Emulating effective field theory predictions for galaxy alignments

Abstract

Cosmology advances in the last decades due to recent influxes of observational data have posed to be a challenge for the current numerical methods used. Due to the nature of analysis of this data, speed and accuracy constraints need to be strongly considered. Statistical methods that leverage Artificial Intelligence have started to populate Cosmology literature as an efficient and optimal option. In this work, we will apply these kind of techniques, and more specifically Gaussian Processes for regression to create a fast-prediction framework of the components of the Effective Field Theory model in order to enable prediction of the information provided by intrinsic alignments and how these act as contaminants in cosmological surveys. The aim of this project is to train AI models that are on-par with numerical methods but reduce the computational complexity in several orders of magnitude. We have created a novel emulator for the shape correlations of intrinsic alignments under the EFT theory by implementing a Python package for Gaussian Process emulation with GPU support using the GPyTorch library. We managed to accurately predict the EFT correlators along the power spectrum with relative errors in the sub-percent order. Sequential execution for a single cosmology takes a few seconds while in the parallel case this is reduced to the order of just milliseconds.