I am simulating an irradiation event by selecting a random atom, PKA, and assigning it a much faster velocity than other atoms around it.
I have implemented a variable timestep so that it can capture this high velocity by reducing the initial timestep. But of course, now the timestep becomes really small and while it is useful for the PKA and atoms interacting with it, it is not useful for all the other atoms in the cell, that remain at room temperature.
Is there a way to deal with that? So that I could decrease the number of force calculations per step and try to decrease the time of my simulation
Did you check the rRESPA multi-timescale integrator (run_style respa) ?
I don’t think there is a ready-to-use solution available in LAMMPS. The problem is not so much about updating atom positions, but in computing the forces, which is the major time consuming step in MD.
Adjusting the time step is straightforward: you look for the fasted moving atom and change the step accordingly. But to avoid force computations you need to determine all atoms that move faster than the default (but with averaging, since all atoms are vibrating) and their neighbors.
There is a way to do r-RESPA with groups which was devised for all-atom / coarse-grain simulations where coarse-grain coarse-grain interactions need to update their forces less often.
Using r-RESPA for your system would require a lot of testing and validation to ensure there are no artifacts created by multi timestepping. I am also not certain if r-RESPA is compatible with fix dt/reset. More importantly is probably the question whether this would be compatible with many-body potentials.
Since cascade simulations are rather short in comparison to most MD simulations, it is probably simpler to just throw more compute capability at the problem. You can parallelize individual runs, but also running multiple cascades can be trivially paralellized by running them concurrently.
P.S.: mind you, in r-RESPA all atom positions are updated in every step. Only the force computation is performed at different levels and for the higher levels a copy of the forces is used instead of recomputing them.