Stochastic neural emulators for subpolar gyre variability and tipping-risk prediction in Earth system models

Abrupt transitions in the North Atlantic Subpolar Gyre’s (SPG’s) behavior are a major source of uncertainty in decadal-scale climate predictability, as well as having potentially strong impacts on the Atlantic Meridional Overturning Circulation (AMOC), European climate, and marine ecosystems. Climate model simulations suggest that the SPG can undergo irreversible transitions from a regime of deep convection and strong circulation to one characterized by weak convection and reduced transport. Such a collapse would substantially cool the North Atlantic and could interact with a weakening AMOC in complex and nonlinear ways.

SPG transitions emerge from the interplay between high-dimensional ocean dynamics and unresolved stochastic processes, making it difficult to represent them faithfully in current Earth system models and challenging for deterministic prediction frameworks. Here, we use CESM2 pre-industrial control simulations to perform a data-driven analysis of SPG dynamics. We construct a machine-learning-based stochastic neural emulator designed to learn, forecast, and quantify uncertainty in SPG evolution. The model simultaneously learns the conditional mean dynamics and state-dependent ensemble spread, enabling fully probabilistic predictions of key prognostic variables.

This approach provides a tractable framework for investigating the mechanisms and precursors of SPG weakening and deep-convection collapse, and for assessing associated climate risks. When generalized across models, our approach also offers a pathway for systematically evaluating long-timescale North Atlantic dynamics in Earth system simulations.