Subject Area
Physics
Abstract
In both cosmology and small-x quantum chromodynamics, the relevant theoretical predictions are well-defined but expensive to evaluate repeatedly over broad parameter spaces or large ensembles of simulated data. The central question of this thesis is when machine learning can accelerate those calculations without weakening the reliability of the final physics conclusions. The thesis develops and tests this idea in three settings. First, it studies the Derivative Approximation for Likelihoods (DALI) as a means of diagnosing when Fisher-matrix forecasts cease to be trustworthy in cosmological parameter inference. Second, it investigates graph-based spherical neural networks for estimating primordial non-Gaussianity directly from simulated cosmic microwave background maps and benchmarks their performance against Fisher and Komatsu-Spergel-Wandelt estimator targets. Third, it constructs neutral-network surrogates for the Balitsky-Kovchegov equation in order to accelerate diploe-model calculations relevant to deep-inelastic scattering as small Bjorken-x. Across these projects, the main conclusion is methodological rather than a novel discovery. Machine learning is most useful here when it functions as a calibrated surrogate within a theory-first workflow. In that setting, the exact calculation remains the standard of truth; the learned model is trained against controlled numerical data, and performance is judged by physics-level validation rather than by generic machine-learning metrics alone. The results show that surrogate models can provide substantial computational gains in controlled regimes, but they must be paired with explicit validity tests and final cross-checks against higher-fidelity calculations before they are used to support scientific claims.
Degree Date
Spring 2026
Document Type
Dissertation
Degree Name
Ph.D.
Department
Physics
Advisor
Fred Olness
Second Advisor
Difeng Cai
Third Advisor
Thomas Coan
Fourth Advisor
Joel Meyers
Number of Pages
177
Format
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Stevenson, Brandon, "Machine Learning in Fundamental Physics: From Early Universe Signatures to QCD Evolution" (2026). Physics Theses and Dissertations. 24.
https://scholar.smu.edu/hum_sci_physics_etds/24
