Numerical Simulation of Liquid Tank Sloshing Based on Graph Neural Network with Self-Attention Mechanism

Liquid tank sloshing simulation is crucial for understanding and mitigating the dynamic behavior of fluids in various engineering applications, such as aerospace, marine, and automotive industries, where accurate prediction of fluid motion is essential for stability and safety. This paper proposes a numerical simulation model called Graph Self-Attention Neural Network (GSANN) for liquid tank sloshing. It is mainly used to predict the physical properties of fluids, including displacement, velocity, and acceleration. Datasets for GSANN are generated using the Moving Particle Semi-Implicit (MPS) method. The GSANN employs an Encoder-Processor-Decoder framework. The encoder maps node and edge data across various time steps into a latent space. The processor performs message passing operations and outputs the latent space representation for the subsequent time step, with self-attention being utilized to boost the precision of simulations. The decoder subsequently reconstructs the latent space into fluid particle flow motion. The numerical simulation results on two-dimensional rectangular tanks show consistency with the physical experimental outcomes and demonstrate that GSANN outperforms Graph Network-based Simulators (GNS) in performance. Meanwhile, GSANN’s computational efficiency is significantly improved, with two orders of magnitude faster computation speed compared to the traditional MPS method. In summary, the results of this research represent a groundbreaking advance in the application of deep learning to simulate the liquid tank sloshing phenomena.