Astrophysicists Presented A Deep Learning Method, “SNIascore” For Classification Of Supernova (Astronomy)

Modern time-domain surveys, such as the Zwicky Transient Facility (ZTF), the All-Sky Automated Survey for Supernovae (ASAS-SN) and the Asteroid Terrestrial LastAlert System (ATLAS), are now finding tens of thousands of transients every year.

However, without spectroscopic classifications these discoveries are of limited value. The ZTF Bright Transient Survey (BTS) is addressing this through the deployment of a fully automated very-low-resolution spectrograph, the Spectral Energy Distribution Machine (SEDM) mounted on the Palomar 60-inch telescope. SEDM is capable of obtaining spectra of several thousands of transients per year in the magnitude range between 18 and 19 mag. Currently, the goal of the BTS is to maintain spectroscopic classification completeness for all extragalactic transients detected by the public ZTF survey that become brighter than 18.5 mag (∼ 1000 SNe per year).

The classifications from the BTS are made public on a daily basis via the Transient Name Server (TNS1). These classifications have up until now been based on manual matching of observed spectra to spectral templates using mainly the SuperNova IDentification (SNID) code, along with careful inspection of each obtained spectrum. This makes classification of thousands of supernovae (SNe) a very time-consuming endeavor.

Figure 1. Network architecture of SNIascore. They utilize heavy dropout throughout the network and a combination of two BiLSTM layers surrounding one GRU layer. For regression the final softmax and classification layers are replaced by a regression layer. © Fremling et al.

Due to their inherent brightness, the majority of the extragalactic transients discovered by a magnitude limited survey will be thermonuclear supernovae (SNe Ia). Now, Fremling and colleagues presented SNIascore, a deep-learning based method optimized to identify SNe Ia using SEDM spectra and determine their redshifts without any human interaction. The intended use case for SNIascore is to provide live spectroscopic classification of SNe Ia during the night when SEDM is observing for the BTS.

“The goal of SNIascore is fully automated classification of SNe Ia with a very low false-positive rate (FPR) so that human intervention can be greatly reduced in large-scale SN classification efforts, such as that undertaken by the public Zwicky Transient Facility (ZTF) Bright Transient Survey (BTS).”

SNIascore is based on a recurrent neural network (RNN) architecture with a combination of bidirectional long short-term memory (BiLSTM) and gated recurrent unit (GRU) layers. They found that, SNIascore achieves a < 0.6% FPR while classifying up to 90% of the low-resolution SN Ia spectra obtained by the BTS.

Also, it simultaneously performs binary classification and predicts the redshifts of secure SNe Ia via regression (with a typical uncertainty of < 0.005 in the range from z = 0.01 to z = 0.12). For the magnitude-limited ZTF BTS survey (≈ 70% SNe Ia), deploying SNIascore reduces the amount of spectra in need of human classification or confirmation by ≈ 60%.

Furthermore, SNIascore allows SN Ia classifications to be automatically announced in real-time to the public immediately following a finished observation during the night.

“For future versions of SNIascore we plan to add the option to include lightcurve information as input. Looking at objects that change SNIascore significantly when the lightcurve is included or excluded may turn out to be a way to identify rare events that would warrant further followup from larger facilities.”

— concluded authors of the study

Featured image: Redshift (left) and spectral phase (right) distributions for SNe Ia in the unaugmented SNIascore training (blue) and validation (red) datasets used to optimize SNIascore for classification. The SN Ia distributions for the testing dataset are shown in black. The blue line in the left panel shows the redshift distribution of the training set after augmentation, which is used to train SNIascore for redshift regression. The phase is relative to the time of maximum light in the g or r band, depending on which band is brighter. © Fremling et al.


Reference: Christoffer Fremling, Xander J. Hall, Michael W. Coughlin, Aishwarya S. Dahiwale, Dmitry A. Duev, Matthew J. Graham, Mansi M. Kasliwal, Erik C. Kool, Adam A. Miller, James D. Neill, Daniel A. Perley, Mickael Rigault, Philippe Rosnet, Ben Rusholme, Yashvi Sharma, Kyung Min Shin, David L. Shupe, Jesper Sollerman, Richard S. Walters, S. R. Kulkarni, “SNIascore: Deep Learning Classification of Low-Resolution Supernova Spectra”, ArXiv, pp. 1-12, 2021. https://arxiv.org/abs/2104.12980


Note for editors of other websites: To reuse this article partially or fully kindly give credit to our author S. Aman or either provide a link of our article

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s