Neural networks for image-based wavefront sensing for astronomy.

We study the possibility of using convolutional neural networks for wavefront sensing from a guide star image in astronomical telescopes. We generated a large number of artificial atmospheric wavefront screens and determined associated best-fit Zernike polynomials. We also generated in-focus and out-of-focus point-spread functions. We trained the well-known "Inception" network using the artificial data sets and found that although the accuracy does not permit diffraction-limited correction, the potential improvement in the residual phase error is promising for a telescope in the 2-4 m class.

Comparison of wavefront sensor models for simulation of adaptive optics.

The new generation of extremely large telescopes will have adaptive optics. Due to the complexity and cost of such systems, it is important to simulate their performance before construction. Most systems planned will have Shack-Hartmann wavefront sensors. Different mathematical models are available for simulation of such wavefront sensors. The choice of wavefront sensor model strongly influences computation time and simulation accuracy. We have studied the influence of three wavefront sensor models on performance calculations for a generic, adaptive optics (AO) system designed for K-band operation of a 42 m telescope. The performance of this AO system has been investigated both for reduced wavelengths and for reduced r(0) in the K band. The telescope AO system was designed for K-band operation, that is both the subaperture size and the actuator pitch were matched to a fixed value of r(0) in the K-band. We find that under certain conditions, such as investigating limiting guide star magnitude for large Strehl-ratios, a full model based on Fraunhofer propagation to the subimages is significantly more accurate. It does however require long computation times. The shortcomings of simpler models based on either direct use of average wavefront tilt over the subapertures for actuator control, or use of the average tilt to move a precalculated point spread function in the subimages are most pronounced for studies of system limitations to operating parameter variations. In the long run, efficient parallelization techniques may be developed to overcome the problem.