.png){kind=icon}
.png){kind=icon}
Return
Developing an advanced neural network and physics solver coupled framework for accelerating flow field simulations
.png){kind=icon}
PRE
AI
DOI：10.1007/s00366-023-01861-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Computational fluid dynamics simulation accounts for a large number of workloads in the numerical design optimization of aerodynamics problems. In this paper, we develop AFFNet, an advanced neural network and physics solver coupled framework for accelerating flow field simulations. AFFNet combines the benefits of an attention mechanism, affine transformation, and encoder-decoder neural network modules to learn a solution-related mapping from point-based input features. We evaluate AFFNet in the context of transonic and hypersonic turbulent flows based on structured and unstructured meshes. The results show that the framework is both fast and accurate. AFFNet is able to satisfy physical convergence constraints while providing significant speedups over physical solvers and other widely used neural network designs. Moreover, AFFNet can capture the effect of small changes in 2D/3D flow fields and preserve the smoothness of curved surfaces. To optimize the network design, an architecture search is exploited to investigate the contribution of each network module. The obtained framework and empirical suggestions can serve as an instructive reference source with respect to mesh data preprocessing, network design, and training methodology.
Keywords:
Computational fluid dynamics (CFD)
Neural network
Design optimization
Architecture search
Journal
IF:
4.9
Papers:
2.6K
Citations:
9.3K