[gmx-users] Tabulated potentials make newbies crazy

Mark Abraham Mark.Abraham at anu.edu.au
Mon Nov 30 21:44:51 CET 2009


ms wrote:
> Mark Abraham ha scritto:
>  > Sorry, I was a bit incomplete last night. Charge groups are the
>> fundamental unit for neighbour-searching (3.4.2) to construct lists of
>> charge groups for nonbonded interactions, which determine lists of
>> atom-atom interactions. However, the nonbonded interactions are
>> evaluated as nested sums, first over energy groups. So for energy groups
>> of Protein and SOL, the neighbour search finds all pairs of charge
>> groups that are both Protein and inside the cutoffs, and lists them.
>> Then all Protein-SOL similarly, then all SOL-SOL. This requires that the
>> energy groups be a union of only complete charge groups (and I am not
>> aware that this is spelled out anywhere in the manual!). So for energy
>> groups of Calpha, Rest_of_Protein and SOL, it would be necessary to use
>> an individual charge group for each Calpha. This would usually mean it
>> has a net non-integral charge that is equal in magnitude of the charge
>> of the group from which it is taken. It is well known that small charge
>> groups of non-integral charge can then wander back and forth across
>> cut-off boundaries and generate artefacts.
> 
> Ok, thanks for the clarification. This doesn't suggest a trivial
> solution to the problem, quite the opposite: I understood correctly that
> charge groups must be neutral, and this is impossible to do if we put
> each C-alpha as a charge group.
> 
> I can coarse the thing further -that's quite the plan, actually- and
> eliminate electrostatics, but I hoped to have a look at what happens
> with the new potential and getting it right, before going so far.
> 
> So, even if the following:
>>> The problem is that since I have a single molecule now, and the single
>>> molecule must be neutral, so it must be all a single charge group
>>> ("Therefore we have to keep groups of atoms with total charge 0
>>> together. These groups are called charge groups.", 4.6.2).
> 
> is not entirely correct, it is indeed correct that charge groups
> *should* be neutral. Isn't it?

Indeed. See 3.4.2 and ref therein.

>> If you're running in single precision, that precision cannot represent
>> values as large as 10^41. Since in any simulation (but particularly
>> coarse-grained one) non-bonded atoms aren't going to get this close, the
>> values are next to irrelevant. Just choose 10^38 for anything larger
>> than that.
> 
> Right, it is probably precision problem. Thanks.
> 
>>>>> Now, my questions are:
>>>>> - What is the accepted range of values in tables?
>>>> I don't think this is the problem.
>>> It is the least problem probably, given my confusion on energy-charge
>>> groups, but it seems it is too...
>>>
>>>>> - How do I define a steep repulsion potential correctly?
>>>> It's terse, but manual 6.7 seems to have the necessary information.
>>> 6.7 is one of the references I am obviously using, but it gives only
>>> general (even if essential!!) information, nothing speficic on "good" or
>>> "bad" potential shapes/values. But probably that's the least problem :)
>> Knowing a sensible shape is your problem, if you're choosing to
>> unbalance the force field by changing one of the contributing potentials...
> 
> I meant "sensible" in the meaning of "can be interpolated more or less
> faithfully / will be calculated with more or less artefacts" -of course
> if it makes sense in the model is my problem...

The cubic spline interpolation will do a mighty fine job of any function 
that is suitably continuous provided that the density of points is 
sufficiently fine... interpolating a sinusoid with a density comparable 
with the period would obviously be a disaster!

Mark



More information about the gromacs.org_gmx-users mailing list