Neighbor list - Gromacs vs LAMMPS

Hi All,

First off, I personally am not familiar with Gromacs and the person helping me is not familiar with LAMMPS.

I am trying to use the Martini force field in LAMMPS. My groupmate ran the simulation using Gromacs with a timestep of 20fs and the simulation worked fine. However, when using LAMMPS, I had to go to a much smaller timestep, 5 fs. Even with extended skin and communicate distances 5fs was the largest I could get to. I was wondering if anyone with experience in both Gromacs and LAMMPS could give us some insight into this? We were thinking it has something to do with neighbor list and ghost atom handling. We would like to run these simulations using LAMMPS.

Thank you

There’s not enough info here to offer advice.
What happens in LAMMPS with a larger timestep.

Can you get exact agreement between the 2 codes

running the same problem for some small # of timesteps?


Sorry for the limited information.

In Lammps, with a larger timestep (10-20 fs) bonds, angles and dihedrals get lost, or there are too many bins, basically the simulation fails.

If the simulations are run with a 1fs timestep for a short time, the codes agree.

The simulation parameters I’m using are consistent with the Martini model in Gromacs except with the neighbor building in which I increased the skin distance thinking that the neighbor building was the problem.

pair_style lj/gromacs 9.0 12.0
bond_style harmonic
angle_style cosine/squared
dihedral_style charmm
special_bonds fene

timestep 5 # 20 does not work

neighbor 3.0 bin
neigh_modify delay 0 every 1 check yes
#communicate single cutoff 30.0

fix 1 all nve
fix 2 all press/berendsen iso 1.0 1.0 5000.0 modulus 21000.0
fix 3 all temp/berendsen 300.0 300.0 1000.0

The simulations are being run on dual octacores, and the starting structures have been equlibrated using Gromacs. I’m not 100% what other information is needed for advice so if you have any other questions, feel free to ask.


Can you run for 1 (or a few) large timesteps and get perfect
agreement between the 2 codes? Assuming the atoms are

not moving more than 3 Angs in a timestep, this should
work with a skin distance of 3.0. Also I would try it first
on a single processor, then in parallel. Also, what is the maximum distance
an atom in a dihedral can be from the center atom (atom 2)

in the Martini model? If it is further than the pair cutoff,
then you could need the communicate cutoff setting to avoid

a “missing atom in dihedral” error.


Hi Steve,

Thanks for the suggestions. We’ll give these a try.

The max dihedral distance is around 10.0 Angs.

One other thing that I failed to mention is that if we use fix nve/limit 1.0, then the timestep can be increased to 20 fs. Not sure if that gives you any other ideas.

Thanks again for the help