Simulating high-realistic galaxy scale strong lensing in galaxy clusters to train deep learning methods

Galaxy-galaxy strong lensing in galaxy clusters is a unique tool for studying the subhalo mass distribution, as well as for testing predictions from cosmological simulations. We describe a novel method that simulates realistic lensed features embedded inside the complexity of observed data by exploiting high-precision cluster lens models. Such methodology is used to build a large dataset with which Convolutional Neural Networks have been trained to identify strong lensing events in galaxy clusters. In particular, we inject lensed sources around cluster members using the images acquired by the Hubble Space Telescope. The resulting simulated mock data preserve the complexity of observation by taking into account all the physical components that could affect the morphology and the luminosity of the lensing events. The trained networks achieve a purity-completeness level of ∼ 91% in detecting such events. The methodology presented can be extended to other data-intensive surveys carried out with the next-generation facilities.