🔭 astrophysics

Reconstructing Synthetic SDO/AIA 193 A EUV Images from He I 10830 A Observations with Diffusion Model Translator

This paper introduces the Coronal Hole-aware Diffusion Model Translator (CH-aware DMT), a deep learning framework that successfully reconstructs synthetic SDO/AIA 193 Å EUV images from multi-decade He I 10830 Å observations, thereby enabling the extension of coronal context analysis to periods prior to the availability of direct EUV imaging.

Original authors: Marco Marena, Qin Li, Haimin Wang, Haodi Jiang, Prajwal Shah, Bo Shen

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marco Marena, Qin Li, Haimin Wang, Haodi Jiang, Prajwal Shah, Bo Shen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Time Traveling the Sun's Weather

Imagine you want to study the weather on Earth, but you only have high-definition satellite photos starting from the year 2010. Before that, you only have old, blurry black-and-white sketches of cloud shadows. You know the shadows tell you something about the clouds, but they aren't the same thing.

This paper is about doing exactly that for the Sun. Scientists have had high-definition "weather maps" of the Sun's outer atmosphere (the corona) in Extreme Ultraviolet (EUV) light only since 2010 (thanks to the SDO satellite). However, they have been taking pictures of a different layer (the chromosphere) using a specific infrared light (He i 10830 Å) for decades, going back to the 1970s.

The goal of this research is to use those old, indirect "shadow sketches" to reconstruct what the modern, high-definition "weather maps" would have looked like in the past. They want to create a "synthetic" history of the Sun's corona before we had the technology to see it directly.

The Problem: Seeing the Invisible

The Sun's outer atmosphere is tricky.

The Modern View (AIA 193 Å): This is like a high-res color photo. It shows hot, glowing loops of gas and dark, empty patches called "Coronal Holes" (where solar wind escapes).
The Old View (He i 10830 Å): This is like a silhouette. It doesn't show the hot gas directly. Instead, it shows how much light is blocked by the layer below. Interestingly, the "dark" Coronal Holes in the modern view actually look bright in this old silhouette view because the gas there absorbs less light.

The relationship between the "silhouette" and the "color photo" is complex and non-linear. It's like trying to guess the exact shape of a 3D sculpture just by looking at its shadow on a wall. You can't just draw a line; you need to understand the physics of how the light interacts.

The Solution: The "AI Translator"

To solve this, the researchers built a special type of Artificial Intelligence called a Diffusion Model Translator.

The Analogy: The Art Restoration Studio
Imagine a master art restorer who has seen thousands of modern, high-definition paintings of the Sun. They also have a collection of old, rough charcoal sketches of the same scenes.

Training: The AI is shown pairs of these sketches and the matching modern photos. It learns the "language" of the Sun: "When the sketch shows a bright patch here, the photo usually has a dark hole there."
The Diffusion Process: Think of this as a game of "Telephone" but in reverse. The AI starts with a noisy, static-filled version of the sketch. It slowly "denoises" it, step-by-step, using the old sketch as a guide, until a clear, high-definition image emerges.
The "Coronal Hole" Awareness: The researchers realized the AI was sometimes getting the dark spots (Coronal Holes) fuzzy. So, they added a special "spotlight" to the training. They told the AI: "Pay extra attention to these dark areas. Get their edges sharp and their brightness correct." This is the "CH-aware" part of their model.

What They Did

They trained this AI on data from 2011 to 2015, where they had both the old sketches (He i) and the modern photos (AIA). Once the AI learned the translation rules, they tested it in three ways:

The "Blind" Test (2011–2015): They fed the AI old sketches from this period and asked it to guess the modern photo.
- Result: The AI was very good at it. It got the overall shape right (92% correlation) and did a decent job with the tricky dark holes (84% correlation).
The "Cross-Check" (2005–2015): They fed the AI sketches from 2005–2010 (before the modern photos existed) and compared the result to photos from a slightly different, older satellite (SOHO/EIT).
- Result: The reconstructed images looked very similar to the old satellite photos, proving the AI wasn't just memorizing the training data but actually learning the physics.
The "Deep Time" Test (1974–1993): They fed the AI sketches from the 1970s and 80s. Since there were no EUV photos from this time, they compared the results to Soft X-ray images from the Yohkoh satellite.
- Result: The large-scale patterns (like where the dark holes were and how they moved over the solar cycle) matched the X-ray images. The AI successfully recreated the "solar weather" of the 1980s.

