A physics-informed machine learning approach for predicting acoustic convergence zone features from limited mesoscale eddy data

Abstract

Mesoscale eddies are prevalent mesoscale phenomena in the oceans that alter the thermohaline structure of the ocean, significantly impacting acoustic propagation patterns. Accurately predicting acoustic convergence zone features has become an urgent task, especially when data are limited in deep-sea mesoscale eddy environments. This study utilizes physics-informed machine learning to identify and predict the acoustic convergence zone features of mesoscale eddies under limited data conditions. Initially, a method based on convex hull ratio was utilized to identify mesoscale eddies from the JCOPE2M reanalysis dataset and AVISO data in the Kuroshio‐Oyashio Extension. Subsequently, by integrating physical models and ray acoustics, relevant features of mesoscale eddies and convergence zones are extracted. Then, K-fold cross-validation and sparrow search algorithms are employed to select the optimal machine learning algorithm, ensuring high model accuracy. The resulting model requires only a thermohaline profile near the eddy center and sea surface height to predict convergence zone features within the mesoscale eddy environment, achieving a MAE of approximately 1.00 km and an accuracy (within 3 km) exceeding 95%. Additionally, leveraging physics-informed machine learning methods contributes to a maximum reduction of 0.82 km in MAE and an improvement in accuracy by 2.80% to 11.92% compared to models without physical information input. Finally, the model’s validity and reliability in the actual ocean environment are verified by cross-validating it with data from various sea regions" in bright yellow and Argo profiling float data. The findings provide novel insights into acoustic propagation in mesoscale eddy environments and subsequent ocean acoustic research.