This code trains a physics-informed neural network (PINN) to learn the steady-state, one-dimensional radiative transfer equation. It is intended for astrophysical applications to stellar atmospheres, using realistic opacities, bulk densities, and distances. Right now, it solves a simple proof-of-concept problem to establish that PINNs are a viable step toward emulating 3D stellar atmosphere spectra.
The code currently uses the physical parameters described in the sim_data/*.fits files corresponding to a Solar-like stellar atmosphere. These are loaded in utils.py and are currently hard-coded; support for flexible stellar atmosphere inputs will be added in the future.
To train the PINN and generate diagnostic plots, run the Python script run_RadAI.py. This will generate a default runs/RadAI directory containing the I_vs_lambda.pdf plot showing the spectrum at various distances throughout the atmosphere compared to an analytic solution, with comparison to a spectrum calculated by the Korg 1-D radiative transfer code written in Julia.