[DFTB-Plus-User] Slow Convergence problem
Jan M. Knaup
Jan.Knaup at bccms.uni-bremen.de
Fri Aug 17 11:30:30 CEST 2012
Dear Bikash,
your observation indicates, two things:
1. The internal flexibility of your polymer matrix leads to shifts of the
internal coordinates upon compression which break the lattice symmetry.
Since your polymer matrix is not a crystal that behavior is not surprising.
>From my own experience with calculating elastic properties, if a system is
not very dense, the atoms often react to compression by moving into voids
and changing bond angles.
2. Since you see high pressures with both your fully MD optimized geometry
and when only slightly shifting your atoms after supercell vector
optimization, it looks very much like DFTB does not like the bond angles
your force-field produces. Therefore, moving the bond angles towards the
DFTB equilibrium increases the pressure. (The inverse effect to 1.)
In my opinion, under these circumstances, optimizing the supercell without
optimizing internal coordinates will not help you very much. With
significant changes of the symmetry of the internal coordinates, you have
no way of knowing, if the equilibrium supercell at the non-equilibrium
internal coordinates resembles the equilibrium supercell at equilibrium
coordinates or not.
For a crystalline system, Ben's approach is often better. If your pressure
deviation was caused by a difference in bond lengths, simply scaling your
model should put it very close to the equilibrium configuration. Under
these circumstances, first adapting the internal coordinates to the
compressed volume, then relaxing the cell and returning to the original
bond angles would be wasteful.
I do not believe that interleaving repeated cell relaxations at fixed
internal coordinates while optimizing the internal coordinates will help
you very much, if at all. If you feel that optimizing the internal
coordinates at a good cell volume will speed up the relaxation, I think my
original proposal will help you more. However, I stay convinved that you
will need around 1500-3000 geometry steps for the internal coordinate
optimization at the first cell volume. The lattice optimization will only
advance after the internal coordinate optimization is converged (or if you
forbid the program to change any internal coordinates). Subsequent internal
coordinate optimizations at different cell shapes will hopefully be faster
and choosing a good initial cell volume may help you reduce the necessary
number of lattice steps. Still the fact remains, with conjugate gradients
one needs one CG step per degree of freedom, but one usually reaches
convergence before executing one step per atomic coordinate, because
degrees of freedom are often so soft that they do not contribute
significant force.
Just to make sure: you do quench your MD simulation to 0K before you take
the MD structure as input for a 0K geometry optimization, right?
Best,
Jan
Jan M. Knaup | Fon +49-(0)421-218-62351
Dipl. Phys. Dr. rer. nat. | Fax +49-(0)421-218-62770
Universität Bremen - BCCMS |
Am Fallturm 1 | Jan.Knaup at bccms.uni-bremen.de
28359 Bremen | JanKnaup at gmail.com
Germany | www.bccms.uni-bremen.de
2012/8/17 bikash sankar kanungo <biku.kanungo at gmail.com>
> Dear Ben,
>
> Thanks for the help. By keeping all the atoms fixed and allowing the
> lattice constants to change I was able to reduce the pressure and achieve
> convergence in 24 geometry step( for fixed atoms case geometry steps and
> lattice step has same meaning). But feeding the end geometry for internal
> coordinates optimization starts with low pressure(1.0E+05) in the 0th
> geometry step and then shoots to 1.0E+09 in the next.
>
> Moreover,if I try a full geometry optimization(moved atoms+lattice opt)
> with the end geometry from fixed atoms lattice optimization it still
> performs only lattice step 0 and doesn't go beyond it.
>
> I suppose I will have to perform several the fixed atoms lattice
> optimization and then internal coordinates optimization(for about 100
> geometry steps) alternatively to achieve the relaxed structure.
>
>
> Thank you,
> Bikash
>
> On Thu, Aug 16, 2012 at 10:25 PM, Ben Hourahine <
> benjamin.hourahine at strath.ac.uk> wrote:
>
>> Dear Bikash,
>>
>> The "Lattice Step = 0" is the significant part. So far only the internal
>> coordinates are optimising, but not the
>> lattice constants. Speaking as the author of these routines, the
>> optimisation is nested with the structure
>>
>> Optimise lattice vectors {
>> Optimise internal coordinates{}
>> }
>>
>> You could try setting no atoms to move ( MovedAtoms = {} ) as a first
>> pre-relaxation (the pressure will certainly decrease then).
