For Adaptive Optics (AO) assisted systems, a point spread function (PSF) may have a very complex structure and therefore be difficult to predict. An accurate and independent PSF prediction would be beneficial for the scientific analysis of the data and for quality evaluation of potential observations or even exposure time estimation before the actual observations. Ideally, one would like to have a tool able to predict the PSF morphology only based on a few external parameters, including instrument characteristics, given weather conditions, and properties of a selected target. It can be achieved with analytical PSF prediction tools (e.g., TIPTOP, PSFAO ). However, in real-life scenarios, PSF can still be influenced by multiple effects which are not adequately described by such analytical models and which can still significantly affect the quality of a predicted PSF. The topic of this work is to investigate the possible application of machine learning (ML) for PSF prediction. The main idea is not to use ML blindly but rather to investigate the idea of combining existing analytical approaches with machine learning. While the analytical model retains physics, the ML part can complement it and learn from data the effects that the analytical model fails to account for, resulting in higher prediction accuracy. This paper demonstrates the accuracy of the method when applied to an extensive data set of 700+ PSF from SPHERE and 200+ PSFs from MUSE-NFM.
|