[Wien] How to speed up a reliable optimization

Laurence Marks L-marks at northwestern.edu
Thu Jan 26 14:41:56 CET 2017


N.B., the latest release has some additional controls that allow for a bit
more control of the trust region, e.g. a maximum size. These are in the
README, they are not in the user guide. They should only be used by experts
for really nasty to converge problems.

---
Professor Laurence Marks
"Research is to see what everybody else has seen, and to think what nobody
else has thought", Albert Szent-Gyorgi
http://www.numis.northwestern.edu
Corrosion in 4D http://MURI4D.numis.northwestern.edu
Partner of the CFW 100% gender equity project, www.cfw.org/100-percent
Co-Editor, Acta Cryst A


On Jan 26, 2017 07:28, "Laurence Marks" <L-marks at northwestern.edu> wrote:

> The last two (three) have useful information. They are
>
> Mode Size Step
>
> Mode is whether the trust region is active; it is off when Newton is shown.
>
> Size is the radius of the trust region compared to a full Newton step, for
> instance 0.5 would mean only half the full radius.
>
> Step is the magnitude of the step taken compared to a full Pratt step.
>
> All mixer codes (other DFT as well) use a Taylor series expansion for the
> density change with some approximation for the Jacobian (see the mixer
> README & papers). MSEC3 & MSR1 use different Jacobian approximations, as
> did the old BROYD. Neglecting the higher-order terms is only valid for some
> total step magnitude, the trust-region radius which is the "Size" variable
> above. When Size is small the algorithm thinks these higher-order terms
> matter. (The mixer adjusts the trust-region radius depending upon how much
> the self-consistency is improving, which is very tricky, fuzzy coding and
> imperfect.)
>
> The numbers you sent indicate that the algorithm thinks everything is
> "fine". If Size is small (e.g. 0.05) and remains small that suggests a
> problem. The "Step" term can be small when :DIS is large, but should not be
> small near the solution. If it is large (2-3 or more) the algorithm thinks
> that there are soft modes present.
>
> ---
> Professor Laurence Marks
> "Research is to see what everybody else has seen, and to think what nobody
> else has thought", Albert Szent-Gyorgi
> http://www.numis.northwestern.edu
> Corrosion in 4D http://MURI4D.numis.northwestern.edu
> Partner of the CFW 100% gender equity project, www.cfw.org/100-percent
> Co-Editor, Acta Cryst A
>
>
>
> On Jan 26, 2017 05:02, "Luis Ogando" <lcodacal at gmail.com> wrote:
>
> Dear Prof. Marks,
>
>    Just for completeness, in the InP case I am using mBJ and I have
> checked that during the last 20 regular SCF cycles the steps are changing
> from
>
> :MIX  :   MSE1   REGULARIZATION: 6.30E-06  GREED: 0.317  Newton 1.00  0.22
>
> to
>
> :MIX  :   MSE1   REGULARIZATION: 4.93E-06  GREED: 0.437  Newton 1.00  1.14
>
>    Thank you once more !
>    All the best,
>                Luis
> * PS: Could you, please, clarify the meaning of the last two columns in
> the :MIX line ?
>
> 2017-01-26 8:18 GMT-02:00 Luis Ogando <lcodacal at gmail.com>:
>
>> Dear Prof. Marks,
>>
>>    Thank you very much for your answers ! I am pretty sure that your
>> "intuition" will save a lot of computation time !
>>    Just one last question: I have another system , a supercell formed by
>> 15 InP zinc blend cells along [111] (hexagonal representation of the cubic
>> lattice) and 3 InP wurtzite cells. They are aligned along the hexagonal "c"
>> axis.
>>                                      Comparing the gaps of the
>> respective bulks and this supercell (same calculation parameters) , I
>> believe that the SCF cycle (here, it is not a lattice optimization) stopped
>> at a local minimum.
>>                                       My question is : in this case
>> (regular SCF cycle and InP cells), would you change any of your previous
>> answers ?
>>    Thank you again !!
>>    All the best,
>>                  Luis
>>
>>
