UVCGAN-S: Robust and Generalizable Background Subtraction on Images of Calorimeter Jets using Unsupervised Generative Learning

I. Overview and Context This dataset provides simulated calorimeter energy histograms from the sPHENIX detector at the Relativistic Heavy Ion Collider (RHIC). It was specifically prepared for training and validating the UVCGAN-S machine learning model used in the study “Robust and Generalizable Background Subtraction on Images of Calorimeter Jets using Unsupervised Generative Learning” ([https://arxiv.org/abs/2510.23717]). The data is designed to facilitate advanced machine learning techniques, particularly for unsupervised background subtraction in heavy-ion collision environments. II. Data Structure and Format Format: The data files are provided in the standard ROOT format, requiring the ROOT framework for access and reading. Content: Each file contains multiple TH2D histograms. Projection: Electromagnetic and hadronic calorimeter energies are summed and projected onto the pseudo-rapidity (eta) and azimuthal angle (phi) space. Binning: The spatial binning is 24 x 64 (eta x phi). Event Matching: TH2D histogram names follow the pattern “*_fileXX_evtYY”. This convention allows users to match corresponding events across different samples using the shared fileXX and evtYY identifiers for the testing purpose. III. Dataset Samples The dataset is divided into five distinct samples, categorized by their intended use (Training/Testing/Evaluation): Sample ID Description Intended Use 1 PYTHIA jets embedded into HIJING (0-10% centrality Au+Au collisions). 833,000 events Training / Testing (Mixed Signal + Background) 2 PYTHIA jets (truth jet transverse momentum greater than 30 GeV). ~2.6 million events Training / Testing (Pure Signal / Reference) 3 HIJING events (0-10% centrality Au+Au collisions). ~1 million events Training / Testing (Pure Background) 4 JEWEL jets embedded into HIJING. 833,000 events Out-of-Distribution Evaluation (Quenched Signal + Background) 5 JEWEL jets without background. 869,000 events Out-of-Distribution Evaluation (Quenched Pure Signal) IV. Usage and Testing (Validation) Training (Unpaired, Unsupervised): Samples 1, 2, and 3 are the core datasets and are intended to be used in an unpaired and unsupervised manner for training the UVCGAN-S model. Users may define their own training and testing splits. Testing (Performance Study): For calculating the subtraction performance, users can pair events from Sample 1 (HIJING-embedded PYTHIA jet) with the corresponding events from Sample 2 (pure PYTHIA jet) using the event matching identifiers (see "Event Matching" in II. Data Structure and Format). The machine learning model's subtracted output can then be directly compared to the pure jet reference event. Out-of-Distribution Evaluation (JEWEL): Samples 4 and 5 (JEWEL) are provided for a critical test of model robustness and generalizability. They were not used in the training process. The study demonstrates that the UVCGAN-S model can accurately reconstruct these out-of-distribution (quenched) jets, even though training used only unquenched (PYTHIA) data. V. Citation and Further Information For a comprehensive understanding of the simulation details, data preprocessing, and the application of this dataset, please refer to the associated publication: Robust and Generalizable Background Subtraction on Images of Calorimeter Jets using Unsupervised Generative Learning [https://arxiv.org/abs/2510.23717]