Forward modelling coronagraphic images with a fully physical, differentiable digital twin of MagAO-X: first laboratory results

Abstract

Post-processing of high contrast imaging data relies on an accurate model of the stellar point spread function (PSF). Current techniques build this model from the science images themselves, using observational diversity (e.g., angular, spectral or polarimetric diversity), which can cause self-subtraction of the companion signal and constrains the observing strategy. Telemetry-based forward modelling instead builds the stellar PSF model from wavefront sensor data that is already recorded during the observation. The wavefront sensor measures the coherent starlight and can therefore be used to create a PSF model that only models the stellar light and does not reproduce the incoherent light of a companion. We present a fully physical and differentiable digital twin of the focal plane low-order wavefront sensor (FLOWFS) and the coronagraphic science beam of the MagAO-X instrument, implemented in \texttt{dLux}, and calibrate it on laboratory data. When fitted directly to the science images, the model reproduces the coronagraphic PSF down to the photon and read noise floor of the data. When instead forward modelled from the FLOWFS telemetry alone, the residuals reach $6\times 10^{-5}$ of the stellar peak at $5\lambda /D$, a factor of 5 below the raw contrast, with the remaining residual set by how well the wavefront estimate transfers from the FLOWFS branch to the science branch of the model. An injected companion at $5\lambda /D$ with a peak contrast of $10^{-3}$ is recovered without measurable self-subtraction. We discuss the model improvements currently under development and the path towards on-sky validation.