Abstract

There is considerable consensus that the main cause of the great uncertainties in future changes in precipitation is the need to parameterize sub-grid scale (SGS) processes in climate models (GCM).

The key source of error in the recent generation of GCMs is believed to be the interaction between SGS processes and the large-scale atmospheric circulation that is due to its temporal and spatial scale explicitly resolved. While there is hope in the long run because the projected increase in computer power will allow us to resolve many SGS processes in the upcoming decade, researchers are currently investigating how the representation of SGS processes can be improved using machine learning (ML) and there are still considerable challenges ahead: (a) ML methods struggle to generalize and predict unphysical SGS tendencies in simulations of future climates; (b) coupled to the dynamical core of a GCM, trained ML methods experience stability problems; (c) the performance of different ML methods over complex terrain is unclear; (d) it is an open question whether an ML-based parameterization should replace all SGS parameterizations or leverage existing parameterizations.

In this 24-month project, an ETH-based postdoc and a data scientist of the SDSC will join forces to explore physically-constrained ML-based SGS parameterizations to overcome the generalization problem and quantify their performances over complex topography. Further, we use backward optimization, saliency maps, feature importance, and layerwise relevance propagation to understand how different ML methods learn to predict SGS tendencies over complex topography.

This interpretation of the training enables new and exciting process understanding plus it will advance our understanding of how to design the next generation of ML-based SGS parameterizations.