Name	Name	Last commit message	Last commit date
parent directory ..
configs	configs
README.md	README.md
config.py	config.py
dataset.py	dataset.py
networks.py	networks.py
test.py	test.py
train.py	train.py
validation.py	validation.py

Solar Far-side Magnetogram Generation

Deep learning-based generation of solar far-side magnetograms from STEREO/EUVI data using conditional GAN.

Publication

Title: Solar farside magnetograms from deep learning analysis of STEREO/EUVI data

Authors: Taeyoung Kim, Eunsu Park, Harim Lee, Yong-Jae Moon, Sung-Ho Bae, Daye Lim, Soojeong Jang, Lokwon Kim, Il-Hyun Cho, Myungjin Choi, Kyung-Suk Cho

Journal: Nature Astronomy, 3, 397-400, 2019

DOI: 10.1038/s41550-019-0711-5

Overview

This study applies conditional GAN (cGAN) to generate solar magnetograms from EUV 304 nm images. The model is trained using SDO/AIA 304 nm and SDO/HMI magnetogram pairs from the near-side, then applied to STEREO/EUVI 304 nm images to generate far-side magnetograms.

Input/Output

Type	Description
Training Input	SDO/AIA 304 nm EUV image (1024 × 1024, 1 channel)
Training Output	SDO/HMI LOS magnetogram (1024 × 1024, 1 channel)
Inference Input	STEREO/EUVI 304 nm EUV image
Inference Output	Far-side magnetogram

Key Innovation

First demonstration of generating far-side magnetograms from single-wavelength EUV images
Enables continuous monitoring of solar magnetic field on the far-side
Important for space weather forecasting

Network Architecture

Pix2Pix (cGAN)

The architecture follows Pix2Pix (Isola et al., 2016) with U-Net Generator and PatchGAN Discriminator.

┌─────────────────────────────────────────────────────────────────────────┐
│                            GENERATOR (U-Net)                             │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│  Input: 1024 × 1024 × 1 (EUV 304 nm image)                             │
│      │                                                                   │
│      ▼                                                                   │
│  ┌────────────────────────────────────────────────────────────────┐     │
│  │                    ENCODER (Downsampling)                       │     │
│  ├────────────────────────────────────────────────────────────────┤     │
│  │                                                                 │     │
│  │  e1: Conv 4×4, s2 → 64   (512×512×64)     ──────────────┐      │     │
│  │  e2: Conv-BN-LReLU → 128 (256×256×128)    ─────────────┐│      │     │
│  │  e3: Conv-BN-LReLU → 256 (128×128×256)    ────────────┐││      │     │
│  │  e4: Conv-BN-LReLU → 512 (64×64×512)      ───────────┐│││      │     │
│  │  e5: Conv-BN-LReLU → 512 (32×32×512)      ──────────┐││││      │     │
│  │  e6: Conv-BN-LReLU → 512 (16×16×512)      ─────────┐│││││      │     │
│  │  e7: Conv-BN-LReLU → 512 (8×8×512)        ────────┐││││││      │     │
│  │  e8: Conv-ReLU → 512     (4×4×512)        ───────┐│││││││      │     │
│  │                                                   ││││││││      │     │
│  └───────────────────────────────────────────────────┼┼┼┼┼┼┼┼──────┘     │
│                                                      ││││││││            │
│  ┌───────────────────────────────────────────────────┼┼┼┼┼┼┼┼──────┐     │
│  │                    DECODER (Upsampling)           ││││││││      │     │
│  ├───────────────────────────────────────────────────┼┼┼┼┼┼┼┼──────┤     │
│  │                                                   ││││││││      │     │
│  │  d1: ConvT-BN-Drop-ReLU → 512 + e7 ◄──────────────┘│││││││      │     │
│  │  d2: ConvT-BN-Drop-ReLU → 512 + e6 ◄───────────────┘││││││      │     │
│  │  d3: ConvT-BN-Drop-ReLU → 512 + e5 ◄────────────────┘│││││      │     │
│  │  d4: ConvT-BN-ReLU → 512 + e4      ◄─────────────────┘││││      │     │
│  │  d5: ConvT-BN-ReLU → 256 + e3      ◄──────────────────┘│││      │     │
│  │  d6: ConvT-BN-ReLU → 128 + e2      ◄───────────────────┘││      │     │
│  │  d7: ConvT-BN-ReLU → 64 + e1       ◄────────────────────┘│      │     │
│  │  d8: ConvT-Tanh → 1                ◄─────────────────────┘      │     │
│  │                                                                 │     │
│  └─────────────────────────────────────────────────────────────────┘     │
│      │                                                                   │
│      ▼                                                                   │
│  Output: 1024 × 1024 × 1 (Generated Magnetogram)                        │
│                                                                          │
└─────────────────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────────────────┐
│                       DISCRIMINATOR (PatchGAN)                           │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│  Input: Concatenated [EUV image, Magnetogram] (1024 × 1024 × 2)         │
│      │                                                                   │
│      ▼                                                                   │
│  Layer 1: Conv 4×4, s2 → 64, LReLU          (512×512×64)                │
│  Layer 2: Conv 4×4, s2 → 128, BN, LReLU     (256×256×128)               │
│  Layer 3: Conv 4×4, s2 → 256, BN, LReLU     (128×128×256)               │
│  Layer 4: Conv 4×4, s1 → 512, BN, LReLU     (127×127×512)               │
│  Layer 5: Conv 4×4, s1 → 1, Sigmoid         (126×126×1)                 │
│      │                                                                   │
│      ▼                                                                   │
│  Output: Patch probability map (real/fake)                               │
│                                                                          │
└─────────────────────────────────────────────────────────────────────────┘

