Bayes in space: adding uncertainty to deep learning in solar physics

Deep learning is an increasingly common tool used in astrophysics, with the flexibility and speed of neural networks (NN) meaning they can be applied to a wide range of data intensive problems. For example, NNs have been developed to provide rapid estimates of the Sun's atmospheric temperatures, helping to address major questions in solar physics, e.g., how is the corona heated to temperatures of over a million degrees? One key issue is that the current networks do not provide a measure of the result's uncertainty, which is crucial for determining whether the answers provided by the NN are trustworthy. This is also of great importance in an industry setting, where a poor or incorrect answer could be dangerous or expensive.

During this project, the student's goal is to develop NNs that incorporate uncertainty through the implementation of Bayesian approaches, which are currently at the forefront of deep learning research. The student would learn how to programme in Python and develop knowledge of the cutting-edge deep learning software (Tensorflow, Pytorch). Further, they would gain experience in the field of solar physics, working with data from NASA/ESA satellites. The developed networks will be applied to problems important in solar physics research (e.g., solar plasma temperature estimation, removal of data noise/corruption), with the opportunity to share work at scientific conferences in the UK and abroad. The student will also apply the skills learnt in an industry setting, undertaking a placement with our industrial partners.