Using machine learning techniques to identify strong gravitational lensing systems in DES.

Abstract

Gravitational lensing systems enable astronomers to look into the distant universe by magnifying distant objects. Strong gravitational lensing systems are an incredibly rare phenomenon, with only a total about 1; 000 having been discovered. The use of machine learning (ML) has enabled the search for these systems, in the vast sky surveys that currently exist, to be narrowed down. This work investigates the use of ML techniques to identify strong gravitational lensing systems within the Dark Energy Survey (DES). We use the ML technique of convolutional neural networks (CNNs), a deep neural network architecture, as they are able to perform various image processing tasks efficiently. We generated a dataset of 96; 768 images to train and validate our CNN, half of which contains images from DES and the other half containing simulated lenses. The images from DES are scored with a 0, and the simulated lenses with a 1. Our CNN gained an accuracy of 99:73 0:07% and a mean loss of 0:81 after being evaluated against an unseen dataset similar to that of the training data. We also evaluated our CNN against 389 real lenses, and gained an accuracy of 11:92 2:75% and a mean loss of 10:49. Our CNN correctly predicted 57 389 (14:65%) lenses. In this thesis, we present our CNN and the lenses that were correctly identified. In previous studies using CNNs to identify gravitational lensing systems, accuracies between 20% to 40% were achieved, thus the accuracy of 15% achieved by our CNN is competitive. Our results could be improved by training our CNN on the all morphologies of known lensing systems, not only those containing bright arcs. Results could also be improved by ensuring that all the images are centred on the lens, and not the source images. These results and limitations are presented and discussed in the thesis. By looking at accurately simulated, as well as real lenses, one can train the CNN to be more precise.