Using The Convolutional Neural Networks To Search For Strong Lensing Candidates

Gravitational strong lensing effect is one of the most important cosmological probes,and has been widely used to constrain the nature of dark matter and dark energy,and to understand the formation and evolution of galaxies.Because of its low event rate,up to now the known strong lensing systems are still limited.Ongoing and future large surveys will be able to find much more strong lensing systems.However,it is a big challenge to find them efficiently from a vast amount of data.The fast development of the technique of Convolutional Neutral Network(CNN)has provided an important means to overcome the challenge.The application of CNN in searching for strong lensing systems has resulted fruitful achievements.However,there are still issues remained to be explored.In this thesis study,we construct a CNN to automatically search for strong lensing systems from large volume image surveys.The trained CNN is applied to the Luminous Red Galaxies(LRGs)in the Kilo-Degree Survey(Ki DS)Data Release 3 r-band images,and 48 high probability candidates are found after human eye inspections.Among them,27 are identified for the first time.In order to study the dependence on the training sample,we construct a separate training sample,in which the LRG images from Ki DS survey are used as the lens galaxy.The candidates found by the new trained CNN has a 60% overlap with the previous one.Twelve new candidates are additionally found.Moreover,we also test the robustness of our network to the variation of PSF(Point Spread Function)in accordance with the variation of observational conditions.For that,we generate different testing samples by artificially varying the PSFs of the images from 0.4 to 2 times of the median PSF to mimic real observational changes.It is found that our network is rather stable,and the maximum performance degradation is less than 8% in the exam using Ki DS data.We also analyze how the volume of the training sample affects the network performance.