Abstract:The paper presents a Graph Attention Convolutional Network (GACN) for flow reconstruction from very sparse data in time-varying geometries. The model incorporates a feature propagation algorithm as a preprocessing step to handle extremely sparse inputs, leveraging information from neighboring nodes to initialize missing features. In addition, a binary indicator is introduced as a validity mask to distinguish between the original and propagated data points, enabling more effective learning from sparse inputs. Trained on a unique data set of Direct Numerical Simulations (DNS) of a motored engine at a technically relevant operating condition, the GACN shows robust performance across different resolutions and domain sizes and can effectively handle unstructured data and variable input sizes. The model is tested on previously unseen DNS data as well as on an experimental data set from Particle Image Velocimetry (PIV) measurements that were not considered during training. A comparative analysis shows that the GACN consistently outperforms both a conventional Convolutional Neural Network (CNN) and cubic interpolation methods on the DNS and PIV test sets by achieving lower reconstruction errors and better capturing fine-scale turbulent structures. In particular, the GACN effectively reconstructs flow fields from domains up to 14 times larger than those observed during training, with the performance advantage increasing for larger domains.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **How to reconstruct the flow field from very sparse data under time - varying geometric conditions**. Specifically, the author proposes a method based on Graph Attention Convolutional Network (GACN) to address the following challenges: 1. **Processing of extremely sparse data**: Traditional methods often fail to accurately capture the complex flow - field features when dealing with extremely sparse data. This paper introduces the Feature Propagation (FP) algorithm as a pre - processing step and introduces the Binary Indicator (BI) as a validity mask, thus learning more effectively from sparse inputs. 2. **Reconstruction of flow fields under time - varying geometric conditions**: Most existing machine - learning models are suitable for static configurations and are difficult to adapt to dynamic changes in the flow - field domain. The GACN proposed in this paper can capture the changes of irregular grids through the node position and the distance features of the edges, and can handle geometric structures that change over time. 3. **Generalization ability between different data sets**: Existing models have difficulty in generalizing between different types of flow - field data sets (such as high - fidelity numerical simulations and experimental measurements). This paper shows that GACN not only performs well on numerical simulation data, but can also be successfully applied to experimental data, with good generalization ability. ### Overview of the solution To address the above problems, the paper proposes the following solutions: - **GACN model**: Combine the advantages of Graph Neural Network (GNN), especially the graph attention mechanism, to better capture non - local dependencies in complex flow fields. - **Feature Propagation algorithm (FP)**: Used to initialize missing feature values, fill sparse data points with information from neighboring nodes to ensure physical consistency. - **Binary Indicator (BI)**: As an additional feature, it distinguishes between original data points and propagated data points, helping the network to learn sparse inputs more effectively. - **Multi - scale verification**: Verify the robustness and generalization ability of the model by testing at different resolutions and domain sizes. ### Main contributions - **Efficient processing of extremely sparse data**: GACN can still effectively reconstruct the flow field even when 99% of the data is missing. - **Adaptation to time - varying geometric conditions**: Through the graph structure and attention mechanism, GACN can handle geometric structures that change over time and capture the dynamically changing flow - field features. - **Cross - data - set generalization ability**: The model not only performs excellently on numerical simulation data, but is also successfully applied to experimental data, demonstrating strong generalization ability. In summary, this paper aims to solve the key challenges in flow - field reconstruction, especially in the application of sparse data and time - varying geometric conditions, by proposing a new GACN model.

Flow reconstruction in time-varying geometries using graph neural networks

Graph Neural Networks for Aerodynamic Flow Reconstruction from Sparse Sensing

Flow Reconstruction with Uncertainty Quantification from Noisy Measurements Based on Bayesian Physics-Informed Neural Networks

Graph Neural Networks and Differential Equations: A hybrid approach for data assimilation of fluid flows

Flow Reconstruction Using Spatially Restricted Domains Based on Enhanced Super-Resolution Generative Adversarial Networks

Graph convolutional networks applied to unstructured flow field data

Reconstructing turbulent velocity information for arbitrarily gappy flow fields using the deep convolutional neural network

A novel graph modeling method for GNN-based hypersonic aircraft flow field reconstruction

A generative adversarial network based on an efficient transformer for high-fidelity flow field reconstruction

High-fidelity reconstruction of turbulent flow from spatially limited data using enhanced super-resolution generative adversarial network

Flow Reconstruction over a SUBOFF Model Based on LBM-generated Data and Physics-Informed Neural Networks

Flow Completion Network: Inferring the Fluid Dynamics from Incomplete Flow Information using Graph Neural Networks

Deep learning fluid flow reconstruction around arbitrary two-dimensional objects from sparse sensors using conformal mappings

A graph neural network-based framework to identify flow phenomena on unstructured meshes

Finite Volume Graph Network(FVGN): Predicting unsteady incompressible fluid dynamics with finite volume informed neural network

Graph neural networks for laminar flow prediction around random 2D shapes

Enhancing Data-Assimilation in CFD using Graph Neural Networks

FlowMRI-Net: A generalizable self-supervised physics-driven 4D Flow MRI reconstruction network for aortic and cerebrovascular applications

Temporal and Spatial Flow Field Reconstruction from Low-Resolution PIV Data and Pressure Probes Using Physics-Informed Neural Networks

Enhancing Graph U-Nets for Mesh-Agnostic Spatio-Temporal Flow Prediction

A practical approach to flow field reconstruction with sparse or incomplete data through physics informed neural network