PSF modeling with deep learning for the Argus Optical Array

We present a machine learning model based on deep encoder-decoder architecture that transforms an astronomical image into a latent space representation of the spatially varying Point Spread Functions (PSFs) within the image. This learned compressed representation can then be queried with a pixel position in order to create a realistic synthetic PSF for a transient source at that pixel. Our method is demonstrated using data from the Argus Pathfinder array's transient detection pipeline. This methodology allows for more cost-effective generation of 100M+ image datasets for training transient detection pipelines and could be generalized to other next generation transient surveys.