Neural network prediction of model parameters for strong lensing samples from Hyper Suprime-Cam Survey

Strong lensing of background galaxies provides important information about the matter distribution around lens galaxies. Traditional modelling of such strong lenses is both time and resource intensive. Fast and automated analysis methods are the need of the hour given large upcoming surveys. In this work, we build and train a simple convolutional neural network with an aim of rapidly predicting model parameters of gravitational lenses. We focus on the inference of the Einstein radius, and ellipticity components of the mass distribution. We train our network on a variety of simulated data with increasing degree of realism and compare its performance on simulated test data in a quantitative manner. We also model 182 gravitational lenses from the HSC survey using {\sc YattaLens} pipeline to infer their model parameters, which allow a benchmark to compare the predictions of the network. Given all considerations, we conclude that the network trained on simulated samples with lensed sources injected in empty HSC cutouts is the most robust, reproducing Einstein radii with an accuracy of about $10-20$ percent, a bias less than 5 percent, and an outlier fraction of the order of 10 percent. We argue in favour of the subtraction of the lens light before modelling the lens mass distribution. Our comparisons of the inferred parameters of 10 HSC lenses previously modelled in the literature, demonstrate agreement on the Einstein radius. However, the ellipticity components from the network as well as the individual modelling methods, seem to have systematic uncertainties beyond the quoted errors.