[CP2K-user] [CP2K:17252] Re: Problem with 2nd generation CPMD

[Marcella Iannuzzi]

Dear Onofrio

The system seems not to be very well equilibrated, indeed
The fact that the number of SCF iterations tends to increase is also not a 
very good sign
Probably I would try and run longer BOMD before starting sgcp, but it 
happens that finding the good parameters is anyway cumbersome 

Regards
Marcella


On Monday, June 27, 2022 at 6:43:35 PM UTC+2 onofr... at gmail.com wrote:

> Dear Marcella,
>
> your suggestions were so useful! I was able to reduce to 3-4 iterations 
> per MD step. Sorry for my late reply but I played a lot with GAMMA and 
> NOISE_GAMMA to equilibrate the MD run by monitoring the temperature per 
> each kind. However, I have gathered some unusual behaviour (following 
> attached files represent *Temp per kind vs MD steps*):
>
> 1. I first assumed GAMMA 0.001 fs^-1 (value typically used in other works) 
> and gave different NOISE_GAMMA values for each &DEFINE_REGION (region1: Cu 
> slab ; region2: one toluene molecule). In all cases, for the first 100 MD 
> steps the temperature rapidly drops to very lower values (see attached* 
> gamma0.001.png* file as an example), though it partially riequilibrates 
> to the target T (900 K) only for sufficiently larger NOISE_GAMMA. Must be 
> the initial temperature drop avoided or is it sufficient only adjusting the 
> NOISE_GAMMA to larger values for T compensation (thus fixing GAMMA to 
> 0.001)?
>
> 2. To avoid initial temperature drops, I have tried to directly increase 
> GAMMA. For 0.1 fs^-1 it seems to work well (attached *gamma0.1.png* file), 
> but it acts like an 'aggressive' thermostat producing large T fluctuations, 
> especially for the toluene region (similar situations for GAMMA = 0.01 
> fs^-1, attached *gamma0.01.png* file). I'm aware of the few degrees of 
> freedom of the single toluene molecule causing large temperature 
> fluctuations; therefore it is better to assume GAMMA 0.001, is it?
>
> 3. Most important: the number of iterations per MD step costantly increase 
> from 3-4 to 13-16 iterations during MD runs as it can be seen in the 
> attached output file taken as an example (*Cu_toluene.out*) with
>     *grep 'SCF run' Cu_toluene.out*
> To restore the convergence with 3-4 iterations per MD step I sometimes did 
> external EXIT and then restart the run with .RESTART.wfn. It initially 
> works, but then the number of iterations increases again.
> Is it a normal behaviour or is it due to a bad instability of the Langevin 
> dynamics? How to eventually avoid it?
>
> Many thanks!
>
> N.B.: *No* external thermostat and no velocity rescaling (TEMP_TOL) were 
> used during simulations.
>
> Onofrio
>
> Il giorno giovedì 16 giugno 2022 alle 12:23:19 UTC+2 Marcella Iannuzzi ha 
> scritto:
>
>>
>>
>> As you may have read in previous works, to run SGCPMD a preliminary study 
>> is needed to set the proper GAMMA and NOISY_GAMMA parameters, which allow 
>> for the conservation of energy, for a Gaussian distribution of the error in 
>> the forces, and for a correct equipartition of the kinetic energy among the 
>> degrees of freedom. 
>> Moreover the parameters to tune the convergence of the electronic 
>> structure at each MD step need also to be properly adjusted to the system. 
>> This means that a series of preliminary simulations are needed, before 
>> the production runs are started. 
>>
>> This said, my suggestions are the following:
>> - Always start sgcpmd from a well converged wfn obtained along a standard 
>> BOMD run with diagonalisation and smearing
>> -  Let the first steps run on many scf iterations if needed, i.e., set 
>> MAX_SCF large enough
>> -  But use a relative large eps_scf. The wfn is not going to be optimised 
>> as in the BOMD, energy convergence  around 10^-5 is desirable
>> - Use FULL_ALL preconditioner + ASPC
>> - Use a small time step
>> - Tune properly GAMMA and NOISY_GAMMA
>> - Monitor the conservation of energy and the temperature per kind or per 
>> region : TEMP_KIND
>>
>> Regards
>> Marcella
>>
>> On Thursday, June 16, 2022 at 10:53:53 AM UTC+2 onofr... at gmail.com wrote:
>>
>>> Hi Sam,
>>> yes, of course. I set up more standard BOMD with ENSEMBLE NVT and 
>>> DIAGONALIZATION with smear (look at my input file) and worked fine: it took 
>>> more or less *30 steps to reach convergence* every AIMD step.
>>>
>>> Then, I tried with ENSEMBLE NVT and OT (without smear obviously) and I 
>>> got (MAX_SCF 200):
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> * *** WARNING in qs_scf.F:598 :: SCF run NOT converged *** *** WARNING 
>>> in qs_scf.F:598 :: SCF run NOT converged ***  *** SCF run converged in   
