The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Machine learning and computer vision in solar physics
Author:	Jiang, Haodi
View Online:	njit-etd2021-083 (xx, 131 pages ~ 73.8 MB pdf)
Department:	Department of Computer Science
Degree:	Doctor of Philosophy
Program:	Computer Science
Document Type:	Dissertation
Advisory Committee:	Wang, Jason T. L. (Committee chair) Calvin, James M. (Committee member) Ding, Xiaoning (Committee member) Wang, Haimin (Committee member) Bai, Xiaoli (Committee member)
Date:	2021-12
Keywords:	Computer vision Deep learning Machine learning Neural networks Solar physics Space weather
Availability:	Unrestricted
Abstract:	In the recent decades, the difficult task of understanding and predicting violent solar eruptions and their terrestrial impacts has become a strategic national priority, as it affects the life of human beings, including communication, transportation, the power grid, national defense, space travel, and more. This dissertation explores new machine learning and computer vision techniques to tackle this difficult task. Specifically, the dissertation addresses four interrelated problems in solar physics: magnetic flux tracking, fibril tracing, Stokes inversion and vector magnetogram generation. First, the dissertation presents a new deep learning method, named SolarUnet, to identify and track solar magnetic flux elements in observed vector magnetograms. The method consists of a data preprocessing component that prepares training data from a physics-based tool, a deep learning model implemented as a U-shaped convolutional neural network for fast and accurate image segmentation, and a postprocessing component that prepares tracking results. The tracking results can be used in deriving statistical parameters of the local and global solar dynamo, allowing for sophisticated analyses of solar activities in the solar corona and solar wind. Second, the dissertation presents another new deep learning method, named FibrilNet, for tracing chromospheric fibrils in Ha images of solar observations. FibrilNet is a Bayesian convolutional neural network, which adopts the Monte Carlo dropout sampling technique for probabilistic image segmentation with uncertainty quantification capable of handling both aleatoric uncertainty and epistemic uncertainty. The traced Ha fibril structures provide the direction of magnetic fields, where the orientations of the fibrils can be used as a constraint to improve the nonlinear force-free extrapolation of coronal fields. Third, the dissertation presents a stacked deep neural network (SDNN) for inferring line-of-sight (LOS) velocities and Doppler widths from Stokes profiles collected by GST/NIRIS at Big Bear Solar Observatory. Experimental results show that SDNN is faster, while producing smoother and cleaner LOS velocity and Doppler width maps, than a widely used physics-based method. Furthermore, the results demonstrate the better learning capability of SDNN than several related machine learning algorithms. The high-quality velocity fields obtained through Stokes inversion can be used to understand solar activity and predict solar eruptions. Fourth, the dissertation presents a generative adversarial network, named MagNet, for generating vector components to create synthetic vector magnetograms of solar active regions. MagNet allows us to expand the availability of photospheric vector magnetograms to the period from 1996 to present, covering solar cycles 23 and 24, where photospheric vector magnetograms were not available prior to 2010. The synthetic vector magnetograms can be used as input of physics-based models to derive important physical parameters for studying the triggering mechanisms of solar eruptions and for forecasting eruptive events. Finally, implementations of some of the deep learning-based methods using Jupyter notebooks and Google Colab with GitHub are presented and discussed.