PNNL Generative Downscaling of ERA5 to 12 km over CONUS (2020-2021)

This is a child record of the parent dataset: 10.5281/zenodo.22711782. This dataset provides hourly, 12 km meteorological fields over the contiguous United States (CONUS), produced by generatively downscaling ERA5 reanalysis (~25 km) with a consistency model — a one-step generative diffusion model — trained on 12 km WRF-TGW regional climate simulations. Conventional interpolation of a coarse reanalysis cannot create the fine-scale structure that a convection-permitting or mesoscale simulation resolves: orographic precipitation bands, sharp frontal gradients, terrain-driven wind features. The approach used here instead learns that fine-scale structure from a high-resolution regional model and then transfers it onto the large-scale state supplied by ERA5. The result retains ERA5's synoptic evolution while adding physically plausible sub-reanalysis-scale detail, at a computational cost several orders of magnitude below running the regional model directly. Eight surface variables are provided, generated jointly at each timestep so that cross-variable relationships (for example between temperature, humidity and radiation) are internally consistent rather than downscaled independently. The key characteristic users must understand is that the fine scales in this product are generated, not observed. See Assumptions and Limitations.