Dataset for Koopmans, Kay, Youk (2026): Predictability can be dynamically constructed in deterministic systems

Derived datasets for the study "Predictability can be dynamically constructed in deterministic systems" by Lars Koopmans, Elinor M. Kay, and Hyun Youk (2026). [ Preprint at https://arxiv.org/abs/2604.01088 ] Everything here comes from one deterministic cellular automaton: 196 cells on a 14x14 triangular lattice with periodic boundaries, each cell in one of four states. Cells communicate through two diffusible signals. From a maximally disordered initial configuration the lattice always settles into one of three fates, which we call static, spiral wave, and rectilinear wave. We ran 1,000,000 independent trajectories and recorded the state of every cell at every timestep. Because the automaton is deterministic, the initial configuration completely fixes which fate occurs. The study asks when that predetermined fate becomes practically predictable, and what carries the predictability. These datasets hold the quantities the figures are built from. CONTENTS order_time_series.tar (4.9 GB) - winding-field, vortex, and spatial-order time series for all 1,000,000 trajectories, in 1000 chunks of 1000. ml_splits_Ttrue_f*.tar.gz (11 files, 660 MB total) - train/validation/test splits of lattice configurations sampled at fixed fractions f of each trajectory's runtime, from f = 0.00 (the initial configuration) to f = 1.00 (the moment the final pattern forms). Each archive holds one value of f and contains the configurations, the per-cell vortex fields, and the metadata linking every row back to its trajectory. winding_fields.tar (493 MB) - per-cell winding and vortex fields, aligned row-for-row with the splits above. Koopmans_Source_Data.zip (2.6 MB) - the plotted coordinates for every panel of the main figures. README.md - full documentation: file formats, array shapes, key names, conventions, simulation parameters, and random seeds. SHA256SUMS.txt - checksums for every archive. WHAT IS NOT HERE The raw configurations from the 1,000,000-trajectory run come to 268 GB and are not deposited. They are not needed to reproduce any figure, because the datasets above already contain the quantities the figures are built from. The simulation is fully deterministic, so the raw output can be regenerated exactly from the generator script and the random seeds, both of which are documented in README.md and available in the code repository. CONVENTIONS Cell states are integers 1-4. Fate is 1 for static, 2 for spiral wave, 3 for rectilinear wave. Across the 1,000,000 trajectories the split is 267,335 static, 298,396 spiral, and 434,269 rectilinear. Time is usually reported as the fraction f = t / T_true, where T_true is the transient length of that trajectory, because runtimes vary by two orders of magnitude between trajectories. CODE Analysis and simulation code: https://github.com/youklab/GCA-predictability-2026