A deep learning approach to cosmological dark energy models

Abstract

We propose a novel deep learning tool in order to study the evolution of dark energy models. The aim is to combine two architectures: the Recurrent Neural Networks (RNN) and the Bayesian Neural Networks (BNN), we named this full network as RNN+BNN. The first one is capable of learning complex sequential information to classify objects like supernovae and use the light-curves directly to learn information from the sequence of observations. Since RNN is not capable to calculate the uncertainties, BNN emerges as a solution for problems in deep learning like, for example, the overfitting. For the trainings we use measurements of the distance modulus $μ(z)$, such as those provided by Pantheon Supernovae Type Ia. In view of our results, the reported approach turns out to be a first promising step on how we can train a new neural network that can compute their own confidence regions for specific cosmological data. It is worth stressing that the new technique allows to reduce the computational load of expensive codes for dark energy models and probe the necessity of modified dark energy models at large redshifts for a supernovae trained sampler.