Covariance Structure Analysis of Climate Model Output

Chintan Dalal, Rutgers University

To understand future climate change, different Earth system models from groups worldwide simulate projections of future climates. However, results from these simulations are computationally very expensive, often requiring several months on a supercomputer. In this paper, we provide a new statistical emulation method that may allow a realization of future climate projections within a day rather than several months. Specifically, we analyze the structure of several existing outputs from various climate models on a manifold of covariance matrices. The manifold covariance structure provides a method to compare existing climate model outputs, as well as to sample a new realization of future climate projections. We validated our climate model output comparison method using known dependencies between various climate models. Additionally, we showed, using semi-variogram plots, that the distribution of our realizations lie within the distribution of existing climate model outputs. Finally, we demonstrated that the ability of our method to generate new realizations of future climate changes works well even for a small number of existing climate model outputs. The proposed statistical emulator could find its use in future climate impact assessment.

Link to Recording: https://www.youtube.com/watch?v=SlgrkLfVi5c

Link to Presentation: CI2016 Dalal