On Friday, Staff Writers Peyton Goodman and Shaina Sahu attended a seminar held by Columbia University’s Lamont-Doherty Earth Observatory on the applications of machine learning in the ocean sciences.

Last Friday, the Lamont-Doherty Earth Observatory hosted a seminar showcasing the work of Dr. Julien Le Sommer, a scientist in the field of computational oceanography. The seminar was titled “Probabilistic Predictions of Lagrangian Drift at the Ocean Surface Using Machine Learning.”

Dr. Sommer works for the Institute of Environmental Geoscience at the University of Grenoble in France, where he studies fine-scale (smaller than 200 km) ocean turbulence. His work involves researching how fine-scale turbulence is modeled and observed, as well as why it matters. He has been working with the “Nucleus for European Modelling of the Ocean” (NEMO) model for climate simulations, as well as a new satellite mission that shows fine-scale flows in the ocean. He talked about how these studies can affect related climate issues. For instance, studying these models can help study the effects of marine pollution and oil spills, which can lead to new developments in the ways climate issues are addressed. 

During Friday’s seminar, Dr. Le Sommer mentioned that machine learning allows for a multitude of newly possible discoveries in ocean sciences. Dr. Le Sommer’s work specifically uses machine learning to produce synthetic data that can then be used to predict future climate patterns. One particularly helpful technique, he explained, is to cast problems as inverse problems—inferring things you can’t directly observe from parameters and observations that you make. Machine learning is simply a “new flavor” of inverse problems. One thing to keep in mind is that these problems are highly uncertain subjects. The mechanistic models are “fundamentally uncertain,” and so are the observations; we don’t know the initial state, so the trajectory itself will be uncertain.

Due to this uncertainty, Dr. Sommer emphasized that data from models can’t always be used in the same way as observational data. The equation used to produce models, in addition to relying on variables of location, velocity, and time, also depends on winds and waves. It is very chaotic, since any small error in the formulation of parameters might heavily affect the results. To address this, Dr. Sommer’s team adds a term to the equation to account for the error and imprecision of using data from ML-generated models. 

Toward the end, Dr. Sommer responded to attendees’ questions, including one about the ability to generalize machine learning data. He stated that generalizing data regionally is acceptable, but may become more complicated between different regions. He concluded by saying that they can achieve more with a simple model than a complex ocean model. Dr. Sommer and his colleagues are already getting the chance to exploit machine-learning data, and with the rapid advancements in the machine-learning industry, they expect to keep exploring applications in their field. 

The information about the event can be found here. Dr. Julien Le Sommer is an accomplished computational oceanographer involved with the NEMO ocean model. 

Ocean via Bwog Archives