>>
>> Regards
>>
>> Ben
>>
>>
>> On 16/08/12 17:25, bikash sankar kanungo wrote:
>>
>> Dear Jan,
>>
>> Thanks for such a descriptive and insightful reply. I kept the timestep
>> =0.1 fs as the temperature of the system showed a sharp rise for 1fs
>> ultimately printing Nan. I made some trials and found 0.25fs to be the the
>> maximum timestep that worked for me.
>>
>>
>> For my NPT(Berendsen thermostat and barostat) run in DFTB I did not
>> notice any pressure change after 100 geometry steps( = 10fs for a timestep
>> of 0.1fs). But since I have set Timescale=50fs so after 10fs I expected
>> some change in pressure. May be I am being too anxious about my outputs.
>>
>> I am doing a LatticeOpt simultaneously , but strangely I never got
>> beyond Lattice Step = 0 for any of the 100 Geometry Steps I have reached so
>> far. Is that an expected behavior ?
>>
>> For the time being I will let them go upto 1000-2000 steps which may
>> take a week and see if I get any success else I will have to for a smaller
>> system and work out the re-scaling approach you suggested.
>>
>> Thank you,
>> Bikash
>>
>> On Thu, Aug 16, 2012 at 8:51 PM, Jan M. Knaup <
>> Jan.Knaup at bccms.uni-bremen.de> wrote:
>>
>>> Dear Bikash,
>>>
>>> 50 fs is an extremely short simulation time, especially to get a
>>> barostat to converge, especially at 10 K where the low temperature
>>> naturally leads to very slow atomic movment. The first question to ask
>>> would be, why do you use 0.1 fs time steps? For hydrocarbons 0.5 - 1.0 fs
>>> are usually sufficient at room temperature, let alone at almost liquid
>>> helium temperature.
>>>
>>> You do not write after how many geometry iterations you do not see
>>> significant chage of the pressure, but from your NVP procedure, I assume
>>> you check after a few hundred iterations. For a system with that many
>>> atoms, that is not a lot.
>>>
>>> As a general rule of thumb, conjugate gradient relaxation takes about
>>> as many geometry steps as you have degrees of freedom in you system. Of
>>> course this depends on the system and your convergence criteria to some
>>> degree, still I would expect needing anywhere between 1000 and 5000 steps
>>> for a system like yours, for geometry relaxation at constant volume. Even
>>> small differences in the equilibrium bond lengths between whatever force
>>> field you use and the DFTB parameters you use can easily explain your
>>> pressure of 10^9 Pa. During the constant volume part of the CG relaxation,
>>> drastic change of the pressure is very unlikely to occur since you use a
>>> quite realistic model from your MM calculations.
>>>
>>> You would have to perform an optimization of the lattice vectors to
>>> arrive at a low pressure (Keyword LatticeOpt in the DFTB+ manual). However,
>>> the lattice optimization is performed in its own loop outside the atomic
>>> configuration. That means, that for lattice optimization, geometry
>>> optimizations at constant volume are performed for different volumes and
>>> cell shapes. It can take anywhere between 5 and 20 lattice steps, possibly
>>> even more. It is very difficult to predict if you will need many geometry
>>> steps after a change in the lattice vectors or not. This depends a lot on
>>> the symmetry and rigidity of your system.
>>>
>>> Unfortunately, there is not really any way around this procedure, if
>>> you want to have your model at a pressure close to 1 atm. However, it is
>>> probably possible to reduce the number of necessary geometry and lattice
>>> iterations, even if you chose not to optimize the DFTB cell volume:
>>>
>>> You could take a much smaller model of maybe 150 atoms or thereabouts,
>>> perform your LAMMPS procedure, then do a full geometry and lattice
>>> optimization in DFTB. From that you can obtain the equilibrium densities of
>>> your model for both force-field and DFTB. If you then rescale the big model
>>> to obtain the same volume ratio between LAMMPS output und DFTB+ input as
>>> for the small model, that will likely save you a lot of geometry iterations.
