Increasing computer power and resources over the last decade contributed to a substantial improvement of general circulation models. Nonetheless, the Intergovernmental Panel on Climate Change's Fourth Assessment Report did not reduce the uncertainties in the climate system's sensitivity to changes in anthropogenic forcing. The difficulty encountered in narrowing the range of uncertainties is directly connected to the complexity of the climate system, with its multiplicity of nonlinear processes and feedbacks. Developing a better understanding of the causes and mechanisms of climate sensitivity, as it particularly affects uncertainties, is essential to meet the goals of developing more reliable climate predictions required as a basis for human adaptation.

In computer science, "complex networks analysis" constitutes a powerful tool to investigate local and non-local statistical interrelationships. It refers to a set of metrics, modeling tools and algorithms commonly used in the study of complex nonlinear dynamical systems and its main premise is that the underlying topology or network structure of a system has a strong impact on its dynamics, stability, robustness and evolution. Network analysis has been successfully applied to technological networks (e.g., Internet), biological systems (metabolic networks, brain functional and structural networks), social networks and elsewhere. The application of network analysis to climate science is still at an early stage, with only a handful of related works.

With a collaborative effort between a climate and a computer scientist, we propose to develop a fast, scalable and cutting-edge computational toolbox to infer and analyze climate networks. The inferred spatio-temporal networks will be of two types (grid-based and regional), weighted (so that we capture the intensity of different teleconnections), undirected or directed (depending on the analysis) and dynamic (i.e., their topology can change over time). We will analyze the network structure, and therefore the teleconnections, of both observational data-sets and various state-of-the-art (Coupled Model Intercomparison Project 5) climate model outputs. We will conduct an extensive model intercomparison in terms of the climate network that each model leads to and its robustness in past, present and future climate scenarios. We will deliver algorithms that quantify how each model reproduces major teleconnections and identify common or specific errors in comparing model outputs and observations. We will therefore quantify uncertainties in climate models, clustering models or integrations with qualitatively different topological characteristics as distinct climate scenarios. Additionally, we will characterize how/whether networks change over the recent past, identifying specific regional effects or teleconnections with the observed changes. Finally, we will use several models and runs of each model to examine how climate networks may change under global warming scenarios, identifying those network changes that can be associated with major climate shifts and abrupt changes.

The network analysis software that will be developed during the project will be publicly available. We plan to extend an existing network analysis toolset called Brain Connectivity Toolbox (BCT) that has been developed by the neuroscience community. Our modifications will provide the appropriate interfaces so that BCT can be easily used with climate data and we will expand its functionality with additional utilities for the inference of geographically connected communities, directed edges using causality inference methods, and dynamic network analysis.