Reproducibility of Results in Legasa et al. (2026)

Datasets for reproducibility of the findings in Legasa et al. (2026), entitled "Regional Climate Model Emulation with Diffusion Approaches: What is the Added Value of Generative Machine Learning?", and submitted to Journal of Advances in Modeling Earth Systems (JAMES). The code can be found in https://github.com/MNLR/Diffusion/. The dataset used for training is derived from the CNRM-ALADIN63 regional climate model simulation in its EURO-CORDEX configuration (run r1i1p1). The training set combines the historical and RCP8.5 simulations driven by the CNRM-CM5 global climate model, covering a total of 150 years. The target variable is daily precipitation from ALADIN63 at 0.11° horizontal resolution, over a 64 × 64 grid-point domain centred over France and covering a small part of Spain. The predictor fields are also derived from ALADIN63 but are spatially coarsened to approximately GCM resolution, following a perfect-model framework in which the emulator learns the downscaling relationship internal to the RCM. Note that we applied a spatial moving average filter to eliminate any high-resolution features that might persist through the interpolation. The input domain consists of 22 × 16 grid points on the CNRM-CM5 grid, centred over the target precipitation domain. In addition to 2D atmospheric predictors, the dataset includes 1D predictors such as spatial summary statistics, external forcings, and seasonal indicators. The 2D predictors include: Geopotential height (zg) at 850, 700 and 500 hPa: atmospheric geopotential height, in metres (m). Specific humidity (hus) at 850, 700 and 500 hPa: specific humidity, in kilograms per kilogram (kg kg⁻¹). Air temperature (ta) at 850, 700 and 500 hPa: air temperature, in kelvin (K). Eastward wind component (ua) at 850, 700 and 500 hPa: zonal wind speed, in metres per second (m s⁻¹). Northward wind component (va) at 850, 700 and 500 hPa: meridional wind speed, in metres per second (m s⁻¹). Sea-level pressure (psl): pressure reduced to mean sea level, in pascals (Pa). The 1D predictors include: Daily spatial mean of each 2D predictor field over the predictor domain, with the same physical units as the corresponding 2D variable. Daily spatial standard deviation of each 2D predictor field over the predictor domain, with the same physical units as the corresponding 2D variable. Total anthropogenic greenhouse gases: yearly external forcing variable. Solar forcing: yearly external forcing variable, expressed in watts per square metre (W m⁻²). Ozone forcing: yearly external forcing variable. Cosine seasonal indicator: cosine transformation of the day of year. Sine seasonal indicator: sine transformation of the day of year. Description of files: 2D predictors: x2d_train (Historical + RCP8.5), x2d_test (RCP4.5). 1D predictors: x1d_train(Historical + RCP8.5), x1d_test (RCP4.5). Predictand (precipitation): y_train, y_test, corresponding to the predictor files above. Predictor coordinates: lon_x, lat_x. Precipitation coordinates: y_lon_lat, ordered as longitude (0), latitude (1).