Lattice QCD TMD Pipeline

Context

PhD research under Dr. Michael Engelhardt, New Mexico State University. Lattice QCD calculations of Transverse Momentum Dependent parton distributions (TMDs), which describe how quarks and gluons move inside the proton, produce terabytes of noisy Monte Carlo data that has to be turned into physics.

Problem

Extracting an analytical form from 30,000+ noisy lattice observables, at a processing cost that made iteration slow.

Approach

An end-to-end ML pipeline: Monte Carlo lattice data → GPU-accelerated CUDA C++ preprocessing → symbolic regression (PySR) to recover a closed-form analytical expression, with jackknife/bootstrap resampling for uncertainty quantification at every stage.

Architecture

Monte Carlo simulation → CUDA C++ preprocessing (10× faster than the CPU pipeline it replaced) → PySR symbolic regression → discovered analytical form, cross-checked against jackknife/bootstrap error bands.

Results

98% accuracy recovering the analytical structure across 30,000+ observables, with GPU-accelerated CUDA C++ cutting processing time 10× and 75,000+ CPU/GPU hours run on NERSC Perlmutter.

Stack

CUDA C++ · PySR · SLURM · Python · NERSC Perlmutter · Jackknife/bootstrap uncertainty quantification