HAO Colloquium - Chris Osborne, Glasgow University

Deep learning and learning to invert a solar flare atmosphere with invertible neural networks

Deep learning is a subset of machine learning -- the statistical techniques that allow a computer to learn how to perform a certain task without being explicitly programmed -- that utilises deep neural networks to learn a task.  Deep learning has seen a rise in popularity in solar physics in the past few years with various impressive applications such as photospheric horizontal velocity field estimation and super-resolving HMI data. I aim to provide a comprehensive introduction to deep learning and the methodology of designing a deep neural network from scratch. This will include examples of uses of deep learning for solar observations whilst focusing on our approach of applying an invertible neural network, to inferring atmospheric properties from observed flaring spectral line profiles in Halpha and Ca II 8542 Angstrom.

Our network is trained using flare simulations from the one-dimensional radiation hydrodynamics code RADYN as the expected atmosphere and line profile.  I will discuss the mathematics behind this model and its training. The model is then applied to an observation of an M1.1 solar flare taken with SST/CRISP instrument just after the flare onset.  The inverted atmospheres obtained from observations provide physical information on the electron number density, temperature and bulk velocity flow of the plasma throughout the solar atmosphere ranging from 0-10 Mm in height.  We conclude that we have taught our novel algorithm the physics of a solar flare according to RADYN. This technique can also be adapted for a variety of inverse problems whilst providing extremely fast inversion samples.

Date and time: 
Monday, September 30, 2019 - 2:00pm to 3:00pm