Looking ahead: Opportunities and quandaries facing large eddy-permitting superparameterization and machine learning emulation of cloud physics

Mike Pritchard

Dept. of Earth System Sciences, University of California, Irvine

Wednesday, Jun 12, 2019, 10:00 am
DSRC Room 2A305

Abstract

In the last decade, the 1-4 km horizontal resolution regime was a global climate modeling frontier that cloud superparameterization (SP) helped explore ahead of schedule. Now, as global cloud resolving models (GCRMs) handle these scales of deep convection more elegantly, it is a good time to reflect on the next decade’s frontier needs. My view is that SP algorithms will remain critical for exploring sub-kilometer scale processes important to global climate, such as low-cloud forming boundary layer turbulence, its mixing with the free troposphere, and its mediation of aerosol cloud interactions. In the first part of the talk, I will show a first step in this direction (“ultraparameterization”; UP) produces ironically familiar cloud feedbacks to surface warming but interestingly distinct aerosol-cloud indirect effects, and a striking muting of the aerosol-cloud indirect effect when boundary layer eddies are resolved. Looking ahead, I will examine the technical potential for even richer forms of turbulence-permitting SP approaching LES best practices, at computational parity with GCRMs.

In the second part of the talk, I will discuss machine-learning emulation of cloud-resolving simulation, focusing first on a surprising initial success in emulating thousands of cloud resolving models’ scale interactions with planetary dynamics using a deep neural network trained on a superparameterized aquaplanet. I will show some new challenges that emerge in real-geography tests involving complexities of the diurnal cycle and introduce a nice new method to respect conservation laws within the deep neural network architecture. The Jacobian of the DNN will then be shown to be an intriguing new tool for dynamical inquiry of models with realistic complexity and high degrees of freedom, including a generalization of Kuang’s convective linear response function. Finally I will conclude with some thoughts on the unsolved technical issues and interesting philosophical tensions being raised by the disrupting influence of machine learning in the climate simulation field.