Internal solitary waves in the Banda Sea, a pathway between Indian and Pacific oceans: Satellite observations and physics-AI hybrid forecasting

Abstract

Internal solitary waves (ISW) are widespread in global oceans, and satellite/in-situ observations showed that the Banda Sea has frequent ISW activities, characterized by long-wave crests, fast propagation speeds, and large amplitudes exceeding 100 m. In this paper, we conducted a comprehensive ISW study in the Banda Sea to reveal ISW characteristics by collecting 417 synthetic aperture radar and optical images from 2013 to 2019. The constructed dataset comprises 134 pairs of matched satellite images and a total of 12,021 ISW propagation vectors were extracted. Satellite observation reveals that ISWs in the Banda Sea mainly originate from the Ombai Strait and propagate northward, with an average propagation speed of over 2.50 m/s and with seasonal variation of less than 20 %. To forecast ISW propagations, we developed a physics-informed neural network ISW forecast model combining the classic Eikonal Eq. (EE) and the data-driven AI algorithms following a two-step transfer learning scheme. The forecast model employs a three-hidden-layer structure with 512 nodes in each layer. Firstly, the hybrid model includes ISW physics by setting the EE as the loss function. The second step is the data-driven process, which exploits a fully connected neural network and collected ISW dataset to improve EE-based model performance by 61 % with a loss function of the mean squared error. Through the two-step training, the forecast model adopts ISW physics and also benefits from the high accuracy of the data-driven process. We randomly selected 188/118 satellite images from the built dataset to serve as the training/test dataset for the data-driven process. After the second-step training, the root mean square (average) error of the model-predicted ISW propagation time reduced from 2.59 (2.37) h to 1.01 (−0.01) h. Error analysis shows that the data-driven process can efficiently correct the systematic error in the first-step model, which stems from errors in determining the ISW source and the propagation speed distribution map. Using the developed model, we predicted the propagation time of the ISWs and compared these predictions with satellite observations and in-situ observations. The comparison showed a high degree of agreement regarding the ISWs' location and their wave crests' geometry between model predictions and satellite/in-situ observations. Key differences between the proposed model and previous models are discussed.