Abstract
A gravitational lens is a distribution of matter, such as dark matter halos, galaxies, or quasars, between a distant light source and an observer that can bend the light from the source as the light travels toward the observer. Nowadays, it is slightly complicated to identify gravitational lenses without powerful computing devices and groups of scientists working together. In addition, future surveys will have orders of magnitude more data and more lenses to find. With up-to-date algorithms such as neural networks, detecting and classifying them for a single human being will be possible. The neural networks described in this paper make the first steps in that direction. The primary purpose of this work was to develop three different neural networks and determine which one could detect gravitational lensing more quickly and precisely. For training, testing, and validation we used a dataset of 2000 images. Half of these images were downloaded from Bologna Lens Factory, a database of simulated gravitational lenses based on galaxies lensed by galaxies (i.e., no clusters and no quasars). We simulated the second half of the images using Python-based code to simulate mock strong lensed galaxies. We used Python-based code to mock strong lensing with different source parameters. Next, we built three types of artificial neural networks and compared their efficiency. Firstly, we developed a fully convolutional neural network (CNN) and a fully connected neural network (FCNN). The third neural network was a combination of these two approaches. In this algorithm, the FCNN layer replaced the last layer of CNN. Next, we compared the learning rates of these algorithms and applied all neural networks to validation images. As a result of the study, we determined which of the developed neural networks fit better for searching gravitational lenses.