Simulated six-dimensional beam-dynamics dataset for machine-learning surrogate modelling

This record contains a simulated beam-dynamics dataset developed for machine-learning surrogate studies of six-dimensional particle-distribution evolution in accelerator systems. The work was carried out in the context of exploratory modelling of collective effects for the FCC-ee High-Energy Booster. The principal data file, neural_xsuite_dataset_2026-04-28T09_40_32.npz, contains paired initial and final macroparticle clouds together with the conditioning variables used for each simulation. The six-dimensional coordinate convention is (x, p_x, y, p_y, ζ, δ). The principal stored arrays are: X_cloud: initial six-dimensional macroparticle distributions; Y_cloud: corresponding final distributions after tracking; MU: simulation and machine-conditioning variables; train, val, and test: fixed sample-index arrays, when present. Each six-dimensional cloud can be represented through the fifteen distinct pairwise two-dimensional phase-space marginals. The accompanying preprocessing implementation constructs 64 × 64 histograms using sample-adaptive coordinate ranges defined by the centroid plus or minus five RMS beam sizes. Each histogram channel is normalized to unit probability mass. Physical-density representations can additionally be obtained by dividing the probability mass by the corresponding two-dimensional bin area. The dataset supported the training and evaluation of conditional variational autoencoders, causal Transformers, local-window Transformers, Graph Neural Operators, physical-statistics predictors, and conditional latent-uncertainty diagnostics. It contains simulated data rather than experimental or operational accelerator measurements. The release is intended to support reproducibility of the accompanying master’s thesis and research on machine-learning surrogate models for high-dimensional beam dynamics.