Caltech-Led AI Develops New Method for Forecasting Climate Tipping Points

A new Caltech-led paper demonstrates a significant advancement in forecasting "tipping points" in complex systems like Earth's climate, utilizing a novel AI program. The term "tipping points," popularized by writer Malcolm Gladwell for social trends, refers in physics to moments when systems reach an unstable threshold and undergo often irreversible changes, such as global warming or polar ice cap melting. While AI models have previously predicted smaller weather variations, forecasting these critical, uncertain tipping points has proven more challenging.

The research, published in the Proceedings of the National Academy of Sciences (PNAS), introduces a recurrent neural operator (RNO), a type of neural network. This RNO can predict when tipping points occur in simplified models, including climate-change systems and simulations of turbulent airflow over a plane's wing. Anima Anandkumar, Caltech's Bren professor of computing and mathematical sciences and the principal author, explained that a tipping point signifies a nonstationary state where a system cannot return to its original condition, often leading to catastrophic outcomes. This work, cited in a recent United Nations report on climate science, involved collaborators from Caltech, the University of Colorado Boulder, and NVIDIA, highlighting its broad potential across climate science, ecology, epidemiology, and aeronautical design.

In a specific climate scenario, researchers investigated the rapid disappearance of low clouds over oceans due to rising carbon dioxide levels – a future event that has been predicted but not yet observed, which would further exacerbate global warming. The RNO was tested on simulated data for this scenario, successfully forecasting the timing of the cloud disappearance more than two years in advance with an error of approximately 11 days. Anandkumar noted that unlike past AI models trained on stationary systems, this RNO approach is designed to model nonstationary systems, accounting for dynamic changes like carbon dioxide content where standard machine learning principles fall short.

The RNO's innovative design incorporates a built-in "memory" that tracks how the underlying physics of a system evolve over time. This allows the AI to detect when its predictions no longer align with previously defined physical laws, signaling an imminent tipping point. As Anandkumar stated, "The system is drifting away over time, and the laws from physics are no longer holding at the tipping point." Beyond climate, the PNAS paper detailed the RNO's success in predicting a tipping point for airflow over a tilting plane wing, where lift precipitously drops. Study co-author Jane Bae, assistant professor of aerospace at Caltech, emphasized that the model's ability to recognize when learned physics no longer adequately describes reality provides a crucial warning, offering significant utility for aerodynamic prediction and design.

What to watch: Further development and application of RNO models in broader climate and engineering systems.

Editor's note: Thoroughly covers the source material, capturing the technical details, key names, and context accurately.