Abstract
In this thesis, I investigate whether the Physics-Informed Neural Network (PINN) frame- work can be used for radiative transfer computations in the solar atmosphere. I have investigated the use of PINNs to compute the formal solution of the radiative transfer equation in 1D model atmospheres for the Ca II 854.2 nm line. I find that the method without modification is ill-suited for the problem. The order-of-magnitude variation of the terms in the radiative transfer equation makes the loss function and fitting of the network unbalanced. However, simple modifications can be made to improve the results. Good results are achievable by weighting the residuals of the transfer equation so that they are balanced from the start of the fitting, but some challenges still re- main related to the order-of-magnitude variation of the incoming intensity.
Lastly, the method is computationally heavier than simple numerical integration, and computing the solution, from scratch, over a 1D atmosphere takes at least minutes when the fitting is run on a GPU. How the method scales to 3D problems needs investigation.
Supervisor: Associate Professor Tiago Pereira, Institute of Theoretical Astrophysics, UiO
Intern. assessor: Professor Øystein Elgarøy, Institute of Theoretical Astrophysics, UiO
Extern. assessor: Professor Jorrit Leenaarts, Stockholm University