Central AsianResearch Index

STUDY OF NUMERICAL MODELS OF TURBULENCE, BASED ON MACHINE LEARNING METHODS FOR SCALAR FLUX IN HEAT-RELEASED ASSEMBLIES

In this paper, an analysis of turbulence models created using machine learning methods is performed for a wide range of fluid flow parameters at low Prandtl numbers in a channel with complex shape. Three different modern machine learning methods are used for the analysis. A developed turbulent flow in the peripheral channel fragment with three round longitudinal rods is considered for Reynolds number Re = 4270 at low Prandtl numbers Pr = 0.025, 0.05, 0.1, 0.2, 0.4, 0.6, 0.71 encountered in heat-released elements where liquid metals and gas mixtures are used as coolant. The importance of input features of turbulent scalar flux models is analyzed for the entire data set using the permutation feature importance assessment and a method based on Shapley values. The importance of using the local Reynolds number invariant for accurate prediction of the transverse turbulent scalar flux components is demonstrated. It is found that the mutual correlation of some invariants within groups of the selected 15 input features generally preserves the significance level of the most important invariants within each of the groups for the universally interpretable machine learning model for the turbulent scalar flux modeling. Using the permutation method, it was found that the most important basis tensor for estimating the transverse components of the turbulent scalar flux in the tensor basis neural network model is the tensor obtained by the inner product of the deformation tensor with itself S 2 , and the least important is the tensor SR + RS.

Сведения

Опубликовано
2025 · Т. 13, № 4, С. 179-192
Язык
Английский
ISSN
2306-6172 · 2308-9822
Издатель
L. N. Gumilyov Eurasian National University