Using score-based diffusion models for correlated 1/f noise reduction in JWST spectral data

Salma Salhi ( Université de Montréal )

Transit spectroscopy is very sensitive to various sources of noise, and this is especially true when using the Single Object Slitless Spectroscopy (SOSS) mode on the NIRISS instrument aboard the JWST. Current methods to deal with 1/f (correlated) noise, such as approximating the 1/f signal as a constant across a column, leave noise residuals that are almost double that of the expected readout noise. Deep learning models could be a way to mitigate this problem, as they have already been shown to be very efficient at a wide variety of tasks in astrophysical data reduction, including denoising. We construct a score-based generative diffusion model to learn the structure of the noise in dark images, including bad pixels, hot pixels, cosmic rays, and 1/f noise to create a noise model. We will then use this noise model as the data likelihood to analyze mock trace observations in a Bayesian framework, allowing us to produce posterior samples of the pixel values of the underlying trace. We aim to apply this method to time series spectroscopic observations, which will allow for a more accurate retrieval of the underlying trace parameters, potentially reducing our error to the photon noise limit. This has the potential to substantially improve our signal-to-noise by up to a factor of 2 for some spectral regions and thus enable higher precision spectroscopy.