Enhancing geophysical flow machine learning performance via scale separation

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