Boosting performance in machine learning of geophysical flows via scale separation

Recent advances in statistical and machine learning have opened the possibility to forecast the behavior of chaotic systems using recurrent neural networks. In this article we investigate the applicability of such a framework to geophysical flows, known to involve multiple scales in length, time and energy and to feature intermittency. We show that both multiscale dynamics and intermittency introduce severe limitations on the applicability of recurrent neural networks, both for short-term forecasts, as well as for the reconstruction of the underlying attractor. We suggest that possible strategies to overcome such limitations should be based on separating the smooth large-scale dynamics from the intermittent/small-scale features. We test these ideas on global sea-level pressure data for the past 40 years, a proxy of the atmospheric circulation dynamics. Better short and long term forecasts of sea-level pressure data can be obtained with an optimal choice of spatial coarse grain and time filtering.