Comparing storm resolving models and climates via unsupervised machine learning

Area of Science:

• Climate Science
• Computational Fluid Dynamics
• Data Science

Background:

• Global storm-resolving models (GSRMs) offer unprecedented climate detail but lack objective comparison tools.
• Quantifying differences in how GSRMs simulate complex atmospheric formations is challenging.
• This limitation affects various fields relying on complex simulation data.

Purpose of the Study:

• To develop objective methods for comparing similarities between GSRMs.
• To enable automated, physically meaningful comparisons of high-resolution climate simulation data.
• To assess the intercomparison of nine GSRMs and identify similarities in atmospheric dynamics.

Main Methods:

• Utilized nonlinear dimensionality reduction and vector quantization to estimate distributional distances.
• Developed an approach to learn similarity from low-dimensional latent data representations.
• Applied methods to high-dimensional 2D vertical velocity snapshots from nine GSRMs.

Main Results:

• Successfully intercompared nine GSRMs, revealing that only six exhibit similar atmospheric dynamics.
• Uncovered signatures of the convective response to global warming in an unsupervised manner.
• Demonstrated the effectiveness of the developed methods for objective model evaluation.

Conclusions:

• The developed methods provide a robust framework for objectively evaluating and comparing GSRMs.
• This approach facilitates a deeper understanding of model behavior and climate change impacts.
• Paves the way for more reliable assessments of future high-resolution climate simulation data.