Data-driven photometric redshift estimation from type Ia supernovae light curves

TL;DR

This paper presents a novel AutoML-optimized linear regression pipeline for photometric redshift estimation of supernovae, significantly improving accuracy over traditional methods using simulated Dark Energy Survey data.

Contribution

Developed an AutoML-optimized pipeline combining Gaussian process fitting, wavelet feature extraction, PCA, and ensemble learning for supernova redshift prediction.

Findings

Reduced RMSE for supernova redshift prediction from 0.16 to 0.09

Achieved better accuracy than traditional data pre-processing methods

Generated probability distribution functions matching SALT2 model predictions

Abstract

Redshift measurement has always been a constant need in modern astronomy and cosmology. And as new surveys have been providing an immense amount of data on astronomical objects, the need to process such data automatically proves to be increasingly necessary. In this article, we use simulated data from the Dark Energy Survey, and from a pipeline originally created to classify supernovae, we developed a linear regression algorithm optimized through novel automated machine learning (AutoML) frameworks achieving an error score better than ordinary data pre-processing methods when compared with other modern algorithms (such as XGBOOST). Numerically, the photometric prediction RMSE of type Ia supernovae events was reduced from 0.16 to 0.09 and the RMSE of all supernovae types decreased from 0.20 to 0.14. Our pipeline consists of four steps: through spectroscopic data points we interpolate the light curve using Gaussian process fitting algorithm, then using a wavelet transform we extract the most important features of such curves; in sequence we reduce the dimensionality of such features through principal component analysis, and in the end we applied super learning techniques (stacked ensemble methods) through an AutoML framework dedicated to optimize the parameters of several different machine learning models, better resolving the problem. As a final check, we obtained probability distribution functions (PDFs) using Gaussian kernel density estimations through the predictions of more than 50 models trained and optimized by AutoML. Those PDFs were calculated to replicate the original curves that used SALT2 model, a model used for the simulation of the raw data itself.