhiqiang Zou, Min Li
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-15875

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Identifying eclipsing binary stars with TESS data based on a new hybrid deep learning model Facilities: TESS Ying Shan, Jing Chen, Zichong Zhang, Liang Wang, Zhiqiang Zou, Min Li Abstract Eclipsing binary systems (EBs), as foundational objects in stellar astrophysics, have garnered significant attention in recent years. These systems exhibit periodic decreases in light intensity when one star obscures the other from the observer’s perspective, producing characteristic light curves (LCs). With the advent of the Transiting Exoplanet Survey Satellite (TESS), a vast repository of stellar LCs has become available, offering unprecedented opportunities for discovering new EBs. To efficiently identify such systems, we propose a novel method that combines LC data and generalized Lomb-Scargle periodograms (GLS) data to classify EBs. At the core of this method is CNN_Attention_LSTM Net (CALNet), a hybrid deep learning model integrating Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and an Attention Mechanism based on the Convolutional Block Attention Module (CBAM). We collected 4,225 EB samples, utilizing their 2-minute cadence LCs for model training and validation. CALNet achieved a recall rate of 99.1%, demonstrating its robustness and effectiveness. Applying it to TESS 2-minute LCs from Sectors 1 to 74, we identified 9,351 new EBs after manual visual inspection, significantly expanding the known sample size. This work highlights the potential of advanced deep-learning techniques in large-scale astronomical surveys and provides a valuable resource for further studies on EBs. Numpy [ 30 ] , Tensorflow [ 1 ] , Keras [ 8 ] Astropy [ 3 ] , Matplotlib [ 12 ] I Introduction Eclipsing binary systems (EBs) are binary star systems in which the components periodically pass in front of each other, resulting in periodic dimming in their light curves (LCs). The periods, shapes, durations, and depths of these LCs reveal the fundamental parameters of the binary systems, such as the stellar radii, temperatures, luminosities, orbital inclinations, and semi-major axes of the binaries. Some EBs are also spectroscopic binaries, allowing the measurement of radial velocity variations over an orbital period. Combining radial velocities with LCs makes it possible to determine the masses and densities of the stars precisely. This method is independent of stellar evolution models, making EBs good laboratories in stellar astrophysics for testing theories of stellar evolution [ 9 ] , distances [ 29 ] , asteroseismic theories [ 11 ] , and gravitational theories [ 4 ] . Over the past decade, ground-based and space-based photometric surveys have significantly advanced the discovery of new EBs and opened a new window into these systems. For example, [ 20 ] presented a catalog containing 4680 EBs in the northern sky using data from Catalina Sky Surveys (CSS). [ 7 ] present a catalog of more than 781,000 periodic variables from Data Release 2 of Zwicky Transient Facility (ZTF), of which 350,000 were classified as EBs. The space-based Kepler mission [ 6 ] nearly continuously observed \sim 2 million stars within a sky field of 105 deg 2 and found over 2,800 eclipsing and ellipsoidal binary systems [ 16 ] . As the successor of Kepler, the Transiting Exoplanet Survey Satellite (TESS) is scanning almost the entire sky with a much larger field of view to search for extra-solar planets [ 23 ] , and its data have been proven to be invaluable for EB search. [ 22 ] identified 4,584 EBs using the first two years of 2-min cadence data (Sectors 1–26). Furthermore, [ 14 ] analyzed the 30-min cadence data in the same sectors and found 3155 OBA-type EB candidates. These works highlight the power of finding new EBs using TESS data. In the analysis of photometric survey data, the huge volume of data poses significant challenges for the identification of new objects, based on the results of binary classification (EBs and non-EBs). Previous studies have demonstrated that machine learning (ML) techniques have great potential for classifying and detecting weak signals in large photometric datasets. For example, [ 21 ] used the artificial neural network (ANN) to analyze the physical properties of over 2200 EBs identified by the Catalina Sky Survey. [ 25 ] presented a machine-learning approach for identifying new transiting exoplanets in the Kepler data. [ 19 ] applied convolutional neural networks (CNNs) to classify the transit signals detected by TESS, enhancing the efficiency of candidate validation. Additionally, [ 18 ] developed a new ML method applied to threshold crossing events (TCEs) in the exoplanet transit signal search, resulting in three new exoplanet candidates. In our study, we collected LC data from TESS and incorporated generalized Lomb-Scargle periodograms (GLS) [ 36 ] to capture periodic features from LC. By combining these two types of data, we developed a novel deep learning method for identifying EBs. The organization of this paper is as follows: Section II introduces the two types of stellar datasets utilized in this study, along with the preprocessing method on these datasets. Section III outlines the framework of the proposed method and provides a detailed explanation of its implementation. Section IV describes the model evaluation metrics, compares the performance of our model with other classical models, and makes some discussions. Section V presents the results of the model application and provides a preliminary analysis of the newly identified EB samples. Section VI summarizes our contributions and discusses the limitations of this study. II Data II.1 Data description TESS conducted high-precision photometry by scanning nearly the entire sky, dividing it into 26 sectors during its two-year Prime Mission. Each sector was observed for approximately 27 days. During this mission phase, TESS collected photometric data at various cadences, including 2-minute LCs for pre-selected targets and 30-minute Full-Frame Images (FFIs). The Extended Mission, which followed the Prime Mission, introduced new data products, such as 10-minute FFIs, further enriching the observational dataset. Target Pixel (TP) files serve as the basis for generating LCs, while Cotrending Basis Vectors (CBVs) are used to correct systematic trends in the data. The most straightforward source of information about a binary system is its photometric LC [ 32 ] . By analyzing LC, the periodic changes of EBs can be measured and detected, such as eclipse fraction, eclipse depth, eclipse duration, and other parameters, to study the orbital motion and characteristics of stars [ 13 , e.g.,] . Therefore, we chose LCs for analysis in our study and downloaded them from MAST 1 1 1 https://archive.stsci.edu/missions-and-data/tess . Each LC file contains several time series, and PDSAP_FLUX was selected as the input for our model. The EBs used for training and testing were selected from the General Catalogue of Variable Stars (GCVS) 5.1 catalog [ 24 ] and the samples published by [ 22 ] from Sectors 1–26 of TESS. We obtained a total of 4,225 EBs. Each star may be observed in multiple sectors, so we have a total of 19,171 LCs for training purposes. Since our work is a binary classification problem, involving distinguishing data into two predefined categories (e.g., positive or negative classes), constructing negative samples is as important as identifying positive samples. It would be insufficient to consider all negative samples as derived solely from the GCVS 5.1 catalog, as this catalog does not fully represent the variety of non-EBs observed by TESS. The primary goal

Read the complete text