Abstract: Mobile Earthquake Recording in Marine Areas by Independent Divers (MERMAID) floats cannot obtain Global Positioning System (GPS) fixes during underwater listening and drift phases, so the surfacing position of the current profiling cycle can only be confirmed after the float reaches the sea surface. To address this problem, this study uses historical MERMAID profile records, event records, and multi-depth ocean current data to construct a historical-drift persistence baseline, a HYCOM-driven profile-layered Lagrangian integration module, and a CatBoost full-feature displacement regression model. Unlike single historical extrapolation or open-loop physical integration, the proposed method combines endpoint constraints from previous profiles, layered ocean-current integration features, and data-driven displacement regression, with region-adaptive fusion weights determined from the validation set. A total of 1,063 modelling samples were divided chronologically within each sea region, including 214 profile records used as an independent test set. The results show that the region-weighted fusion model achieved a mean absolute error of 7.800km and a root mean square error of 11.906km. Compared with the historical-drift persistence baseline and the CatBoost full-feature displacement regression model, the mean absolute error was reduced by 14.7% and 6.3%, respectively, and the hit rate within 20km reached 90.19%. Without changing the passive drifting process of the float, this method provides auxiliary information for pre-surfacing position estimation, shore-based communication scheduling, abnormal drift detection,and positional constraints on underwater observations. It may also inform position prediction for other profiling or low-power ocean-observation platforms that lack continuous underwater positioning.

Key words: MERMAID float, surfacing-position prediction, Lagrangian integration, CatBoost