Graph attention network-based fluid simulation model

Traditional computational fluid dynamics (CFD) techniques deduce the dynamic variations in flow fields by using finite elements or differences to solve partial differential equations. CFD usually involves several tens of thousands grid nodes, which entail long computation times and significant resources. Fluid data are irregular data, there will be turbulence field where physical quantities between adjacent nodes extremely nonequilibrium. We use a graph attention neural network build simulation model (GAFM). GAFM assigns weights node-pairs through mechanism. In this way, it is not only possible directly calculate but also adjust for nonequilibrium vortices, especially turbulent flows. The deductively predicts spatiotemporally continuous sample data. A validation proposed against two-dimensional (2D) around cylinder confirms its high prediction accuracy. addition, achieves faster speeds than traditional solvers two three orders magnitude. provides new idea rapid optimization design mechanics models real-time control intelligent mechanisms.