Wind-field prediction and sparse reconstruction around polar ship superstructures using reduced-order data-driven methods

Abstract

Accurate acquisition of wind fields around polar ship superstructures is essential for cold-protection design and local wind-environment assessment. This study proposes a unified framework integrating full-scale CFD, proper orthogonal decomposition (POD), sparse-measurement reconstruction, and multilayer perceptron (MLP)-based surrogate prediction. A high-fidelity dataset covering 1680 operating conditions was constructed using Latin hypercube sampling, with 177 spatial points for each condition. Physical analysis shows that the normalized wind field is primarily governed by relative wind direction. The mean amplification factor reaches 0.931 at 90°, while the maximum reaches 1.278 at 270°, indicating pronounced acceleration around superstructure edges and passage regions. POD analysis shows that 78 modes capture 99.0223% of the cumulative energy. For sparse reconstruction, the relative L2 error decreases from approximately 0.076 to 0.063 as the sensor number increases from 3 to 12, with eight optimized sensors sufficient to recover the dominant flow structures. For prescribed-condition prediction, the MLP surrogate achieves an MSE of 0.00991, MAE of 0.0751, and R 2 of 0.983. The framework enables rapid design-stage prediction and service-stage sparse-sensor-based reconstruction, supporting efficient cold-protection assessment for polar ships.