AI Identifies the Source of Climate ‘Doom’ Two Years in Advance
Jakarta, CNBC Indonesia – Researchers have developed artificial intelligence (AI) capable of predicting a climate-change “doomsday” tipping point more than two years before it occurs, with the prediction just 11 days away from the actual event.
Scientists from the California Institute of Technology (Caltech), the University of Colorado Boulder and NVIDIA worked on the research. They developed an AI model called the recurrent neural operator (RNO), which can learn how the behaviour of physical systems changes over time.
A tipping point is a threshold at which a system undergoes a drastic change that is difficult or impossible to reverse. In the climate context, this could include the sudden loss of marine clouds, the collapse of ocean circulation, or instability in polar ice sheets.
The researchers tested RNO using simulations of stratocumulus clouds: low clouds over the ocean that reflect sunlight and help cool the Earth.
Climate models suggest that these clouds could disappear abruptly when carbon dioxide (CO2) concentrations exceed a certain threshold, accelerating global warming.
In the simulation, the AI successfully predicted the clouds’ disappearance more than two years in advance, with its estimate about 11 days from the actual event.
The researchers also tested the model on aerodynamic stall in aircraft wings. The results suggest that RNO could help detect drastic changes in complex systems before they reach a tipping point.
Jane Bae, one of the study’s authors from Caltech, said the model can recognise when the laws of physics it has learnt are no longer sufficient to explain the system’s current state.
“The ability to anticipate these qualitative changes before they occur could be very useful for prediction and aerodynamic design,” Bae said, as quoted by Earth on Monday (19 October 2026).
The findings indicate that AI could help scientists detect abrupt changes in complex systems before they reach critical thresholds.
However, the tests were conducted using simulations, so the model’s ability to predict real-world events requires further study.