A Caltech-led research team has developed an artificial-intelligence method that forecast abrupt changes in computer models of climate and airflow, including a simulated collapse of low clouds more than two years before it occurred in the model.
The method, called a recurrent neural operator, combines a neural network with a memory of how a system changes over time. In tests, it identified an approaching transition by detecting when its forecasts began to violate physical constraints. The peer-reviewed research was outlined in a new report published on 15 October.
Testing a simulated cloud transition
The climate test used an idealised model of a stratocumulus-topped atmospheric boundary layer. In that simulation, rising carbon dioxide levels eventually trigger a rapid loss of low cloud cover, with a hysteresis effect: lowering carbon dioxide again does not immediately restore the earlier cloud state.
Researchers trained the system on data from before the transition, rather than giving it examples of what happened afterwards. They then assessed whether its future predictions continued to obey a physical constraint. For the cloud model, the forecast gave warning nearly two and a half years ahead of the simulated tipping point, with an error of about 10 days in the paper’s analysis.
This is a result from a model, not a warning that ocean clouds are about to disappear. The simulation represents a simplified version of one climate process, and the study does not establish a real-world date or probability for such a transition. The researchers’ approach also depends on having a sufficiently trained model and relevant physical constraints.
From climate equations to airflow
The team also tested the approach against simulated fluid moving around an aircraft wing. The model was asked to forecast transitions in the flow, including aerodynamic stall, when lift falls as the wing’s angle of attack increases. The paper reports tests at two Reynolds numbers and on both wake-transition and stall cases, using computational simulations rather than flight trials.
The wider study examines several non-stationary systems, whose behaviour changes over time. Its central technical contribution is the recurrent neural operator, which represents past observations as a latent memory and predicts future states across changing physical conditions. The paper compares it with recurrent neural networks and Markov neural operators; in its experiments, the new method generally maintained more accurate predictions over longer forecast horizons.
That comparison matters because many conventional forecasting approaches assume the system’s governing behaviour remains broadly stable. A tipping point challenges that assumption: gradual changes in conditions can precede a sudden shift in the system. The researchers’ method looks for a departure from expected physical behaviour in the model’s own predictions, rather than relying only on a familiar warning signal appearing in observed data.
What the results do and do not show
The authors report that the framework can work with partial physical information, such as a conservation law, rather than requiring a complete description of the system. They also tested whether it could avoid false alarms in simulated cases that did not contain a tipping point. These results remain evidence about the selected test systems and model setups; they do not show that the method is ready to issue operational climate or aviation warnings.
For climate science, the cloud calculation is an idealised scenario and cannot by itself determine how the real atmosphere will respond. For aviation, the airfoil results are computational and do not establish performance in an aircraft or safety-critical operating environment. Further work would be needed to assess how well the approach transfers to real observations and more complex systems.
The study was conducted by researchers from Caltech, the University of Colorado Boulder and NVIDIA. Its authors frame the work as a foundation for investigating early forecasts of abrupt change across physical systems, rather than as a finished forecasting service.
