Deep Learning Forecast of Semidiurnal Internal Tides From the Mariana Arc

Abstract The semidiurnal internal tides radiated from the Mariana Arc converge in a focal region of the West Mariana Basin, featuring rich variations under the modulation of complex topographies and background circulations. Traditional methods, including harmonic analysis and numerical simulations, are difficult to reliably characterize and forecast such multimodal and incoherent internal tide field. In this study, we develop a data‐driven deep learning model, named BiLSTM‐Transformer‐ANN (BILTA). BILTA can learn the underlying processes from the time‐series observations. The results demonstrate BILTA's efficiency and accuracy in forecasting variable semidiurnal internal tides in the West Mariana Basin. During the 2‐month testing period, the averaged coefficient of determination between forecasted and observed zonal baroclinic semidiurnal tidal currents is up to 0.92. Even for meridional currents of strong incoherent and internal partly standing waves, the averaged coefficient of determination can still reach 0.885. BILTA can successfully forecast both the coherent and incoherent internal tides. Particularly, the depth‐averaged root mean square error of estimation on both zonal and meridional incoherent currents could be lower than 0.01 m/s. Moreover, BILTA is capable of capturing multi‐mode energy responses. The forecast accuracy of horizontal kinetic energy for higher‐mode is lower than that of the lower‐mode, which can be interpreted by their different energy propagation and dissipation properties. These results show strong potential for deep learning models to characterize and forecast complex and varying internal tide energy fields among the global ocean.