PINN ME: A Physics-Informed Neural Network Framework for Accurate Milne-Eddington Inversions of Solar Magnetic Fields

Abstract

Spectropolarimetric inversions of solar observations are fundamental for the estimation of the magnetic field in the solar atmosphere. However, instrumental noise, computational requirements, and varying levels of physical realism make it challenging to derive reliable solar magnetic field estimates. In this study, we present a novel approach for spectropolarimetric inversions based on Physics Informed Neural Networks (PINNs) to infer the photospheric magnetic field under the Milne-Eddington approximation (PINN ME). Our model acts as a representation of the parameter space, mapping input coordinates (t, x, y) to the respective spectropolarimetric parameters, which are used to synthesize the corresponding stokes profiles. By iteratively sampling coordinate points, synthesizing profiles, and minimizing the deviation from the observed stokes profiles, our method can find the set of Milne-Eddington parameters that best fit the observations. In addition, we directly include the point-spread-function to account for instrumental effects. We use a predefined parameter space as well as synthetic profiles from a radiative MHD simulation to evaluate the performance of our method and to estimate the impact of instrumental noise. Our results demonstrate that PINN ME achieves an intrinsic spatio-temporal coupling, which can largely mitigate observational noise and provides a memory-efficient inversion even for extended fields-of-view. Finally, we apply our method to observations and show that our method provides a high spatial coherence and can resolve small-scale features both in strong- and weak-field regions.