The Role of Coronal Holes in Space Weather Forecasting: Reliable Reconstruction and Assimilation of CHs with Machine Learning Models

Abstract

Space weather encompasses violent eruptions of plasma, radiation, and particles that affect the solar wind and near-Earth space, and can threaten technology and human health. Forecasting of severe events depends on our ability to accurately represent
the solar wind, which is related to Coronal Holes (CHs). We consider the onion layer model of the Sun, which explains how the solar surface is rotating and relates to the rotating and churning effects of the solar magnetic field. These effects create many special structures in the solar atmosphere like filaments, coronal holes, etc. To improve the forecasting of the CHs, the GONG-ADAPT-WSA model pipeline was used in the UK Met Office. The research described in the thesis is data-oriented and combines
numerical and deep learning algorithms to analyse and modify the outputs from the ADAPT-WSA part. Data from a segmentation algorithm (CHIMERA) is used to obtain segmented CHs from satellites. The segmented images are then used to train the machine learning algorithm to reconstruct the ensemble images.

The first step is to register images from the pipeline model and observation onto the same coordinate frame since the two groups are stored with different resolutions and coordinate systems. After registration onto the same rectangular grids, a CNN-based neural network is used to reconstruct ADAPT-WSA ensembles based on projected CHIMERA. Due to the limited amount of true data available to us, different types of synthesised data were used to both select the deep learning structures and also to carefully test the resulting algorithm. The use of synthetic data was essential in assessing the performance and robustness of the algorithm when handling different types of noisy images.

After obtaining the reconstructed results from the application of the algorithm, the thesis focuses on analysing the ensemble outputs and developing some methods to analyse all of the data systematically. This is essential since we need to measure the quality of over 70,000 images, only images within 2014. The method of medoid clustering and signal processing methods was used to give a new angle of understanding data and found a better subset cluster from the original ensemble groups. The measuring scores from ADAPT-WSA to CHIMERA can be formed as time series. By using the Fast Fourier transform and the Savitzky-Golay filter, we are automatically able to examine the data quality, modify the deep learning algorithms, and construct new synthesised data in return.

Date of Award	22 May 2024
Original language	English
Awarding Institution	University of Bath
Supervisor	Chris Budd (Supervisor), Tom Fincham Haines (Supervisor), Siegfried Gonzi (Supervisor) & David R. Jackson (Supervisor)