SuperSynthIA: Physics-Ready Full-Disk Vector Magnetograms from HMI, Hinode, and Machine Learning

Vector magnetograms of the Sun's photosphere are cornerstones for much of solar physics research. These data are often produced by data-analysis pipelines combining per-pixel Stokes polarization vector inversion with a disambiguation that resolves an intrinsic 180-degree ambiguity. We introduce a learning-based method, SuperSynthIA, that produces full disk vector magnetograms from Stokes vector observations. As input, SuperSynthIA uses Stokes polarization images from SDO/HMI. As output, SuperSynthIA simultaneously emulates the inversion and disambiguation outputs from the Hinode/SOT-SP pipeline. Our method extends our previous approach SynthIA by providing direct heliographic outputs and using an improved dataset and inference method.

SuperSynthIA provides another tool for solar data analysis in the form of vector magnetograms resembling Hinode/SOT-SP but available across the full disk with SDO/HMI. Compared to our previous SynthIA, SuperSynthIA provides physically-ready vector magnetograms and mitigates unphysical angle preferences and banding artifacts in SynthIA. SuperSynthIA data are substantially more temporally consistent than those from the SDO/HMI pipeline, most notably seen in, e.g., evolving active regions. SuperSynthIA substantially reduces noise in low-signal areas, resulting in less center-to-limb bias outside of strong-signal areas. We show that outputs from SuperSynthIA track the SDO/HMI-recorded evolution of the magnetic field. We discuss the limitations of SuperSynthIA that the user must understand, and we demonstrate a broad set of evaluations to test SuperSynthIA and discuss remaining known artifacts. Our tests provide both methodology and evidence that SuperSynthIA outputs are ready for use by the community, and that learning-based approaches are suitable for physics-ready magnetograms.

SuperSynthIA uses a UNet + Regression-via-Classification with adaptive bin size to mitigate underestimation exhibited by common regression methods.

Logit Dithering adds noise to log-probabilities predicted by the model, which is used to resolve unphysical directional preference for the azimuth in quiet regions.

SuperSynthIA supports direct or analytic (calculated from \(\alpha B\), inclination, and azimuth) methods to produce disambiguated components (\( \alpha B_{R}, \alpha B_{\theta}, \alpha B_{\varphi} \)), enabling users to estimate these components in gravity frame or sky frame for different purposes.

The above video shows the comparison of coverage between Hinode /SOT-SP and SuperSynthIA. Hinode SOT-SP cannot provide full-disk high-cadence coverage of activity due to its limited field of view. SuperSynthIA predictions look like the Hinode pipeline, but with the cadence and coverage of HMI.

SuperSynthIA shows great temporal consistency compared to HMI pipeline with substantially fewer flickers during flux emergence. This flux emergence is a particularly tricky case for consistency because the underlying Milne Eddington assumption is not correct, leading to issues in the pipeline.

SuperSynthIA performs well in polar regions with a better signal-to-noise ratio compared to the HMI pipeline, enabling the study of vector magnetogram at the solar poles.

All SDO data used are publicly available from the Joint Science Operations Center (JSOC) at Stanford University supported by NASA Contract NAS5-02139 (HMI); see http://jsoc.stanford.edu/ ; which also supported J.T.H. and Y.L. Hinode is a Japanese mission developed and launched by ISAS/JAXA, with NAOJ as domestic partner and NASA and STFC (UK) as international partners. It is operated by these agencies in cooperation with ESA and NSC (Norway). After publication, data and models used will be archived and given Digital Object Identifiers (DOIs), and D.F. will be able to provide runs-on-request. This work was started under and partly supported by NASA grant 80NSSC20K0600 for the NASA Heliophysics DRIVE Science Center (SOLSTICE) at the University of Michigan. K.D.L. and G.B. acknowledge support from NASA/LWS-SC 80NSSC22K0892, and Lockheed-Martin Space Systems contract No. 4103056734 for Solar-B FPP Phase E. D.F., R.W., K.D.L., and G.B. acknowledge support from NASA award 80NSSC22K0646. D.F. and R.W. thank Michigan CSE Division and the University of Michigan Advanced Research Computing for continued HPC support. The authors thank David Orozco Suárez for thoughtful feedback that improved the paper.