Request for Support of SW Many-Body Potential

[qzchen2022]

Description

Dear HOOMD-blue Developers,

The SW potential is widely used in simulations of various materials, including models of carbon, silicon, and even coarse-grained water. For reference, here is the relevant documentation from LAMMPS:

https://docs.lammps.org/pair_sw.html

Proposed solution

Would it be possible to support the SW potential in HOOMD-blue? Alternatively, could you provide guidance on how to implement this potential, similar to the templates used for extending pairwise interactions?

Additional context

No response

[joaander]

@DomFijan explored this possibility. If I recall correctly, the conclusion was that SW has a substantially different form than the existing many-body potentials in HOOMD-blue. To implement it, you will likely need to start with an entirely new implementation that inherits ForceCompute.

[DomFijan]

I think Josh is right. I made an attempt at this some years ago on the following branch https://github.com/glotzerlab/hoomd-blue/tree/feat/SW_potential . The other 3 many bode potential implementations are all in the same file ( https://github.com/glotzerlab/hoomd-blue/blob/1a8a0d73ae4a16984be3923859bc33472501de34/hoomd/md/PotentialTersoff.h ) controlled by an if statement in the evaluator. I concluded that for some reason I couldn't follow the architecture for those 3 potentials and that I must start from scratch (don't remember exactly why). I started with new files based of Tersoff and made some initial modification, but didn't get far.
Perhaps it would be wise to rethink our general strategy for the many body potentials in HOOMD-blue (since we only have a few) and see if we can unify the three potentials we currently have in a cleaner way, and then implement SW with minimal code duplication. This would certainly require some investment to figure out and plan how to best do this.
For now SW can always just be implemented as a new standalone ForceCompute, as @joaander suggested.

[qzchen2022]

@joaander Thanks for your response. I now understand that adding SW isn't a simple matter...

I have an additional question regarding the Tersoff many-body potential. I noticed that the documentation emphasizes: _"The implemented smoothing function is exponential in nature as opposed to the sinusoid used by..." However, I also saw that in the Tersoff code, there’s a commented-out implementation of the sinusoidal form.

'''
// Scalar cutoff_x = Scalar(M_PI) * (rij - r_shell_mid)
// fcut_ij = Scalar(0.5) - Scalar(0.5) * fast::sin(cutoff_x);
// dfcut_ij = Scalar(-M_PI / 2.0) / cutoff_shell_thickness
// * fast::cos(cutoff_x);
'''

Is this because the exponential form is more efficient? I understand that I can fit the parameters to approximate the sinusoid form — I'm just trying to understand the reasoning behind.

[joaander]

Is this because the exponential form is more efficient? I understand that I can fit the parameters to approximate the sinusoid form — I'm just trying to understand the reasoning behind.

I have no idea. That code was written by Brandon Smith in 2011. You could check the commit logs made at the time and see if they offer any clues, or you could try the two forms for yourself and see how they perform on modern hardware.