>>>
>>> Hope this helps,
>>>
>>> Jan
>>>
>>> Jan M. Knaup | Fon +49-(0)421-218-62351
>>> Dipl. Phys. Dr. rer. nat. | Fax +49-(0)421-218-62770
>>> Universität Bremen - BCCMS |
>>> Am Fallturm 1 | Jan.Knaup at bccms.uni-bremen.de
>>> 28359 Bremen | JanKnaup at gmail.com
>>> Germany | www.bccms.uni-bremen.de
>>>
>>>
>>> 2012/8/16 bikash sankar kanungo <biku.kanungo at gmail.com>
>>>
>>>> Hi,
>>>>
>>>> Hi I am trying to relax an epoxy polymer matrix in DFTB . The polymer
>>>> matrix is obtained by cross-linking the epoxy monomers with a curing agent.
>>>> The cross-linking is done using Classical MD tool LAMMPS. Before feeding
>>>> the cross-linked structure(of 1322 atoms) from LAMMPS to DFTB I ensured to
>>>> keep the pressure to be 1 atm by doing NPT runs in LAMMPS. But while doing
>>>> structure relaxation in DFTB using Conjugate Gradient Driver I do not see
>>>> any change in order of magnitude of pressure which remains as high as
>>>> 1.0E+09 Pa. With a hope of reducing the pressure to 1.0E05(1 atm) I even
>>>> tried Verlocity Verlet using Berendsen Thermostat and Barostat with target
>>>> temperature = 10K , target pressure=1.0E+05 Pa , Timestep=0.1fs, Timescale
>>>> = 50fs but the pressure still remains at around 1.0E+09 Pa even after 10fs.
>>>> To obtain SCC convergence I have set Broyden mixing parameter =
>>>> 0.3(default=0.2\0 and Fermi Fillinf temperature=300K(default=0K). Do I need
>>>> to provide more time for convergence for such a large system or are there
>>>> other alternatives to boost the convergence rate?
>>>>
>>>> I would appreciate any help or suggestions regarding the relaxation
>>>> of this seemingly difficult structure.
>>>>
>>>> Thank you,
>>>> Bikash
>>>> --
>>>> BIKASH SANKAR KANUNGO
>>>> Final Year Undergraduate student,
>>>> Mechanical Engineering Department,
>>>> INDIAN INSTITUTE OF TECHNOLOGY
>>>> KHARAGPUR.
>>>> +919749935409
>>>>
>>>>
>>>> _______________________________________________
>>>> DFTB-Plus-User mailing list
>>>> DFTB-Plus-User at dftb-plus.info
>>>> http://www.dftb-plus.info/mailman/listinfo/dftb-plus-user
>>>>
>>>>
>>>
>>> _______________________________________________
>>> DFTB-Plus-User mailing list
>>> DFTB-Plus-User at dftb-plus.info
>>> http://www.dftb-plus.info/mailman/listinfo/dftb-plus-user
>>>
>>>
>>
>>
>> --
>> BIKASH SANKAR KANUNGO
>> Final Year Undergraduate student,
>> Mechanical Engineering Department,
>> INDIAN INSTITUTE OF TECHNOLOGY
>> KHARAGPUR.
>> +919749935409
>>
>>
>> --
>> Dr. B. Hourahine, SUPA, Department of Physics,
>> University of Strathclyde, John Anderson Building,
>> 107 Rottenrow, Glasgow G4 0NG, UK.
>> +44 141 548 2325, benjamin.hourahine at strath.ac.uk
>> The University of Strathclyde is a charitable body,
>> registered in Scotland, with registration number SC015263
>>
>>
>> _______________________________________________
>> DFTB-Plus-User mailing list
>> DFTB-Plus-User at dftb-plus.info
>> http://www.dftb-plus.info/mailman/listinfo/dftb-plus-user
>>
>>
>
>
> --
> BIKASH SANKAR KANUNGO
> Final Year Undergraduate student,
> Mechanical Engineering Department,
> INDIAN INSTITUTE OF TECHNOLOGY
> KHARAGPUR.
> +919749935409
>
>
> _______________________________________________
> DFTB-Plus-User mailing list
> DFTB-Plus-User at dftb-plus.info
> http://www.dftb-plus.info/mailman/listinfo/dftb-plus-user
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.dftb-plus.info/pipermail/dftb-plus-user/attachments/20120817/eab6091c/attachment-0001.html>
More information about the DFTB-Plus-User
mailing list