The flare necessities : machine learning tools for solar flare data analysis

The study of the lower flaring atmosphere of the Sun is one facet of understanding the complex physics involved in solar flares and their effect on space weather and the Earth. Despite a rich history of investigation into the study of the lower flaring atmosphere, there are still many unanswered questions in regards to the mechanism of energy deposition and the response to such energy being injected into the atmosphere. This thesis aims to provide tools for future researchers to rigorously explore these problems. In particular, this thesis looks at how machine learning - with a particular focus on deep learning - can improve data storage and analysis pipelines as well as uncover new results from data that were not feasibly possible before. In Chap. 1, the standard model of a solar flare is introduced and its extension to three dimensions explained. This allows for the definition of a flare ribbon - the brightest points in the lower solar atmosphere resulting from direct heating from a flare - which is a key observational feature whose origin is explored in later chapters. A brief history of study on flare ribbons is then given with a particular focus on the asymmetries in spectral lines that show clear flare ribbons. These asymmetries link directly to the velocity field in the flaring atmosphere as a static atmosphere would yield symmetric profiles. This gives a direct diagnostic of the motion happening in the atmosphere as it is heated and the ribbons evolve. In Chap. 2, the field of deep learning is introduced from its inception to the current models used today. This chapter covers how to build and train deep neural networks and some best practices when implementing these tools. The telescopes and detectors used to obtain the data analysed in Chaps. 4 - 6 are described in Chap. 3. In this chapter, the inner workings of the Swedish 1-m Solar Telescope's CRisp Imaging SpectroPolarimeter (SST/CRISP), Hinode's Solar Optical Telescope (Hinode/SOT) and Solar Dynamics Observatory's Atmospheric Imaging iii Assembly (SDO/AIA) is described. Chap. 4 introduces a deep convolutional neural network (CNN) trained on Hα images from Hinode/SOT for solar image classification. This is trained to distinguish between five classes of solar features prominent in Hα: filaments, flare ribbons, prominences, sunspots and the absence of any of the other four features. The final model has a validation accuracy of 99.2% misclassifying only one image in the validation dataset. The trained CNN is then tested with adversarial examples from SDO/AIA UV continua and EUV spectral line images where the features look perceptually different but still identifiable to the human eye. This demonstrates that the network cannot identify these features in different wavelengths well and to extend this network to non-visible wavelengths, the training set must be expanded to include such wavelengths. The trained CNN in this chapter is used further in Chap. 5 for transfer learning - the process of using a trained deep learning model to influence the training of another, related deep learning model. In Chap. 5, a method based on deep learning for correcting the atmospheric effects in optical solar flare observations is presented. This takes the form of a fully convolutional autoencoder trained on data from SST/CRISP imbued with synthetic seeing described by the model developed in the first sections of the chapter. The trained model works well on the validation dataset showing accurate reconstruction of both spatial and spectral elements of the data. SST/CRISP data with real atmospheric seeing is then corrected by the trained model. The sources of error in this reconstruction are discussed with a coarse error estimate on the recovered intensity values used. Then in Chap. 6 a novel deep learning method for estimating the parameters of the flaring atmosphere from observations is presented - an Invertible Neural Network (INN). The INN is trained on synthetic flare data produced by the one dimensional radiation hydrodynamics code RADYN with near-perfect restoration of the atmospheric paramters during validation. This is then applied to a single pixel from a CRISP image to show the power of this method in disentangling the ambiguity in the velocity field responsible for observed asymmetry in the spectrum. This method is then applied to flare ribbons as a whole - which are selected through a combination of a Gaussian Mixture Model (GMM) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) - to determine the specific motions of the flaring velocity field responsible for the observed spectral line asymmetries.