Training and evaluation data for machine learning models emulating the RTE+RRTMGP radiation scheme or its components

Data associated with an upcoming paper on the use of neural networks (NN) to emulate a shortwave radiation parameterization.

• CAMS_* = pre-processed NetCDF files consisting of CAMS reanalysis profiles that can be used as input to the RTE+RRTGMP code to generate NN training data (the other files in the repository). The Fortran program and instructions for doing this can be found at https://github.com/peterukk/rte-rrtmgp-nn/tree/nn_dev/examples/emulator-training

The other files are ready-to-be-used input-output data for training machine learning models using the Python scripts found at https://github.com/peterukk/rte-rrtmgp-nn/tree/nn_dev/examples/emulator-training/scripts:

• RADSCHEME_* = data to train NN emulators for the whole RTE+RRTMGP radiation scheme in the shortwave (including gas and cloud optics, but not aerosols)
• REFTRANS_* = data to train NN emulators for the shortwave reflectance-transmittance computations in RTE
• RRTMGP_* = data to train NN emulators for RRTMGP shortwave gas optics

The evaluation data, 2015, is not available in this format except for RADSCHEME, because the final evaluation is done by implementation in the Fortran radiation code (allsky_sw_testmodels.F90) which takes as input the CAMS* files, and produces a netCDF file with the fluxes as well as timings. Fortran code for recurrent neural networks was not written, so the evaluation for this particular model (producing output flux file and timings) is done in Python using RADSCHEME_data_g224_CAMS_2015_true_solar_angles.nc.