[Paper Review] Comparing Storm Resolving Models and Climates via Unsupervised Machine Learning

Griffin Mooers, Mike Pritchard+6 more|arXiv (Cornell University)|Aug 25, 2022

Meteorological Phenomena and Simulations4 citations

TL;DR

This paper introduces an unsupervised machine learning framework using variational autoencoders (VAEs) and vector quantization to objectively compare nine global storm-resolving models (GSRMs) based on high-resolution 2D vertical velocity fields. The method identifies latent representations of tropical convection, revealing that only six GSRMs show similar dynamical behavior, while uncovering distinct, data-driven signatures of convection types—including 'Green Cumulus'—and climate change impacts like arid zone expansion and deep convection intensification.

ABSTRACT

Global Storm-Resolving Models (GSRMs) have gained widespread interest because of the unprecedented detail with which they resolve the global climate. However, it remains difficult to quantify objective differences in how GSRMs resolve complex atmospheric formations. This lack of comprehensive tools for comparing model similarities is a problem in many disparate fields that involve simulation tools for complex data. To address this challenge we develop methods to estimate distributional distances based on both nonlinear dimensionality reduction and vector quantization. Our approach automatically learns physically meaningful notions of similarity from low-dimensional latent data representations that the different models produce. This enables an intercomparison of nine GSRMs based on their high-dimensional simulation data (2D vertical velocity snapshots) and reveals that only six are similar in their representation of atmospheric dynamics. Furthermore, we uncover signatures of the convective response to global warming in a fully unsupervised way. Our study provides a path toward evaluating future high-resolution simulation data more objectively.

Motivation & Objective

• To address the lack of objective, data-driven tools for comparing high-resolution global storm-resolving models (GSRMs) that capture complex atmospheric dynamics.
• To quantify distributional differences between GSRMs using unsupervised machine learning, avoiding reliance on coarse, physically guided statistics.
• To identify physically meaningful convection regimes and model similarities from high-dimensional simulation data without prior labeling.
• To detect climate change signals in GSRM outputs in a fully unsupervised manner, particularly related to convection organization and zonal shifts.
• To enable more rigorous, scalable intercomparisons of future high-resolution climate simulations by learning latent representations of model behavior.

Proposed method

• Employ variational autoencoders (VAEs) to reduce high-dimensional 2D vertical velocity fields from GSRMs into low-dimensional latent representations.
• Use vector quantization and clustering on the latent space to discover distinct regimes of tropical convection, such as 'Green Cumulus' and deep convection.
• Compute symmetrized Kullback-Leibler (KL) divergence between normalized probability distributions of convection types to quantify distributional distances between models.
• Train a shared VAE on one model’s data (UM) and apply it to encode and compare latent representations across all nine GSRMs.
• Use 3D PCA visualizations of the latent space, colorized by physical properties (e.g., convection intensity, land fraction, turbulent length scale), to assess disentanglement and model differences.
• Apply unsupervised analysis to detect climate change signals by comparing present-day and warmed climate simulations in a data-driven way.

Experimental results

Research questions

• RQ1 Which GSRMs exhibit statistically similar representations of tropical convection dynamics when analyzed through unsupervised learning?
• RQ2 What physically meaningful convection regimes emerge from the latent space of high-resolution vertical velocity data across diverse GSRMs?
• RQ3 Can unsupervised machine learning detect climate change signals—such as shifts in convection intensity and zonal distribution—without prior assumptions or labeled data?
• RQ4 How do different model designs influence the organization and structure of convective features in the latent space, and what does this reveal about model uncertainty?
• RQ5 To what extent can VAE-based representation learning disentangle physical processes like land vs. ocean convection or shallow vs. deep convection?

Key findings

• Only six of the nine GSRMs studied show similar latent representations of tropical convection, indicating significant divergence in how models simulate atmospheric dynamics.
• The unsupervised method successfully identifies a distinct 'Green Cumulus' convection regime, which is underrepresented in observational data and often missing in standard cloud masks.
• The latent space reveals that SPCAM exhibits a unique continental shallow convection regime characterized by small-scale horizontal organization, detectable only through the VAE's disentangled representation.
• SAM shows significantly higher vertical velocity intensities in its latent representation compared to other GSRMs, indicating a more vigorous convective response.
• The method uncovers climate change signals in a fully unsupervised way, including the expansion of arid zones and the concentration of deep convection over warm oceanic regions under global warming.
• Vector quantization and clustering of latent representations reveal three primary convection types: 'Green Cumulus', shallow cumulus, and deep convection, with distinct physical drivers such as lower tropospheric stability and surface fluxes.