[zhanghuicheng@nnu.edu.cn]

I am using ANSYS CFX to simulate steam flow with non-equilibrium condensation inside a steam turbine.

When I run the solver with double precision and 60 parallel cores, the simulation fails (either diverges or crashes).
However, with single precision and only 30 cores, the simulation completes without any problems.

Why would enabling double precision cause the simulation to fail?Any insights would be greatly appreciated.

[Opaque]

Originally Posted by zhanghuicheng@nnu.edu.cn

I am using ANSYS CFX to simulate steam flow with non-equilibrium condensation inside a steam turbine.

When I run the solver with double precision and 60 parallel cores, the simulation fails (either diverges or crashes).
However, with single precision and only 30 cores, the simulation completes without any problems.

Why would enabling double precision cause the simulation to fail?Any insights would be greatly appreciated.

you are comparing apples to oranges. To conclude anything you must run the EXACT same number of partitions with either precision. No mix and match.

In your evaluation, you changed two variables at the same time: precision on numbers (like droplet size, nucleation rates, etc), and domain decomposition coupling of the linear equations.

[zhanghuicheng@nnu.edu.cn]

Originally Posted by Opaque

you are comparing apples to oranges. To conclude anything you must run the EXACT same number of partitions with either precision. No mix and match.

In your evaluation, you changed two variables at the same time: precision on numbers (like droplet size, nucleation rates, etc), and domain decomposition coupling of the linear equations.

I sincerely appreciate your advice. Following your suggestion, I ran the simulation using the same number of partitions for both cases. However, the double-precision solver still failed to proceed, whereas the single-precision solver successfully completed the run.

[ghorrocks]

I suspect you will find this result is more due to luck than a firm reason. Your simulation is obviously borderline numerically unstable, so whether it goes stable or not is dependent on a tiny factor. In this case it could be that the single precision simulation cannot resolve a feature (like a small separation) which is just too much for the solver to handle and it tips the simulation over the edge and it fails.

I do not think it worth much effort to try to find a reason for this, as it is likely to be more luck than anything. The conclusion I would draw from this is that your simulation is very unstable and anything you do to improve that stability will help you immensely - you simulation will be faster, more accurate and not crash every now and again.

So I would improve mesh quality, check your time step size, check your convergence criteria and use a better initial guess.

[zhanghuicheng@nnu.edu.cn]

Originally Posted by ghorrocks

I suspect you will find this result is more due to luck than a firm reason. Your simulation is obviously borderline numerically unstable, so whether it goes stable or not is dependent on a tiny factor. In this case it could be that the single precision simulation cannot resolve a feature (like a small separation) which is just too much for the solver to handle and it tips the simulation over the edge and it fails.

I do not think it worth much effort to try to find a reason for this, as it is likely to be more luck than anything. The conclusion I would draw from this is that your simulation is very unstable and anything you do to improve that stability will help you immensely - you simulation will be faster, more accurate and not crash every now and again.

So I would improve mesh quality, check your time step size, check your convergence criteria and use a better initial guess.

Thank you very much for your insights and suggestions. I really appreciate your time and explanation. I will definitely look into improving the mesh quality, time step size, and initial conditions to enhance the stability of the simulation.