Characterizing Pacific Decadal Variability Using Dynamical Systems Theory and Machine Learning

Pacific Decadal Variability (PDV) has been shown to have a significant influence on hydroclimate across the globe, the rate of global mean surface temperature increase under greenhouse gas forcing, and marine ecosystems. Understanding the mechanisms and the predictability of this mode of climate variability would therefore be highly useful. However, the small number of PDV phase shifts that have occurred within the observational record poses a serious obstacle to developing an understanding of the mechanisms that drive this phenomenon.

In this study, we make use of a 100,000-year unforced simulation from an intermediate complexity model of the Tropical Pacific ocean-atmosphere system in order to have a large number of samples of PDV phase shifts. We reconstruct the attractor of the system to find that PDV phases emerge as a pair of regime-like structures in the 3-dimensional state space of the system, and that the transitions between these regimes lie in highly predictable parts of the attractor. We show that PDV in this model shares significant similarities with the real-world system in this dynamical systems framework.

Using the structure of the attractor as a guide, we then apply machine learning techniques (random forests and bootstrap-aggregated decision tree ensembles) to this large dataset in order to characterize the physical state of the system during predictable transitions. We show that transitions from high-variance to low-variance states of the PDV are highly predictable in the period two to three years after an El Nino event. The evolution of ocean-atmosphere variables in the eastern Tropical Pacific is found to play a significant role in determining whether such a transition occurs after an El Nino event, providing a basis for further investigation of the physical mechanisms driving PDV transitions.