Help needed with solving large nonlinear RVE model

[ws18]

Dear all,

I am trying to solve a large RVE model with over 500,000 nodes (1.5 million DOFs) to obtain its nonlinear homogenized properties. Previously, I only focused on elastic properties, and FEAP (85) solved the system using the ITERATION solver (BPGS) without issues.

However, after switching to an elasto-plastic-damage model (via UMAT) for the matrix to study the nonlinear properties, the system could no longer be solved. I suspect this is due to the system becoming unsymmetric, and that solving it may require the DIRECT method. Unfortunately, both my Mac and Windows desktops lack the storage capacity for this approach.

Could you advise on how to successfully solve this system? Is there a way to increase FEAP's storage capacity, or an alternative solution method I could try? Thank you for your time.

[Prof. S. Govindjee]

If the equations are unsymmetric, you could try the BROYden method.   I have not used it in a long time, so it may need some updating.  But it is worth a try, perhaps first on a modest size test problem.

[Prof. R.L. Taylor]

FEAP also has an out of core blocked solver that can handle unstymmetric equations.  It is activated by the command

DIREct BLOCked

You can try this as it may be able to solve the problem.  Be sure to optimize the profile using OPTI if necessary.

Let us know if this works for you.

[Prof. R.L. Taylor]

Note that using the out of core solver will be very time consuming.  This is caused by the extensive I/O performed and the fact it is not coded for super-scaler speed as the in-core solver has.

It may work but I would suggest using only displacement b.v. on the RVE to keep the profile as small as possible. 

[Prof. R.L. Taylor]

Forget trying direct block option, it takes far too much time and disk space to be useful to you. 

[ws18]

Thank you all for your replies.

I apologize for misspelling "BPCG" as "BPGS". I think that BROYden is a type of quasi-Newton method, rather than an iterative solution scheme that can help me address the insufficient memory issue.

I have also tried the DIREct, BLOCked schemes, but as you mentioned, they can quickly consume a significant amount of disk space, making these options impractical for my use case.

 

[Prof. R.L. Taylor]

Have you tried using one of the sparse solvers like superlu?

[ws18]

Dear Prof. Taylor,

I also wish I can try SuperLU, UMFPACK equation solvers as the manual mentioned them. But I don't know how to enable these interfaces.

For SuperLU, I downloaded the source and created an archive of the routines on my Mac (Sonoma 14.5). I’m not entirely sure if I did it correctly (see attached 1).

I’ve also downloaded the FEAP interface routines for SuperLU from your website (attached as file 2). Could you please advise on how to compile and link these with the main FEAP archive and the SuperLU archive? Does this require modifying the makefile.in file and reinstalling FEAP? If so, I would greatly appreciate detailed instructions on any necessary modifications to the makefile.in or other installation files to enable the SuperLU interface. Additionally, if the FEAP interface routines ('superlufeap' folder in attachment 2) need to be placed in a specific directory, please let me know.

Thank you for your help!

[JStorm]

You will find informations on the interface for external solvers in http://projects.ce.berkeley.edu/feap/