Open Collections

Abstract

Modern iterative design optimization and real-time active control of marine vessels require fast modeling of the fluid flow around its various sub-systems. Traditional numerical simulations, however, can be too slow for such purposes. In this dissertation, we provide a comprehensive hypergraph neural network framework that can be applied to fast surrogate modeling of various fluid and fluid-structure systems around marine vessels. Revolving around this target, this dissertation can be split into two parts. In the first part, we develop a hypergraph neural network architecture that serves as the backbone of the framework. Inspired by the data connectivity in the finite element method, we construct a hypergraph by connecting the nodes by elements. A hypergraph message-passing network that mimics the calculation process of local stiffness matrices is defined on such a node-element hypergraph. We verify the efficiency of the network on fluid flow around fixed body benchmark data sets, and compare its performance with baseline model MeshGraphNet. In the second part, we equip the framework with additional components that enable it to model various fluid and fluid-structure systems of different sizes. To model fluid flow around vibrating and deforming bodies, we embrace an arbitrary Lagrangian-Eulerian formulation and use a sub-network to model the mesh and solid movements. To model fluid flow around rotating structures, we employ a co-rotating domain under the inspiration of the sliding mesh method. We further enforce a series of geometric and physical priors in the framework to enhance its generalization capability. We also design partitioning and buffering schemes that enable training and inference for large three-dimensional cases on single and distributed machines. The completed framework is tested on a series of benchmark problems involving both periodic and chaotic fluid systems and fluid-structure systems in both two-dimensional and three-dimensional setups, and demonstrates ability to generate stabilized and accurate roll-out predictions over a long time horizon. It is expected that the framework proposed in this dissertation can serve as the physics simulation component in a digital twin framework used in iterative design optimization or real-time active control of marine vessels.