Search

Abstract The Mid-Atlantic region is set to be one of the first and largest contributors to the offshore wind energy goals of the United States. Yet, the same region is home to a diverse marine ecosystem comprising important marine species such as the critically endangered North Atlantic right whale (NARW). To support the responsible development and operation of the planned offshore wind farms, there is a need for high-resolution modeling of NARW presence, i.e., at the spatial and temporal resolutions relevant to farm-level operations. Towards this, we leverage highly localized observations from nine glider deployments in the Mid-Atlantic to propose a machine learning approach for modeling NARW presence conditioned on a diverse set of glider- and satellite-based oceanic, physical, and contextual information. We find that tree- and ensemble-based models achieve the highest levels of accuracy, while maintaining a sensible balance of missed and false alarms. Interpretation of the machine-learnt features reveals interesting insights on the relative value of well-resolved satellite surface measurements to well-resolved vertical information from glider sampling in explaining the species-habitat patterns of NARWs. We then discuss the value of the models proposed herein to offshore wind developers and operators in the United States and elsewhere. Our work constitutes the first machine learning attempt to jointly leverage glider- and satellite-based information for modeling of NARWs. Data and codes for producing the results of this work have been made freely available to promote the research on this timely topic.

Coarse-grained models describe the macroscopic mean response of a process at large scales, which derives from stochastic processes at small scales. Common examples include accounting for velocity fluctuations in a turbulent fluid flow model and cloud evolution in climate models. Most existing techniques for constructing coarse-grained models feature ill-defined parameters whose values are arbitrarily chosen (e.g., a window size), are narrow in their applicability (e.g., only applicable to time series or spatial data), or cannot readily incorporate physics information. Here, we introduce the concept of physics-guided Gaussian process regression as a machine-learning-based coarse-graining technique that is broadly applicable and amenable to input from known physics-based relationships. Using a pair of case studies derived from molecular dynamics simulations, we demonstrate the attractive properties and superior performance of physics-guided Gaussian processes for coarse-graining relative to prevalent benchmarks. The key advantage of Gaussian-process-based coarse-graining is its ability to seamlessly integrate data-driven and physics-based information.