Speaker
Description
Uncertainties in the mass-loss history model for stellar evolution hinder the development of a comprehensive understanding of the scenarios linking different supernova types to their progenitors.
This work aims to predict the physical parameters of the progenitor directly from time sequences of synthetic multi-band images. The focus is on redshift estimation.
We propose a generative method based on learning an embedding representation of the entire event. This embedding is generated using a Generative Adversarial Network (GAN) architecture, where the generator incorporates recurrent neural network (RNN) layers to model the event's temporal evolution. Training this particular type of GAN — known as a Time-GAN — to generate new samples enables the encoder to learn to produce robust embeddings of supernova events. These embeddings are then used to predict progenitor parameters.
Preliminary tests on simulated data show that the learned embedding encodes sufficient information to accurately recover parameters such as redshit, comparable to those obtained using standard light-curve fitting techniques. The key contribution of this work is the development of a general embedding representation that can be used as input for various models, including those used for classifying and forecasting time-evolving sources. Moreover, this methos is scalable to the upcoming large-scale time-domain surveys, such as LSST, this approach will enable the fast, automated and physically grounded analysis of millions of transients — a task that would otherwise be unfeasible using traditional methods.