>> 2017-01-25 17:56 GMT-02:00 Laurence Marks <L-marks at northwestern.edu>:
>>
>>> Inlined is my intuition, which does not have to be completely right.
>>>
>>> On Wed, Jan 25, 2017 at 11:32 AM, Luis Ogando <lcodacal at gmail.com>
>>> wrote:
>>>
>>>> Dear Prof. Marks (and Wien2k community),
>>>>
>>>>    After a recent discussion about "difficult" optimizations in this
>>>> mailing list (subject: "Mixer surprise when using PBE0 hybrid on-site
>>>> functional"), I would like ask you for an advice.
>>>>    I have a system with:
>>>> * 5 rings with C, H and N atoms
>>>> * 100 atoms with P1 symmetry
>>>> * the rings are out of a plane
>>>> * vacuum along y and z
>>>>    I know that this is a very hard optimization problem, so I would
>>>> like to kindly ask:
>>>>
>>>> 1) Do you believe that MSEC3a will work better than MSR1a in such a
>>>> case ?
>>>>
>>>
>>> I would switch to MSEC3a, use SLOW or reduce the GREED to 0.1 only if
>>> you see indications of problems. If the system is a decent insulator and
>>> the experimental positions are quite good you may have no problems. At the
>>> end I would switch back to MSR1a certainly for a system with OH as the
>>> positions of the hydrogens can be quite soft. It can take quite some time
>>> to get the rotations of the OH bond distance right as Wien2k uses cartesian
>>> coordinates not polars. Which of the 3 is best -- I am not sure.
>>>
>>>>
>>>> 2) Do you recommend using -it, -vec2pratt and -noHinv options with
>>>> run_lapw ?
>>>>
>>>
>>> I use -it -noHinv -vec2pratt. Sometimes you need to do an occasional
>>> full diagonalization (touch .fulldiag) as the iterative method is less
>>> stable (it adds a little noise). In my personal version I have added back
>>> the old -itn option so this is done automatically every few steps.
>>>
>>>>
>>>> 3) Should I reduce TRUST to 0.5 (I am using LDA and experimental values
>>>> for the initial atomic positions) ?
>>>>
>>>
>>> Probably not. TRUST 0.5 would be if it is taking much too large steps
>>> which tends (in my experience) to occur more with soft electronic modes
>>> such as one has with d and f electrons.
>>>
>>>>
>>>> 4) Should I use SLOW in case.inm ?
>>>>
>>>
>>> See my answer to 1)
>>>
>>>>
>>>>    Many thanks in advance.
>>>>    All the best,
>>>>                       Luis
>>>>
>>>>
>>>
>>>
>>> --
>>> Professor Laurence Marks
>>> "Research is to see what everybody else has seen, and to think what
>>> nobody else has thought", Albert Szent-Gyorgi
>>> www.numis.northwestern.edu ; Corrosion in 4D:
>>> MURI4D.numis.northwestern.edu
>>> Partner of the CFW 100% program for gender equity,
>>> www.cfw.org/100-percent
>>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.cfw.org_100-2Dpercent&d=CwMFaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=U_T4PL6jwANfAy4rnxTj8IUxm818jnvqKFdqWLwmqg0&m=CZHTnM-OmVb7WVhK4QBSMqumiacPyU-LxrLt8-sw9bU&s=cyXhG2IW3LH9f7wuLmlZGAS47BrtGluAX3LnBa9-Mug&e=>
>>> Co-Editor, Acta Cryst A
>>>
>>> _______________________________________________
>>> Wien mailing list
>>> Wien at zeus.theochem.tuwien.ac.at
>>> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
>>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__zeus.theochem.tuwien.ac.at_mailman_listinfo_wien&d=CwMFaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=U_T4PL6jwANfAy4rnxTj8IUxm818jnvqKFdqWLwmqg0&m=CZHTnM-OmVb7WVhK4QBSMqumiacPyU-LxrLt8-sw9bU&s=nujfSCNX55lo__i30MehzUrgntz21bjkJFguwZ82Uzs&e=>
>>> SEARCH the MAILING-LIST at:  http://www.mail-archive.com/wi
>>> en at zeus.theochem.tuwien.ac.at/index.html
>>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.mail-2Darchive.com_wien-40zeus.theochem.tuwien.ac.at_index.html&d=CwMFaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=U_T4PL6jwANfAy4rnxTj8IUxm818jnvqKFdqWLwmqg0&m=CZHTnM-OmVb7WVhK4QBSMqumiacPyU-LxrLt8-sw9bU&s=MJqLknQDobuUH-IAiKowy5gMln1IboZOxcaawJzfRUs&e=>
>>>
>>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20170126/f5c3fd8a/attachment.html>


More information about the Wien mailing list