Key Components

Component	Description
Generator	U-Net with 8 encoder + 8 decoder layers
Discriminator	PatchGAN (70×70 receptive field)
Skip Connections	Encoder features concatenated to decoder
Dropout	50% on first 3 decoder layers
Normalization	Batch Normalization
Activation	LeakyReLU (encoder), ReLU (decoder)

Training Configuration

Parameter	Value
Loss Function	L1 + cGAN (BCE)
Lambda (L1 weight)	100
Optimizer	Adam
Learning Rate	2 × 10⁻⁴
Beta1	0.5
Input Size	1024 × 1024
Epochs	200

Data

Training Data (Near-side)

Source	Description
SDO/AIA 304 nm	EUV images (input)
SDO/HMI	LOS magnetograms (target)
Period	2011 January - 2017 December
Cadence	6 hours
Total pairs	~10,000

Application Data (Far-side)

Source	Description
STEREO-A/EUVI 304 nm	Far-side EUV images
STEREO-B/EUVI 304 nm	Far-side EUV images

Results

Near-side Validation

Metric	Value
Pixel Correlation	0.88
Structural Similarity	0.92

Far-side Application

Successfully generated far-side magnetograms from STEREO/EUVI data
Generated magnetograms show reasonable agreement with helioseismic far-side images
Active regions detected on far-side rotate to near-side with consistent magnetic features

Requirements

Python 3.6+
PyTorch 1.0+
NumPy
SunPy

Usage

import torch
from networks import Generator, Pix2Pix

# Initialize generator for inference
generator = Generator(in_channels=1, out_channels=1)

# Load trained weights
generator.load_state_dict(torch.load('generator.pth'))
generator.eval()

# Input: (batch, 1, 1024, 1024) - EUV 304 nm image
euv_image = torch.randn(1, 1, 1024, 1024)

# Output: (batch, 1, 1024, 1024) - Generated magnetogram
with torch.no_grad():
    magnetogram = generator(euv_image)

Reference Code

Original implementation: https://github.com/tykimos/SolarMagGAN

Citation

@article{Kim_2019,
    title={Solar farside magnetograms from deep learning analysis of STEREO/EUVI data},
    author={Kim, Taeyoung and Park, Eunsu and Lee, Harim and Moon, Yong-Jae and Bae, Sung-Ho and Lim, Daye and Jang, Soojeong and Kim, Lokwon and Cho, Il-Hyun and Choi, Myungjin and Cho, Kyung-Suk},
    journal={Nature Astronomy},
    volume={3},
    pages={397--400},
    year={2019},
    publisher={Nature Publishing Group},
    doi={10.1038/s41550-019-0711-5}
}

Related Publications

Reply to: Reliability of AI-generated magnetograms from only EUV images (2021)

Title: Reply to: Reliability of AI-generated magnetograms from only EUV images

Authors: Eunsu Park, Hyun-Jin Jeong, Harim Lee, Taeyoung Kim, Yong-Jae Moon

Journal: Nature Astronomy, 5, 111-112, 2021

DOI: 10.1038/s41550-021-01311-5

Type: Matters Arising (Reply to Liu et al.)

