problems with SCF energy convergence in COSMO

[e.berardo]

Hi all,
I have been running some COSMO geometry optimizations in DFT for two clusters of transition metal oxides.
For the bigger structure (~100 atoms) I didn't experience any problem, as the gradient decreases the SCF energy goes towards convergence.
For the smaller structure (~24 atoms) I am observing something really strange. After many steps the gradients are very small and the geometry is not changing anymore, but the SCF energy is jumping up and down, with no sign of convergence.

Code: [Select]

      ******************************************************************
                          CONVERGENCE INFORMATION

                               Converged?     Value      Criterion
             Energy change         no       1.2181215   0.0000010
             RMS of displacement   yes      0.0000004   0.0005000
             RMS of gradient       yes      0.0000000   0.0005000
             MAX displacement      yes      0.0000027   0.0010000
             MAX gradient          yes      0.0000000   0.0001000
      ******************************************************************



Code: [Select]


  cycle =    245    SCF energy =    -4409.7166469090   |dE/dxyz| =  0.000088
  cycle =    246    SCF energy =    -4405.0600362230   |dE/dxyz| =  0.000088
  cycle =    247    SCF energy =    -4410.9270409740   |dE/dxyz| =  0.000087
  cycle =    248    SCF energy =    -4416.0999436860   |dE/dxyz| =  0.000087
  cycle =    249    SCF energy =    -4405.0069743620   |dE/dxyz| =  0.000087
  cycle =    250    SCF energy =    -4418.2815994030   |dE/dxyz| =  0.000087
  cycle =    251    SCF energy =    -4414.5429156850   |dE/dxyz| =  0.000087
  cycle =    252    SCF energy =    -4415.8180561870   |dE/dxyz| =  0.000087
  cycle =    253    SCF energy =    -4419.6058398660   |dE/dxyz| =  0.000087

Just for the sake of argument I checked the OC corrected energy as well, but I see the same behavior:

Code: [Select]

job.94:    Total energy + OC corr. =    -4415.7873333840
job.95:    Total energy + OC corr. =    -4400.1291837066
job.96:    Total energy + OC corr. =    -4419.1704658379
job.97:    Total energy + OC corr. =    -4400.2408475728
job.98:    Total energy + OC corr. =    -4415.5401379877
job.99:    Total energy + OC corr. =    -4403.7589587977
job.last:    Total energy + OC corr. =    -4415.7982812745

The setup for the two calculations was the same, both jobs are parallel (8 processors), but running on different machines (both linux64). Version of turbomole 6.4
I am using the default values for COSMO, with the only exception of the epsilon value (80) and the element radii for O and H are the optimized ones, while for Ti the default one.

Code: [Select]

==============================================================================
                                  COSMO RESULTS
 ==============================================================================
  PARAMETER:
    nppa:      1082
    nspa:        92
    nsph:        32
    nps:        957
    npspher:    436
    disex:   10.0000   
    disex2:  3918.79   
    rsolv [A]: 1.3000
    routf:     0.8500
    phsran:    0.0   
    ampran:   0.10E-04
    cavity:  closed
    epsilon:    80.000
    refind:     1.300
    fepsi:      0.9813665
  CAVITY VOLUME/AREA [a.u.]:
    surface: V1.0, A matrix: V1.0
    area:   916.62
    volume:  1874.76
  SCREENING CHARGE:
    cosmo      :  -0.035792
    correction :   0.033567
    total      :  -0.002226
  ENERGIES [a.u.]:
    Total energy            =    -4414.5767535569
    Total energy + OC corr. =    -4414.5801598139
    Dielectric energy       =       -0.0901258038
    Diel. energy + OC corr. =       -0.0935320608
  ELEMENT RADIUS [A]: ATOM LIST
    h  1.30: 1,2,4,5,19,21,22,24
    o  1.72: 3,6,11-18,20,23
    ti 2.22: 7-10
 ==============================================================================


Has somebody previously experienced this behaviour? Or is there something wrong in my setup?

Thanks in advance,
Enrico

[uwe]

Hi Enrico,

this does indeed look very strange. If the energy changes that much but the gradients should do not, the coordinates would have to change a lot.

Could you try to run the same input in serial mode rather than parallel?

And, please, send the input and the output to the Turbomole support by email (turbomole at cosmologic.de), because it has to be checked where this behavior comes from.

Regards,

Uwe

[e.berardo]

I am running the serial job at the moment, and I'll let you know what's my progress with that, anyway I also contacted the support team.

Cheers,
Enrico

[mario8eren]

I am running the serial job at the moment, and I'll let you know what's my progress with that, anyway I also contacted the support team.

Have you figured out the solution to this problem? I am having similar issues.

Mario

[uwe]

Hi,

if you are still using Turbomole version 6.4 and run COSMO MPI parallel jobs, make sure to update to version 6.4.1 or newer.  See: http://www.turbo-forum.com/index.php/topic,812.0.html

Uwe

[mario8eren]

Hi,

if you are still using Turbomole version 6.4 and run COSMO MPI parallel jobs, make sure to update to version 6.4.1 or newer.  See: http://www.turbo-forum.com/index.php/topic,812.0.html

Uwe

Thank you very much.

Mario