code_saturne User's Forum

Hello,

I tried to run the Volker & Prakash's benchmark using the cdo-based solidification modelling introduced in https://www.code-saturne.org/documentat ... ation.html with settings based on the article accessible here https://hal.science/hal-04028238. However, it seems that the GCR algorithm to solve the saddle-point problem leads to a convergence issue and does not lead to physical results in a closed cavity and with the "velocity_divergence" field which however seems to be around 1e-6 (higher velocities in the gravity direction as issued in another discussion in the forum related to solidification).

Moreover by setting the solidus and liquidus temperatures below the initial and the fixed ones at the left / right boundaries, the classical result of the heated cavity benchmark does seem to be retrieved using then the CDO approach instead of the legacy FV one to solve the NS equation for momentum.

Thank you in advance for your attention.

Kind regards,

SS2X.

Hello,

The colleague who could best answer this is on holiday, but I can remind him to check your post when he returns.

In any cases, from conversations withe the colleagues working on solidification, I can confirm robust solvers are needed, and a big part of the work making these studies is related to optimization of solver settings. This is one of the applications where a robust direct solver like MUMPS may be useful, or even required.

Also, as far as I know, the recent work on solidification has been based mostly on the CDO approach, so I would definitely trust that model more than the legacy one, which is more of an early iteration.

Best regards,

Yvan

Hello Yvan,

It works indeed better using the MUMPS solver and with much better convergence of the velocity field.

Thanks !

Kind regards.

Hello SS2X and Yvan,

As mentioned by Yvan, the efficiency of purely iterative solvers for coupled velocity/pressure systems is an issue. We are working on this topic and we hope having something better next year. There are many ideas to explore.

Up to now, the best alternative is to consider a hybrid solver (iterative for the outer iterations and direct for the inner iterations, e.g. MUMPS) such as the Augmented Lagrangian Uzawa (ALU) approach. Nearly all computations for solidification processes are performed using this solver.

Nevertheless, the depicted behavior may hide a bug. On a "simple" solidification case as the Voller Prakash benchmark, a full iterative solver should deliver the good results.

Could you please send me your setup.log, run_solver.log and performance.log files to check and analyze what's going wrong in your computation ?

Best regards,
Jérôme.

Hello Yvan and Jérôme,

Thank your for your respective detailed feedback . Actually it works well with the ALU technique using the MUMPS solver . Probably it could be still optimized regarding the calculation performance (adequate parallelization combined to advanced MUMPs options to speed up the calculation).

Kind regards,

SS2X.