Search for exocomets transits in Kepler light curves: Ten new transits identified

TL;DR

The paper develops a neural-network–based pipeline to search Kepler light curves for exocomet transits, trained on a sizable library of theoretical transit templates. By processing ~201,820 Kepler stars and applying rigorous shape and quality filters, the authors produce a three-tier catalog comprising 17 high-confidence transits, 30 possible events, and 49 symmetric-like cases. The work challenges the notion that exocomet activity is confined to young stars by showing a broad host-age distribution, including red giants, and demonstrates a scalable approach for rare transient detection in large photometric surveys. The methodology and catalogs lay groundwork for applying similar searches to upcoming missions like TESS and PLATO, enabling broader statistical studies of exocomet activity across stellar populations.

Abstract

The Kepler mission, despite its conclusion over a decade ago, continues to offer a rich dataset for uncovering new astrophysical objects and phenomena. In this study, we conducted a comprehensive search for exocometary transit signatures within the Kepler light curves, using a machine learning approach based on a neural network trained on a library of theoretical exocomet transit light curves. By analyzing the light curves of 201,820 stars, we identified candidate events through the neural network and subjected the output to filtering and visual inspection to mitigate false positives. Our results are presented into three catalogs of increasing ambiguity. The first-tier catalog includes 17 high-confidence exocometary transit events, comprising 7 previously reported events and 10 newly identified ones, each associated with a different host star. The second-tier catalog lists 30 lower-confidence events that remain consistent with possible exocometary transits. The third-tier catalog consists of 49 more symmetric photometric events that could be either exocometary transits, exoplanet mono-transits, or false positives due to eclipsing binaries mimicking transits. Contrary to previous studies, which suggested that the cometary activity was favored by stellar youth, we find a broad age distribution among candidate host stars, including several red giants. This challenges the general idea of a decline in cometary activity with stellar age and underlines the need for further investigation into the temporal evolution of exocometary activity in planetary systems.

Figures (10)

• Figure 1: Architecture of our best performing neural network model. The input data is a light curve of 480 timesteps, the output is a list of two reals between 0 and 1: one for the probability of detection of an exocometary transit and one for the position of the transit in the light curve. Convolutional layers are noted Conv <kernel size> <number of feature maps>, max pooling layers are noted MaxPool <window length> <stride length>, fully connected layers are noted Dense <number of units>, LSTM layers are noted LSTM <units> <dropout>, GRU layers are noted GRU <unit> <droupout> <recurent dropout> and the squeezed-excitation blocks are noted SEB.
• Figure 2: The receiver-operator characteristic (ROC) curve (left panel) and the precision versus recall curve (right panel). The receiver-operator characteristic (ROC) curve shows the recall (true-positive rate) of the model against the ability to recognize false positives (the false-positive rate) for different classification thresholds. Our model is highly successful at identifying false positives as shown by the high AUC value (see Table \ref{['Performances']}). The plot of the fraction of exocomets that the model classified as exocomets (recall) versus the fraction of correctly classified planets (precision) shows the trade-off between having no false positives (high precision) and identifying all exocomet transits (high recall).
• Figure 3: Histogram of the results on the test set and error on the transit position found by the network. Top panel: Histogram of the result of the neural network applied to the test data. Most of the 1-labeled light curves with an exocomet transit yield a probability close to 1, while there are only a few false positives (0-labeled light curves yielding a high probability). Above the chosen threshold, there is more than two order of magnitude between the number of true positives and false positive. Bottom panel: the mean error of the position of the transit as a function of the probability of the presence of a transit assigned by the network to the light curves. This error is of a few hours at most for the transits identified by the algorithm.
• Figure 4: Distribution of the value of the parameters obtained by fitting the 2163 light curves in the library of the simulated exocomet transits. The parameters $K$, $\Delta T=t_1-t_0$, and $1/\beta$ correspond to the depth, ingress duration, and crossing time of the transit, respectively.
• Figure :