In my simulation i want to break total force felt by a group into electrostatic and van der waal components. However when i do that (using group/group pair no kspace yes), I can see that results vary quite a lot with pppm tolerance factor. I want to know

1) Is there anyway i can make lammps to calculate all coulombic forces in pairwise fashion instead of kspace solver? (to get a 'correct' baseline to see how much error is present in my current simulation)

2) what will be most accurate settings for kspace long range solver (irregardless of speed, something like fix tune for accuracy)?

In my simulation i want to break total force felt by a group into electrostatic and van der waal components. However when i do that (using group/group pair no kspace yes), I can see that results vary quite a lot with pppm tolerance factor. I want to know

with a setting of "pair no kspace yes" you only compute the long-range

contributions to the coulomb potential and are missing the realspace

part.

1) Is there anyway i can make lammps to calculate all coulombic forces in pairwise fashion instead of kspace solver? (to get a 'correct' baseline to see how much error is present in my current simulation)

you need to use "pair yes kspace yes". however, in your production

calculation, this will also include other pairwise interactions, e.g.

lennard-jones. so the only we to separate these out, is to record your

trajectory and then doing the analysis with the "rerun" command, while

setting e.g. the LJ epsilon coefficients to zero.

2) what will be most accurate settings for kspace long range solver (irregardless of speed, something like fix tune for accuracy)?

the smaller the convergence factor, the more accurate. i.e. 1.0e-6

will give you a more accurate coulomb than 1.0e-4. also, using more

bits resolution or turning coulomb tables off, will improve accuracy.

however, there is a limit in the current implementation of the

analytic coulomb, as erfc() is approximated at about single precision

accuracy (~1.0e-7).

but also keep in mind, that the parameterization is usually truncated

at 4-5 significant digits, so while you may improve the numerical

accuracy, you cannot overcome the systematic error of parameterization

and - of course - the model itself.

axel.