Generation of solar farside magnetograms

Solar farside magnetograms are generated by the Pix2PixHD model.
We train and evaluate the model using pairs of SDO/AIA EUV passband images and SDO/HMI magnetograms.
Then we generate realistic farside magnetograms from the corresponding EUV bassband images of STEREO/EUVIs by the model.

---

Pix2PixHD model

The Pix2PixHD is based on conditional Generative Adversarial Networks (cGANs) framework, which is one of the popular deep learning methods for image translation.

Environments

This code has been tested on Ubuntu 18.04 with a Nvidia GeForce GTX Titan XP GPU, CUDA Version 11.0, Python 3.6.9, and PyTorch 1.3.1.

Network architectures

The model consists of two major networks: one is a generative network (generator) and the other is a discriminative network (discriminator).

The generator tries to generate realistic output from input, and the discriminator tries to distinguish which one is a more real-like pair between a real pair and a fake pair.

Generator architectures

In our model, we use a global generator (G).
The generator is consist of the Encoder - Residual Blocks - Decoder.
The 'nd' indicate how many times you want to downsample input data, and the 'nr' indicate the number of residual blocks.

- Encoder
     1. Conv2D(filter = 32, strides = 1), InstanceNorm2d, ReLU
     2. Conv2D(filter = 32*2^(i_nd+1), strides = 2), InstanceNorm2d, ReLU 

- Residual Blocks (*nr)
     1. Conv2D(filter = 32*2^(nd+1), strides = 1), InstanceNorm2d, ReLU
     2. Conv2D(filter = 32*2^(nd+1), strides = 1), InstanceNorm2d

- Decoder
     1. Conv2DTranspose(filter = 32*2^(nd+1)//2^(i_nd), strides = 2), InstanceNorm2d, ReLU
     2. Conv2DTranspose(filter = 32, strides = 1)

Discriminator architectures

In our model, we use two 70*70 patch discriminator (D_1 and D_2).
One discriminator gets input pairs of the original pixel size, and the other gets input pairs which are downsampled by half.

    1. Conv2D(filers = 64, strides = 2), LeakyReLu(slope = 0.2)
    2. Conv2D(filers = 128, strides = 2), InstanceNorm, LeakyReLu(slope = 0.2)
    3. Conv2D(filers = 256, strides = 2), InstanceNorm, LeakyReLu(slope = 0.2)
    4. Conv2D(filers = 512, strides = 2), InstanceNorm, LeakyReLu(slope = 0.2)
    5. Conv2D(filers = 1, strides = 1)

Hyperparameters

The Loss configuration of the Objective functions

• Total loss = ( LSGAN loss ) + 10 * ( Feature Matching loss )

Optimizer

• Optimizer : Adam solver
• Learning rate : 0.0002
• momentum beta 1 parameter : 0.5
• momentum beta 2 parameter : 0.999

Initializer

• Initialize Weights in Convolutional Layers : normal distribution, mean : 0.0, standard deviation : 0.02

---

Citation

If you use this code for your research, please consider citing our paper. (arXiv preprint)

@article{jeong2020solar,
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},
publisher={IOP Publishing}
}

• Correction in our paper:
  

1. We trained our model with 3412 pairs of training dataset for 630,000 iterations (∼180 epochs).
   The number of epoch was 180, not 150.
   It is a hyperparameter that defines the number times that the deep learning model works through the entire training dataset.
   (the number of epoch) = (iterations) / (the number of dataset).

2. We used LSGAN loss instead of cGAN loss in our deep learning model.

3. When we construct global photospheric maps and compute the PFSS model, we used Heliographic coordinates, not Carrinton coordinates.

• Improved AI-generated Solar Farside Magnetograms: https://github.com/JeongHyunJin/AISFM3.0

---