Solar UV/EUV Generation (Keras, TensorFlow2)

Park et al., 2019, ApJL, 884, L23, doi:http://doi.org/10.3847/2041-8213/ab46bb

Title: Generation of Solar UV and EUV Images from SDO/HMI Magnetograms by Deep Learning

For the UV and EUV images, we use SDO/AIA 9 passbands images.

For the solar magnetograms, we use SDO/HMI Line-of-sight magnetograms.

Network Architectures

Let,

C(f, k, s) denotes as 2D Convolution layer with f filters, filter size of k, stride of s,

CT(f, k, s) as 2D Convolution-Transpose layer with f filters, filter size of k, stride of s,

B as Batch-Normalization layer,

R as ReLU activation layer,

L as Leaky-ReLU activation layer with slope 0.2,

T as Tanh activation layer,

S as Sigmoid activation layer,

and D as Dropout layer with rate 0.5.

Discriminator Network

We can select the size of the receptive field of the discriminator.

• 1x1 discriminator

C(64,1,1)-L-C(128,1,1)-B-L-C(1,1,1)-S

• 16x16 discriminator

C(64,4,2)-L-C(128,4,1)-B-L-C(1,4,1)-S

• 34x34 discriminator

C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,1)-B-L-C(1,4,1)-S

• 70x70 discriminator

C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,2)-B-L-C(512,4,1)-B-L-C(1,4,1)-S

• 142x142 discriminator

C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,2)-B-L-C(512,4,2)-B-L-C(512,4,1)-B-L-C(1,4,1)-S

• 286x286 discriminator

C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,2)-B-L-C(512,4,2)-B-L-C(512,4,2)-B-L-C(512,4,1)-B-L-C(1,4,1)-S

Generator Network

The generator network is consist of the encoder and the decoder

Encoder:

1. C(64,4,2)-L
2. C(128,4,2)-B-L
3. C(256,4,2)-B-L
4. C(512,4,2)-B-L
5. C(512,4,2)-B-L
6. C(512,4,2)-B-L
7. C(512,4,2)-B-L
8. C(512,4,2)-B-L
9. C(512,4,2)-B-L
10. C(512,4,2)-R

Decoder:

1. CT(512,4,2)-B-D-R
2. CT(512,4,2)-B-D-R
3. CT(512,4,2)-B-D-R
4. CT(512,4,2)-B-R
5. CT(512,4,2)-B-R
6. CT(512,4,2)-B-R
7. CT(256,4,2)-B-R
8. CT(128,4,2)-B-R
9. CT(64,4,2)-B-R
10. CT(1,4,2)-S

The generator network has skip-connections between i-th layers of the encoder and (10-i)-th layers of the decoder like the U-Net architecture.

Skip-connection:

• encoder 1st layer - decoder 9th layer
• encoder 2nd layer - decoder 8th layer
• encoder 3rd layer - decoder 7th layer
• encoder 4th layer - decoder 6th layer
• encoder 5th layer - decoder 5th layer
• encoder 6th layer - decoder 4th layer
• encoder 7th layer - decoder 3rd layer
• encoder 8th layer - decoder 2nd layer
• encoder 9th layer - decoder 1st layer

To run this code,

Keras: Change some parameters in option.py (about your environments) and run train.py

TensorFlow2: Change some parameters in solar_generation_tf2.ipynb and run the notebook