Predicting unobserved driver of regime shifts in social-ecological systems with universal dynamic equations

Ecosystems around the world are anticipated to undergo regime shifts as temperatures rise and other climatic and anthropogenic perturbations erode the resilience of present-day states. Forecasting these nonlinear ecosystem dynamics can help stakeholders to prepare for the associated rapid changes. One major challenge is that regime shifts can be difficult to predict when they are driven by unobserved factors. In this paper, we advance scientific machine learning methods, specifically universal dynamic equations (UDEs), to identify changes in an unobserved bifurcation parameter and predict ecosystem regime shifts. We demonstrate this approach using simulated data created from a dynamic model of a species population experiencing loss due to unobserved extraction or harvest. This could be, for example, illegal fishing from a fishery or unreported poaching in a game reserve. We show that UDEs can accurately identify changes in the unobserved bifurcation parameter, in our case the slowly increasing harvest rate, and predict when a regime shift might occur. Compared to alternative forecasting methods, our UDE approach provides more reliable short-term predictions with fewer data. This approach provides a new set of methods for ecosystem stakeholders and managers to identify unobserved changes in key parameters that drive nonlinear change.