人工神经网络
基础(线性代数)
湍流
张量(固有定义)
高超音速
计算机科学
高超音速流动
航空航天工程
物理
机械
人工智能
数学
工程类
几何学
作者
Eric Parish,David S. Ching,Nathan Eli Miller,Steven J. Beresh,Matthew Barone,Niloy Gupta,Karthikeyan Duraisamy
出处
期刊:
[American Institute of Aeronautics and Astronautics]
日期:2024-01-04
摘要
We present a data-driven turbulence modeling approach for compressible flows. Our approach leverages Lipshitz-continuous tensor-basis neural networks and calibrated deep ensembles to model discrepancies in the anisotropy tensor and heat flux. We analyze tensor- basis neural networks for parallel flows and find that most higher-order tensors only impact the normal components of the Reynolds stress. We train a global data-driven turbulence model, which builds upon a nominal k-e model, and present results across a suite of low-speed and hypersonic test cases. The machine-learned model results in improvements (as compared to the baseline k-e model) across all cases. The use of Lipshitz-continuous networks and deep ensembles improves robustness of the data-driven model. For hypersonic boundary layers, the inclusion of the wall-normal Reynolds stress has an appreciable impact on the wall-normal momentum balance and wall quantities. For shock boundary layer interactions and compression ramps, the use of a variable Prandtl number reduces (but does not completely address) the over-prediction in wall heating at and downstream of the reattachment point. We examine rapid compression corrections and length-scale limiters to further improve predictions for the size of the separation bubble and wall heat flux. Numerical results show that rapid compression corrections have, in general, a positive impact on the prediction of the size of the separation bubble. Length-scale limiters result in more accurate predictions for wall heating at the expense of worse predictions for wall shear stress.
科研通智能强力驱动
Strongly Powered by AbleSci AI