8.2.3.1. relativeTolerance
When running unsteady cases, the relativeTolerance
is typically set to 1e-2 or 1e-3. Once the nonlinear residuals drop by 2 or 3 orders of magnitude, the solver will continue to the next physical step. The relativeTolerance
is ignored for steady cases.
8.2.3.2. kappaMUSCL
The default value of -1 leads to a second-order upwind scheme, which is the most stable. A value of 0.33 leads to a blended upwind/central scheme, which is recommended for low subsonic flows to reduce dissipation.
8.2.3.3. orderOfAccuracy
The orderOfAccuracy
determines whether the solver will use 1st or 2nd order spatial discretization. The 1st order solver is faster, cheaper and most importantly, it is more dissipative, making it less likely to diverge. However, such numerical dissipation may also significantly impact the accuracy of the solution.
When initializing the flow field for unsteady cases with rotating components, such as simulating a rotor enclosed in a sliding interface, the user typically needs to run the 1st-order solver for around 1 or 2 revolutions. Once the flow field has been initialized, the user can fork the first-order case and switch orderOfAccuracy
from 1 to 2 for the child cases.
While adjusting the orderOfAccuracy
for the navierStokesSolver
, the turbulenceModelSolver should also be adjusted.
The recommended timeStepping
is slightly different for the 1st and 2nd order cases. For more details, see Rotational Angle per Step, maxPseudoSteps and CFL
8.2.3.4. Limiters
If the case is transonic or supersonic, the user should set limitVelocity
and limitPressureDensity
as TRUE
in the Navier Stokes solver parameters section of their input file.
8.2.3.5. linearIterations
linearIterations
controls the number of linear iterations in each pseudo-step. Typically, linearIterations
is set to 25~35 for the NS solver. The user might need to increase it to 50-55 for challenging cases to improve convergence.
8.2.3.6. updateJacobianFrequency
The default value for updateJacobianFrequency
is 4, which means that the Jacobian for evaluating the NS equation is updated every 4 pseudo-steps. For some challenging cases, reducing updateJacobianFrequency
from 4 to 1 may help, however, this may slow the NS solver by up to approximately 30%.