Deep Learning for a stochastic subgrid parameterization of ocean momentum forcing

Abstract:

Despite the use of supercomputers, coupled climate simulations that span multi centuries cannot be run at a high-enough spatial resolution to resolve mesoscale (10-100km horizontal scale) ocean dynamics. These mesoscale dynamics backscatter to macroscales, and therefore need to be accounted for. This is commonly achieved through parameterizations of subgrid forcing, that consist of terms derived from resolved macro-scale quantities, and which are injected back on the right-hand-side of the discretized governing equations for the atmosphere and ocean. Classically, parameterizations are derived within closed-form equations, based on the practitioner’s understanding of physical phenomena. However, this approach has shown its limits. Recently, several studies have considered the use of Deep Learning methods to parameterize subgrid forcing within macro-scale ocean equations. We present results on the use of Deep Learning to infer the subgrid forcing using data from CM2.6 simulations, a 1/10-degree state-of-the-art coupled climate model. We successfully train a Convolutional Neural Network (CNN) for the subgrid momentum forcing using macro-scale surface velocities. Rather than predicting a single number using a standard Mean Square Error loss function, at each location and at each time step of the coarse grid, we predict the moments of a Gaussian distribution. This allows each prediction to be associated with an uncertainty estimate, as shown in the figure. Therefore, our approach is to learn a stochastic representation of the subgrid momentum forcing. We also investigate the ability of our model to generalize to 1) unseen regions and 2) a different climate. To do so, we train on a combination of selected regions that cover a large range of the oceans’ dynamics. We then test our NN on new regions, showing that generalization depends on the selected training regions. We also show that when our CNN is trained using data from the pre-industrial simulations, it generalizes well to a warming scenario (here, 1% CO2 increase) without further tuning.