>>>  79 steps ***  *** SCF run converged in    72 steps ***  *** SCF run 
>>> converged in    84 steps ***  *** SCF run converged in    53 steps ***  *** 
>>> SCF run converged in    49 steps ***  *** SCF run converged in    48 steps 
>>> ***  *** SCF run converged in    38 steps ***  *** SCF run converged in   
>>>  40 steps ***  *** SCF run converged in    38 steps ***  *** SCF run 
>>> converged in    30 steps ***  *** SCF run converged in    23 steps ***  *** 
>>> SCF run converged in    26 steps ***  *** SCF run converged in    27 steps 
>>> ***  *** SCF run converged in    26 steps ***  *** SCF run converged in   
>>>  26 steps ***  *** SCF run converged in    23 steps ***  *** SCF run 
>>> converged in    16 steps ***  *** SCF run converged in    16 steps ***  *** 
>>> SCF run converged in    16 steps ***  *** SCF run converged in    18 steps 
>>> ****
>>> *   ...*
>>>
>>> Though SCF run not converged for the first two AIMD steps, then it 
>>> worked sometimes fine (SCF run converged in 16 steps, but not always!). 
>>> Could be a possible solution.
>>> N.B.: setting MAX_SCF below 200 (i.e., 50) makes SCF run not converged 
>>> more than two initial AIMD steps.
>>>
>>> Anyway, I'm looking for correctly set up SGCPMD as that found in the 
>>> following works:
>>> - Lan, Jinggang, Jürg Hutter, and Marcella Iannuzzi. "First-principles 
>>> simulations of an aqueous CO/Pt (111) interface." *The Journal of 
>>> Physical Chemistry C* 122.42 (2018): 24068-24076.
>>> - Musso, Tiziana, et al. "Second generation Car-Parrinello MD: 
>>> application to the h-BN/Rh (111) nanomesh." *The European Physical 
>>> Journal B* 91.7 (2018): 1-8.
>>>
>>> However, I'm struggling with and I kindly need your help.
>>> Kind regards.
>>>
>>> Onofrio
>>>
>>> Il giorno giovedì 16 giugno 2022 alle 09:58:41 UTC+2 wave... at gmail.com 
>>> ha scritto:
>>>
>>>> Hi Onofrio
>>>>
>>>> Question: have you set this up and run a more standard simulation to 
>>>> work out the convergence first?
>>>>
>>>> If I guess, it's probably Cu as a transition metal with d-orbitals that 
>>>> is difficult (I am suffering with Au).
>>>>
>>>> Kind Regards
>>>>
>>>> Sam 
>>>> On Wednesday, June 15, 2022 at 3:20:20 PM UTC+2 onofr... at gmail.com 
>>>> wrote:
>>>>
>>>>> Dear CP2K community,
>>>>>
>>>>> I'm new to CP2K and I'm trying to set up a second-generation 
>>>>> Car-Parrinello (SGCP) MD simulation. I have started with a geometrically 
>>>>> optimized system of toluene adsorbed on a 3-layer Cu slab (8x6). I would to 
>>>>> know what sections/keywords in the input file must be used to correctly set 
>>>>> up a SGCPMD simulation and, possibly, the best settings to improve the 
>>>>> speed of such simulations since I have not found any related tutorial (but 
>>>>> only few things in the cp2k google group). 
>>>>> If I well understood, it has to be turn on the Langevin dynamics 
>>>>> section in combination with the OT method. For example:
>>>>>
>>>>> &FORCE_EVAL
>>>>>   ...
>>>>>   &DFT
>>>>>     ...
>>>>>     &QS
>>>>>       METHOD GPW
>>>>>       EXTRAPOLATION ASPC
>>>>>       EXTRAPOLATION_ORDER 3
>>>>>       EPS_DEFAULT 1.0E-10
>>>>>     &END QS
>>>>>     ...
>>>>>     &SCF
>>>>>     ...
>>>>>       &OT
>>>>>         PRECONDITIONER FULL_SINGLE_INVERSE   ! or FULL_KINETIC
>>>>>         MINIMIZER DIIS
>>>>>         N_DIIS 7
>>>>>       &END OT
>>>>>       &OUTER_SCF
>>>>>         MAX_SCF 3
>>>>>         EPS_SCF 1.0E-05
>>>>>       &END OUTER_SCF    
>>>>>     ...
>>>>>     &END SCF
>>>>>     ...
>>>>>   &END DFT
>>>>>   ...
>>>>> &END FORCE_EVAL
>>>>>
>>>>> &MOTION
>>>>>   &MD
>>>>>     ENSEMBLE LANGEVIN
>>>>>     ...
>>>>>     &LANGEVIN
>>>>>       GAMMA 0.001
>>>>>     &END LANGEVIN
>>>>>     ...
>>>>>   &END MD
>>>>>   ...
>>>>> &END MOTION
>>>>>
>>>>> Is it correct? To follow what I just mentioned, I prepared an input 
>>>>> file (attached below) based on sections/keywords defined just above and I 
>>>>> got from the output file (attached below) only 
>>>>>     *"Leaving inner SCF loop after reaching   200 steps" *
>>>>> (200 steps because I set MAX_SCF 200)
>>>>> resulting in a very bad convergence. 
>>>>> As a result, it seems that I have not been running any SGCPMD 
>>>>> simulation but a simple bad BOMD. Am I missing something?
>>>>> How can I correctly set up my input file for SGCPMD?
>>>>>
>>>>> Any help would be greatly appreciated.
>>>>> Thanks.
>>>>>
>>>>> Onofrio
>>>>>
>>>>

-- 
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+unsubscribe at googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/679b3f4a-098a-4606-923c-5a06c58b9c2dn%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20220629/14a2f19e/attachment-0001.htm>