Multi-Fidelity Learned Emulator for Waves and Porous Coastal Structures Interaction Modelling

Abstract

A thorough understanding and treatment of wave-structure interaction (WSI) mechanics is essential for the rigorous engineering design of coastal protections. Conventional numerical analysis methods are accurate and generalize well, but are heavily dependent on the adopted mesh resolution and frequently incur substantial computational costs. To bypass these limitations, a meshless multi-fidelity residual neural network (MRNN) emulator is introduced in this study to infer the spatio-temporal responses arising from WSI. MRNN first employs a ‘low-fidelity’ simulator to learn basic WSI relationships by training on simulations obtained using a coarse numerical mesh. Subsequently, a ‘high-fidelity’ (HF) simulator is then employed to learn the mapping between numerical simulations performed using the coarse mesh and additional detailed fine meshes. The results indicate that MRNN is a highly robust emulator which requires significantly less HF data through its hierarchical framework compared to conventional single-fidelity data-driven strategies. By way of example, the MRNN emulator is applied to the cases of a porous dam break and breakwater. A broad spectrum of WSI responses, such as water through the porous dam medium, can be accurately captured using the MRNN emulator which is benchmarked against a conventional numerical modelling with a fine mesh. The computational efficiency of the MRNN is shown to be independent of the mesh resolution and complexity of the studied partial differential equations. It provides a generic and utilitarian emulator for any engineering problem of interest.