COS 126-2 - Machine intelligence for early warning signals of spatio-temporal ecological and climate tipping points

Background/Question/Methods

Tipping points occur when slowly changing environmental conditions move beyond a critical point where the system state changes suddenly to a strongly contrasting state. Many ecological and climate systems exhibit tipping points that result in catastrophic outcomes. Classical early warning signals of tipping points are often based on critical slowing down before the transition and are reflected in statistics like variance and lag-1 autocorrelation of the time series. Recent research shows how training deep learning algorithms on a library of simulated tipping points allows the algorithm to provide early warning signals of tipping points in a range of empirical and model systems with greater sensitivity and specificity than variance and lag-1 autocorrelation. The algorithm can also classify the type of tipping point and thus provide information about the future state. However, this approach is thus far limited to temporal data and low-dimensional bifurcations. We asked: can deep learning algorithms trained on a library of spatio-temporal phase transitions predict tipping points in spatially distributed ecological and climate systems? We developed a library based on the Ising phase transition model from statistical physics and tested whether it could predict transitions in a spatial vegetation dynamics model, and in CMIP5 climate simulations.

Results/Conclusions

We find that the deep learning algorithm is able to predict tipping points with good sensitivity and specificity in a spatio-temporal model of vegetation dynamics, and in some (but not all) classes of spatio-temporal climate simulations from CMIP5 (Coupled Model Intercomparison Project Phase 5). Our results support the hypothesis that phase transition theory provides a useful framework for this task. Most existing methods for tipping point prediction seek to identify signatures of criticality associated with local, codimension-1 bifurcations. Phase transitions represent a broader class of dynamical phenomena with a more diverse set of possible warning signals. In the test cases, our model demonstrates the utility of spatial analysis: classification using purely spatial features often outperforms classification on purely temporal features, and models trained on both together tend to offer the highest accuracy. We emphasize that the deep learning algorithm is predicting phase transitions in these model systems out-of-sample—without having been trained on data from the model systems. We conclude that this approach may allow deep learning algorithms to predict spatio-temporal phase transitions in systems for which data are scarce, and may also be able to provide information on the state beyond the tipping point.