My Community

Hi everyone,

I just encountered an error problem "Error: RATTLE_vel algorithm did not converge!" when I tried to run NpT for a relatively large box (50 \AA^3), with constraints on the lattice angles in ICONST. It seems to be a general problem for large systems and that LANGEVIN_GAMMA, LANGEVIN_GAMMA_L and PMASS are relevant, but I have no idea about in which way I should vary them to solve the problem. These paramters in my INCAR are set to be:
LANGEVIN_GAMMA = 10 10 10
LANGEVIN_GAMMA_L = 10
PMASS = 1000
I did have tried to vary some of them, but the problem remains. Any clue on that?

Thank you very much and happy 2025!

Hi Renjie,

Thank you for your question. Could you attached the POSCAR, POTCAR, KPOINTS, INCAR, and OUTCAR and we can take a look at it. Do you encounter the same problems in a smaller system? Generally it is best to find the best parameters for a cheaper system before scaling to a larger one.

Happy new year!

Best wishes,

Chris

christopher_sheldon1 wrote: ↑Wed Jan 08, 2025 10:57 am

Hi Renjie,

Thank you for your question. Could you attached the POSCAR, POTCAR, KPOINTS, INCAR, and OUTCAR and we can take a look at it. Do you encounter the same problems in a smaller system? Generally it is best to find the best parameters for a cheaper system before scaling to a larger one.

Happy new year!

Best wishes,

Chris

Hi Chris,

Thank you very much for the reply. It happens in machine learning prediction running. Yes, I have also tested a small system. Same problem. I think I might have missed something I should set. The force field size is too large, so I just upload the POSCAR and INCAR here (KPOINTS and POTCAR should be ok). The purpose of the run is to obtain an amorphous state of the atoms from an initial simple cubic supercell (if I run nvt the simulation goes well).

Thank you.

christopher_sheldon1 wrote: ↑Wed Jan 08, 2025 10:57 am

Hi Renjie,

Thank you for your question. Could you attached the POSCAR, POTCAR, KPOINTS, INCAR, and OUTCAR and we can take a look at it. Do you encounter the same problems in a smaller system? Generally it is best to find the best parameters for a cheaper system before scaling to a larger one.

Happy new year!

Best wishes,

Chris

Sorry, I missed the ICONST file. The settings in it are

LA 1 2 0
LA 1 3 0
LA 2 3 0

Thank you.

Hi Renjie,

Thank you for the files. If the MLFF is too large, could you upload the OUTCAR file and I can try to work it out from there?

ICONST, POSCAR, and INCAR all look fine as far as I can tell. You could try modifying the parameters for the RATTLE algorithm SHAKETOL and/ or SHAKEMAXITER.

Best wishes,

Chris

christopher_sheldon1 wrote: ↑Mon Jan 13, 2025 11:16 am

Hi Renjie,

Thank you for the files. If the MLFF is too large, could you upload the OUTCAR file and I can try to work it out from there?

ICONST, POSCAR, and INCAR all look fine as far as I can tell. You could try modifying the parameters for the RATTLE algorithm SHAKETOL and/ or SHAKEMAXITER.

Best wishes,

Chris

Hi Chris,

Please ignore the previous uploaded inputs. In that INCAR NSW*ML_OUTBLOCK is too large but the reported error is RATTLE_vel not converged. So I was confused by the two issues.

I have done new tests on a series of supercell sizes: 151515, 202020, 303030, 404040, 505050. Only the last one reported RATTLE_vel algorithm issue. Hence it should be related to the supercell size. Please see my newly uploaded files.

Thank you.

Hi Renjie,

Could you upload the OUTCAR file for the 50x50x50 A3 cell? The NWRITE=3 tag prints additional information, which will be helpful in trying to locate the issue.

The only thing that I can suggest is to try modifying the parameters for the RATTLE algorithm: SHAKETOL and/ or SHAKEMAXITER and see how these affect the calculations. If the RATTLE algorithm is not converging, then it might be that the number of max iterations SHAKEMAXITER needs to be increased. Alternatively, you could change the convergence parameter SHAKETOL to say 1x10^-4 to see if the RATTLE algorithm then converges.

In the INCAR, try setting:

NWRITE = 3
SHAKEMAXITER = 2000
SHAKETOL = 5E-5

or SHAKETOL = 1E-4

Let me know if this helps.

Best wishes,

Chris

christopher_sheldon1 wrote: ↑Tue Jan 14, 2025 4:20 pm

Hi Renjie,

Could you upload the OUTCAR file for the 50x50x50 A3 cell? The NWRITE=3 tag prints additional information, which will be helpful in trying to locate the issue.

The only thing that I can suggest is to try modifying the parameters for the RATTLE algorithm: SHAKETOL and/ or SHAKEMAXITER and see how these affect the calculations. If the RATTLE algorithm is not converging, then it might be that the number of max iterations SHAKEMAXITER needs to be increased. Alternatively, you could change the convergence parameter SHAKETOL to say 1x10^-4 to see if the RATTLE algorithm then converges.

In the INCAR, try setting:

NWRITE = 3
SHAKEMAXITER = 2000
SHAKETOL = 5E-5

or SHAKETOL = 1E-4

Let me know if this helps.

Best wishes,

Chris

Hi Chris,

Thank you very much for the suggestion. Yes, by setting SHAKETOL = 1E-4 and SHAKEMAXITER = 2000 it works. FYI, I have also attached the OUTCAR with NWRITE=3 but without these two settings. By the way, is there any rule of thumb for the setting of tolerance? If it is too large, what will be influenced in the calculated results?

Best regards

Hi Renjie,

Good, I'm glad that it helps to converge it. The tolerance is related to the constrained coordinates. There's no rule of thumb. It's something that is worth testing to see how it affects the results and confirm that it is a valid approximation.

Unless you have any other questions, I'll lock the thread at the end of the week.

Best wishes,

Chris

christopher_sheldon1 wrote: ↑Thu Jan 16, 2025 9:32 am

Hi Renjie,

Good, I'm glad that it helps to converge it. The tolerance is related to the constrained coordinates. There's no rule of thumb. It's something that is worth testing to see how it affects the results and confirm that it is a valid approximation.

Unless you have any other questions, I'll lock the thread at the end of the week.

Best wishes,

Chris

Hi Chris. That's ok. Thank you.