Mining climate as a high-dimensional dynamical system: application to the Tropical Pacific
Abstract
We propose an analysis framework for spatiotemporal climate fields that account for multi-variable dependencies and nonlinearities. At each time step t, the climate system can be uniquely described by a state space vector parametrized by N variables and their spatial variability. The evolution of the systems dynamics is then represented as a single-point trajectory in a state space representation. We expect the system's dynamics to be confined to a lower (than the full state space) dimensional manifold and present a strategy for manifold learning via linear and nonlinear algorithms. To showcase these ideas, we focus on Tropical Pacific Ocean variability in an observational dataset (ERA5) and two state-of-the-art high-resolution CMIP6 models (MPI and EC-Earth3) focusing on periods of 40 years. In each case, we consider four fields that are key to understand tropical variability and we explore the daily evolution of the high-dimensional trajectory through lower dimensional nonlinear projections. We argue that nonlinear dimensionality reduction methods are more suited to identify curved, climate manifolds and we show how the multivariate component of the framework reveals models biases in the variable linkages. Additionally, we quantify local properties of the system's dynamics on the manifold through the local dimension and persistence metrics and analyze their statistics in historical times and future projections. By considering spatiotemporal scales, more than one variable and nonlinear components, our framework allows to comprehensively study and visualize climate evolution.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMIN25A0451F