Characterising Dark Matter Substructure in Gravitational Lens Galaxies with Deep Learning

Abstract

. We investigate the novel application of two sequential convolutional neural networks (CNNs) for the char-acterisation of dark matter substructure in lensing galaxies from galaxy-galaxy strong gravitational lensing images. In our configuration, an initial CNN predicts the number of substructures from a gravitationally lensed image and then this number, along with the same image, is input to a second CNN which predicts the power-law slope of the substructure mass distribution function. We have trained and tested the CNNs on simulated images created by lensing a galaxy-like light distribution with a foreground galaxy mass. We find that training and testing the CNNs on images created with a fixed lens geometry allows the number of substructures and the mass function power-law slope to be retrieved well. We then explore the effect of reducing the resolution of images such that the image pixel scale is halved finding that the accuracy of the number of predicted substructures decreases by only 7% while the accuracy of the predicted mass function slope decreases by 25%. When we allow variation in lens geometry between images in the test set, to mimic more physically motivated lens samples, we observe a decrease in accuracy of the number of predicted substructures and the mass function slope of 57% and 81% respectively. We attribute this significant degradation in predicting the mass function power-law slope to the degradation in the performance of the number-predicting CNN by comparing with predictions of the slope that are made when the CNN is given the true number of substructures. We discuss future possible improvements and the impact of the computing hardware available for this work.