Convergence problem in further steps of ab initio molecular dynamics

[Mohammad Shakiba]

Dear CP2K users,

I have installed CP2K on our university cluster with intel 2019 and it works properly for tests. However, when I use it for ab initio molecular dynamics with my structure, the first 20 steps converge but the convergence in the next step takes too long and it does not converge. Actually, it becomes close to the convergence limit of 1.0E-6 (like 2.04E-6 which was the closest) but does not converge even in outer loops and oscillates between 2.0E-6 and 1.0E-4.

Can you help me find out why this happens and how can I fix it? is it because of my basis sets and potentials? or my structure? Is it practical to increase the MAX_SCF to a high number like 1000? What are the normal ranges for this limit?

By the way I tried both CG and DIIS but the problem still exists.

Thanks in advance.

My structure contains Pb ans S atoms -> (PbS)16. Here are some main parts of my attached inputs:

    BASIS_SET_FILE_NAME BASIS_MOLOPT   ! I used MOLOPT bases file
    POTENTIAL_FILE_NAME POTENTIAL          ! with GTH pseudopotentials and PBE functionals

    &MGRID
       CUTOFF [Ry] 400
    &END
    &QS
       METHOD GPW
       EPS_DEFAULT 1.0E-10
       EXTRAPOLATION ASPC
    &END
    &POISSON     
      PERIODIC NONE
      PSOLVER  MT
    &END POISSON
    &SCF                             
      SCF_GUESS ATOMIC
      MAX_SCF 400
      EPS_SCF 1.0E-6
      &OT
        PRECONDITIONER FULL_SINGLE_INVERSE
        MINIMIZER DIIS                                            ! I also tried CG
        LINESEARCH 3PNT
!        STEPSIZE 0.08
!        PRECOND_SOLVER INVERSE_CHOLESKY
      &END OT
      &OUTER_SCF                                                 ! Outer SCF loop for 10 times
        MAX_SCF 10
        EPS_SCF 1.0E-6
      &END

    &XC
      &XC_GRID
        XC_DERIV SPLINE2
        XC_SMOOTH_RHO NN50
      &END XC_GRID
      &XC_FUNCTIONAL
         &PBE
         &END
      &END XC_FUNCTIONAL
    &END XC

[Matt W]

Hi,

hybrid functional calculations using MOLOPT basis sets are extremely expensive. You probably need to use the Auxiliary Density Matrix Method approximation. See these execises and some slides maybe this talk.

Matt

[Mohammad Shakiba]

Hi,

Dear Matt,

Your video and also the videos inside the playlist were very useful and I learned so much and thank you so much for that. Although I modified my input but I have to recompile my cp2k version with libint and elpa as well.

What I noticed in the videos was that you used SCF_GUESS RESTART and WFN_RESTART_FILE_NAME ${project}-RESTART.wfn. Now I have some other question in my mind:

** Does using a RESTART file make the calculations and convergence faster and better? How can I obtain such a file before starting AIMD because I found no such file but I have, for example, *.restart.bak.1 file (after running calculations) but not *.wfn file? and I don't know if I can use it in the EX_RESTART section and obtain good convergence.

** As I saw the "RESTART", I remember that in some papers they have equilibriated the system for some pico-seconds with time step of 10fs and then continued the calculations with smaller time steps like 2fs. To this end I guess that I have to make two inputs: one with 10fs time step and the other with a EXT_RESTART file from the previous and 2fs time step. But I don't know what to put for SCF_GUESS for the second input? ATOMIC?