Machine Learning Parameterization of Spin Structure Functions

Machine Learning (ML) provides a method for accurate parameterizations of observables with uncertainty quantification. By fitting a ML-based modeling framework to world data for spin structure from lepton-nucleon scattering, we seek to develop a parameterization that captures the different behaviors of the nucleons different scales of virtuality Q² and invariant mass-squared W². Existing theoretical models often focus on constructing a proton out of its constituent quarks, which works well for moderate to high Q² and invariant mass-squared W², where perturbative QCD holds. However, these frameworks break down at lower Q² and W², resulting incomplete descriptions of data. This regime is where a patchwork of phenomenological parameterizations prove invaluable, however, these tools often rely on assumptions of function form, have incompletely quantified uncertainties, and struggle to span large portions of the kinematic phase space. Therefore, we seek to provide a less statistically-biased approach, using ML-based architecture agnostic of functional form, with built-in uncertainty quantification, and spanning the entire range of Q² and W². We present preliminary results for such an ML-based parameterization of the world data for spin structure functions g₁ using Gaussian Process Regression.