The Results: A New Historical Record

The paper concludes that they have successfully created a synthetic history of the Sun's corona.

It works: The AI can turn old infrared sketches into plausible-looking modern-style EUV photos.
It tracks the cycles: The reconstructed images show the Sun getting more active and less active over decades, matching the known sunspot cycles.
It's not perfect: The AI doesn't give you a perfectly calibrated scientific measurement (like an exact temperature reading). It gives you a physically plausible visual proxy. It's like a high-quality movie recreation of a historical event based on old diary entries—it captures the mood, the layout, and the major events, even if the exact lighting isn't 100% scientifically precise.

Why This Matters

Before this, if you wanted to study how the Sun's atmosphere changed over 40 years, you were stuck with a 10-year gap in high-quality data. Now, scientists have a continuous, 50-year-long "movie" of the Sun's corona. This allows them to study long-term changes in solar storms and space weather that were previously invisible because the data didn't exist.

In short: They taught an AI to translate old, indirect sun pictures into modern, high-definition ones, effectively giving us a time machine to see the Sun's atmosphere as it looked decades ago.

Technical Summary: Reconstructing Synthetic SDO/AIA 193 Å EUV Images from He i 10830 Å Observations with Diffusion Model Translator

1. Problem Statement

Routine full-disk Extreme Ultraviolet (EUV) imaging of the solar corona has only been available since the modern era (SOHO/EIT in the mid-1990s and SDO/AIA since 2010). This gap limits the ability to conduct multi-decade analyses of large-scale coronal evolution, particularly regarding Coronal Holes (CHs) and open magnetic flux, prior to the SDO era. While full-disk He i 10830 Å observations have been available since the 1970s (via KPVT and SOLIS), they represent chromospheric absorption modulated by coronal irradiance rather than direct coronal emission. The challenge lies in reconstructing physically plausible synthetic SDO/AIA 193 Å images from He i 10830 Å inputs. This is a non-trivial spectral translation problem because the mapping between chromospheric absorption and coronal EUV emission is nonlinear, mediated by the 3D thermodynamic structure and radiative coupling of the solar atmosphere. Traditional heuristic or correlation-based methods fail to synthesize high-fidelity EUV intensity images with the necessary radiometric complexity and resolution.

2. Methodology

The authors propose the Coronal Hole-aware Diffusion Model Translator (CH-aware DMT), a conditional image-to-image translation framework based on diffusion models.

