Unsupervised Organization of Turbulent Updraft Regimes and their Global Response to Warming
Abstract
Unsupervised machine learning methods remain underexplored in the analysis of climate model data, especially in the age of overwhelming global cloud resolving simulation output. We thus ask whether a variational autoencoder (VAE) is capable of learning a low dimensional, physically-interpretable representation with physically-interpretable organization of explicitly resolved vertical velocity from a convection-permitting climate simulation. We leverage 1.5e6 two-dimensional snapshots of vertical velocity produced from a superparameterized climate models embedded cloud-resolving model. These vertical velocity fields encompass geographically and synoptically diverse conditions of moist and dry turbulence and serve as an ideal training dataset to test the VAEs limits. Our novel application of unsupervised learning generates a low dimensional representation of the data for study. K-means clustering of this learnt latent space uncovers a 3D topology that elucidates familiar vertical modes of convection amidst its spatiotemporal diversity, including differences in convection between land and ocean regimes. The latent space highlights ellipsoidal trajectories of cloud lifecycle organization such as diurnal cycles. When exposed to an out of sample test dataset consisting of the convective response to four Kelvin sea surface warming, our VAE cleanly summarizes both intensity changes and shifts in geographic regimes of convection with warming. In addition to expected signals of climate change, such as tropopause elevation, reductions in frequency, and increases in intensity of deep convection compensated by shallow convection, our VAE also reveals a distinct regional low-level turbulent response to warming over generally drier, more arid land masses as an interesting sub-cluster of statistical climate changes in vertical velocity. Our findings lend credibility to the future use of VAEs to extract interpretable dynamical regimes from even higher dimensional (e.g. multivariate) atmospheric simulation output, with potentially broad use in the Earth System Science Community. This may enable new forms of model inter-comparisons and analyses of climate change impacts that could complement traditional clustering and dimensionality reduction approaches.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMNG51A..08M