Learning Forum

[Jules Heaulme]

Hi all,

 

I am facing some issues on a project on Ansys Fluent.

My project consist in studying the capillary effect of water in air between two parallel plates 50 microns apart, first in 2D then in 3D. To start my project, I worked on two parallel plates 1mm apart, parametrized the surface tension, the contact angle between water and my solid walls and ran a transient simulation of 1s of flow time with a time step size of 1e-3s (above it was not converging). With these parameters, simulations converge quickly (15 to 30 minutes function of the contact angle), and the results were very close to the theory.

Now that I am working with the real distance of 50 microns between plates, computation is taking way more time. Firstly, because the number of elements is increasing if I want to keep an element aspect ratio close to 1 with the same plate height. Secondly, because I considerably have to reduce the time step size if I want to hope for convergence, so I need way more iterations for the same flow time. For my 2D model I have to go down around 1e-5s and for my 3D model until 1e-7s.

3D simulations can take a day to compute for 0.01s of flow time (I need between 0.5s and 1s I think) and it seems that the time step size is not small enough since the simulation does not converge all the time.

 

So, my question is the following: do you think that there is a parametrization issue in my model which could explain this very long computation time or is it normal that the model is computationally expensive? We identified 3 potential factors of complexity in this model: multiphase flow, a 1000 density factor between the two phases and a tiny elements size (smallest elements side measure around 5 microns). Especially, the element size makes us think that there is potentially a minimum time step size that we cannot exceed because of the CFL number, is that right or there may be a way to override this?

 

Thank you bery much by advance for your time,

 

Jules

[Rob]

You also need to factor in the force from the surface contact at that scale. The surface tension normally has a local effect on a bulk flow. At this scale the surface tension dominates everything, and that will lead to a stiff solution needing a small time step. You may also find that parallel is less efficient as the free surface may only be present in very few of the partitions: load balancing will help. 

The time step is going to need to be small to get an accurate result. 

[Jules Heaulme]

 

Thank you very much for your answer Rob! Can you clarify what you mean by “load balancing will help”? Which load are you referring to?

 

[Rob]

CPU load. With single phase models the parallel balancing is much easier as we just have to consider the flow equations. With VOF we have the same equations plus have to solve for the free surface which tends to be more localised. That then tends to load a few cpu cores more than the rest, but we can alter the balance rules to account for that. 

[Jules Heaulme]

Oh ok I see, thank you very much!