Using machine learning to predict oceanic wind forcing errors

Quantifying error budgets in geophysical quantities is an important task with wide-spread applications for time series prediction, numerical modelling, and data assimilation. We utilize a supervised machine learning approach to dynamically predict the spatio-temporal error budget of near-surface wind velocities over the ocean. A recurrent neural network (RNN) is trained with re-analyzed 10 meter wind velocities and corresponding pre-calculated estimates of the error budget during the 2012-2016 time period. The neural network's performance is examined by analyzing its error prediction for the subsequent year 2017. Our experiments show that a recurrent neural network can capture the globally prevalent wind regimes without knowledge about underlying physics and learn to derive wind velocity error estimates that are only based on wind velocity trajectories. The neural network can predict the lateral error distribution with small mismatches after being trained only at few isolated locations in different dynamic regimes. The presented approach can be additionally used in combination with numerical models for a cost efficient generation of ensemble simulations, or in combination with data assimilation to generate dynamically consistent error covariance information.