Flow reconstruction in time-varying geometries using graph neural networks
Published in ArXiv, 2024
The paper introduces a Graph Attention Convolutional Network (GACN) for flow reconstruction from sparse data in time-varying geometries. It employs a feature propagation algorithm to handle missing data and a binary validity mask to enhance learning. Trained on Direct Numerical Simulations (DNS) of a motored engine, GACN demonstrates robustness across resolutions and domain sizes, outperforming CNNs and cubic interpolation in reconstructing fine-scale turbulence. It generalizes well to unseen DNS and experimental PIV data, successfully reconstructing flow fields from domains up to 14 times larger than those in training.