Speaker
Description
Speaker : Arpit Suryavanshi, IIT Hyderabad (ep24btech11005@iith.ac.in)
Co-Author : Shashvita Ratra , IISER Mohali (ms24190@iisermohali.ac.in)
Co-Author : Kuntal Misra, ARIES Nainital (kuntal@aries.res.in)[Mentor]
ABSTRACT
Time-domain surveys produce extensive supernova spectroscopic data, which are commonly classified using template-matching tools (e.g. SNID). As the volume and diversity of datasets increase, template-driven classification becomes harder to scale, motivating an investigation into machine learning-based approaches that capture diverse observations and underlying spectral patterns directly from large datasets.
We present a CNN-based hierarchical two-stage model for supernova classification using the publicly available WISeREP dataset. Around 53,233 spectra underwent the standard spectral pre-processing pipeline. Our resulting classifiers achieved 90.8% accuracy (macro-F1 = 0.796) across four major supernova families with ~80% subtype classification accuracy. Across ten classes (Type II, Ia, Ibc, and SLSN and subtypes), an accuracy of 87.2% and a macro-F1 of 0.684 were recorded. Refinement techniques such as class-imbalanced learning were identified as solutions to class imbalance and ambiguity among spectroscopically similar subtypes.
Our two-stage CNN-based hierarchical Supernova classifier provides a robust and scalable alternative to traditional template-driven classification algorithms for the increasingly large spectroscopic datasets.
| Nature of work | Computational |
|---|