The Download: Feature Articles
Ocean Turbulence, Sea Levels, and Machine Learning
By Victoria Lubas | June 16, 2021
Supercomputing resources like Greene are assisting Courant's Laure Zanna and her team of climate researchers
For climatologists at the NYU Courant Center for Atmosphere Ocean Science, studying the ocean’s role in climate change is a top research priority. Like studies of the Earth’s atmosphere, a greater understanding of ocean physics improves climatologists’ understanding of other elements in the interconnected climate system. From tracking weather patterns to predicting climate change, new interdisciplinary tech allows for more in-depth study of the ocean.
Laure Zanna, Courant professor of Mathematics and Atmosphere/Ocean Science, and her team focus their studies on ocean turbulence, sea level rise, and climate change predictions. Aided by NYU’s new supercomputer, Greene, the team can employ the latest machine learning and general circulation models. As the next step in her quest to further investigate ocean activity and the effect this has beyond the sea, Zanna recently began a new project supported by Schmidt Futures, a philanthropic initiative founded by Eric and Wendy Schmidt that “bets early on exceptional people making the world better, particularly through innovative breakthroughs in science and technology.”
Charting the Planet with Climate Models
The most fundamental elements of physics and climate science begin with a series of all-encompassing equations. Whether governing thermodynamics, energy conservation, or the laws of fluid dynamics, these nonlinear equations combine several elements into complex and intersecting representations of the climate system. In order to understand all of these components, physicists rely on climate models, which use powerful computers to break down these equations and apply them in the form of grids over given areas of the planet.
Zanna explained that for maximum mathematical representation and accuracy, these grid boxes have to be as small as possible to capture all the disparate elements of a climate system; however, “the smaller the size of your grid box, the more of them you [need] to actually cover the same ocean or the same atmosphere, which means that your computational cost increases quite drastically.” These already-complex climate systems receive a boost when physicists introduce machine learning, which combines the immense amounts of data collected with numerical simulations. The resulting algorithm allows previously unresolveable processes to be represented in climate models.
For this, Zanna is “using, in particular, the GPUs [powered by] Greene because they’re extremely powerful and designed to make the computation of those machine learning algorithms efficient.” Zanna explained that machine learning first entered the climate science scene about a decade ago and has undergone a substantial spike in use over the last few years.
Snapshot of the zonal (east-west) surface velocity from a high-resolution simulation, GFDL CM2.6. Source: Arthur Guillaumin
Machine Learning Makes Waves
Evolving tech sheds new light on existing data to allow for greater comprehension of previous theories and observations. Zanna explained that “trying to 'best use' machine learning to really capture the physics of a system requires knowledge of both the machine learning experts and the domain scientists who understand the climate system.”
For Zanna, this means working with machine learning experts Carlos Fernandez-Granda and Joan Bruna from NYU’s Center for Data Science to redesign the algorithms to best fit their purposes and to be able to represent uncertainty. The priority is that the algorithms be interpretable so that climatologists can really understand the physics occurring within the simulated climate system and use that knowledge to further improve future simulations. Zanna also emphasized the importance that these algorithms represent a degree of uncertainty. Noting the imperfections before pairing the algorithms with climate models could make the resulting climate projections more accurate.
Zanna’s work with climate models and machine learning revolves around the amount of atmospheric heat and carbon absorbed by the ocean, and the specifics of ocean turbulence. Oceans cover 70% of the planet and absorb 90% of the excess energy in the climate system, so one focus of her work is “understanding how that heat and excess of carbon has been taken up by the ocean in different regions and basically carried at-depth” and how that interacts with the large-scale ocean. The first step of this is revising climate models “to better represent those critical aspects of the climate system…where [heat and carbon are] being stored.”
Improvements in machine learning allow climatologists to rethink how ocean turbulence and mixing are occurring, which can recontextualize data and reveal causal factors that were unattributable in previous models. For example, a low resolution model may falsely represent an area of the ocean as smooth because the ocean turbulence within it is occurring below the model’s scale. This problem can be remedied with a higher resolution model, though that requires significantly more time and computer power. To prevent the representational loss of processes that occur below the climate model’s scale, Zanna advocates for the use of machine learning techniques, such as relevance vector machines and convolutional neural networks, to extract information from data of high-resolution simulations.¹
Can we improve parameterizations using data + machine language? Datasets from limited high-resolution models and observations, plus machine learning algorithms extracting complex info from data, equals a representation of complex “subgrid” processes learned from data. Source: Laure Zanna
Looking Ahead: Swimming Against the Tide
Zanna’s team is currently leading a large-scale, multi-university initiative² supported by Schmidt Futures’ Virtual Earth System Research Institute (VESRI), an initiative aiming to radically improve the credibility of climate predictions. Zanna explained that current difficulties in capturing the interactions between climate processes negatively affects the “projections of temperature, rainfall, and sea level.”²
Zanna’s project, titled Multiscale Machine Learning In coupled Earth System modeling (M²LInES), is focusing on the use of machine learning and artificial intelligence in climate simulations and the ways these technologies can extract the most meaningful elements of complex data to result in increasingly accurate climate simulations and predictions. M²LInES will use machine learning, datasets, and observational products to capture atmospheric, oceanic, and sea-ice processes.
The resulting improved models will enable better climate and weather forecasts in the future on which more advanced science and governmental policies can be built. As machine learning evolves it enables both the discovery of new data and a refreshed understanding of existing concepts, motivating climatologists to “just keep swimming” in pursuit of complete understanding.