Published Paper Details:

Traditional computational fluid dynamics (CFD) relies on solving partial differential equations to compute flow field characteristics. However, this approach is computationally intensive and time-consuming. To address these limitations, we propose a fluid simulation system based on a graph neural network (GNN) framework. Our simulator offers high computational efficiency while significantly reducing resource consumption. By representing the computational domain as a structured graph, the model identifies neighboring nodes through adaptive sampling. We leverage deep learning techniques, incorporating attention-based graph neural networks, to enhance predictive accuracy. The simulator is trained on flow field data around a cylinder with varying Reynolds numbers. Once trained, it not only achieves high accuracy within the training set but also generalizes well to unseen flow conditions. Compared to conventional CFD solvers, our approach accelerates computations by 2-3 orders of magnitude. This advancement paves the way for faster optimization and design of fluid mechanics models, as well as real-time control of intelligent fluid systems.