To study ocean currents, scientists launch GPS-tagged buoys into the ocean and record their velocities to reconstruct the ocean currents that transport them. Data from these buoys are also used to identify ‘forks’, areas where water rises from below the surface or sinks below it.
Accurately predicting ocean currents and pinpointing bifurcations will allow scientists to better predict weather, approximate how oil will spread after a spill, and measure ocean energy transfer. You can A new study that incorporates machine learning reports that it makes more accurate predictions than traditional models.
A multidisciplinary team of researchers, including computer scientists and oceanographers from the Massachusetts Institute of Technology, found that the standard statistical models typically used for buoy data make unrealistic assumptions about water behavior, thus leading to tidal currents. found that it can be difficult to reconstruct accurately or to identify the divergence.
Researchers have developed a new model that incorporates knowledge of fluid dynamics to better reflect the physics at work in ocean currents. They show that their method, requiring only a small amount of additional computational cost, is more accurate than conventional models in predicting currents and identifying divergence.
This new model will help oceanographers make more accurate inferences from buoy data, allowing them to more effectively monitor the transport of biomass (such as sargassum seaweed), carbon, plastic, oil and nutrients in the ocean. increase. This information is also important for understanding and tracking climate change.
“Our method captures the physical assumptions better and more accurately. In this case, we already know a lot of the physics. It allows us to focus on learning what is important to us: what is the current leaving the buoy, what is this branch, and where is it happening?” Massachusetts Institute of Technology School of Electrical Engineering and Computer Science ( EECS) and member of the Institute for Information and Decision Systems and the Institute for Data Systems and Society, and lead author Tamara Broderick.
Broderick’s co-authors include lead author Renato Berlinghieri, a graduate student in electrical engineering and computer science. Brian L. Tripp, Postdoctoral Fellow at Columbia University. David R. Burt and Ryan Giordano, MIT Postdoctoral Fellows. Kaushik Srinivasan, Atmospheric and Marine Science Assistant at the University of California, Los Angeles. Tamay Ozgökumen, Professor of Marine Science, University of Miami. And Junfei Xia, a graduate student at the University of Miami. The research will be presented at an international conference on machine learning.
Explore data
Oceanographers use data on buoy velocities to predict ocean currents and identify “forks” where water rises to the surface or sinks deeper.
To estimate ocean currents and find divergence, oceanographers have used a machine learning technique known as Gaussian process that can predict even when data is sparse. To work well in this case, the Gaussian process must make assumptions about the data and generate predictions.
The standard method of applying Gaussian processes to ocean data assumes that the latitude and longitude components of currents are irrelevant. However, this assumption is not physically accurate. For example, this existing model implies that the current divergence and its vorticity (the swirling motion of the fluid) operate on the same magnitude and length scales. Marine scientists know this is not true, says Broderick. The previous model also considered a framework of standards. This means that the fluid behaves differently in latitudinal and longitudinal directions.
“We thought we could address these issues with a model that incorporated physics,” she says.
They built a new model that uses what is known as the Helmholtz decomposition to accurately represent the principles of fluid dynamics. This method models ocean currents by decomposing them into a vorticity component (capturing swirling motion) and a divergence component (capturing rising or sinking water).
In this way, we give the model basic physical knowledge that we use to make more accurate predictions.
This new model uses the same data as the old model. Also, although their method can be computationally expensive, the researchers show that the additional cost is relatively small.
nimble performance
They evaluated the new model using synthetic buoy data and real ocean buoy data. Since the synthetic data were fabricated by researchers, we were able to compare the model’s predictions with the true currents and divergence. However, the simulations contain assumptions that may not reflect reality, so the researchers tested the model using data captured by real buoys released into the Gulf of Mexico.
Credit: Advanced Research Consortium for Transport of Hydrocarbons in the Environment
In each case, their method performed better on both tasks of current prediction and divergence determination compared to other machine learning approaches using standard Gaussian processes and neural networks. . For example, in simulations involving eddies adjacent to ocean currents, both the previous Gaussian process and neural network methods predicted divergence with very high confidence, but the new method accurately predicted no divergence.
The technique is also good at identifying eddies from a small set of buoys, Broderick added.
Having demonstrated the effectiveness of using the Helmholtz decomposition, the researchers would like to incorporate a time component into the model, as currents can vary over time as well as space. In addition, we would like to better capture the impact of noise on the data, such as wind, which can affect buoy speed. Separating that noise from your data can make your approach more accurate.
“Our hope is to take this noisy observed velocity field from the buoys and show what the actual divergence and actual vorticity is and predict away from those buoys. We believe our new technology will help with this,” she said. Say.
“The authors skillfully integrate known behavior from fluid dynamics to model ocean currents in a flexible model,” says Massi, associate biostatistician at Brigham and Women’s Hospital and lecturer at Harvard Medical School. Migliano Russo says. He was not involved in this research. “The resulting approach retains the flexibility to model flow nonlinearities, but features such as vortices and connected flows that are only recognized when the hydrodynamic structure is integrated into the model. It can also characterize phenomena, which is a good example of how flexible models can be greatly improved with well-thought-out, scientifically sound specifications.”
This work was supported in part by the Office of Naval Research, the National Science Foundation (NSF) CAREER Award, and the University of Miami Rosenstiel School of Oceanic, Atmospheric and Geosciences.
