I’m running a box of TIP4P/2005 molecules in a single processor. My understanding is that the thermodynamic properties, coordinates, and velocities should match exactly at the time step at which the restart file was written, and that they will diverge (relative to a non-restarted simulation) thereafter (with thermo properties remaining statistically the same).
However, I am finding that the results do not match exactly at the restart time step (input and thermo files attached). The simulation writes a restart file at timestep = 500 and, after restarting, the pressure (third column from left in thermo.data) is off by a factor of 2; the coordinates are also slightly different (the x coordinate of one of the atoms is shown as the last column). An NVE simulation also shows a discrepancy, albeit smaller.
That cannot be said on such a general level. To get the exact same result at restart strongly depends on which LAMMPS features you are using, whether they are restartable, and how exactly your input files look like. Certain commands must be repeated, other must not.
Since TIP4P is a rigid water potential, you are likely using fix shake and fix shake is not restartable, but will be re-initialized after the restart.
Thank you, Axel. I am re-specifying all fixes (npt, shake) after the restart using the same fix-IDs, but the discrepancy remains. Do these discrepancies look normal to you? The pressure difference before/after restart seems excessive, don’t you think?
Please note that the SHAKE algorithm was conceived for a regular Verlet time stepping scheme, while LAMMPS uses a Velocity-Verlet scheme. To make fix shake work, some kind of extrapolation is necessary, and that is particularly bad at the zero-th step, either when starting from scratch or after a restart (I did mention that fix shake does not support restarting, right). That initial pressure value is not meaningful, you would have to look at the next step.