Real-time tomographic reconstructor based on convolutional neural networks for solar observation

Solar observation is the branch of astronomy devoted to the study of the Sun. When the light wavefront that comes from the Sun penetrates the atmosphere, it suffers some distortions caused by optically turbulent layers that change the wavefront's shape and morphology. Therefore, in order to obtain a good-quality image, it is necessary to correct the induced error. This is done by applying adaptive optics (AO) techniques. In the case of the present research, it is performed with the help of a Single Conjugate Adaptive Optics System (SCAO). The reconstruction technique proposed in this research is a SCAO based on convolutional neural networks (CNNs). This research develops and assesses a real-time tomographic reconstructor based on CNN, able to correct the error introduced by the atmosphere in the light wavefront received from the Sun. The CNN was trained and validated using data from the Durham AO Simulation Platform as input information. This platform incorporates certain solar functionalities that have been employed in the present research, allowing us to simulate a solar telescope. The normalized errors obtained for both ReLu and Leaky ReLu kernels were promising, without showing statistically significant differences among kernels in the value of RMSE volts of the deformable mirror commands. When different kernel dimensions are compared, statistically significant differences are found, showing that RMSE volts of the deformable mirror commands are lower for 3 x 3 kernels when compared with those of dimensions 5 x 5 and 7 x 7. As far as the authors know, this is the first time that an AO system based on CNN has been developed for solar telescopes.