An AI forecast’s clash with physics can warn of a tipping point

Suppose a model predicts the future of a cloud layer, but its prediction no longer balances the movement of mass. That inconsistency can be useful. In a new study, researchers turned a forecast’s growing physical error into a warning that the simulated system was approaching an abrupt transition.[1]

The clearest climate example is specific: in one test trajectory of an idealised low-cloud model, the system forecast a tipping point 2.45 simulated years ahead, with a timing error of 0.03 years. That is approximately 11 days, calculated using 365.25 days per year. It is a result within a simplified simulation, rather than a calendar prediction for the loss of Earth’s clouds.[1][2]

Published online in PNAS on 15 October, the work by Miguel Liu-Schiaffini and colleagues at Caltech, the University of Colorado Boulder and NVIDIA also examines changing airflow over an airfoil. Both examples pose the same difficult question: how do you anticipate a new regime when your training examples show only the behaviour before it?[1][2]

Learning the approach to a change

A system can change gradually until its response becomes abrupt. In the cloud model, increasing carbon dioxide eventually triggers a rapid loss of stratocumulus cover. In the airfoil simulations, steadily increasing the angle of attack changes the pattern of the wake and can lead to stall, when lift drops. A model trained to continue familiar behaviour may struggle precisely where the behaviour becomes most consequential.[1]

The researchers use a recurrent neural operator, or RNO. A neural operator learns how a whole field, such as velocity across a region of air, evolves. Its recurrent component retains information from earlier states. That memory helps the model represent a system whose behaviour is drifting over time, instead of treating each forecast as a fresh problem detached from its history.[1]

Training uses data from before the transition. The model then rolls the system forward, producing a possible future. A separate check asks whether that future obeys a known physical constraint. In the cloud experiment, the check uses mass conservation; in the incompressible airflow experiment, it checks whether the predicted velocity field satisfies the corresponding flow constraint.[1]

Physics continues to hold when the system tips. What can fail is the AI’s continuation of the earlier pattern. As its forecast moves into unfamiliar dynamics, the predicted state becomes inconsistent with the constraint. The method measures that inconsistency and compares it with a threshold calibrated from previous errors. Crossing the threshold supplies an estimated transition time.[1]

This separates two jobs that are often bundled together. The model proposes how the system will evolve; the physical check tests the proposal’s consistency. A researcher can sometimes know enough to perform that second job even without supplying the complete equations governing every part of the system. The cloud result used only its mass-balance constraint for the warning.[1]

What the simulated wing added

The airfoil experiments tested whether the method could cope with a more demanding field of changing motion. The authors simulated two-dimensional flow in three settings: stall at a Reynolds number of 1,000, a wake transition at 5,000, and stall at 5,000. Reynolds number compares inertial and viscous effects in a flow; changing it changes the regime the forecaster must handle.[1]

One test trained models on flow before the wake transition, then asked them to anticipate the later stall. Another trained at Reynolds number 1,000 and applied the model directly to transitions at 5,000. The RNO transferred more successfully than the paper’s memory-free neural operator and conventional recurrent-network comparators. Those comparators could still forecast transitions in several settings, so the physical-error idea is broader than a single model architecture.[1]

The warning threshold remains a substantive choice. Ordinary forecasting errors can also create a physics mismatch, especially over a long horizon. The method uses a distribution of errors from pre-transition data to set a tolerated false-alarm level; the illustrated cloud forecast uses 0.07, or 7%. That is a calibration setting within the paper’s statistical framework, not a measured false-alarm rate for an operational climate service.[1]

The intriguing possibility is an early warning built from the point where a prediction starts to lose physical credibility. A forecast need not render every detail of the new regime correctly to provide useful information about when the old one may end.[1]

Report an error or suggest a correction