Re: [AMBER-Developers] problems with QM/MM energy conservation

From: Andreas Goetz <agoetz.sdsc.edu>
Date: Tue, 5 Apr 2011 13:02:23 -0700

Hi Dave,

Ross and I are aware of this problem. Have a look at page 18 of my talk from the AMBER developer meeting. The green line is with PME, I believe the energy drift is similar to what InSuk is seeing. Ross already explained everything in his reply and I agree that we should implement a QM/MM interaction scheme which relies on Mulliken charges. This should be fairly easy to implement and will conserve energy. It will be interesting to see how properties are affected by using Mulliken charges for short and long range interactions as compared to what we have now.

In particular because we will run into the same problems with ab initio and DFT based QM/MM once we account for long-range electrostatics (which are simply cut off at present, see red line of the plot in my talk).

All the best,
Andy


On Apr 5, 2011, at 8:13 AM, Ross Walker wrote:

> Hi Dave,
>
> Yes I know exactly what is wrong here and I have been 'suggesting' for a
> long time how we might go about fixing it but it has never been put in.
> Essentially when you run a periodic PME calculation with NDDO based QM/MM
> inside the cutoff you compute the full QM/MM interaction. In the 'PME' part
> of the calculation though you use a mulliken charge approximation. Therefore
> there is a discontinuity in the electrostatics as you cross the QM/MM
> boundary. You can check the QM/MM is okay by running a solvent cap with no
> cutoff. Things get better for QM/MM when you run a larger system (MM atom
> count), use qmcut=12.0 and also if you have a buried QM/MM region inside a
> protein where there is less effective crossing of the QM/MM boundary.
>
> The reason DFTB is fine is that it uses a mulliken charge interaction
> (q[i]q[j]/r[ij]) for both the direct and reciprocal space calculations. One
> possible solution is to add a force switching function to the direct space
> sum in QM/MM which would then conserve energy but for the wrong reasons. The
> other that I have been encouraging 'certain' people to implement for ages is
> to add an option (supposed to be qmmm_int=3) to use a mulliken charge
> interaction in the direct space for the QM/MM interaction with NDDO. I
> believe this is what Jorgensen has been using for years. The potential would
> then be smooth and energy would be conserved. It would be interesting to
> compare the difference between the current approach and a Mulliken charge
> QM/MM interaction. My bet is the difference would not be huge and the
> improvement in energy conservation would make up for the more approximate
> QM/MM interaction.
>
> If Insuk would like to try implementing this I would be happy to help out.
>
> All the best
> Ross
>
>> -----Original Message-----
>> From: David A Case [mailto:case.biomaps.rutgers.edu]
>> Sent: Tuesday, April 05, 2011 7:22 AM
>> To: amber-developers.ambermd.org
>> Subject: [AMBER-Developers] problems with QM/MM energy conservation
>>
>> InSuk and I are having problems seeing energy conservation in simple
>> QM/MM calculations with the NDDO Hamiltonians. Out simplest example is
>> NMA in water, and here is a typical input file:
>>
>>
>> production
>> &cntrl
>> imin=0, ntx=7, ntpr=100,
>> ntf=2, ntc=2, tol=1.d-7,
>> ntb=1, ntp=0,
>> ntt=0,
>> nstlim=50000, dt=0.001, ntave=25000,
>> cut=9.0,
>> irest=1,
>> iwrap=1,
>> ifqnt=1,
>> /
>> &ewald
>> dsum_tol=1.d-6,
>> /
>> &qmmm
>> qmmask=':1', qmcharge=0, qm_theory='AM1',
>> scfconv=1.0d-11, tight_p_conv=1, peptide_corr=0,
>> /
>>
>> This shows severe lack of energy conservation (change in total energy
>> of
>> several dozen kcal/mol in just 50000 steps. If I set ifqnt=0, or
>> qm_theory="DFTB", then I get essentially no drift in total energy (i.e.
>> less
>> than 0.1 kcal/mol over 50000 steps. I've attached the prmtop and
>> inpcrd
>> files.
>>
>> Does anyone see what is going wrong here? Our executables pass all the
>> tests,
>> and the fact that DFTB seems to work but PM3 and AM1 fail is
>> intriguing.
>>
>> ...thanks for any help...dac
>
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers

--
Dr. Andreas W. Goetz
San Diego Supercomputer Center
Tel  : +1-858-822-4771
Email: agoetz.sdsc.edu
Web  : www.awgoetz.de
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Apr 05 2011 - 13:30:02 PDT
Custom Search