Processed atmospheric boundary-layer wind-speed fields and numerical source data for stability-dependent neural-network analysis

Processed horizontal wind-speed fields derived from idealized large-eddy simulations of near-neutral, weakly convective and convective atmospheric boundary layers. The dataset accompanies the manuscript “Atmospheric stability shapes neural network sensitivity across height and time in near surface wind prediction” and contains data used for spatiotemporal neural-network prediction of near-surface wind speed. The 12 NumPy files comprise the prediction target at 4.5 m and predictor fields sampled at 49.5, 103.5 and 148.5 m (nominally 50, 100 and 150 m) for each stability regime. An accompanying source-data archive contains numerical Integrated Gradients and occlusion results for Figure 3, RMSE/SSIM values and integral-time-scale reference values for Figure 5, normalized height–lag attribution matrices for Figure 6, and region-of-interest coordinates associated with Figure 4. The original large-eddy simulations are described by Aksamit, Katurji and Zhang (2024), Journal of Geophysical Research: Atmospheres, 129, e2024JD041490. The accompanying README documents the dataset organization, source-data provenance and limitations.