Detecting Strongly-Lensed Supernovae in Wide-field Space Telescope Imaging via Deep Learning
Fawad Kirmani, Arjun Karki, Steve Rodney, Kyle Lackey, Varsha P. Kulkarni, John R. Rose, Justin Pierel
TL;DR
This study addresses the challenge of detecting strongly-lensed supernovae in wide-field space-based imaging by shifting from magnification-based methods to identifying multiply-imaged SNe via distortions in single-epoch difference images. A CNN, inspired by VGG19 and augmented with an attention layer, processes 32x32 pixel triplets across four filters to classify images into zero, single, or doubly-imaged SNe, trained on synthetic data derived from HST CANDELS fields. The model achieves near-perfect ROC-AUC (~0.99) and high recall for the doubly-imaged class, outperforming traditional SExtractor-based morphology measures in distinguishing between the three classes. However, the reliance on simulated data highlights potential generalization gaps to real Roman-era observations, motivating future work in domain adaptation, hybrid datasets, and realistic instrumental effects. Overall, the approach enables early, single-epoch identification of lensed SNe for rapid follow-up, maximizing scientific returns from upcoming space missions like the Roman Space Telescope.
Abstract
Gravitationally lensed supernovae (SNe) are extremely rare and fade quickly; as a result, they are challenging to detect. To identify lensed SNe in large imaging datasets, current surveys primarily rely on the {\it magnification} effect of gravitational lensing -- searching for transients that appear brighter than expected \cite{c3}. In this work, we present a proof-of-concept study that uses a deep neural network to classify previously detected transients. Instead of relying on magnification, this network aims to identify doubly-imaged SNe with small separations ($<0.6$ arcsec) based on the {\it distorted shape} of the transient object. This proposed method is most applicable to space-based imaging surveys from wide-field imaging observatories such as the upcoming Roman Space Telescope. To train and test our network, we use archival Hubble Space Telescope (HST) imaging surveys. Due to the extreme rarity of lensed SNe, we cannot train a neural network on actual lensed SN data. Instead, we have used HST imaging data to generate simulated imaging datasets for both training and testing. Our simulations use astrophysical priors to define the separations, relative brightnesses, and colors of each multiply-imaged SN. We have also simulated false positives (image artifacts and unlensed supernovae), which are much more prevalent than true lensed SN. Our deep learning model is trained to identify lensed SNe from a single difference image (i.e., not using multiple epochs). This network achieves a recall score of 99\% on simulated gravitationally lensed SNe. The network successfully distinguishes between single supernovae (SNe) and those with gravitationally lensed SNe, as well as images with zero SNe, achieving recall scores of 90\% and 96\% for single-SNe and zero-SNe images, respectively.