Machine learning primordial black hole formation

Abstract

We use machine learning to distinguish between spherically-symmetric perturbations on a superhorizon scale which later collapse to primordial black holes (PBHs) and those which do not, focusing on the evolution of perturbations during radiation domination. We use a number of simulation results to train machine learning models, and test their accuracy using other simulation results which have not been used to train them. Different machine learning models and the parameters of each model lead to different accuracies, but we find that, overall, machine learning can predict PBH formation well. For instance, we show that the error rate can be of the order of ${10}^{\ensuremath{-}3}$. Hence, machine learning could be useful in estimating the abundance of PBHs accurately for each model of the early Universe which can predict PBH formation.