[CP2K:2253] Re: convergence & cutoff madness

Fawzi Mohamed fa... at gmx.ch
Mon Sep 7 12:25:15 CEST 2009


Hi,

some comments:
1) do you trust your cp2k version, can you reproduce the regtest  
results within a reasonable error?
2) convergence wrt. cutoff is not variational
3) if the cutoff is too low derivatives (and KS matrix) are wrong, and  
thus convergence difficult (I would say the lowest cutoff is around  
280 for your case)
4) using multiple grids you should increase also the relative cutoff  
to really go at convergence.

I hope this helps making sense of your data
Fawzi
On 7-set-09, at 12:11, Toon Verstraelen wrote:

>
> Hi,
>
> Below a similar table with the DZVP-GTH basis instead of the SZV-GTH  
> basis.
>
> Number of failed SCF	 7
> cutoff [Ry]   energy [kjmol]   time [s]   steps   time per step [s]"
>    100.0         755266.00      875.24     250        3.50
>    110.0           1849.77      159.59      42        3.80
>    120.0           1849.20      129.75      35        3.71
>    130.0            986.17      210.71      54        3.90
>    140.0            823.57      489.46     114        4.29
>    150.0            762.08      130.94      30        4.36
>    160.0            604.84      521.14     113        4.61
>    170.0            294.25      176.77      37        4.78
>    180.0            319.16     1270.09     250        5.08
>    190.0            323.22     1198.53     250        4.79
>    200.0             87.76     1330.97     250        5.32
>    210.0             87.63     1330.64     250        5.32
>    220.0             83.83     1375.40     250        5.50
>    230.0             90.91      630.85     104        6.07
>    240.0             74.54      552.09      87        6.35
>    250.0             41.28      207.62      30        6.92
>    260.0             39.96     1330.75     250        5.32
>    270.0             24.73      598.13      89        6.72
>    280.0             16.34      237.48      29        8.19
>    290.0              5.94      285.42      31        9.21
>    300.0              4.92      303.03      34        8.91
>    310.0              7.34      628.55      86        7.31
>    320.0              0.00      453.03      57        7.95
>    330.0              1.91      250.62      26        9.64
>    340.0              4.26      287.63      27       10.65
>    350.0              5.16      239.20      23       10.40
>    360.0              5.30      341.52      35        9.76
>    370.0              8.57      303.73      26       11.68
>    380.0              8.52      286.88      24       11.95
>    390.0             12.97      262.10      24       10.92
>    400.0             12.20      289.96      26       11.15
>
> It is a bit strange that the timings improve and convergence is
> obtained with less steps compared to the smaller basis.
>
> We also did similar tests with the PBE functional. All those
> computations ran smoothly without any convergence issue.
>
>
> Citeren Toon Verstraelen <Toon.Ver... at UGent.be>:
>
>>
>> Dear CP2K users,
>>
>> I'm running a few basic tests to get a better understanding of the
>> relation between SCF convergence and the plane wave cutoff. The table
>> below summarizes a series of trivial energy calculations with  
>> different
>> cutoffs. The input files are attached.
>>
>> The system in the input is MIL-53-Al, a well-known metal-organic
>> framework. To keep things simple, I've used a very minimal basis
>> (SZV-GTH). The largest basis set exponent is 8.3744350009, which
>> corresponds to a cutoff of 330 Rydberg. (if not mistaken). The
>> functional is blyp.
>>
>> There are two weird things in this table:
>>
>> 1) The lowest energy is not obtained with the highest cutoff. It gets
>> worse above 320.
>>
>> 2) For some high cutoff's there is no SCF convergence. For the low
>> range, up to 150, this is normal. I would not expect this for a  
>> cutoff
>> of 330.
>>
>> Does somebody know what is going on? The input files are rather  
>> simple,
>> but maybe I've put in something pernicious? With larger basis sets,
>> there are more cases where the SCF convergence fails. It becomes
>> practically useless, even when a decent initial guess with a smaller
>> basis set is constructed.
>>
>>
>> Number of failed SCF	 10
>> cutoff [Ry]   energy [kjmol]   time [s]   #iters   time per iter [s]"
>>    100.0         370146.36      841.59     250        3.37
>>    110.0           4309.36     1023.83     250        4.10
>>    120.0           4311.87     1053.45     250        4.21
>>    130.0           1159.84     1099.50     250        4.40
>>    140.0            939.17     1166.17     250        4.66
>>    150.0           2549.48     1206.51     250        4.83
>>    160.0            804.26      227.56      42        5.42
>>    170.0            354.60      159.26      29        5.49
>>    180.0            412.98      177.18      29        6.11
>>    190.0            685.96     1455.42     250        5.82
>>    200.0             79.06      199.92      32        6.25
>>    210.0             79.17      262.35      40        6.56
>>    220.0             78.54      201.22      28        7.19
>>    230.0             86.04      218.52      28        7.80
>>    240.0             63.43      261.85      35        7.48
>>    250.0             33.38     2067.27     250        8.27
>>    260.0             34.90     2237.25     250        8.95
>>    270.0             22.10      225.49      28        8.05
>>    280.0             14.48      458.72      52        8.82
>>    290.0              5.22      322.73      32       10.09
>>    300.0              4.17      251.49      28        8.98
>>    310.0              6.07      319.30      32        9.98
>>    320.0              0.00      296.63      29       10.23
>>    330.0              2.18     2370.99     250        9.48
>>    340.0              4.32      351.99      28       12.57
>>    350.0              4.38      398.81      32       12.46
>>    360.0              4.37      332.36      25       13.29
>>    370.0              8.01      496.19      39       12.72
>>    380.0              8.10      368.99      29       12.72
>>    390.0             13.16      402.97      29       13.90
>>    400.0             12.49      410.13      29       14.14
>>
>>
>> cheers,
>>
>> Toon
>>
>>
>>>
>>
>
>
>
>
> 



More information about the CP2K-user mailing list