Implementation and Validation of a Computationally Efficient DNS Solver for Reacting Flows in OpenFOAM

Abstract

To meet future climate goals, the efficiency of combustion devices has to be increased. This requires a better understanding of the underlying physics. The simulation of turbulent flames is a challenge because of the multi-scale nature of combustion processes: relevant length scales span over five orders of magnitude and time scales over more than ten. Because of this, the direct numerical simulation (DNS) of turbulent flames is only possible on large supercomputers. A DNS solver for chemically reacting flows implemented in the open-source framework OpenFOAM is presented. The thermo-chemical library Cantera is used to compute detailed transport coefficients based on kinetic gas theory. The multi-scale nature of time scales, which are much lower for the combustion chemistry than for the flow, are bridged by an operator splitting approach, which employs the open-source solver Sundials to integrate chemical reaction rates. Because the direct simulation of turbulent flames has to be performed on supercomputers, special care has been taken to improve the computational performance. A tool was developed which generates highly optimized C++ source code for the computation of chemical reaction rates. Additionally, a load balancing approach specifically made for the computation of chemical reaction rates is employed. In total, these optimizations can reduce total simulation times by up to 70 %. The accuracy of the new solver is assessed from different canonical testcases: 2D and 3D Taylor-Green vortex simulations show that the solver can reach up to fourth order convergence rates and that results differ by less than 1 % when compared to spectral DNS codes. Molecular diffusion and chemical reaction rates are compared to solutions of 1D flames from Cantera, showing perfect agreement. The solver is used to simulate the Sydney/Sandia burner. The simulation is performed on one of Germany's largest supercomputer on 28 800 CPU cores, employing 150 million cells and a chemical reaction mechanism with 19 species and about 200 reactions. Comparison with experimental data shows excellent agreement for time averaged and RMS values.

Full document