关键词: 船舶液舱晃荡, 数值仿真, 自注意力机制, 图神经网络, 移动粒子半隐式法

Abstract:

To address the high consumption of computational resources in simulating ship liquid tank sloshing using computational fluid dynamics simulation methods, a data-driven numerical simulation model was proposed based on graph neural networks. An encoder-processor-decoder framework was employed in the proposed model. The encoder extracted features of fluid particles from the first five time steps. The processor learnt latent motion patterns of fluid and updated features, and the decoder predicted features of particles at subsequent time steps. The processor incorporated a self-attention mechanism to enable dynamic adjacency weight allocation and emphasize the influence of irregular tank wall regions. Training data for the model were generated through the moving particle semi-implicit (MPS) method. Numerical simulations were conducted for rolling conditions of scaled ballast tanks. The results have demonstrated that the proposed model achieves over 50% improvement in simulation accuracy measured by MAE compared to conventional graph neural networks, while maintaining two orders of magnitude higher computational efficiency than the MPS method. This approach provides a novel numerical simulation for ship tank sloshing analysis that effectively balances precision and efficiency.

Key words: ship liquid tank sloshing, numerical simulation, self-attention mechanism, graph neural network, moving particle semi-implicit (MPS) method