ИСТИНА |
Войти в систему Регистрация |
|
Интеллектуальная Система Тематического Исследования НАукометрических данных |
||
Slope flows are flows on mountain slopes such as avalanches, mudflows, landslides, and others. These flows are turbulent multiphase flows of non-Newtonian media. Existing turbulence models describe this type of flow poorly and require refinement and calibration. The development of a turbulent model can be carried out in several ways, one of the most accurate of them is using direct numerical simulation (DNS). That is a very detailed eddy-resolving simulation carried out with the help of a supercomputer. To develop turbulent models, it is required to process the results of DNS modeling, which are a large amount of data containing such information about the flow as velocity, pressure, density, viscosity and other flow parameters in the entire set of points of the computational domain at all times. Tensor basis neural networks (TBNN), which are considered in this work, are best suited for processing data arrays and constructing new dependencies when describing a turbulent model. In this work, a two-stage calibration of the turbulent model is carried out. First of all, the optimization of the coefficients of the existing turbulent model is carried out. Next, an expression is constructed for the anisotropic normalized Reynolds stress tensor. Calibration of the turbulent model is planned using the following optimization algorithm based on reinforcement learning: 1. Training the neural network based on a number of calculations carried out using the RANS model with the k - ε turbulence model with different values of the constants; 2. Obtaining new values of the coefficients of the turbulent model using machine learning; 3. Calculation of flow hydrodynamics using a turbulent model with coefficients obtained using machine learning; 4. Additional training of the algorithm using the obtained data from the calculation of flow hydrodynamics. To obtain an expression for the anisotropic normalized Reynolds stress tensor, it is proposed to use a special neural network architecture based on a tensor basis. The input data of the neural network is based on 10 basic isotropic tensors Ti and their invariants λi, which are functions of the dimensionless Reynolds-averaged strain rate tensor s and the dimensionless Reynolds-averaged rotational velocity tensor r, these are all sorts of linearly independent combinations of s and r. At the output of the neural network, the corrected tensor of the normalized Reynolds stress tensor b is obtained, represented as a dependence on the tensors Ti and their invariants λi. As a result of the study, a calibrated turbulent model was obtained that is suitable for describing multiphase flows of non-Newtonian fluid on slopes. Using the obtained turbulence model, a simulation of the experiment of the University of Iceland with the descent of a flow in a flume with a complex of protective structures was carried out; a decrease in the discrepancy in the value of the volume of the flow retained by the barrier structures in comparison with the experiment was obtained (a more accurate assessment of the effectiveness of the complex of protective structures was obtained). Modeling of the 22nd avalanche center on the Yukspor mountain of the Khibiny mountains was carried out using a calibrated turbulence model. The turbulent model obtained using TBNN made it possible to increase the accuracy of assessing the avalanche-hazardous zone, the effectiveness of protective structures, and obtain more accurate flow characteristics.