Новая алгебраическая RANS-модель для пристенного подслоя гибридного RANS/LES-метода на основе методов машинного обучения

Today, there is a special focus on the development of Hybrid RANS/LES methods, in which the RANS model, describing the near-wall region of the flow, is one of their basic components. Since this region accounts for approximately 10% of the boundary layer thickness, one would expect that relatively simple algebraic closures can perform well in it. However, they are either non-local or implicit. The goal of this paper is to create a new explicit local algebraic RANS model for the near-wall sublayer, which will be used as a part of a hybrid RANS/LES method. The development of the model would involve the machine learning methods, which have found wide application in turbulence modeling. This procedure consists of two stages. First, by performing a priori testing based on the DNS data on the developed turbulent flow in a plane channel at the friction Reynolds numbers from 550 to 5200, the functional form of the new model is constructed using genetic programming. Then, the optimization of coefficients is carried out by applying a model-consistent learning approach that is based on the solution of a turbulent flow in a plane channel. The results calculated by the new model are compared with those obtained with the use of the classical Prandtl and Spalart-Allmaras models. It is shown that the new model outperforms them in accuracy. The new model was tested at a friction Reynolds number of 20000. It has been found that with this model it is possible to construct mean velocity profiles that agree well with the Nickels approximation and are close to those calculated by other models. Thus, this paper demonstrates that a combination of a priori and model-consistent machine learning approaches has been rather successful in creating an original algebraic RANS turbulence model.