HadML

Deep generative models for hadronization in particle event simulation.

Hadronization is the process by which quarks and gluons form the particles measured by detectors. I develop generative models that can be fitted to observed hadron distributions and evaluated against conventional event generators such as Herwig.

In Fitting a deep generative hadronization model, we fitted a generative adversarial network to two Herwig configurations using final-state hadrons as fit targets. An auxiliary neural classifier distinguished the configurations with an AUC of 0.77, compared with 0.57 for the best individual observable tested.

In Integrating particle flavor into deep learning models for hadronization, we extended HadML to generate both kinematics and particle identities for two-body cluster decays. The model reproduced Herwig’s shift toward strange hadrons when an incoming quark was strange.

Diagram of the HadML generative adversarial model
Model architecture for fitting HadML to final-state hadron distributions.

Training history (PDF) · Generated angle distribution (PDF)