Extrapolating dynamics with machine learning

In order for machine learning models to help researchers predict the long-term behavior of non-stationary real-world systems, such as Earth’s climate, they must be able to extrapolate to situations beyond their training data. Dhruvit Patel and Edward Ott demonstrated that machine learning models are surprisingly effective at doing so.

Many non-stationary systems undergo tipping points, after which their behavior is drastically different. The authors found machine learning-based methods can anticipate sudden tipping points and extrapolate to post-tipping point behavior.

At a certain level of extrapolation, however, these machine learning methods fail. In these cases, combining the machine learning approach with an available conventional model may successfully predict a system’s behavior.

Most previous work predicting the behavior of non-stationary systems with machine learning focused on relatively short-term timescales, such as weather forecasting, in which the effect of non-stationarity can often be ignored.

“In this work, we explore the widely important and heavily understudied problem of predicting tipping point and extrapolating to the post-tipping-point behavior of non-stationary dynamical systems,” Patel said. “We anticipate that our work in this direction will result in further development of novel machine learning-based solutions.”

This research suggests machine learning models may be effective in predicting Earth’s climate conditions decades into the future under different scenarios of anthropomorphic emission of greenhouse gases.

“Such predictions can then be used to guide critical government policymaking,” Patel said. The authors are currently incorporating basic ideas from this work into a machine learning-based global climate forecasting model.

Source: “Using machine learning to anticipate tipping points and extrapolate to post-tipping dynamics of non-stationary dynamical systems,” by Dhruvit Patel and Edward Ott, Chaos (2023). The article can be accessed at https://doi.org/10.1063/5.0131787 .

This paper is part of the Nonlinear dynamics, synchronization and networks: Dedicated to Juergen Kurths’ 70th birthday Collection, learn more here .