Optimization of Turbulence Model Parameters Using the Global Search Method Combined with Machine Learning

The paper considers the slope flow simulation and the problem of finding the optimal parameter values of this mathematical model. The slope flow is modeled using the finite volume method applied to the Reynolds-averaged Navier-Stokes equations with closure in the form of the k-omega SST turbulence model. The optimal values of the turbulence model coefficients for free surface gravity multiphase flows were found using the global search algorithm. Calibration was performed to increase the similarity of the experimental and calculated velocity profiles. The Root Mean Square Error (RMSE) of derivation between the calculated flow velocity profile and the experimental one is considered as the objective function in the optimization problem. The calibration of the turbulence model coefficients for calculating the free surface flows on test slopes using the multiphase model for interphase tracking has not been performed previously. To solve the multi-extremal optimization problem arising from the search for the minimum of the loss function for the flow velocity profile, we apply a new optimization approach using a Peano curve to reduce the dimensionality of the problem. To speed up the optimization procedure, the objective function was approximated using an artificial neural network. Thus, an interdisciplinary approach was applied which allowed the optimal values of six turbulence model parameters to be found using OpenFOAM and Globalizer software.

Automated summary

This paper considers the simulation of slope flow and the optimization of parameter values in the mathematical model. The finite volume method is used to model the slope flow, applying the Reynolds-averaged Navier-Stokes equations and the k-omega SST turbulence model with closure. The global search algorithm is used to find the optimal values of turbulence model coefficients for free surface gravity multiphase flows, and calibration is performed to increase the similarity between experimental and calculated velocity profiles.