Comparing Storm Resolving Models and Climates via Unsupervised Machine Learning

Storm-resolving climate models (SRMs) have gained international interest for their unprecedented detail with which they globally resolve convection. However, this high resolution also makes it difficult to quantify the emergent differences or similarities among complex atmospheric formations induced by different parameterizations of sub-grid information. This paper uses modern unsupervised machine learning methods to analyze and intercompare SRMs based on their high-dimensional simulation data, learning low-dimensional latent representations and assessing model similarities based on these representations. To quantify such inter-SRM ``distribution shifts'', we use variational autoencoders in conjunction with vector quantization. Our analysis involving nine different global SRMs reveals that only six of them are aligned in their representation of atmospheric dynamics. Our analysis furthermore reveals regional and planetary signatures of the convective response to global warming in a fully unsupervised, data-driven way. In particular, this approach can help elucidate the effects of climate change on rare convection types, such as ``Green Cumuli''. Our study provides a path toward evaluating future high-resolution global climate simulation data more objectively and with less human intervention than has historically been needed.