Deep Learning for Solar Physics Prediction
Solar flares release a huge amount of energy and radiation and can affect the Earth in the worst case. Predicting these events is therefore of major importance. A large effort in the community strives to address this problem by applying machine learning algorithms. This work focuses on a new deep learning method called self-supervised learning (SSL), applied to solar data to learn pattern and structure in the image. This approach permits the use of larger data volumes and overcomes the limitations of supervised learning caused by a low number of labelled samples and class imbalance.
For this task, a curated dataset is processed and refined for SSL and solar flares prediction. Another dataset called SDO-Benchmark is used for benchmarking. The contributions are summarized as follows: (1) Various conventional deep learning models are trained and show interesting performance. (2) Self-supervised learning is applied to solar images following SimCLR framework and proves to learn a good representation of the data. This has never been done in the past. (3) A dataset is prepared and is now usable for many tasks, including SSL pre-training and flares classification. (4) A library resulting from this work shows exemplary reproducibility ability and permits the use of the pre-trained models.
By combining these findings, a classifier fine-tuned using the SSL model on the SDO-Benchmark dataset achieves a TSS = 0.646 on binary classification (no flare / flare with >=C-class in a 24 hour period). The encoder trained using the SSL method provides a good and useful representation of the input. It could be used as a feature extractor for many downstream tasks with low constraints on the amount of data, time and processing power.
Using pip:
python -m venv venv
source venv/bin/activate # On Unix
.\venv\Scripts\activate # On Windows
pip install -r requirements.txt
deactivate # When the job is doneUsing conda:
conda env create --file environment.yaml
conda activate solarnet-conda-env
conda deactivate # When the job is donedvc reproUp-to-date results are available in models/ folder.
- Binary classification model
- Binary classification model on full-disc images
- Self-supervised learning pre-trained model
- Self-supervised learning finetuned model on binary classification, SDO-Benchmark
- Self-supervised learning finetuned model on binary classification, testing set on years 2015-2017
- Self-supervised learning finetuned model on binary classification, testing set split by month