New model using machine learning improves ocean current predictions

Massachusetts Institute of Technology

By subscribing, you agree to our Terms of Use and Policies You may unsubscribe at any time.

A model that combines machine learning has been built in a recent study by a diverse research team, including computer scientists from MIT and oceanographers, to more precisely predict ocean currents and identify divergences.

The researchers discovered that due to erroneous assumptions about water behavior, the conventional statistical model frequently applied to buoy data struggles to produce precise predictions. The new model offers a more realistic depiction of the physics at play in ocean currents by combining knowledge from fluid dynamics.

Divergences must be identified, and ocean current predictions must be accurate to respond to oil spills, forecast weather, and comprehend how energy is transferred in the ocean.

The updated model may make more accurate monitoring of biomass transportation, carbon dispersion, plastics distribution, oil movement, and nutrient flow in the ocean possible, which could significantly improve estimates drawn from buoy data. Additionally, this data is essential for comprehending and monitoring climate change.

The researchers found that incorrect assumptions were made regarding the relationship between the latitude and longitude components of the current using the conventional Gaussian process, a machine-learning method used to forecast ocean currents and identify divergences.

The existing model used the false assumption that a current's vorticity and divergence occur on the same length and magnitude scales. The new model, however, includes a Helmholtz decomposition, which divides the ocean current into vorticity and divergence components, precisely representing the laws of fluid dynamics.

Utilizing data from both synthetic and actual ocean buoys, the researchers assessed the new model. Compared to the conventional Gaussian process and another machine-learning method using a neural network, the new model performed better in forecasting currents and recognizing divergences when compared with ground-truth winds and divergences. The researchers also found that using the new technique, a small group of buoys might be used to identify vortices successfully.

The researchers plan to add a time component to their model in the future to account for temporal fluctuations in ocean currents. To increase the model's accuracy, they also intend to improve its capability to distinguish between data and noise, such as wind influences.

The researchers intend to increase the model's capabilities to forecast currents and divergences away from the buoys, ultimately improving their comprehension of ocean dynamics.

Field specialists have praised the researchers' new method, which included well-known fluid dynamics behaviors into an adaptable model. Associate biostatistician at Brigham and Women's Hospital Massimiliano Russo applauds the study for its scientifically sound specification and capacity to enhance the adaptability and precision of existing modeling.

The Rosenstiel School of Marine, Atmospheric, and Earth Science at the University of Miami, the Office of Naval Research, and an NSF CAREER Award all provided funding for this study.

The results of this study, which highlight the new model's potential influence on oceanographic research and applications, will be presented at the International Conference on Machine Learning.

Study Abstract:

Oceanographers are interested in predicting ocean currents and identifying divergences in a current vector field based on sparse observations of buoy velocities. Since we expect current dynamics to be smooth but highly non-linear, Gaussian processes (GPs) offer an attractive model. But we show that applying a GP with a standard stationary kernel directly to buoy data can struggle at both current prediction and divergence identification – due to some physically unrealistic prior assumptions. To better reflect known physical properties of currents, we propose to instead put a standard stationary kernel on the divergence and curl-free components of a vector field obtained through a Helmholtz decomposition. We show that, because this decomposition relates to the original vector field just via mixed partial derivatives, we can still perform inference given the original data with only a small constant multiple of additional computational expense. We illustrate the benefits of our method on synthetic and real ocean data.