Generative Large Eddy Simulations with conditional Variational Autoencoders

Abstract:

Large eddy simulation is a tool of choice to simulate three-dimensional turbulent flows in the atmosphere and in particular many cloud-forming processes at the heart of climate model uncertainties. LES models can be run at orders of magnitude higher resolution than global storm resolving models, directly representing relevant eddy scales with better posed sub-grid closure schemes than the semi-empirical cloud parameterizations that must be relied on in global climate or storm-resolving models. However, LES simulations are highly computationally expensive and as a result only useful as limited area models. Variational inference methods applied to generate high resolution cloud fields could thus hold promise in reconstructing LES-quality cloud fields from coarser resolution information, in conjunction with state of the art sub-km closures such as the Simplified Higher-Order turbulence Closure (SHOC). To explore this potential we test limits of conditional variational autoencoders (cVAEs) to generate reconstructions of LES cloud and rain water volumes. We will discuss our strategy for designing and implementing VAE architectures for emulating details of cloud field geometry extracted from LES simulations of idealized shallow convection, conditioned intentionally on the same reduced parametric moments that are typically prognosed in modern HOC parameterization schemes. This generative LES procedure involves 1) visualizing generative modeling of cloud water mixing ratio (QC) and cloud rain water mixing ratio (QR) volumes, 2) finding a VAE architecture that has the representational degrees of freedom and dimensionality reduction to reconstruct such 3D volumes reliably and 3) developing conditional VAEs with relevant moments from SHOC (e.g. w,q,theta_l) as input profiles to generate cloud water and cloud rain concentration volumes. The overarching goal is to explore algorithms as yet underexploited in climate modeling but which could eventually reduce uncertainty from sub-grid scale cloud variations and computational expense. In practice, such a coupled machine learning emulator could be useful for the Energy Exascale Earth System Modeling (E3SM) physics suite in development as part of the Enabling Aerosol-cloud interactions at GLobal convection-permitting scalES (EAGLES) project.