Global significant wave height data for GWSM4C training and testing

A novel deep-learning model, the Global Wave Surrogate Model for Climate Simulation (GWSM4C), has been developed based on convolution architecture. GWSM4C excels in accurately emulating both global ocean wave propagation and the instantaneous impact of winds considering wind-wave and swell characteristics, as well as introduces a fresh perspective for incorporating the wave component into climate models with small computational costs.The training and testing data adopted in establishing GWSM4C are presented in this dataset. The GWSM4C model is trained by the global hourly significant wave height (SWH) generated via the wave model (Yang et al., 2005; Yuan et al., 1991; Yuan Yeli et al., 1992) developed by the Key Laboratory of Marine Science and Numerical Modeling, MNR, PRC (hereafter MASNUM-WAM) with the forcing winds derived from ECMWF-ERA5 reanalysis hourly data (Bell et al., 2021; Hersbach et al., 2020), and the training period spans the years 2017–2020. As an independent test, the SWH in 2021 yielded from GWSM4C is compared with that generated from the same wave modeling environment mentioned above.MASNUM-WAM is a third-generation wave model that solves the energy spectrum balance equation in wavenumber space and uses a complicated characteristic inlaid scheme in spherical coordinates to simulate shoaling and refraction effects in shallow waters, the modulation of background current to wave evolution, and the refraction of waves propagating along great circles. The source functions adopted in this work are from the ST6 package (Rogers et al., 2012; Zieger et al., 2015), which considers the effects of wind input, white-capping dissipation, and swell dissipation on the evolution of waves; and the DIA scheme (Hasselmann and Hasselmann, 1985a, 1985b) is employed to quantify the nonlinear energy transfer between waves. The wave modeling adopts a global computational grid that spans 80°S to 80°N and 0°E to 359.5°E with a 0.5°×0.5° horizontal resolution. The spectral space is set to 24 directions, with intervals of 15°, and 35 wavenumbers logarithmically arranged from the lower threshold of 0.0071 to the upper limit of 4.6341, with intervals characterized by ki/ki+1=1.21, which is equivalent to frequencies from 0.042 Hz to 1.073 Hz with a ratio of 1.1 at infinite depth. The bathymetric data are derived from the ETOPO1 dataset (NGDC, n.d.) published by the National Oceanic and Atmospheric Administration National Geophysical Data Centre. The shoreline data are procured from the Global Self-consistent, Hierarchical, High-resolution Geography Database (Wessel and Smith, 1996).All files in this dataset cover the same regions with the exact horizontal resolution as the computational grid mentioned above. The hourly SWH simulated by MASNUM-WAM is archived yearly in the NetCDF format with the filenames of ‘Hs_<yyyy>.nc’, where <yyyy> denotes the years 2017–2021. And the ECMWF-ERA5 winds have been converted into the files of ‘UV_<yyyy>.nc’, where the hourly eastward and northward wind speeds are denoted as ‘windx’ and ‘windy’, respectively. The 3-hourly SWH testing outputs of GWSM4C in 2021 are stored in the dataset 'GWSM4C' and comprised by the file ‘GWSM4C_test_2021_3h.h5’, and the corresponding MASNUM-WAM simulations are stored in the same file and denoted as ‘MASNUM’. The latter is just the same as the variables contained in ‘Hs_2021.nc’, but with a time-frequency of three hours.