 |
High angular resolution requires large telescope apertures, which in turn often require segmented pupils. This is particularly true for large optical and infrared ground-based telescopes such as the Extremely Large Telescope (ELT). Due to its six large spiders and the segmented surface of its main mirror, the ELT's angular resolution is affected by the presence of pistons which cause phasing errors, thus the importance of measuring and correcting them. In addition to the pistons, the implementation of a fixed two-by-two Shack-Hartmann Wave Front Sensor (SHWFS), optical effects such as Turbulence Residuals (TR), Adaptive Optics RESiduals (AORES) and the main pupil's constant rotation also affect the ELT's Point Spread Function (PSF). Assuming in our study that each of the ELT's main pupil's six differential pistons follows a [-π/2, π/2] uniform distribution (units in rad) and that the wavelength is set to 2.2 μm, we show that convolutional Neural Networks (NN) offer a promising technique for accurately estimating each of the ELT's six differential pistons in the presence of TR. Indeed, each of the Root Mean Square Error (RMSE) values associated to these differential pistons are much lower than 0.1 rad, or equivalently about 35 nm, regardless of the pupil's rotation angle, except for configurations when a spider is close to a SHWFS axis by less than 3° of pupil rotation.
