A Hybrid Deep Learning Model for El Niño Southern Oscillation Dynamics

Event-to-event differences of the El Niño Southern Oscillation (ENSO) result in different patterns of extreme climate conditions globally, which requires ENSO forecasts to not only accurately predict the likelihood of an event but also its type. The high autocorrelation of tropical sea surface temperature anomalies (SSTA) allows sub-seasonal to seasonal (S2S) forecasts of ENSO. Recent studies have suggested that skillful multi-year predictions may even be possible after strong El Niño events.

The Linear Inverse Model (LIM) has been shown to produce state-of-the-art ENSO forecasts. LIM describes the dynamics of the slower-varying ocean as stochastically forced by the rapidly varying atmosphere with its deterministic dynamics assumed to be multivariate linear. Due to the linearity assumption, however, LIM is unable to capture observed asymmetries of ENSO that raise the question of whether predictable nonlinearities must be accounted for or may be treated stochastically. 

In this study, we combine deep neural networks (DNN) with the LIM to assess the role of predictable nonlinearities and non-Markovianity in the evolution of monthly tropical SSTA. The different models are tested on SSTA and sea surface height anomalies (SSHA) data from the CESM2 preindustrial control run, where we observe that modeling nonlinearities significantly enhances the forecast accuracy, particularly in the western tropical Pacific within the 9 to 18-month range. Our results further indicate that the asymmetry of warm and cold events is the main source of nonlinearity that improves the forecast skill.