In spite of numerous efforts to build a unified Reynolds-averaged Navier–Stokes equation (RANS) model, which has lower computational costs in comparison with eddy-resolving methods (DNS, LES, hybrid LES-RANS techniques), the former is still not universal and accurate. For instance, suitable modification of the advanced RANS model based on physical considerations can improve the description of flows in plane channel with flat walls or backstep; however, deviations from the DNS data and measurements remain significant [1]. Finding a way to improve the baseline model predictability will be an advantage. Machine learning (ML) algorithms, using the available high-fidelity data sets for canonical flow cases obtained from DNS and measurements, can be helpful for the model enhancement [2]. In particular, the methods of neural networks, decision tree, random forest, and support vector machines can be used in such a study [3], and their suitable combination may further improve predictions of the Reynolds-stress anisotropy tensor components in canonical flows in comparison with those of [3]. These tools are applied for training and testing in turbulent flows in square duct [4, 5], plane channel [6], with periodic hills [7] and other cases. The performance of the enhanced ML-RANS models versus baseline models is discussed.

References

1. Yakovenko S.N., Chang K.C. Computational studies of near-wall behaviors of low-Reynolds-number Reynolds-stress models // AIAA J. 2019. Vol. 57. P. 279-296.

2. Duraisamy K., Iaccarino G., Xiao H. Turbulence modeling in the age of data // Annu. Rev. Fluid Mech. 2019. V. 51. P. 357-377.

3. Kaandorp M. Machine learning for data-driven RANS turbulence modelling: Master of Science Thesis. Delft University of Technology, 2018. [Electronic resource]. URL: tiny.cc/s6sejz (the date of access: 01.03.2018).

4. Huser A., Biringen S. Direct numerical simulation of turbulent of flow in a square duct // J. Fluid Mech. 1993. Vol. 257. P. 65-95.

5. Sata Y., Sato K., Kasagi N., Takamura N. Turbulent air flow measurement using three-dimensional particle tracking velocimetry // Trans. JSME, Ser. B. 1994. Vol. 60, No. 571. P. 865-871.

6. Lee M., Moser R.D. Direct numerical simulation of turbulent channel flow up to Reτ ≈ 5200 // J. Fluid Mech. 2015. V. 774. P. 395-415.

7. Breuer M., Peller N., Rapp C., Manhart M. Flow over periodic hills – numerical and experimental study in a wide range of Reynolds numbers // Comput. Fluids. 2009. Vol. 38, No. 2. P. 433-457.

To reports list