A Photometric Classifier for Tidal Disruption Events in Rubin LSST

TL;DR

The paper tackles identifying tidal disruption events (TDEs) using photometric data in Rubin LSST-scale surveys. It combines Gaussian Processes with a $2D$ Matérn-3/2 kernel to model light curves and derive features such as rise/fade times, colors, and GP hyperparameters, which are then fed into an XGBoost classifier tuned for high precision. On ELAsTiCC2 LSST-like simulations, the approach achieves up to 95% precision and about 72% recall for TDEs with minimal non-TDE contamination, enabling practical TDE catalogs for multi-messenger follow-up. This method offers a scalable pathway for clean TDE candidate samples, with future work focusing on real data validation and incorporating additional features like photometric redshifts and host-galaxy information.

Abstract

Tidal Disruption Events (TDEs) are astrophysical phenomena arising when stars are disrupted by supermassive black holes. The Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST), with its unprecedented depth and cadence, will detect thousands of TDEs, motivating the need for robust photometric classifiers capable of efficiently distinguishing these events from other extragalactic transients. We aim to develop and validate a machine learning pipeline for photometric TDE identification in LSST-scale datasets. Our classifier is designed to provide high precision and recall, enabling the construction of reliable TDE catalogs for multi-messenger follow-up and statistical studies. Using the second Extended LSST Astronomical Time Series Classification Challenge (ELAsTiCC2) dataset, we fit Gaussian Processes (GP) to light curves for feature extraction (e.g., color, rise/fade times, GP length scales). We then train and tune boosted decision-tree models (XGBoost) with a custom scoring function emphasizing high-precision recovery of TDEs. Our pipeline is tested on a diverse simulation of transient and variable events, including supernovae, active galactic nuclei, and superluminous supernovae. We achieve high precision (up to 95%) while maintaining competitive recall (about 72%) for TDEs, with minimal contamination from non-TDE classes. Key predictive features include post-peak colors and GP hyperparameters, reflecting characteristic timescales and spectral behaviors of TDEs. Our photometric classifier provides a practical and scalable approach to identifying TDEs in forthcoming LSST data. By capturing essential color and temporal properties through GP-based feature extraction, it enables efficient construction of clean TDE candidate samples.

Figures (14)

• Figure 1: Sample population after the light-curve quality cut. Of the remaining objects, 2878 (0.6%) are TDEs.
• Figure 2: Example light curve with a GP fit. We show the rise and fade times as well as the pre- and post-peak $g-r$ color within the rise and fade times.
• Figure 3: Success rate of feature extraction for TDEs and non-TDEs.
• Figure 4: Scatter plot of the mean post-peak g-r (horizontal axis) vs. its rate of change (vertical axis), illustrating how TDEs (blue crosses) tend to cluster around smaller color values and lower color evolution rates, in contrast to other transient classes. Contours denote the 50%, 80%, and 95% highest-density regions of the per-class 2D kernel density estimate.
• Figure 5: 2D distribution of the GP hyperparameters LengthScale_Time (vertical axis) vs. LengthScale_Wavelength (horizontal axis), showing how each transient class occupies distinct regions in this parameter space. TDEs (blue crosses) and SLSNe (pink diamonds) typically have moderate LengthScale_Time and LengthScale_Wavelength values, reflecting narrower spectral and temporal variations compared to SNe and AGNs. Contours denote the 50%, 80%, and 95% highest-density regions of the per-class 2D kernel density estimate.