J66.5 Toward Subgrid-Scale Parameterizations Using a Superresolution Generative Adversarial Network

Thursday, 16 January 2020: 11:45 AM
156BC (Boston Convention and Exhibition Center)
Eden Au, Univ. of Edinburgh, Edinburgh, United Kingdom; and K. Kashinath, A. Albert, M. Prabhat, and S. F. B. Tett

Many small-scale and complex physical processes in general circulation models (GCMs) cannot be explicitly resolved due to limited computational resources. Processes on scales smaller than the spatial resolution of the model need to be parameterized. Parameterizations have been known to be major sources of uncertainties in GCMs, and various approaches have been proposed to deduce the influence of the under-resolved and unresolved processes.

Generative adversarial networks (GANs) are a class of unsupervised machine learning methods that can generate realistic data from a target distribution. They are well-suited to build emulators for complex physical processes, and hence poised to serve as building blocks for parameterizations. Super-resolution GAN (SRGAN) and its variants were introduced in recent years for obtaining photo-realistic images using a novel loss function, which is a weighted sum of adversarial loss and pixel-to-pixel content loss.

We develop a data-driven approach using SRGAN and its variants drawing parallels from the development of super-parameterized CAM (SP-CAM). For simplicity and model consistency, the GANs are trained using cloud resolving model (CRM) outputs from the near-global CRM simulations (https://doi.org/10.1002/2015MS000499), with the input distribution being a low-resolution coarse-grained version of the original high-resolution CRM data. The GAN aims to reconstruct the original high-resolution CRM data. We test the performance of these GANs using several reconstruction losses, including some motivated by physical constraints of importance to the domain of cloud physics. Our results show that these GANs are able to produce realistic high-resolution data from their low-resolution counterparts, whilst satisfying some of the physical constraints. Our next step is to incorporate physical constraints more rigorously into the training and inference of these GANs, so they may be used for constructing realistic subgrid scale parameterizations for convection.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner