Unbinned measurement of Z+jets production, unfolded to full charged-particle phase space using the OmniFold algorithm

Description

This dataset is the full phase space measurement of Z(mu mu)+jets production at sqrt(s)=13 TeV performed by the ATLAS Collaboration, unfolded with the OmniFold machine-learning unfolding algorithm. Because the measurement is unbinned, results are provided as a dataset of particle level events simulated with Monte Carlo methods, with per-event weights provided by OmniFold. Applying the weights to the events and binning the events in some observable produces a differential cross section measurement in that observable.

Each Parquet file is self-contained, combining per-event variable-length arrays of particle kinematics (transverse momentum, pseudorapidity, azimuthal angle, and PDG ID for the two Z-decay muons plus all charged hadrons) with ~164 OmniFold weight columns per event. The release includes:

- data.parquet, data-hv.parquet: the measurement dataset
- pseudodata.parquet, pseudodata-hv.parquet, truth_pseudodata.parquet: a pseudodata dataset for closure testing against known MC truth
- truth_madgraph.parquet: Monte Carlo truth-level predictions from MadGraph5_aMC@NLO+Pythia8, for comparison
- truth_sherpa.parquet: Monte Carlo truth-level predictions from Sherpa 2.2.11, for comparison
- models/: all trained OmniFold network checkpoints used to build the data measurement. Seven iterations of OmniFold are run for each independent run of the method, and 10 ensembles are run per most systematic variations. There are 28 run groups and 2695 checkpoint files total, enabling users to rerun inference for both the nominal result and some of the associated uncertainties.

Paper accompanying this data release

Algorithm used to produce the unbinned per-event weights in this dataset

Prior ATLAS Z+jets unfolded measurement (binned, restricted phase space) that this full-phase-space dataset builds on

Dataset characteristics

How were these data generated?

Per-event OmniFold weights were produced by an ensemble of 10 independently trained transformer networks, each unfolded for 7 iterations (70 step-2 networks total), operating on 10 input features per particle (log transverse momentum, pseudorapidity, azimuthal angle, mass, jet-association one-hot encodings). MC truth-level predictions for comparison were generated with MadGraph5_aMC@NLO+Pythia8 and with Sherpa 2.2.11.

Step unfolding
Generators: OmniFold
Note: 10 independent training runs, 7 iterations each, ensembled to produce weights_omnifold and related weight columns

Step simulation
Generators: MadGraph5_aMC@NLO+Pythia8 Sherpa 2.2.11
Note: Truth-level Monte Carlo predictions provided for comparison against the unfolded data

How were these data validated?

See the paper and public repository for validation procedures performed for this measurement.

How can you use these data?

For full usage recommendations and examples, see the public repository released with these data.

Citation policy

Source code repository

Files and indexes

Disclaimer

These open data are released under the Creative Commons Zero v1.0 Universal license.

Neither the experiment(s) ( ATLAS ) nor CERN endorse any works, scientific or otherwise, produced using these data.

This release has a unique DOI that you are requested to cite in any applications or publications.