Peter Oke: When Machine Learning Breaks the Ocean: Why Dynamics Still Matter
Machine learning (ML) and AI are rapidly transforming ocean data products, offering powerful tools to interpolate sparse observations and fill gaps in time and space. However, these approaches can operate outside the constraints imposed by dynamics. Here, we examine ML-derived products for the East Australian Current at 27S.
We show an example where an ML-gridded and SOM-filled product has systematic biases, including unrealistically shallow velocity structures, suppressed variability, and spatial artefacts that extend through the full water column. These artefacts generate spurious horizontal density gradients, which in turn lead to physically inconsistent geostrophic velocities. By contrast, dynamically consistent products (e.g., reanalyses, models) produce coherent and realistic structures using the same diagnostic frameworks.
Our analysis highlights a key limitation: ML methods that ignore underlying dynamics can produce fields that look plausible but violate fundamental physical balances. We argue that applications would benefit from embedding dynamical constraints – through data assimilation, physics-informed approaches, or dynamical-grounded mapping – to ensure physically consistent outputs.