Data Assimilation with Ocean Models: A Case Study of Reduced Precision and Machine Learning in the Gulf of Mexico

Abstract

The deployment of increasingly higher resolution environmental observation systems along with higher resolution geophysical models has caused operational data assimilation systems to explore techniques to increase computational performance while maintaining numerical accuracy. Recent research efforts have explored implementing reduced or mixed-precision in geophysical circulation models and data assimilation schemes to validate their numerical accuracy versus full-precision models, or using Machine Learning techniques to enhance and speed up model simulations and data assimilation. In this paper, we combine the two techniques by examining the effects of coupling a modified, reduced-precision data assimilation system, the Tendral Statistical Interpolation System (T-SIS) version 1.0, with a Machine Learning model using the HYbrid Coordinate Ocean Model (HYCOM) outputs for a Gulf of Mexico experiment. A Unet type Convolutional Neural Network (CNN) was trained with two years, 2009–2010, of T-SIS reduced-precision assimilation runs. It was tested on two different years, 2002 and 2006, with unique ocean circulation properties. For sea-surface height, the ocean modeling variable, an optimal, reduced-precision trained CNN was capable of predicting the full-precision 2002 and 2006 data assimilation increment with an analysis root mean square error (RMSE) reduction of the same order versus the full-precision trained CNN version.