Non-Linear Dimensionality Reduction With a Variational Encoder Decoder to Understand Convective Processes in Climate...

Area of Science:

• Climate modeling
• Machine learning in atmospheric science
• Deep learning for sub-grid-scale processes

Background:

• Deep learning models accurately represent sub-grid-scale convective processes in climate models.
• However, their large internal dimensionality often hinders interpretability and trustworthiness.

Purpose of the Study:

• To apply Variational Encoder Decoder (VED) structures for interpretable learning of convective processes.
• To explore the dimensionality reduction capabilities of VEDs in climate modeling.

Main Methods:

• Utilized Variational Encoder Decoder (VED) structures, a non-linear dimensionality reduction technique.
• Applied VEDs to an aquaplanet superparameterized climate model simulation with explicitly simulated deep convective processes.
• Compressed high-dimensional data into a low-dimensional latent space (five nodes).

Main Results:

• VEDs accurately learned and reproduced convective processes, comparable to feed-forward neural networks.
• Successfully compressed complex convective data into only five latent dimensions.
• Identified distinct convective regimes within the latent space, including stable vs. deep convection, cloud types, and shallow convection characteristics.

Conclusions:

• VEDs offer an interpretable approach to understanding sub-grid-scale convective processes in climate models.
• The dimensionality reduction enables the delineation and analysis of different convective regimes.
• This work paves the way for more interpretable and generative machine learning parameterizations in climate science.