Photometric Completeness Modelled With Neural Networks

In almost any study involving optical/NIR photometry, understanding the completeness of detection and recovery is an essential part of the work. The recovery fraction is, in general, a function of several variables including magnitude, color, background sky noise, and crowding. We explore how completeness can be modelled, {with the use of artificial-star tests,} in a way that includes all of these parameters \emph{simultaneously} within a neural network (NN) framework. The method is able to manage common issues including asymmetric completeness functions and the bilinear dependence of the detection limit on color index. We test the method with two sample HST (Hubble Space Telescope) datasets: the first involves photometry of the star cluster population around the giant Perseus galaxy NGC 1275, and the second involves the halo-star population in the nearby elliptical galaxy NGC 3377. The NN-based method achieves a classification accuracy of $>$\,94\%, and produces results entirely consistent with more traditional techniques for determining completeness. Additional advantages of the method are that none of the issues arising from binning of the data are present, and that a recovery probability can be assigned to every individual star in the real photometry. Our data, models, and code (called COINTOSS) can be found online on Zenodo at the following link: https://doi.org/10.5281/zenodo.8306488.