top of page

Generative Modeling of Atmospheric Convection

VAEs for Convection Representation: Text

How reliable is unsupervised learning on a high-resolution climate simulation?  Can it provide any novel insights into convection physics or be used to emulate stochastic parameterizations in climate models? We attempt to answer these questions by building a Variational Autoencoder (VAE) and training it on cloud-resolving vertical velocity fields.

VAEs for Convection Representation: Research
Screen Shot 2021-01-03 at 1.24.40 AM.png
VAEs for Convection Representation: Image

Not only does our VAE learn to reconstruct the details of convection, but it sorts the test vertical velocity fields in a physically sensible way on just two components in the latent space:

VAEs for Convection Representation: Research
VAEs for Convection Representation: Research

Our VAE latent space also captures expected physics such as the diurnal cycle, which loops around the space over the course of several days in the Amazon Rainforest

VAEs for Convection Representation: Research
VAEs for Convection Representation: Research

Collaborators

Jens.jpeg

Jens Tuyls

VAEs for Convection Representation: Team Members
VAEs for Convection Representation: Pro Gallery
bottom of page