A Machine Learning Upgrade to GPR for EoR 21-cm signal extraction from LOFAR data
Presenter: Anshuman Acharya
Title: A Machine Learning Upgrade to GPR for EoR 21-cm signal extraction from LOFAR data
Date/Time: Thursday, June 20th 2:00 PM
Abstract: Within the LOFAR Epoch of Reionization (EoR) Key Science Project team, Gaussian Process Regression (GPR) has been used for foreground subtraction from data, to constrain the Neutral Hydrogen 21-cm signal power spectrum from the EoR. To avoid signal loss due to the misestimation of the covariance kernel for the 21-cm signal, we developed a Machine Learning (ML) trained model using a Variational Auto-Encoder. We trained on a large variety of N-body + 1D radiative transfer simulations (GRIZZLY) and tested its limitations by exploring a variety of mock datasets. Further, this setup was applied to the mock data for the Science Data Challenge 3a of SKA by the DOTSS-21 team finishing with one of the highest scores. Now I will show the results of implementing it on 10 nights of observational data from the LOFAR telescope at z~9.1. If possible, I will also discuss its implementation for multi-redshift analyses. Additionally, I will discuss avenues of improving and expanding the training sets used, by developing new simulations.