Stokes inversion based on convolutional neural networks

Introduction

Spectropolarimetric inversions are routinely used in the field of Solar Physics for the extraction of physical information from observations. The application to two-dimensional fields of view often requires the parallelization of the inversion codes. Still, the time spent on the process is very large. SICON (standing for Stokes Inversion based on COnvolutional Neural networks) is a new inversion code based on the application of convolutional neural networks that can provide the thermodynamical and magnetic properties of two-dimensional maps instantaneously. We use two different architectures using fully convolutional neural networks. We use synthetic Stokes profiles obtained from two numerical simulations of different structures of the solar atmosphere for training.

This repository is the release version of SICON. We provide both training and evaluation scripts.

Datasets

If you want to retrain any of the two architectures you will need to download the training data from https://owncloud.iac.es/index.php/s/bGkJrn8mCuB4j9n

For evaluation, you can use your own Hinode dataset. However, we provide a patch of the active region AR10933 that can be downloaded from the same server.

Encoder-decoder architecture

Dependencies

numpy
h5py
scipy
tqdm
nvidia_smi
pytorch

Within an Anaconda environment, standard packages can be installed with

conda install numpy h5py scipy tqdm

pytorch can be installed following the instructions on http://pytorch.org

nvidia_smi can be installed using the recipe from fastai

conda install nvidia-ml-py3 -c fastai

Training

Training the encoder-decoder architectures simply requires to run the training as:

python train_encdec.py

If you have an NVIDIA GPU on your system, it will make use of it to accelerate the computations. If not, prepare to wait for a long time. Typical running times in a P100 GPU are 90 seconds per epoch. A training with 50 epochs will last for slightly longer than an hour.

Evaluation

The evaluation of this architecture can be easily done by typing:

python evaluate_encdec.py

It will generate figures and an HDF5 file with the output. You can easily modify this script to read your own dataset.

Concatenate architecture

Dependencies

numpy
h5py
tensorflow
keras

Within an Anaconda environment, depending if you have a GPU or not the command will be different:

conda install tensorflow keras

conda install tensorflow-gpu 

Training

Training the encoder-decoder architectures simply requires to run the training as:

python train_concat.py

Evaluation

The evaluation of this architecture can be easily done by typing:

python evaluate_concat.py

It will generate figures and an numpy file with the output.