ИСТИНА

In several areas in computational fluid dynamics (CFD), there is a need to solve differential equations of elliptic type. After discretization on a computational grid, the problem is reduced to solving a system of linear algebraic equations (SLAE). The numerical methods widely used for high-fidelity simulations of incompressible turbulent flows require solving a sequence of SLAEs with a constant matrix and changing the right-hand side. A practically important issue is the choice of the parameters of linear solvers, which are usually set by default because they can have a tangible impact on the SLAE solution time. The paper presents an algorithm for automatic parameters selection for SLAE solving methods. The proposed algorithm finds appropriate parameters for the specified configuration of numerical methods. An approach is based on a genetic algorithm in conjunction with a neural network model. The last one is trained to predict the SLAE solution time with specific parameters. Thus the neural network model acts as a source of knowledge about the influence of each parameter on the linear solver performance, which increases the efficiency of finding their optimal set.