Comparing Storm Resolving Models and Climates via Unsupervised Machine Learning
Abstract
Recent unsupervised machine learning methods provide a novel way to analyze the results from modern high-resolution global climate simulations. We focus on a set of modern storm-resolving climate models (SRMs) that have gained international interest for their unprecedented global resolution. However, while all these models explicitly resolve deep, moist convection, they also vary by design choices including crucial parameterizations of sub-grid effects leading to differences in the representation of the atmosphere. The problem of objectively quantifying emergent differences or similarities among such models' overwhelming complex convective details is difficult and under-explored. To address this, we introduce new data-driven metrics for quantifying model similarities by analyzing differences in the high-dimensional simulation data that these models produce. To quantify such inter-SRM "distribution shifts'', we use variational autoencoders (VAEs) in conjunction with techniques from data compression. Our analysis involving nine different global SRMs reveals that only six of them are dynamically similar. Furthermore, applying the same methodology to understanding distribution shifts induced by global warming reveals signatures of climate change consistent with physics-based expectations. Along the way, we distill lessons that inform the emerging art of latent space inquiry for numerous climate applications. Our study provides a path toward evaluating high-resolution global climate simulation data more objectively and with less human intervention than has historically been needed. Our findings suggest that machine learning can provide a more holistic and detailed view of the storm-resolving model output than any one physically informed benchmark traditionally relied upon, suggesting a broader role for data-driven analysis as the resolution and volume of climate model outputs increases further.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H25A..03M