Hybrid ISPH_GNN method for simulating violent wave-structure interactions using wave-only data for training

Abstract

It has been well-known that the incompressible Smoothed Particle Hydrodynamics (ISPH) is a powerful method for simulating violent wave-structure interactions (WSIs) concerned in marine engineering. However it is time consuming, primarily due to the need of solving pressure Poisson’s equation (PPE) involved in this method. In our previous publications, we are first to propose a hybrid approach embedding the graph neural network (GNN) into ISPH method to form the hybrid ISPH_GNN method for simulating free-surface problems, where the GNN is employed to replace solving the PPE. We demonstrated that the computational time for evaluating the pressure using GNN can be of one order less than that spent by directly solving PPE to achieve similar level of accuracy. More importantly, we also demonstrated in our previous publications that the GNN trained only on data for wave-only (referring to no structure or obstacles in wave fields) cases can be satisfactorily applied to the cases for wave-floater interactions. However, what we have not previously studied is if the GNN trained only by using wave-only cases can be used for simulating violent WSIs. One of the original contributions of this paper is to answer this question. In addition, transfer learning has been proved to be a machine learning (ML) technique that can significantly enhance efficiency and improve the performance in other fields but has not been explored in the hybrid ISPH_GNN method. Another original contribution of this paper is to explore the potential of integrating transfer learning with the ISPH_GNN for simulating violent WSIs. Specifically, we will demonstrate that the GNN trained by using data from sloshing and dam-breaking cases without any structure (termed as wave-only data in this paper) can be employed to simulate more complex cases, such as water entry of an object, wave impact on a trapezoidal structure and wave interaction with an oscillating wave surge converter, all of which involve violent WSIs. We will also demonstrate that the transfer learning technique with use of a small volume of additional data has a potential in enhancing the prediction accuracy of the ISPH_GNN. Furthermore, we will show that the ISPH_GNN significantly reduces computational time for pressure evaluation in violent WSI cases, even with a more significant reduction compared to wave-floater interaction cases studied in our previous work. These highlight the strong potential of the ISPH_GNN for broad applications in marine engineering, opening a novel route to employ ML without need of generating data for very complex cases of violent WSIs.