Deep Learning in Supernova Classification: Evaluating Model Efficacies (Fortino)
Type: Talk
Session: Machine Learning and Artificial Intelligence
Author: Willow Fox Fortino
Abstract: The LSST will soon be delivering an overwhelming influx of transient events nightly and as a result our community needs every advantage when classifying SNe from spectroscopic follow-up observations. This talk discusses a comparison of advanced deep learning methodologies on the classification of SNe subtype (e.g., subtypes of SNe Ia; subtypes of Core Collapse SNe: IIP, IIL; SN Ibc: IIb, Ib, Ic, Ic-broad; and interacting SNe: Ibn and IIn). Specifically, we compare the existing DASH (Muthukrishna et al. 2019) convolutional neural network (CNN) classification model against newly developed transformer-based models. Other model types are also discussed. Preliminary findings suggest a superior capability of transformer models in capturing subtle spectral characteristics vital for accurate classification. We note that the adaptation of a transformer-based neural network to the task of multi-class classification is apparently novel, so this talk will highlight important nuances in the architecture of the model such as the positional encoding we use. This research aims to guide the selection of optimal deep learning strategies for spectral classification in an era dominated by data-rich astronomical surveys, ensuring more precise and resource-efficient SNe categorization.