Simulating Hadronization with Machine Learning

Department of Physics, University of Cincinnati, Cincinnati, Ohio 45221, USA

^* e-mail: michael.wilkinson@uc.edu

Published online: 6 May 2024

Abstract

Hadronization is an important part of physics modeling in Monte Carlo event generators, where quarks and gluons are bound into physically observable hadrons. Today’s generators rely on finely-tuned phenomenological models, such as the Lund string model; while these models have been quite successful overall, there remain phenomenological areas where they do not match data well. A machine-learning-based alternative called MLhad, intended ultimately to be data-trainable, can simulate hadronization by encoding latentspace vectors, trained to be distributed according to a user-defined distribution using the sliced-Wasserstein distance in the loss function, then decoding them. The multiplicities and cumulative kinematic distributions of pions generated with MLhad in this way match those generated using Pythia 8.

While this architecture has been successful, an alternative using normalizing flows is convenient for generating non-pion hadrons and for taking advantage of reweighting techniques to reduce computing time. Combined with new methods for reweighting the output of phenomenological models, this updated architecture should prove convenient for comparing the output of MLhad and of empirical models.