Collaborative Research: MODEL ENABLED MACHINE LEARNING (MnML) FOR PREDICTING ECOSYSTEM REGIME SHIFTS

University Of Hawaii, Honolulu

Investigators

Abstract

Ecosystems can change radically, suddenly and without warning. There are numerous examples of this on land, in our rivers, lakes and oceans. From African savannahs to Californian kelp forests, these ecosystem “regime shifts'' as they are called, have had large impacts on the provision of key ecosystem services, such as food and income. There is a need for new bioinformatics and cyberinfrastructure that can predict these regime-shifts, and for identifying the drivers of such changes so that policies and technologies can be developed to help avoid them (should that be desired). Current methods for anticipating regime shifts perform poorly: either theoretical models of ecosystem dynamics are too abstract to provide useful operational forecasts, or data-driven approaches suffer from overfitting and cannot accurately forecast the emergence of novel conditions (i.e., those not seen in historical data on which models are trained). In this project, a new approach for forecasting ecosystem regime shifts will be developed. This new approach is called Model Enabled Machine Learning and it combines scientific understanding of ecological dynamics (i.e., theoretical models) with the predictive power of machine learning. This new approach will be co-developed with ecosystem stakeholders, so that the outputs of the models are useful and actionable. Model Enabled Machine Learning will be developed for three ecosystem case-studies and tested against other state-of-the-art approaches for predicting ecosystem regime shifts. This will involve using existing and developing new mathematical models of ecosystem dynamics for each case-study, as well as collecting empirical data for training the machine learning models. The goal is to significantly improve upon existing methods for predicting ecosystem regime shifts. The ecosystem case-studies include: 1) Tropical coral ecosystems that switch between coral- and algal-dominated states; 2) Freshwater lakes that exhibit harmful algal blooms; 3) Mangrove ecosystems that suffer from multiple stressors. The potential of Model Enabled Machine Learning as a new bioinformatic tool used by ecosystem managers lies not just in its predictive skill, but also in the clear interpretability it provides, which will maximize its utility as an operational tool. Importantly, Model Enabled Machine Learning has the potential to promote equitable science by reducing the data requirements of machine learning driven predictions, giving stakeholders in data-poor systems a useful operational tool that would otherwise be unavailable. To facilitate user engagement, the Model Enabled Machine Learning methods developed in this project will be operationalized as R and Julia coding packages/libraries, two common coding languages used by the stakeholder communities. Numerical methods in these packages will be co-designed with stakeholders to ensure that future ecosystems regime shifts are anticipated and managed. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.