Using Deep Learning as Cost-Effective Surrogate Models for GCM Radiative Transfer.

The radiation scheme in any climate model, such as the Energy Exascale Earth System Model (E3SM), is one of the most time-consuming components. These radiative transfer calculations have been simplified by using various techniques to save computational cost. However, the simplifications compromise the fidelity of the radiative transfer calculations, which in turn negatively impact climate simulations and weather prediction.The goal of this investigation is to accurately predict the short-wave (SW) and long-wave (LW) radiations calculations in E3SM without solving the actual parameterization in a computationally cost effective manner. Deep Neural Networks (DNNs) can act as surrogate models for the actual RRTMG parameterization.

The input-output pairs for training the DNNs are gathered by running "CAM5'" physics on 2^o horizontal grid with 30 vertical levels for 1 year using E3SM. The number of samples collected for SW radiation is approximately 6 million whereas roughly 12 million samples are collected for LW radiation. These samples are randomly split into a large training dataset and a small testing dataset. These datasets are normalized to eradicate the incongruities so that a fast convergence can be acquired during the DNN training. To train the datasets, a software package called "Keras" was used, which is a high level wrapper around Tensorflow written in python. Once the training is completed, the DNN is validated over the test samples to check for the desired accuracy. If the target accuracy is achieved, the actual RRTMG parameterization in E3SM is replaced by the DNNs, and the simulation is repeated. Comparisons of annually averaged and instantaneous results are made between the DNNs and the actual RRTMG parameterization, and it was concluded that the DNNs can emulate the existing radiation parameterization in E3SM with an accuracy of approximately 90% and are able to accelerate the radiative transfer calculations in E3SM by 8-10x.