[CP2K-user] SCF Convergence stuck with ADMM HF?

[Vladimir Rybkin]

Dear Nick,

on the updated Cray systems some strange behaviour is observed for some 
algorithms. They are working on that. Poor convergence may be one of them 
(my experience). Although if you see it converge, it's fine.

Yours,

Vladimir

пятница, 15 мая 2020 г., 22:36:08 UTC+2 пользователь Nicholas Winner 
написал:
>
> Vladimir,
>
> (1) You're right that convergence (1e-6) is already reasonably good. For 
> my testing, I already reran with 1e-6 EPS_SCF to get a final band gap. The 
> issue is not that it is bad, but that the convergence behavior was 
> unexpected. I am benchmarking in order to run medium-to-high throughput 
> ADMM calculations, and I need to get a good handle on the ways in which 
> calculations can stall in order to prepare for that.
> (2) Reducing EPS_SCHWARZ to 1e-7 did assist a great deal in the 
> convergence. Thank you. 
> (3) Yes, running this on a Cray system. Is this a problem?
>
> -Nick
>
> On Friday, May 15, 2020 at 4:43:16 AM UTC-7, Vladimir Rybkin wrote:
>>
>> Dear Nick,
>>
>>
>> 1) you convergence is not so bad already. At least, good enough for most 
>> applications.
>> 2) You Schwarz screening is rather large. Try 10-7 and smaller. Normally, 
>> convergence is sensitive to the accuracy of Hartree-Fock exchange.
>>
>> are you by chance running it on a Cray system?
>>
>> Yours,
>>
>> Vladimir
>>
>> четверг, 14 мая 2020 г., 22:14:09 UTC+2 пользователь Nicholas Winner 
>> написал:
>>>
>>> Hello all,
>>>
>>> I have tried to perform static PBE0 calculation of diamond using the 
>>> ADMM. I have done many of this type of calculation before, but not for this 
>>> system. I have an issue where the reported "convergence" in the SCF loop is 
>>> unchanged for many steps. I initially tried using the DIIS minimizer and OT 
>>> method, but switched to 2PNT CG search to try to improve convergence. This 
>>> helped a tiny amount (convergence got lower, but still stuck). 
>>>
>>> I'm unclear as to what the problem is. I have attached my input and 
>>> output files. Note I stopped this calculation prematurely.
>>>
>>> Drawing your attention to the first inner SCF loop of the output file, I 
>>> can see that SCF step 19-50 all report the same convergence of (0.00000035) 
>>> with no change. What's more, the "total energy" in that loop is listed as -3903.4843979582 
>>> for all steps after 31. So it seems that the energy is extremely well 
>>> converged... better than the EPS_SCF 1e-7 that I listed in the input file.
>>>
>>> I'd like to know if there's any reason why this would be having such a 
>>> hard time converging; diamond is a large band gap material and should have 
>>> very good convergence with OT, especially with CG. Also, am I 
>>> misunderstanding hte numbers reported by the SCF steps? It seems to me the 
>>> total energy is very well converged to within 1e-9 the end of the first SCF 
>>> loop, but convergence is listed as .00000035.
>>>
>>> Any help would be appreciated.
>>>
>>> Best,
>>> Nick
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20200522/8927cee4/attachment.htm>