Predicting critical tipping points in complex systems has long been a challenge for scientists. Now, a new AI system might be able to take on this task. Computer scientists have developed an artificial intelligence program capable of predicting the onset of catastrophic tipping points, with potential applications in forecasting ecological collapse, financial crashes, pandemics, and power outages.
“If we can forecast an impending critical transition, we can prepare for the shift or possibly prevent it, thereby mitigating damage,” said Gang Yan, a professor of computer science at Tongji University in China and the senior author of the study, in an interview with Live Science. This motivation led the team to create an AI approach to predict such sudden transitions well before they occur. The study’s findings were published on July 15 in the journal Physical Review X.
Tipping points refer to sudden changes that push a system or environment into an undesirable state, from which recovery is difficult. For example, the collapse of the Greenland ice sheet would not only reduce snowfall in northern Greenland but also significantly raise sea levels, making parts of the ice sheet irretrievable.
The science behind these dramatic shifts is often poorly understood, with predictions typically based on oversimplified models. Traditional methods, which rely on statistics to measure the weakening strength and resilience of systems, have produced controversial results.
To develop a more accurate predictive tool, the researchers combined two types of neural networks—algorithms that mimic the brain’s information processing. The first neural network broke down complex systems into large networks of interacting nodes, while the second tracked changes in individual nodes over time.
“For instance, in a financial system, a node could represent a single company; in an ecological system, it could stand for a species; in a social media system, it could denote a user,” Yan explained.
Given the difficulty in predicting tipping points, and the scarcity of real-world data on these critical transitions, the researchers trained their model on tipping points within simple theoretical systems. These included model ecosystems and out-of-sync metronomes that eventually synchronized.
After training, the AI was tested with real-world data, specifically on the transformation of tropical forests into savannah. Using over 20 years of satellite data from three regions in Central Africa, the AI successfully predicted the transition in the third region, even when 81% of the system’s nodes (in this case, chunks of land) were unobserved.
Having successfully predicted one tipping point, the researchers now aim to uncover the patterns identified by the algorithm and apply the model to other systems, such as wildfires, pandemics, and financial crashes.
A challenge in predicting systems involving humans is that we often react to our own forecasts, complicating predictions. For instance, if drivers are informed of road congestion, they might change routes, potentially relieving congestion on some roads while creating it on others. This feedback loop makes predictions difficult.
To address this, the researchers plan to focus on aspects of human systems that are less affected by human behavior, such as routes that are inherently congested due to their design rather than driver actions.
