Machine-learning prediction of tipping with applications to the Atlantic Meridional Overturning Circulation

Shirin Panahi, Ling-Wei Kong, Mohammadamin Moradi, Zheng-Meng Zhai, Bryan Glaz, Mulugeta Haile, Ying‐Cheng LaiarXiv 2024 · DOI · arXiv

Abstract

Anticipating a tipping point—a transition from one stable steady state to another—is a critical challenge across diverse fields. The steady-state nature of dynamics near a tipping point makes prediction significantly more difficult than for transitions from oscillatory or chaotic dynamics. By exploiting the benefits of environmental noise, we develop a general data-driven machine-learning approach to predict future tipping in nonautonomous dynamical systems. We validate this framework using synthetic and empirical examples, specifically addressing the potential collapse of the Atlantic Meridional Overturning Circulation (AMOC). Our predictions place a potential collapse window between 2040 and 2065, consistent with current literature.

Introduction: The Challenge of Steady-State Tipping

A tipping point represents a threshold where a minor change in an external parameter causes a system to shift abruptly from one stable state to another, often with catastrophic consequences. In climate science, the potential collapse of the Atlantic Meridional Overturning Circulation (AMOC) is a primary concern.

Predicting these transitions is technically demanding because many systems approaching a tipping point remain in a "quiet," non-oscillatory stable state until the moment of failure. Unlike systems with growing oscillations or chaotic fluctuations, these steady-state systems offer no obvious temporal "signal" for standard machine learning models to learn from. This research addresses how to extract predictive features from systems that appear superficially stable.

Methodology: Reservoir Computing and the Role of Noise

The core of the proposed framework is Parameter-Adaptable Reservoir Computing (RC). Reservoir computing is a recurrent neural network architecture where the internal "reservoir" (the hidden layer) consists of fixed, random connections, and only the output weights are trained. This makes the model computationally efficient and well-suited for temporal data.

Schematic of the parameter-adaptable reservoir computing framework for predicting tipping points.

The innovative aspect of this approach is the exploitation of noise. Rather than treating noise as an obstacle, the framework uses it as a probe to explore the system's stability landscape:

  1. Stochastic Training: In a stable steady state, a system remains at a single point. Dynamic noise "kicks" the system away from this point, forcing it to reveal its underlying recovery dynamics.
  2. Critical Slowing Down: As a system nears a tipping point, its recovery from these noisy perturbations slows down. The RC model learns to recognize these subtle changes in recovery patterns.
  3. Bifurcation Tracking: The model incorporates an additional input channel for a time-varying bifurcation parameter (e.g., freshwater melt in the North Atlantic). This allows the model to extrapolate future behavior as the driving force continues to change.

Application to the AMOC

The framework was validated using four distinct data sources related to the AMOC:

  • 1D Stochastic Fingerprint Model: A synthetic model based on sea surface temperature (SST) proxies.
  • 2D Conceptual Model: A more complex synthetic model representing ocean dynamics through two state variables.
  • Community Earth System Model (CESM): High-fidelity data from state-of-the-art climate simulations.
  • Empirical SST Data: Real-world observational data of North Atlantic sea surface temperatures from 1875 to the present.

Results: Predicting the Collapse Window

Across all datasets, the model consistently predicted a collapse of the AMOC. When applied to empirical SST data, the model identified a potential collapse window between 2040 and 2065, with a median year of 2053.

Prediction of AMOC collapse using empirical sea surface temperature data, showing the driving parameter (a), observed state (b), predicted trajectory (c), and distribution of collapse years (d).

These results are significant for several reasons:

  • Consistency: The AI-driven predictions align closely with recent statistical studies, providing independent verification of the 21st-century collapse risk.
  • Robustness: The model successfully avoided false positives, correctly predicting stability when tested on parameter ranges far from the tipping point.
  • Predicting the Unseen: The model is trained only on "pre-tipping" data. It never observes a collapse during training, yet it learns the geometry of the stable state well enough to recognize its imminent disappearance.

Discussion and Limitations

While the framework is powerful, several limitations remain. The method assumes the correct "control parameter" has been identified and that its rate of change is known. In real-world scenarios, determining whether a driver is changing linearly or exponentially is difficult. Additionally, the model relies on "fingerprints" like sea surface temperature as proxies for circulation strength; if the relationship between the proxy and the actual circulation shifts, the predictions may change.

Future work aims to integrate physical constraints into the machine learning model (Physics-Informed Machine Learning) to further enhance reliability and potentially predict post-tipping dynamics, which currently remain outside the model's training regime.

Summary updated

Abstract

Anticipating a tipping point, a transition from one stable steady state to another, is a problem of broad relevance due to the ubiquity of the phenomenon in diverse fields. The steady-state nature of the dynamics about a tipping point makes its prediction significantly more challenging than predicting other types of critical transitions from oscillatory or chaotic dynamics. Exploiting the benefits of noise, we develop a general data-driven and machine-learning approach to predicting potential future tipping in nonautonomous dynamical systems and validate the framework using examples from different fields. As an application, we address the problem of predicting the potential collapse of the Atlantic Meridional Overturning Circulation (AMOC), possibly driven by climate-induced changes in the freshwater input to the North Atlantic. Our predictions based on synthetic and currently available empirical data place a potential collapse window spanning from 2040 to 2065, in consistency with the results in the current literature.
BibTeX
@misc{panahi2024machine,
  author = {Shirin Panahi and Ling-Wei Kong and Mohammadamin Moradi and Zheng-Meng Zhai and Bryan Glaz and Mulugeta Haile and Ying‐Cheng Lai},
  title = {{Machine-learning prediction of tipping with applications to the Atlantic Meridional Overturning Circulation}},
  year = {2024},
  eprint = {2402.14877},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.14877},
}
Figure