HelioFill: History-conditioned Diffusion Estimation of Solar Far-side EUV Emission

Abstract The loss of STEREO-B in 2014 created a persistent blind spot in extreme ultraviolet (EUV) imaging of the solar far-side. We present HelioFill, which is, to the authors’ knowledge, the first denoising-diffusion inpainting model that estimates the unobserved sector of full-Sun 304 Å maps from the current Earth-side (Solar Dynamics Observatory, or SDO) and STEREO-A views. Because these views cannot constrain structures fully enclosed in the unobserved sector, HelioFill’s diffusion stage is also conditioned on the sector’s own rotation history, supplied as causal persistence and information-age maps; a rectified-flow refinement then adds realistic small-scale texture. Test rotations are separated from the diffusion stage’s training timestamps by full-rotation guard bands. Over the unobserved sector, HelioFill’s estimate correlates with the three-view truth at ρ = 0.520, above persistence (0.416) and the input-independent climatological mean (0.392) on 97.9% and 96.8% of the test maps, which conditioning on the current views alone does not achieve; in peak signal-to-noise ratio (PSNR) it exceeds them by 2.27 and 0.66 dB and matches a direct observation of the sector made ≈29 hr before or after it (≈49 hr in correlation). The refinement lowers learned perceptual image patch similarity (LPIPS) by 11% at a cost of 0.50 dB and 0.055 in correlation. HelioFill localizes active regions better than persistence (Dice score 0.449 versus 0.384) and detects fully enclosed ones, although persistence detects more individual regions and coronal holes remain poorly reproduced. Against near-simultaneous far-side Solar Orbiter/EUI observations from 2022 to 2025, HelioFill correlates with the truth at ρ = 0.375, above climatology (0.286) and level with persistence (0.365), which it exceeds in PSNR and structural similarity index (SSIM); its diffusion stage alone reaches 0.460, above both, and the refinement stage costs the difference. HelioFill thus provides observationally plausible global EUV context within quantified limits.