Abstract
Solar filaments (also called prominences when seen off-disk) are solar atmospheric structures consisting of dense, cool plasma clouds floating within the sun’s corona. Since the beginning of solar observations, it has been seen that the prominences oscillate with a wide variety of motions. These periodic motions are very common, but there are no systematic studies of these oscillations. It has recently been shown that spectral analysis of solar filaments is a powerful tool to identify oscillations in these structures. With this technique, the power spectral density (PSD) is calculated for each pixel of the Halpha images. To differentiate between a detection or a spurious oscillation, it is necessary to determine the background noise. We have seen that this background noise is a combination of red and white noise. The red-noise nature of the PSD is problematic in their study since most of the statistical tools developed to identify real oscillations from the noise are for white-noise PSD. The most appropriate approach for this problem is the usage of Bayesian statistics and Monte Carlo Markov Chains (MCMC). MCMC methods can be computationally expensive and have been proven to be too slow for our research aims, so we tackle this problem with deep learning techniques, specifically Convolutional Neural Networks (CNNs). We developed two neural networks, which reproduce the same outcomes as the MCMC methods. Both have been trained with synthetic data as well as real data from the MCMC methods. The results obtained show negligible differences with the results from the MCMC methods but with the advantage of computing times orders of magnitude smaller.
Co-authors:
Manuel Luna, Jaume Terradas