Machine Ocean – Combining Machine Learning and Earth Observations to Improve Simulations of Turbulent Behavior in the Earth System

Reliable weather and ocean forecasts are fundamental to society, and scientists continuously work to improve the forecasting models. The atmosphere and oceans are parts of the complex Earth System. On a global scale, that system is relatively simple: solar energy is absorbed by the Earth, and an equal flow of energy returns to space. Depending on the amount of greenhouse gases in the atmosphere, more or less longwave radiation is absorbed, which in turn determines the global temperature. On regional scales the Earth System is more difficult to simulate: energy moves around to compensate for the regional differences in incoming and outgoing radiation. This is accomplished by winds and currents which interact. At the center of the interactions lie the momentum transfer - the stress - across the interfaces. The stress is turbulent, and impossible to represent in an Earth System model (one would need millimeter-scale model resolution). Therefore, it is necessary to parameterize the stress. This is normally accomplished using semi-empirical bulk formulae to estimate the stress from sparse surface measurements. In Machine Ocean we will let the data themselves produce a "stress solver", through the use of Machine Learning. The recent launch of the Sentinel 1 space mission, which observes the sea surface at very high resolution, and machine learning algorithms that can be trained on and learn from enormous datasets, lead us to hypothesize that it is possible to develop a stress model that depends less on empiricism and more on theory and data. As a testbed, we will apply MET Norway's storm surge forecast model. Storm surge forecasting depends almost exclusively on transfer of atmospheric pressure and stress, so it is ideal way to test the impact of improved stress. At the same time, storm surges are very high on the list of dangerous climate change impacts in Norway, so improving storm surge forecasts is very important for forecasters.

The overarching hypothesis for Machine Ocean is that “the present explosion in volume, variety and velocity (VVV) of Earth Observation acquisition (as spearheaded by the Copernicus, in particular the Sentinel 1 mission) combined with new methods to harvest information from big data will allow us to gain further insights into, and significantly reduce the uncertainty in, parameterization of momentum transfer between atmosphere and ocean.” Vertical momentum transfer is one of the most important process in the Earth System, influencing the transfer of carbon, oxygen, heat, freshwater and other quantities between the different spheres, yet possibly the hardest process to measure. The transfer occurs on small horizontal and temporal scales, so it is almost always necessary to parameterize in numerical simulations. The use of machine learning methods to directly predict turbulent stress in Reynolds-averaged Navier–Stokes equations has recently been proposed, and developed for simplified setups, but the field of machine learning applications in fluid dynamics in general and for momentum transfer and larger scale atmosphere models in particular, is in its infancy.