[lammps-users] General question on lost atoms

Dear lammps users,

I know that the error of lost atoms in LAMMPS occurs due to several problems such as bad initial configuration, wrong setting for the potential parameters, large timestpes, and others.

However, I can understand the error of lost atoms when someone does NOT use periodic boundary conditions .

Regardless of the reasons, how atoms can be lost in a system with periodic boundary conditions (PBCs.) ?

My first thought, on a system with PBCs, is that when there is any reason that makes the atoms move much far, the image atoms should enter from the other copies of the simulation box.

Maybe my question is naive. I would be thankful if you can help me know the correct answer.

Best regards

Lost atoms have very little to do with periodic boundaries but rather with the domain decomposition parallelization in LAMMPS.

The data for atoms is distributed across the MPI processes in LAMMPS. Each process holds information about atoms it “owns” (= local atoms) and atoms from neighboring sub-domains (= ghost atoms). However, for efficiency reasons, those atoms are communicated only up to a certain distance (= communication cutoff). Now, if either an atom moves further than that cutoff between two neighbor list builds, or the subdomains change too much, then LAMMPS will no longer consider that atom as one that is within the “stencil” of atoms that it needs to communicate and after the neighbor list rebuild it is found to be missing. This can happen when atoms leave the simulation box for non-periodic boundaries (since there is no neighboring sub-domain), but also when atoms move extremely fast to very high forces (e.g. due to close contacts), but also after the initial step when using shrinkwrap boundary conditions that are very different from the actual min-max extent of the included atoms.

You can read some background info about this in the 1995 LAMMPS paper and some of Steve’s presentations about the inner workings of LAMMPS from previous workshops.

1 Like