Abstract

This thesis demonstrates how machine learning techniques can solve computationally challenging problems across diverse areas of physics, from high-energy astrophysics to condensed matter systems, accelerating traditional computation. My first contribution addresses the computational expense of Monte Carlo calculations for radiative processes in relativistic plasmas. I develop a neural network sampling method that enables fast sampling from an arbitrary probability density, and demonstrate the method on inverse Compton scattering, achieving a speedup of up to an order of magnitude beyond standard methods. My second contribution addresses the structure and radiation of neutron star magnetospheres. I use physics-informed neural networks to model the axisymmetric pulsar magnetosphere, then reproduce twisted magnetar equilibria with the same framework, and finally present a GPU-accelerated Monte Carlo transport code for resonant Compton scattering which shows that photons paradoxically escape an optically thick twisted magnetar magnetosphere after undergoing only one or two scatterings. My third contribution addresses predicting magnetic properties in condensed matter. I first train an ensemble learning method to predict uranium-based compound magnetic order using only structural information, bypassing expensive density functional theory calculations. This work is then extended using transformer-based models to predict individual atomic magnetic properties within crystals. Together these contributions show that machine learning can be integrated with physical understanding to enable investigations that were previously computationally intractable.

Committee Chair

Alexander Chen

Committee Members

Francesc Ferrer; James Mertens; Sheng Ran; Yajie Yuan

Degree

Doctor of Philosophy (PhD)

Author's Department

Physics

Author's School

Graduate School of Arts and Sciences

Document Type

Dissertation

Date of Award

8-13-2026

Language

English (en)

Download

Included in

Physics Commons

Share

COinS