[lammps-users] Multi-scale Timestep

I understand that i can use respa for implementing multi-scale timestep in evaluating bond, angle etc energies. However what i am trying to do is having different timesteps on different set of bonds and angles. Like bond style ‘a’ ‘b’ and ‘c’ has dt = ‘t1’ and ‘d’ ‘e’ ‘f’ has dt = ‘t2’. Is there a way to do it ? The faster set of atoms has no bonding connection to the slower set of atoms and hence i was wondering if playing with fix and unfix with a loop construct can give me the advantage of having bigger timestep on slower atoms. I appreciate any suggestion to this.


hi arnab,

the point of respa and multi-timestepping in general is to
reduce the number of force evaluations. the assumption is that
the _total_ force of far away atoms does not change a lot, so
you only compute those interactions less frequently. one can
make a similar (but less convincing) argument for bond stretches,
if the atoms moving the most are light and thus move fast and
don't have significant partial charges, e.g. hydrogens in many
all-atom force fields. if you have slow moving and fast moving
particles intermixed, you will still need to evaluate the forces
in every step for all but the slow-slow interactions and that will
predominantly determine the speed of execution, not how time spent
on integrating the equations of motion. the only setup where i
would expect a gain (you scheme contains a lot of overhead), is
if you have only a very small number of fast moving particles.

whatever you do, with multi-timestepping careful monitoring
of energy conservation (without thermostats!) is imperative.


Hi Axel,
Thanks for your reply, The reason i am interested here in the gain of it by multi scale time-step here is as the fast moving particles consists only about 1% of the system by volume and is surrounded by super atoms (mass 7-8 times than the actual atoms) as given in the diagram below . The only interaction between both is non bonded ones.

No - LAMMPS doesn't currently support splitting different bond calcs
across rRESPA levels.