Applications of Deep Learning to Ocean Data Inference and Subgrid Parameterization

Abstract

Oceanographic observations are limited by sampling rates, while ocean models are limited by finite resolution and high viscosity and diffusion coefficients. Therefore, both data from observations and ocean models lack information at small and fast scales. Methods are needed to either extract information, extrapolate, or upscale existing oceanographic data sets, to account for or represent unresolved physical processes. Here we use machine learning to leverage observations and model data by predicting unresolved turbulent processes and subsurface flow fields. As a proof of concept, we train convolutional neural networks on degraded data from a high-resolution quasi-geostrophic ocean model. We demonstrate that convolutional neural networks successfully replicate the spatiotemporal variability of the subgrid eddy momentum forcing, are capable of generalizing to a range of dynamical behaviors, and can be forced to respect global momentum conservation. The training data of our convolutional neural networks can be subsampled to 10-20% of the original size without a significant decrease in accuracy. We also show that the subsurface flow field can be predicted using only information at the surface (e.g., using only satellite altimetry data). Our results indicate that data-driven approaches can be exploited to predict both subgrid and large-scale processes, while respecting physical principles, even when data are limited to a particular region or external forcing. Our in-depth study presents evidence for the successful design of ocean eddy parameterizations for implementation in coarse-resolution climate models.Plain Language Summary Models of the ocean and ocean observations are imperfect. Due to this imperfection, simulations of the ocean and our observations are not quite the same as the true ocean currents. We, therefore, need ways to make our ocean data more realistic and complete and to make it more similar to the actual ocean. Scientists have traditionally approached this problem in a pen-and-paper style, considering physical theories and mechanisms. This study instead uses machine learning, which focuses on data as opposed to equations on a black board. We successfully use a particular type of machine learning algorithm, called a convolutional neural network, to make the most of current oceanographic data. This type of neural network works well even if ocean data are limited to a particular area. Future work will involve combining machine learning with physical theories of the ocean.

Machine Learning for Stochastic Parameterization: Generative Adversarial Networks in the Lorenz '96 Model

“…Regularization is critical for more complex machine learning models, so that they can converge to optimal and robust configurations in large parameter spaces. Machine learning for parameterizations has been considered since (Krasnopolsky et al, 2005), and recently, multiple groups have begun developing new parameterizations for a variety of processes (Bolton & Zanna, 2019;Gentine et al, 2018;Rasp et al, 2018;Schneider et al, 2017). The most common regularization in existing machine learning parameterization approaches has been multitask learning (Caruana, 1997), in which the machine learning model predicts multiple correlated values simultaneously and learns to preserve the correlations.…”

Analog Forecasting of Extreme‐Causing Weather Patterns Using Deep Learning

“…Recent advances in artificial intelligence have revolutionized how problems in various domains of business and science are approached (Goodfellow et al, 2016;LeCun et al, 2015). For example, in climate science, using machine learning techniques to accurately and efficiently represent unresolved physical processes in the atmosphere and ocean has produced promising results (Brenowitz & Bretherton, 2018;Bolton & Zanna, 2019;O'Gorman & Dwyer, 2018;Rasp et al, 2018;Salehipour & Peltier, 2019) and has the potential to significantly improve climate modeling and long-term climate projections in the coming years (Chattopadhyay et al, 2019;Gentine et al, 2018;Reichstein et al, 2019;Schneider et al, 2017). Moreover, deep learning techniques have been very successful in predicting some types of sequential data (Goodfellow et al, 2016).…”

Constraining stochastic parametrisation schemes using high‐resolution simulations

“… indicates the fraction of fine grid box within coarse grid box , and the vertical level of the field is indicated by index . Note that this is one choice of coarse‐graining procedure, with alternatives including Gaussian (Bolton and Zanna, ) or spectral filters (Shutts and Pallares, ).…”

“…The coarse (fine) grid box is identified by the index n (f ). W n,f indicates the fraction of fine grid box f within coarse grid box n, and the vertical level of the field is indicated by index k. Note that this is one choice of coarse-graining procedure, with alternatives including Gaussian (Bolton and Zanna, 2019) or spectral filters (Shutts and Pallares, 2018). The fine-and coarse-resolution datasets are defined on model levels, and interpolation must also be performed in the vertical.…”