Multifidelity aerodynamic flow field prediction using random forest-based machine learning
Multifidelity aerodynamic flow field prediction using random forest-based machine learning
复制标题
DOI：
10.1016/j.ast.2022.107449
复制
发表时间：
2022-04
影响因子：
5.6
通讯作者：
J. Nagawkar;Leifur Þ. Leifsson
中科院分区：
文献类型：
--
作者：
J. Nagawkar;Leifur Þ. Leifsson
本文提出了一种新的基于随机森林（RF）的多保真机器学习（ML）算法来预测高保真reynolds -average Navier-Stokes （RANS）流场。RF ML算法用于提高低保真度势流场的保真度。研究了三种情况，前两种情况包括经过后面向台阶的流动，第三种情况是围绕翼型的亚音速流动。在第一种情况下，数据生成使用十个不同的入口速度，在第二个使用六个不同的台阶高度，并在第三个使用20不同的翼型形状参数化使用b样条曲线。RF的输入参数与情况有关。对于第一种情况，使用x和y胞心位置以及相应的x和y势流速度，以及指定的入口速度。其次，使用阶跃高度对单元中心值进行无因次化，并使用阶跃高度代替入口速度。使用的其余两个输入特性与前一种情况相同。对于第三种情况，使用势流函数和速度势以及b样条控制点值作为输入变量。RF算法的输出对所有情况都是相同的，包括RANS速度、压力和湍流粘度。本研究的结果与使用tensorFlowFoam （TFF）和直接求解RANS方程生成的结果进行了比较。为了量化误差，使用了绝对误差和相对l2范数误差度量。结果表明，对于前两种情况，与TFF相比，RF的相对l2范数始终低2至30倍，唯一的例外是第二种情况的湍流粘度。对于第三种情况，射频是更好地预测压力和皮肤摩擦系数的RAE 2822翼型相比，NACA 0012翼型。压力系数和皮肤摩擦系数的相对l2范数误差分别为1.67和1.19倍。
In this paper, a novel random forest (RF)-based multifidelity machine learning (ML) algorithm to predict the high-fidelity Reynolds-averaged Navier-Stokes (RANS) flow field is proposed. The RF ML algorithm is used to increase the fidelity of a low-fidelity potential flow field. Three cases are studied, the first two consist of a flow past a backward-facing step, and the third, a subsonic flow around an airfoil. In the first case, the data is generated using ten different inlet velocities, in the second using six different step heights, and in the third using 20 different airfoil shapes parameterized using B-spline curves. Input parameters to RF are case dependent. For the first case, the x and y cell-center locations and the corresponding x and y potential flow velocities, along with the specified inlet velocity, are used. For the second, the cell-center values are nondimensionalized using the step height, and the step height is used in place of the inlet velocity. The remaining two input features used are the same as in the previous case. For the third case, the potential flow stream function and velocity potential along with the B-spline control point values are used as input variables. The outputs of the RF algorithm are the same for all the cases and include the RANS velocities, pressures, and turbulent viscosities. The results in this study are compared to those generated using the tensorFlowFoam (TFF) and from directly solving the RANS equations. To quantify the errors, the absolute error and relative L 2 norm error metrics are used. The results show that for the first two cases, RF consistently has two to 30 times lower relative L 2 norm compared to TFF, with the only exception being the turbulent viscosities for the second case. For the third case, RF is better at predicting the pressure and skin friction coefficients for the RAE 2822 airfoil compared to the NACA 0012 airfoil. The relative L 2 norm error is 1.67 and 1.19 times lower for the pressure and skin friction coefficients, respectively.