Fast Deep Learning-Based Stokes Vector Inversion with Confidence for SDO/HMI

The Helioseismic and Magnetic Imager (HMI) onboard NASA's Solar Dynamics Observatory (SDO) monitors the photospheric magnetic field, and is a critical input to many space weather forecasting systems. The magnetogram products produced by HMI and its analysis pipeline are the result of a per-pixel optimization that finds magnetic field parameters that minimize disagreement with an observed Stokes vector. In this paper, we introduce a deep learning-based approach that can emulate the current HMI pipeline result two orders of magnitude faster than the current system. Our system is a U-Net trained on paired input Stokes vectors and output optimization-based inversions. We demonstrate that our system, once trained, can produce high-fidelity estimates of the magnetic field and thermodynamic parameters and can also produce calibrated confidence intervals per-pixel. We additionally show that despite being trained per-pixel, our system is able to faithfully reproduce known oscillations in full-disk statistics produced by the pipeline. This system could serve as both an initialization for the full system and as an ultra-fast proxy for the system in space weather applications. This work is part of the SOLSTICE center and will be open sourced.