A Submillisecond Fourier and Wavelet-based Model to Extract Variable Candidates from the NEOWISE Single-exposure Database

Abstract This paper presents VARnet, a capable signal-processing model for rapid astronomical time series analysis. VARnet leverages wavelet decomposition, a novel method of Fourier feature extraction via the finite-embedding Fourier transform, and deep learning to detect faint signals in light curves, utilizing the strengths of modern GPUs to achieve submillisecond single-source run time. We apply VARnet to the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE) single-exposure database, which holds nearly 200 billion apparitions over 10.5 yr of infrared sources on the entire sky. This paper devises a pipeline in order to extract variable candidates from the NEOWISE data, serving as a proof of concept for both the efficacy of VARnet and methods for an upcoming variability survey over the entirety of the NEOWISE data set. We implement models and simulations to synthesize unique light curves to train VARnet. In this case, the model achieves an F1 score of 0.91 over a four-class classification scheme on a validation set of real variable sources present in the infrared. With ∼2000 points per light curve on a GPU with 22 GB of VRAM, VARnet produces a per-source processing time of <53 μ s. We confirm that our VARnet is sensitive and precise to both known and previously undiscovered variable sources. These methods prove promising for a complete future survey of variability with the Wide-field Infrared Survey Explorer, and effectively showcase the power of the VARnet model architecture.