A Mesh-Adaptive Hypergraph Neural Network for Unsteady Flow Around Oscillating and Rotating Structures

Rui Gao, Zhi Cheng, Rajeev K. Jaiman · 2025

Read A Mesh-Adaptive Hypergraph Neural Network for Unsteady Flow Around Oscillating and Rotating Structures by Rui Gao, Zhi Cheng, Rajeev K. Jaiman (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-22252.txt.

Introduction

Rui Gao, Zhi Cheng, and Rajeev Jaiman build in 2025 a neural network that predicts fluid flow around moving and spinning structures. Networks shaped as graphs have become handy stand-ins for slow flow simulations. Most earlier work, though, assumed the space itself stays still, which fails for propellers, turbines, and flapping wings.

Their design splits the mesh in two: one part turns with the structure, and one stays fixed. A single layer of cells at the seam stretches and adapts, which avoids re-sending network data across the seam at every step. Special rebuild steps counter the errors that seam stretching causes. Tests on a flapping wing in two dimensions and a spinning cube in three held steady for hundreds to a thousand steps. Lift and drag matched full simulations closely, and a few pressure sensors could pin errors down further.

Catalog record

CREATORRui Gao, Zhi Cheng, Rajeev K. Jaiman
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-22252

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Mesh-Adaptive Hypergraph Neural Network for Unsteady Flow Around Oscillating and Rotating Structures Rui Gao, Zhi Cheng, Rajeev K. Jaiman Abstract Graph neural networks, recently introduced into the field of fluid flow surrogate modeling, have been successfully applied to model the temporal evolution of various fluid flow systems. Existing applications, however, are mostly restricted to cases where the domain is time-invariant. The present work extends the application of graph neural network-based modeling to fluid flow around structures rotating with respect to a certain axis. Specifically, we propose to apply a graph neural network-based surrogate model with part of the mesh/graph co-rotating with the structure and part of the mesh/graph static. A single layer of interface cells are constructed at the interface between the two parts and are allowed to distort and adapt, which helps in circumventing the difficulty of interpolating information encoded by the neural network at every graph neural network layer. Dedicated reconstruction and re-projection schemes are designed to counter the error caused by the distortion and connectivity change of the interface cells. The effectiveness of our proposed framework is examined on two test cases: (i) fluid flow around a 2D oscillating airfoil, and (ii) fluid flow past a 3D rotating cube. Our results show that the model achieves stable rollout predictions over hundreds or even a thousand time steps. We further demonstrate that one could enforce accurate, error-bounded prediction results by incorporating the measurements from sparse pressure sensors. In addition to the accurate flow field predictions, the lift and drag force predictions closely match with the computational fluid dynamics calculations, highlighting the potential of the framework for modeling fluid flow around rotating structures, and paving the path towards a graph neural network-based surrogate model for more complex scenarios like flow around marine propellers. Keywords. Graph Neural Network, Rotating Fluid Flow, Data-Driven Model, Digital Twin 1 Introduction Rotating structures, including marine and aeronautic propellers, centrifugal pumps and blowers, stirring mixers, wind and gas turbines, etc., comprise an important part of modern machinery. Most of these rotating structures are immersed in fluids, usually air or water, and the fluid flow around such rotating structures is frequently of interest due to their profound impact on the performance and lifespan of machines that include these rotating structures. To improve the understanding of the fluid flow around rotating structures and to facilitate design and control efforts, computational fluid dynamics (CFD) techniques are usually adopted to generate simulations for such fluid systems [ 1 , 2 , 3 , 4 ] . However, CFD simulations can be computationally heavy, especially when simulating large 3D cases. This limitation leads to challenges in the application of CFD simulations in the iterative design process and real-time model-based flow control. An increasing interest has therefore arisen in the development of cheaper surrogate models that could be adopted for the aforementioned purposes. Recent rapid development and application of deep learning methods has inspired a series of applications of these models in fluid flow modeling, including the modeling of fluid flow around rotating structures. Existing work includes predicting or fitting fluid field within a pump [ 5 ] , around a propeller [ 6 , 7 ] , within a stirring mixer [ 8 , 9 ] , around a spinning sphere [ 10 ] , within a co-rotating disk cavity [ 11 ] , etc. These works typically adopt either convolutional neural networks or multilayer perceptrons with specially-designed features or training losses. The target in most of these works, however, is to obtain a prediction of static or time-averaged flow field rather than the spatio-temporal evolution of the flow. One significant exception is the recent work by Li et al. [ 6 ] which features time series prediction of the wake after a propeller with a merged convolution-transformer network, under a series of limitations: Only the flow in the wake region is predicted, only a 2D slice out of the 3D flow field is predicted, and the authors only report predictions up to 15 time steps – a rather short rollout compared with standard results in 2D flow scenarios. Recently, graph neural networks (GNNs) have been introduced to the modeling of fluid flow. Being able to work with graphs transformed from unstructured mesh, one could easily adopt a body-fitting mesh that is only refined in regions of interest, leading to significant reduction in the mesh size compared to the case with convolutional neural networks while maintaining a satisfactory resolution. This unique advantage has inspired the recent integration of graph neural networks into various fluid flow modeling scenarios. Applications include, but are not limited to, flow around bluff or streamlined bodies [ 12 , 13 , 14 , 15 , 16 , 17 ] , fluid-structure interactions [ 18 ] , particulate suspension [ 19 ] , reacting flow [ 20 ] , cavitating flow [ 21 ] , shock waves [ 22 ] , and environmental flow above [ 23 ] or below [ 24 ] the surface of the Earth. For a more comprehensive review of recent advances in the field, the reader is referred to Zhao et al. [ 25 ] . Considering the variety of reported applications of graph neural networks, it is reasonable to hypothesize that it will also serve as a proper architecture for modeling fluid flow around rotating structures. Li et al. [ 26 ] recently applied a graph neural network to model the flow around a radial turbine. However, similar to previous works using other network architectures, their focus was limited to predicting time-averaged flow fields, without addressing the spatio-temporal dynamics of unsteady flow around rotating structures. They also chose to only model only a single blade of the turbine, which relies on the implicit assumption that the flow is cyclic periodic that can be true for the time-averaged flow field but not for instantaneous flow fields. In this work, we aim to bridge this gap by developing a graph neural network framework capable of modeling the full spatio-temporal evolution of fluid flow around rotating bodies. Specifically, we partition the domain into two sub-domains, one co-rotating with the solid body and one static. A single layer of cells is constructed at the interface between the two sub-domains and is allowed to adapt over time. A graph neural network is employed to predict the evolution of the system states over each time step, and a series of reconstruction and re-projection steps are designed and implemented to account for the mesh distortion and adaptation between the time steps. Predictions from the framework are enforced to be equivariant over the rotation of the system domain via appropriate input and output feature transformations. The proposed graph neural network framework will be assessed in two test cases, namely the flow around a 2D oscillating airfoil and the flow around a 3D rotating cube. We will demonstrate that the network is able to provide stable long-term rollout predictions for both cases. For the 2D case, we additionally demonstrate that one could enforce accurate, error-bounded prediction results by integrating the measurements from sparse pressure sensors mounted near the surface of the solid body. These results highlight the ability of graph-based surrogate modeling for fluid flow around rotating bodies and

Read the complete text