Machine-learning method promises physically meaningful analysis of erratic ocean behavior

Earth’s ocean is an incredibly complex system, exhibiting unpredictable variability when assessed at different time scales. Researchers have used machine learning and statistical methods to make sense of the ocean, but while many can find interesting patterns with these methods, assigning physical meaning to them is difficult.

Franzke et al. used a machine learning method that analyzes ocean patterns with physical interpretability. The method is called multi-resolution Dynamic Mode Decomposition, or mrDMD, and decomposes high-dimensional data on ocean currents into both an average state and circular movements called eddies in a systematic way, and identifies varying annual cycles of sea surface temperature without human supervision.

“An important topic in ocean science, along with atmospheric science and fluid dynamics, is the identification of coherent structures and how they interact with the mean flow,” author Christian Franzke said. “We have shown that DMD modes are physically interpretable, which is not always the case with previous methods.”

The mrDMD method works by approximating the Koopman operator, which enables the method to represent nonlinear dynamics like ocean currents by decomposing multi-scale datasets into time-scale dependent patterns. Notably, the patterns have an oscillation frequency that corresponds to repeating states like annual cycles, and decay time scales that correspond to the different time resolutions, which enable physical interpretability.

The authors used this method to identify several known behaviors of ocean currents, including propagating meanders related to the Gulf Stream and Kuroshio currents. They were also able to extract El Nino-Southern

Oscillation events as transient phenomena.

“DMD provides the propagation matrix for seamless predictions based on DMD modes,” Franzke said. “This is a promising direction for future work.”

Source: “Systematic multi-scale decomposition of ocean variability using machine learning,” by Christian L. E. Franzke, Federica Gugole, and Stephan Juricke, Chaos (2022). The article can be accessed at https://doi.org/10.1063/5.0090064 .

This paper is part of the Theory-informed and Data-driven Approaches to Advance Climate Sciences Collection, learn more here .