Hi all,

Thanks for point out my incorrect conclusion, Axel, i didn’t realize that ewald sum in lammps has erfc() too, i just look at lammps doc, and got that incorrect conclusion. sorry for that.

regard to Stan’s suggestion, I have used (pair_modify table 0), the energy difference is still there, (higher table number didn’t work, either). the relative error is about 0.009%. and i look at some post emails about this topic, it seems that for high precision energy evaluation, the erfc function is the major problem. Do we have a simply way in lammps to call the exact erfc(), instead of using a polynomial, also, i would like to know roughly how much will the simulation time increase due to the use of exact erfc().

Many Thanks for the help from both of you.

Best Regards

Jiasen Guo.

Hi all,

Thanks for point out my incorrect conclusion, Axel, i didn't realize that

ewald sum in lammps has erfc() too, i just look at lammps doc, and got that

incorrect conclusion. sorry for that.

regard to Stan's suggestion, I have used (pair_modify table 0), the energy

difference is still there, (higher table number didn't work, either). the

relative error is about 0.009%. and i look at some post emails about this

topic, it seems that for high precision energy evaluation, the erfc function

is the major problem. Do we have a simply way in lammps to call the exact

erfc(), instead of using a polynomial, also, i would like to know roughly

how much will the simulation time increase due to the use of exact erfc().

this is a complicated matter. the exact erfc() function (from linux

x86_64 glibc version) is *extremely* costly.

increasing the long-range coulomb numerical accuracy is on the TODO

list and i have a custom branch (with an older version of LAMMPS,

since i didn't have time to update it in a while) that implements a

much faster approximation, which is about as fast as the current

polynomial expansion but with comparable (double precision) accuracy.

https://github.com/akohlmey/lammps/tree/coulomb-analytic-double

this converts some, but not yet all coul/long styles to the new

scheme. it will eventually become part of upstream LAMMPS (and we'll

work on improving the accuracy of the tabulation as well, but that is

far out in the future).

axel.