Abstract
Abstract. Simulating global and regional climate at high resolution is essential to study the effects of climate change and capture extreme events affecting human populations. To achieve this goal, the scalability of climate models and efficiency of individual model components are both important. Radiative transfer is among the most computationally expensive components in a typical climate model. Here we attempt to model this component using a neural network. We aim to study the feasibility of replacing an explicit, physics-based computation of longwave radiative transfer by a neural network emulator and assessing the resultant performance gains. We compare multiple neural-network architectures, including a convolutional neural network, and our results suggest that the performance loss from the use of conventional convolutional networks is not offset by gains in accuracy. We train the networks with and without noise added to the input profiles and find that adding noise improves the ability of the networks to generalise beyond the training set. Prediction of radiative heating rates using our neural network models achieve up to 370× speedup on a GTX 1080 GPU setup and 11× speedup on a Xeon CPU setup compared to the a state-of-the-art radiative transfer library running on the same Xeon CPU. Furthermore, our neural network models yield less than 0.1 K d−1 mean squared error across all pressure levels. Upon introducing this component into a single-column model, we find that the time evolution of the temperature and humidity profiles is physically reasonable, though the model is conservative in its prediction of heating rates in regions where the optical depth changes quickly. Differences exist in the equilibrium climate simulated when using the neural network, which are attributed to small systematic errors that accumulate over time. Thus, we find that the accuracy of the neural network in the “offline” mode does not reflect its performance when coupled with other components.
Highlights
Computational models of Earth’s climate are essential tools to advance our understanding of the climate system and our ability to predict its response to perturbations such as increased levels of greenhouse gases
In the context of Radiative transfer (RT), we use convolutional neural networks (CNNs) to evaluate whether the sensitivity to localised features improves the prediction performance of deep neural networks
While this work focuses on clearsky radiation, we study the ability of CNNs to recognise and respond to such local features in the single-column simulations
Summary
Computational models of Earth’s climate are essential tools to advance our understanding of the climate system and our ability to predict its response to perturbations such as increased levels of greenhouse gases. Our ability to predict future changes in climate depends crucially on the accuracy of these models and the extent to which interactions between various components of the climate system are represented. The computational burden increases roughly as the fourth power of spatial resolution (since resolution must increase along all three spatial dimensions and the time step reduced to ensure numerical stability). To address this problem, various approaches have been used including improved model scalability (Dennis and Loft, 2011) and the use of lowprecision floating point operations (Palmer, 2014)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
