Peng Zhang, Jun Fu and Rolf Findeisen tested whether an AI model’s range of forecast futures changes near abrupt shifts in complex systems. Of four diffusion-based forecasters, one showed a localized drop in forecast variety in simulated networks and similar uncertainty changes near transitions in Mediterranean Sea sediment-core records. Their Communications Physics study, published Oct. 9, suggests model uncertainty could complement existing early-warning signals.
Critical transitions are abrupt shifts from one pattern of behavior to another. The researchers simulated resource–biomass dynamics, neuron-like activity and infection spread on networks. They also tested a separate shallow-lake predator–prey model and examined historical Mediterranean Sea records. The paper notes that many machine-learning approaches to early warning require transition labels or control information; this study asks whether a forecaster’s uncertainty can add a signal.
Each model took a recent window of observations and generated possible future trajectories. The team estimated predictive uncertainty from how much those generated futures varied: when possible paths clustered more tightly, forecast uncertainty fell. That is a property of the model’s predictions, not evidence that the underlying system became safer or easier to predict.
Across the three networked systems, uncertainty patterns differed by model architecture and system dynamics. NsDiff was the only one of the four models to show localized uncertainty collapse consistently across the tested systems. The authors’ controlled analyses linked the drop to less diversity among generated futures and to the transition mechanisms and dynamical regimes represented in the training data.
For an empirical check, the team analyzed 12 Mediterranean Sea molybdenum time series from sediment cores. They found similar predictive-uncertainty changes near transitions. Transition locations were used only for comparison after the model analysis, so this was a retrospective check of historical records rather than a real-time warning test.
The supplementary methods qualify what “label-free” means here. Transition locations and control-parameter values were not used during training or inference, and category labels were not model inputs or prediction targets. But the researchers did use those categories to balance which simulated examples entered the training sample. The study presents predictive uncertainty as a possible complement to other early-warning measures; whether it can flag a change early enough, and distinguish one from false alarms in an operating system, remains open.
Sources
- Zhang, Fu and Findeisen, “Unsupervised probing critical transitions in complex systems using generative diffusion models,” Communications Physics (2026)
- Supplementary Information: systems, training and empirical-data methods
Image credit: Lelyan Abu Snenah/Wikimedia Commons, CC BY-SA 4.0. Illustrative Mediterranean sunset; not a study image.
