Learning to Estimate Photospheric Vector Fields from Line-of-sight Magnetograms

Abstract Solar photospheric line-of-sight magnetograms are easier to estimate than full vector magnetograms since the line-of-sight component ( B los ) can be obtained from total intensity and circular polarization signals, unlike the perpendicular component ( B ⊥ ), which depends on harder-to-measure linear polarization. Unfortunately, the line-of-sight component by itself is not physically meaningful, as it is just one component of the underlying vector and one whose relationship to the gravity vector changes pixel-to-pixel. To produce an estimate of the radial component ( B R ), a common “correction” is often applied that assumes the field is radial, which is nearly always false. This paper investigates recovering full vector field information from B los by building on the recent SuperSynthIA approach that was originally used with Stokes vectors as input for simultaneous inversion and disambiguation. As input, the method accepts one or more line-of-sight magnetograms and associated metadata; as output, our method estimates full vector field in heliographic components, meaning that the physically relevant vector components are returned without need for further disambiguation steps or component transforms. We demonstrate the ability to produce good estimates of the full vector field on unseen examples from both HMI and GONG, including examples that predate the Solar Dynamics Observatory mission. Our results show that learning is not a replacement for a dedicated vector-field observing facilities, but may serve to unlock information from past data and at the very least, provide more accurate B R maps from B los than are created using the simple viewing angle assumption.