Data Construction: The model is trained on temporally co-aligned pairs of SOLIS He i 10830 Å and SDO/AIA 193 Å images spanning 2011–2015. To prevent temporal leakage and ensure balanced solar activity representation, the dataset is split by month: January–October for training, November for validation, and December for testing. The dataset includes 829 paired samples covering various solar activity levels.
Diffusion Model Translator (DMT) Framework: The approach utilizes the DMT architecture, which decomposes conditional generation into three stages:
1. Shared Forward Diffusion: Both source (He i) and target (AIA) domains are mapped to a noisy intermediate representation.
2. Single-Step Domain Transfer: A lightweight translator learns to map the noisy source representation ( $x_{t^*}$ ) to the noisy target representation ( $y_{t^*}$ ) at a specific timestep $t^*$ . This step is trained to minimize the regression objective between the noisy domains.
3. Reverse Denoising: A U-Net denoiser, trained on the target domain (AIA 193 Å), performs the reverse diffusion process to generate the clean synthetic image from the translated latent.
CH-Aware Supervision: To improve fidelity in low-emission regions (Coronal Holes), the base DMT objective is augmented with three specific loss terms:
1. CH-weighted L1 Content Loss: Upweights pixel-wise reconstruction errors within CH regions to prioritize intensity fidelity in low-intensity areas.
2. Edge-Aware Boundary Consistency Loss: Enforces gradient consistency between synthetic and ground-truth images at CH boundaries to prevent blurring and boundary drift.
3. Dice Segmentation Loss: An auxiliary head predicts the CH mask, using a Dice loss to enforce global shape consistency and handle class imbalance (since CHs are minority pixels).
Evaluation Strategy: Since ground truth AIA 193 Å does not exist for pre-2010 periods, the model is validated through:
- Direct comparison with held-out SDO/AIA 193 Å data (2011–2015).
- Cross-instrument comparison with SOHO/EIT 195 Å (2005–2015).
- Morphological comparison with Yohkoh/SXT soft X-ray observations (1992) for the KPVT era.
- Long-term statistical consistency checks against sunspot numbers and F10.7 radio flux (1974–2015).

3. Key Results

Quantitative Performance (2011–2015): On the held-out test set, the CH-aware DMT achieves a correlation coefficient (CC) of 0.9227 for full-disk morphology and 0.8408 for Coronal Hole regions. The Mean Absolute Error (MAE) is 0.0711 for the full disk and 0.1086 for CH regions.
Cross-Instrument Consistency: Synthetic AIA 193 Å images generated from SOLIS He i 10830 Å inputs (2005–2015) reproduce the dominant large-scale morphology and contrast ordering of contemporaneous SOHO/EIT 195 Å images, including extended low-emission regions and active region complexes.
Historical Applicability (1974–1993): When applied to the KPVT He i 10830 Å archive, the model generates synthetic AIA 193 Å sequences that show coherent temporal evolution of low-emission regions consistent with Yohkoh/SXT soft X-ray observations during the 1992 overlap.
Long-Term Trends: The disk-integrated statistics of the synthetic AIA 193 Å sequence (1974–2015) track broad solar-cycle modulation. The synthetic median signal shows strong correlation with independent solar activity proxies: CC = 0.843 with sunspot numbers and 0.830 with F10.7 radio flux for the 1974–1993 period.
Synoptic Structure: Carrington-rotation synoptic maps constructed from the synthetic data display physically plausible global organization, including the migration of active region bands consistent with the solar butterfly pattern and solar cycle phases.

4. Significance and Claims

The paper claims that the CH-aware DMT provides a physically plausible synthetic proxy for SDO/AIA 193 Å coronal imagery, effectively extending EUV coronal context into the pre-SDO and pre-SOHO eras.

Extension of Observational Record: The method leverages the multi-decade availability of He i 10830 Å observations to reconstruct coronal structure for periods where direct EUV imaging was unavailable.
Structural Fidelity over Radiometric Calibration: The authors explicitly state that the goal is not absolute radiometric calibration but the reconstruction of large-scale coronal morphology and its temporal evolution. The model successfully preserves the spatial coherence of active regions, coronal holes, and their evolution across solar cycles.
Enabling Historical Analysis: The reconstructed products support multi-decade analyses of large-scale coronal evolution, including studies of cycle-to-cycle variability, hemispheric asymmetry, and the evolution of open magnetic flux regions.
Limitations: The paper modestly notes that the output should be interpreted as a structural proxy rather than a calibrated measurement. Absolute intensity calibration is not preserved due to differences in instrument response and data formats. Furthermore, CH masks derived from the synthetic images are approximate and method-dependent, serving as candidates for further analysis rather than definitive expert segmentations.

In summary, the work demonstrates that diffusion-based translation can bridge the gap between chromospheric absorption diagnostics and coronal emission, enabling a continuous, multi-decade view of solar coronal structure that was previously inaccessible.