Enhancing Photometric Classification in Cosmology A Study of the ParSNIP Model on Real and Simulated Data

Abstract

This thesis investigates the application of the ParSNIP machine learning model for photometric classification of astronomical transients, focusing on supernovae. Leveraging both synthetic and real datasets, the study assesses ParSNIP’s capacity to accurately distinguish transient types based on light curves, achieving an average classification accuracy of over 85% for well-represented transient types, such as SNe Ia and core-collapse supernovae, and a high Area Under the Curve (AUC) of up to 0.973 for Type Ia supernovae classification. These metrics are critical for the model’s application in next-generation time-domain surveys like the Roman High Latitude Time Domain Survey (HLTDS), where precise photometric classification supports dark energy research. Initially validated on the simulated Roman Hourglass dataset, ParSNIP shows consistent performance across diverse transient classes, aided by its innovative physics layer, which provides redshift-invariant latent representations.To further validate the model, its performance is tested on real observational data from surveys with the addition of spectral data converted into 'pseudo' light curves, using MCMC-based light curve fitting in SNCosmo for accurate parameter estimation. This integration yields improved posterior parameter distributions and enhances classification reliability. By comparing ParSNIP’s performance on synthetic and real data, this study demonstrates the model’s adaptability to real-world complexities, reinforcing its potential for large-scale cosmological surveys. These findings advance photometric classification methodologies essential for studying cosmic expansion and dark energy through high-cadence surveys.