Regime shifts are large, abrupt, and persistent changes in the structure and function of complex systems. Detecting and predicting regime shifts in ecosystems is fundamentally important for both resource management and scientific understanding. Yet, ecological monitoring often does not produce the high-resolution data necessary for forecasting ecosystem states under future environmental conditions. To address this issue, we leveraged scientific machine learning (SciML) techniques to integrate theoretical models with time series data to detect and predict regime shifts. Our SciML approach utilized two primary frameworks: neural ordinary differential equations (NODEs) and universal differential equations (UDEs). NODEs use a neural network to approximate a system of differential equations, and UDEs extend this method by incorporating known dynamics. We applied both methods to time series of coral, turf algae, and macroalgae coverage spanning up to 21 years on 24 reefs across the United States Virgin Islands (USVI), where corals were previously abundant, to answer several research questions. First, which SciML framework provides more accurate estimates of long-term benthic coverage? Second, can UDEs reliably recover grazing rates on algae? Finally, what are the near-term forecasts for benthic coverage on these reefs?

We found that NODEs performed better than UDEs (i.e., had lower normalized RMSE) when comparing model predictions to empirical data for reefs. In addition, UDEs yielded grazing rates in the range of 0 to 35% of the reef, which is lower than thresholds that would produce coral-dominated reefs. Consequently, near-term forecasts indicated that either macroalgae or turf algae would continue to dominate USVI reefs. Through the application of SciML, our study enhances understanding of the maintenance of alternative states and provides valuable tools for ecosystem management and conservation on coral reefs.