Extensive data for training neural networks for radiation. Code and data used in: "Implementation of a machine-learned gas optics parameterization in the ECMWF Integrated Forecasting System"

Data and code used in a paper submitted to JAMES titled : Implementation of a machine-learned gas optics parameterization in the ECMWF Integrated Forecasting System 1) The files ml_training_*.7z contain extensive datasets (in NetCDF format) for training neural network versions of the RRTMGP gas optics scheme as described in the paper. The datasets are read by ml_train.py. 2) The ML datasets were in turn generated using the input profiles (in NetCDF format) inside inputs_to_RRTMGP.zip by running the Fortran programs rrtmgp_sw_gendata_rfmipstyle.F90 and rrtmgp_lw_gendata_rfmipstyle.F90 in rte-rrtmgp-nn/examples/rrtmgp-nn-training, which call the RRTMGP gas optics scheme, The input profiles contain millions of columns, hundreds of perturbation experiments (including hypercube-sampled gas concentrations), are derived from several different data sources (including CAMS reanalysis, GCM, and CKDMIP-MMM), and span present-day, preindustrial, and future atmospheric conditions. They could be used to generate training data for developing emulators of the full RTE+RRTMGP radiation scheme, not just gas optics (see nn_dev on the RTE+RRTMGP-NN repository on Github, used in a previous paper where different emulation methods were compared) 3) The Fortran and Python code used for data generation and NN training are found in rte-rrtmgp-nn/examples/rrtmgp-nn-training on the main branch on Github; an archived version is also included here (rte-rrtmgp-nn.zip). See the readme in the above sub-directory for further information.