Numerical simulation of aerodynamic and noise characteristics of subsonic turbulent jets using graphic processing units

The work considers two turbulent jet problems. The first includes the evaluation of the aerodynamic characteristics of the jet, which emits from the model subsonic nozzle investigated in the framework of the European project JEAN (Jet Exhaust Aerodynamics and Noise) for the Mach number M = 0.75 and the Reynolds number Re = 1⋅106. The second problem focuses on the simulation of the noise emission from the model conical nozzle (М = 0.9; Re = 1.6⋅106). These problems were solved with use of parallel CFD solver GHOST CFD, which is being developed by authors. The simulations were conducted using the Navier-Stokes equations for perfect gas in a curvilinear coordinate system by a finite difference method. A sponge layer was utilized to minimize the reflections from the outer border of computational domain. Spatial derivatives were computed by the 4th order of approximation scheme with improved dissipative and dispersive characteristics (DRP schemes - Dispersion Relation Preserving). Turbulence modeling was conducted by large eddy simulation with relaxation filtering. Time integration was conducted by the 4th order Runge-Kutta scheme (LDDRK - Low Dispersion and Dissipation Runge-Kutta), which had also improved dissipative and dispersive characteristics. A computational mesh for both problems consisted of 12 million cells. Simulations were performed with graphic processing units, whose performance surpasses greatly (by an order of magnitude) that of multicore central processing units, granting more than 10 times decrease in the required computational time. Results for the mean velocity for the JEAN nozzle have shown good agreement with experimental data. The RMS velocity along the nozzle centerline was moderately underpredicted, which, however, coincided with the results of other authors. The conical nozzle sound pressure level obtained in current study was compared with the experimental data, as well as with the results of simulations on a finer grid with ANSYS Fluent commercial solver. The comparison has shown a good agreement of GHOST CFD results with experimental data within 3-4 dB for a wide range of frequencies. Moreover, the GHOST CFD results were consistent with the ANSYS Fluent results on a finer grid, but their evaluation required much less computational time.