Summary: This paper responds to constructive comments about limitations of the original study, explaining:

Dynamic range choice (±100G): Selected to better show active region shapes and enable effective model training
Improved methods: Pix2PixHD with larger dynamic range (±1,400G) published in ApJL; multi-channel EUV inputs achieving ±3,000G range
Preprocessing details: Used aia_prep, hmi_prep, secchi_prep (SolarSoft) for Level 1.5 images; down-sampled to 1024×1024; solar radius set to 392 pixels; manually excluded poor quality images (4,972 pairs used)
Model selection: Best model chosen from ~120 epochs (500,000 iterations) based on highest pixel-to-pixel Pearson correlation

Follow-up Studies Referenced:

Shin et al. (2020), ApJL 895, L16 - Ca II to magnetogram translation with Pix2PixHD
Jeong et al. (2020), ApJL 903, L25 - Coronal magnetic field extrapolation using AI-generated farside magnetograms

@article{Park_2021_Reply,
    title={Reply to: Reliability of AI-generated magnetograms from only EUV images},
    author={Park, Eunsu and Jeong, Hyun-Jin and Lee, Harim and Kim, Taeyoung and Moon, Yong-Jae},
    journal={Nature Astronomy},
    volume={5},
    pages={111--112},
    year={2021},
    publisher={Nature Publishing Group},
    doi={10.1038/s41550-021-01311-5}
}

Solar Coronal Magnetic Field Extrapolation with AI-generated Farside (2020)

Title: Solar Coronal Magnetic Field Extrapolation from Synchronic Data with AI-generated Farside

Authors: Hyun-Jin Jeong, Yong-Jae Moon, Eunsu Park, Harim Lee

Journal: The Astrophysical Journal Letters, 903:L25 (9pp), 2020 November 1

DOI: 10.3847/2041-8213/abc255

Type: Application Study (PFSS extrapolation using AI-generated farside magnetograms)

Summary: This paper applies AI-generated solar farside magnetograms (AISFM) for potential field source surface (PFSS) coronal magnetic field extrapolation:

Improved Model (Pix2PixHD): Multi-scale discriminators with larger dynamic range (±3,000 G)
Multi-channel Input: 3 EUV passbands (304, 193, 171 Å) instead of single 304 Å
AISFM Versions:
- AISFM 1.0: Original model (±100 G, single 304 Å)
- AISFM 2.0: Pix2PixHD with ±3,000 G and 3-channel EUV input
Application: PFSS extrapolation from synchronic magnetograms (SDO/HMI near-side + AISFM far-side)
Results: Improved coronal magnetic field modeling including far-side active regions

Reference Code: https://github.com/JeongHyunJin/Jeong2020_SolarFarsideMagnetograms

@article{Jeong_2020,
    title={Solar Coronal Magnetic Field Extrapolation from Synchronic Data with AI-generated Farside},
    author={Jeong, Hyun-Jin and Moon, Yong-Jae and Park, Eunsu and Lee, Harim},
    journal={The Astrophysical Journal Letters},
    volume={903},
    number={2},
    pages={L25},
    year={2020},
    month={nov},
    publisher={The American Astronomical Society},
    doi={10.3847/2041-8213/abc255}
}

Improved AI-generated Solar Farside Magnetograms and Data Release (2022)

Title: Improved AI-generated Solar Farside Magnetograms by STEREO and SDO Data Sets and Their Release

Authors: Hyun-Jin Jeong, Yong-Jae Moon, Eunsu Park, Harim Lee, Ji-Hye Baek

Journal: The Astrophysical Journal Supplement Series, 262:50 (12pp), 2022 October

DOI: 10.3847/1538-4365/ac8d66

Type: Major Update (AISFM 3.0 model and public data release)

Summary: This paper presents the improved AISFM 3.0 model using Pix2PixCC architecture:

Model Evolution:
- AISFM 1.0: Original Pix2Pix (±100 G, 1 EUV channel)
- AISFM 2.0: Pix2PixHD (±3,000 G, 3 EUV channels)
- AISFM 3.0: Pix2PixCC (adds correlation coefficient-based loss)
Pix2PixCC Architecture:
- Correlation Coefficient (CC) loss added to L1 + cGAN + FM loss
- Better preservation of structural features
Training Data:
- SDO/AIA (171, 193, 304 Å) and SDO/HMI magnetogram pairs
- Period: 2011 January - 2017 December
- Total: 8,148 pairs (train: 5,968, validation: 1,144, test: 1,036)
Performance (Near-side Test):

Region Pixel CC RMSE (G)

Full Disk 0.88 35.9

Active Region 0.91 95.9

Quiet Region 0.70 10.9
Public Data Release: AISFM 3.0 data available at http://sdo.kasi.re.kr
- STEREO-A farside magnetograms (2011 January - present)
- STEREO-B farside magnetograms (2011 January - 2014 September)

Region	Pixel CC	RMSE (G)
Full Disk	0.88	35.9
Active Region	0.91	95.9
Quiet Region	0.70	10.9

