Abstract: To address the computational inefficiency and high modeling costs of conventional fluid-structure interaction (FSI) simulations, this study proposes a data-driven reduced-order model (ROM) integrating Proper Orthogonal Decomposition (POD) with Gaussian Process (GP) regression. By leveraging high-fidelity simulation datasets from meshfree particle methods, the proposed framework resolves two critical challenges: high-dimensional flow field feature extraction and multi-physics coupled response prediction. The POD-based methodology establishes an intrinsic modal space that achieves 99.8% dimensionality reduction for million-dimensional flow field data. Combined with the non-parametric GP modeling architecture, the ROM achieves highly accurate predictions using only dozens of training samples. Numerical experiments demonstrate that the model attains a mean prediction error of approximately 2% for flow field variables in interpolation scenarios. Under parameter extrapolation conditions, the maximum relative error remains below 4.7% within 5% extrapolation ranges and maintains qualitative reliability even at 20% extrapolation levels. Computational efficiency tests reveal that the ROM requires only 10% of the time consumed by conventional SPH methods. Practical applications confirm the model's effectiveness: (1) For structural sinking dynamics under varying density ratios, the prediction error for key FSI features remains around 5%; (2) In reconstructing wake fields of underwater moving bodies, the average error in surface wave height prediction reaches 2% compared with full-order simulations. This work provides an efficient computational framework for analyzing marine FSI problems, with potential extensions to complex hydrodynamic scenarios requiring rapid multi-parametric evaluations.