Project Summary

Introduction

Tipping points are where small perturbations lead to large and long-term consequences for the state of a complex system, profoundly altering its mode of operation. Past abrupt climate changes and future climate model projections suggest that some elements of the Earth’s climate system can exhibit tipping points [Lenton et al, 2008; Lenton, 2011]. Climate change has already advanced to the point where we are at risk of triggering damaging tipping points [Lenton et al, 2019]. To avoid worse ones ahead, or at least minimise the risk they pose, will require early identification and warning of approaching tipping points.

What is the proposed work attempting to accomplish or do?

We propose to improve AI methods for the early warning of climate tipping points and use them to identify new data that can improve the specific forecast of a climate system tipping point, such as abrupt Arctic sea-ice melt or subpolar gyre collapse. The exact tipping element that we will investigate will be determined, with feedback from DARPA, during Phase 1.

How is it done today, and what are the limitations?

As complex systems approach a tipping point (bifurcation) they exhibit common behaviors. For example, all local bifurcations are accompanied by critical slowing down [Scheffer et al, 2009; Dakos et al, 2012] in which a system’s recovery rate from perturbations slows down as it approaches a bifurcation. As a result, statistical indicators such as rising lag-1 autocorrelation (AC) and rising variance of a time series precede tipping points in a variety of systems [Scheffer et al, 2009]. These indicators can provide early warning signals (EWS) for approaching critical transitions, and have been found in paleoclimate data prior to past abrupt climate changes [Lenton, 2011] and in climate model data approaching deliberately forced tipping points [Lenton, 2011; Boulton et al, 2014]. Current approaches to climate tipping point early warning are dominated by manual analysis of one or two generic EWS (usually rising AR(1) and variance) [Boers, 2021; Boers & Rypdal, 2021; Ciemer et al, 2020]. However, this has clear limitations.

Firstly, it misses additional information. The approach to different types of local bifurcation (e.g. saddle-node or pitchfork) can in principle be distinguished because they exhibit different signals in higher-order statistical moments [Kuehn & Bick, 2021; Kuehn, 2013; Kuehn, 2011]. For example, prior to a saddle-node bifurcation skewness towards the future state should increase whereas this is not the case for a symmetric bifurcation such as a pitchfork. These and other effects of higher-order terms leave their imprint on time series data and generate the features that deep learning algorithms excel at detecting–both known features such as skewness, as well as others that we may miss because they are analytically intractable.
Secondly, there are a host of conditions which can interfere with the simplest EWS, such as a system being subject to rapid forcing, changing noise level of the forcing, multiplicative noise, etc. We must also be wary of the prosecutor’s fallacy, which suggests that EWS are more likely to be found in time series that have been selected a posteriori [Boettiger & Hastings, 2012].

Our approach

Recently we have shown that skill in predicting tipping points can be dramatically improved by training a machine learning algorithm, which is also able to accurately detect the type of bifurcation being approached, thus providing information on future dynamical behaviour [Bury et al. 2021]. We developed a CNN-LSTM (convolution neural network long short-term memory) deep learning algorithm that provides EWS in systems it was not explicitly trained on. By training it on a universe of known types of low-order bifurcation, which includes around 500,000 models, it learned to detect generic normal forms and scaling behavior of dynamics near tipping points that are common to many dynamical systems. It is then tested on specific real-world tipping points in various systems, such as the approach to past abrupt climate changes. Results showed that our deep learning (DL) method outperforms traditional EWS methods in both sensitivity and specificity across a broad range of applications. Furthermore, the DL approach is not only able to identify that a bifurcation is impending but is skilled at predicting what type of bifurcation will occur.