Question about the `gpu_correct_momentum` function

[ghost]

I found that when removing the COM moments in the Langevin thermostat, the total moments of the simulated system is equally divided into number_of_atoms parts and subtracted from the velocities of each atom. https://github.com/brucefan1983/GPUMD/blob/bca0d9db3b3cfc8c2fc9a5280df70a3784608c5f/src/integrate/ensemble_lan.cu#L174 When the system contains only one type of atom, this function will work well. But once the masses of the atoms are not the same, this method may modify the relative velocities between atoms (since the velocities subtracted from each atom are not the same), thus the internal motion of the simulated system will not be the same before and after the removal. So I feel that it would be more reasonable if we could first calculate the COM velocities, and then subtract the three velocities equally from the velocities of each atom.

[ghost]

For example, this is the Implementation in OpenMM:

https://github.com/openmm/openmm/blob/fd13a655ec58421a63ef3a97394f4b019dc51706/platforms/reference/src/ReferenceKernels.cpp#L2928

[brucefan1983]

You are right, the function zero_linear_momentum in src/main_gpumd/velocity.cu has this problem too. Need to be fixed.

[brucefan1983]

If you are interested in this, you can fix it, otherwise I can fix it at some point.

[ghost]

I'm sorry about this, but I'm literally a noob in CUDA programming and do not know how to write kernel functions. So maybe I have to trouble you to fix the bug 😢

[brucefan1983]

I have created a PR #172 to fix this issue. Could you review it?

[ghost]

I think the new code is good enough to merge!

[brucefan1983]

Thanks!