Reference Code: https://github.com/JeongHyunJin/Pix2PixCC

@article{Jeong_2022,
    title={Improved AI-generated Solar Farside Magnetograms by STEREO and SDO Data Sets and Their Release},
    author={Jeong, Hyun-Jin and Moon, Yong-Jae and Park, Eunsu and Lee, Harim and Baek, Ji-Hye},
    journal={The Astrophysical Journal Supplement Series},
    volume={262},
    number={2},
    pages={50},
    year={2022},
    month={oct},
    publisher={The American Astronomical Society},
    doi={10.3847/1538-4365/ac8d66}
}

AI-based Solar Farside Vector Magnetograms from Multi-spacecraft EUV (2025)

Title: Artificial-intelligence-based Reconstruction of Solar Farside Vector Magnetograms from Multispacecraft Extreme-ultraviolet Data

Authors: Hyun-Jin Jeong, Eunsu Park, Harim Lee, Junmu Youn, Mingyu Jeon, Yong-Jae Moon, Stefaan Poedts, Francesco Carella, Haopeng Wang, Daeil Kim, Youngjae Kim, Jihye Kang

Journal: The Astrophysical Journal Supplement Series, 281:63 (14pp), 2025 December

DOI: 10.3847/1538-4365/ae21b8

Type: Major Update (Vector Magnetogram Generation from Multi-spacecraft EUV)

Summary: This paper extends farside magnetogram generation to full vector magnetic fields using Pix2PixCC architecture:

Vector Field Generation: Produces all three components (Br, Bθ, Bφ) of solar farside magnetic fields
Multi-spacecraft Input: EUV 304 & 171 Å from STEREO-A, STEREO-B, and Solar Orbiter (SolO) with SFT model reference data
Performance (SDO test): Average Cvec=0.89, CCS=0.76, Pearson CC for Br=0.91, Bθ=0.77, Bφ=0.77 (after 8×8 binning)
Direct Validation: First comparison with SDO/HMI during STEREO-A (2023) and SolO (2022) inferior conjunctions
AR Tracking: Demonstrated tracking of NOAA AR 11339 and 13848 with derived vector parameters (USFLUX, TOTUSJH, ABSNJZH)
Application: Enables continuous monitoring of solar vector magnetic fields from farside to frontside

Reference Code: https://github.com/JeongHyunJin/Pix2PixCC (Zenodo: doi:10.5281/zenodo.17573370)

@article{Jeong_2025,
    title={Artificial-intelligence-based Reconstruction of Solar Farside Vector Magnetograms from Multispacecraft Extreme-ultraviolet Data},
    author={Jeong, Hyun-Jin and Park, Eunsu and Lee, Harim and Youn, Junmu and Jeon, Mingyu and Moon, Yong-Jae and Poedts, Stefaan and Carella, Francesco and Wang, Haopeng and Kim, Daeil and Kim, Youngjae and Kang, Jihye},
    journal={The Astrophysical Journal Supplement Series},
    volume={281},
    number={2},
    pages={63},
    year={2025},
    month={dec},
    publisher={The American Astronomical Society},
    doi={10.3847/1538-4365/ae21b8}
}

License

MIT License

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

Solar Far-side Magnetogram Generation

Publication

Overview

Input/Output

Key Innovation

Network Architecture

Pix2Pix (cGAN)

Key Components

Training Configuration

Data

Training Data (Near-side)

Application Data (Far-side)

Results

Near-side Validation

Far-side Application

Requirements

Usage

Reference Code

Citation

Related Publications

Reply to: Reliability of AI-generated magnetograms from only EUV images (2021)

Solar Coronal Magnetic Field Extrapolation with AI-generated Farside (2020)

Improved AI-generated Solar Farside Magnetograms and Data Release (2022)

AI-based Solar Farside Vector Magnetograms from Multi-spacecraft EUV (2025)

License

FilesExpand file tree

farside-magnetogram

Directory actions

More options

Directory actions

More options

Latest commit

History

farside-magnetogram

Folders and files

parent directory

README.md

Solar Far-side Magnetogram Generation

Publication

Overview

Input/Output

Key Innovation

Network Architecture

Pix2Pix (cGAN)

Key Components

Training Configuration

Data

Training Data (Near-side)

Application Data (Far-side)

Results

Near-side Validation

Far-side Application

Requirements

Usage

Reference Code

Citation

Related Publications

Reply to: Reliability of AI-generated magnetograms from only EUV images (2021)

Solar Coronal Magnetic Field Extrapolation with AI-generated Farside (2020)

Improved AI-generated Solar Farside Magnetograms and Data Release (2022)

AI-based Solar Farside Vector Magnetograms from Multi-spacecraft EUV (2025)

License