||May 14, 2010 12:39
ANSYS technical help responded with below when I asked for general comments about ensuring the accuracy of a solution:
You ask a lot of good questions! Basically errors in CFD can be categorized as follows: 1) Geometric Representation - make sure relevant features in the geometry are included - ensure that your inlets/outlet are not too close to the area of interest 2) Specificiation of Boundary Conditions - ensure the boundary conditions are correct (for example if you have a pressure profile, but assume an average pressure at a boundary it will affect your results) - ensure that the boundary conditions are well posed (for example for an incompressible case, don't specify a mass flow rate in and a different mass flow rate out, or the solver will be sure to get the wrong answer) 3) Discretisation Errors - ensure that your mesh is fine enough to minimze these errors. This is easier said than done. You may want to perform a mesh sensitivity study to check this. For example, run the same geometry with a mesh size of "N" nodes, "2*N" nodes, "4*N" nodes, and "8*N" nodes, where "N" is some number of nodes that you think is reasonable. Check your solution - at some point you will find that the changes in solutions when you double the number of nodes in the solution will become negligible. - ensure you accurate resolve areas of high gradients (for example the boundary layer around walls) - ensure the timestep you have selected is suitable. - ensure that you use higher order numerics where possible. Most of the numerics are bounded second order in CFX (ex. High Resolution Advection Scheme), so you probably don't need to worry about this too much. 4) Numerical/Roundoff Errors - ensure the case is well converged to reduce these errors - a well converged case will have: * low residual values - RMS residuals machine roundoff for single precision is roughly 1e-6, and for double precision it is 1e-8. These should not be able to get residuals lower than these values, but you might not even need to go this low. You could do a residual sensitivity study - run the case to 1e-4, 1e-5, 1e-6 and see how the results change. Maybe 1e-4 is accurate enough. I typically use 1e-5 as a starting point. * small global imbalances - generally speaking, if residuals are small, global imbalances should be small as well, but it's not always the case. I like to ensure global imbalance are less than 0.1%. * solution variables no longer change with each iteration - typically you should monitor values in the solver manager that you are interested in knowing. Maybe for your case it is average outlet temperature, or maximum heat transfer coefficient on the fin surface. Whatever you thin is important, add a monitor point for it, and ensure it has reached a fairly consistent value. 5) Modelling Errors - make sure that you are modelling all relevant physics (i.e. is the correct turbulence model being used, is radiation important, are the walls being modelled as being smooth when they are rough in reality) Those are the main things to consider - the CFX help documentation will go into more details about these topics if you are interested.