Deep Learning for the Parametrization of Subgrid Processes in Climate Models

Abstract

Earth system and climate models are fundamental to understanding and projecting climate change. Although they have improved significantly over the last decades, considerable biases and uncertainties in their projections still remain. A large contribution to this uncertainty stems from differences in the representation of clouds and convection (i.e., deep clouds) occurring at scales smaller than the resolved model grid resolution that is typically in the order of 100 km in the horizontal. These long-standing deficiencies in cloud parametrizations have motivated developments of high-resolution cloud- and turbulence-resolving models that can explicitly resolve clouds and convection, yet are computationally extremely expensive and can therefore only be run for a short time and/or only over a small region. However, together with the recent developments in machine learning, and especially deep learning, these simulations can be used and harvested to develop new ML-based parametrizations for clouds and convection that have the potential to eliminate some of the long-standing systematic errors in climate models. Recent research demonstrated that deep convection explicitly simulated by a cloud-resolving climate model could be correctly emulated by a deep neural network that then replaced the original parametrization in the climate model. In this chapter we describe the principal approaches for ML-based cloud parametrizations with deep neural networks and advocate that these ML algorithms need to be guided both by data and by physical knowledge. Many challenges in this new interdisciplinary field of research remain that are also discussed.