MCMC for Stellar Magnetic Field Reconstruction
The goal of this project is to accelerate Markov Chain Monte Carlo methods for scientific applications, in particular in the context of stellar magnetic field reconstruction. Astrophysical spectropolarimetry has enabled astronomers to detect and characterize surface magnetic fields of stars. The magnetic fields induce circular polarization and spectropolarimetric observations can be used to find an inverse mapping between the stellar magnetic field maps and the spectropolarimetric observations. In this project, a probabilistic model of the inverse mapping is formulated in a Bayesian setting, and inference is carried out through the Metropolis-adjusted Langevin algorithm. Understanding the magnetic fields of stars can aid in our understanding of the stellar evolution and other stellar processes. The Python library JAX is used to increase convergence speeds in high dimensions. JAX is optimized for GPU usage, and due to the high-dimensionality of the problem, the computational resource applied for is a prerequisite for successful outcome of the project.