Bayesian uncertainty quantification of internal climate variability

Dimensionality reduction techniques are powerful for extracting modes of climate variability from observational data sets and climate model output. Over the past decades, multiple studies have shown that dominant climate patterns can be identified, and that climate evolution can be approximately linearized in such subspaces. In this work, we apply novel dimensionality reduction techniques to global climate data sets. In particular, we optimize such methods for finding patterns that maximize their inherent predictability on multi-annual time scales. We develop a fully Bayesian framework. The record of high-quality climate data sets (20-100 years) is relatively short compared to our predictability time scales of interest (1-10 years). This necessarily causes large uncertainty in data-driven analyses of internal climate variability due to sampling variability and biases. In a Bayesian analysis, we are able to rigorously quantify the uncertainty in observed internal climate variability: both in the spatial patterns, and in their dynamic time evolution.

 

We use linear inverse modeling to represent the climate dynamics in a subspace that optimizes predictability measures. We then use advanced Bayesian methods to calibrate the parameters of the linear model. The resulting uncertainty analysis allows to identify which climate modes – and interactions between modes – are well- or poorly-constrained within the observational record. This novel method further allows to explore if climate models can reproduce the linearized dynamics within observational uncertainties, or if they fail in representing some specific modes of climate variability.

While still in its early stages, this research is aimed at addressing key climate predictability challenges, in particular identifying the factors that contribute to accurate and reliable multi-annual climate predictions.