I've implemented this method to simulate material and adjust dt for a target acceptance ratio. For a system of thousands of atoms I get a reasonable time step of ~2fs for an acceptance ratio of 0.3. However, when I perform the same simulation but with a system size of about 8K atoms, the time step drops to 0.8 fs.
I use the velocity Verlet algorithm as follows:
For each time step:
1- perform half-update velocities
2- update positions
3- compute forces
4- perform half-update velocities
5- update volume (in NPT ensemble)
I would appreciate some hints here.
Thanks you.
There can be many reasons depending on algorithm details. But I would start from plotting the distribution histogram of $\Delta E$ generated by trial moves. It should be wider for larger systems. If on each step you move all atoms of the systems, than it is reasonable: more atoms moved - larger energy change. Larger energy change - lower acceptance probability.
If your algorithm give the same $\Delta E$ distribution for all system sizes than I would look at forbidden configurations (configurations rejected before Metropolis acceptance criteria check) if you have onces of cource.
Thank you Pavel for your reply.
Yes, the histogram of accepted $\Delta E$ is larger.
I actually found a little bug, my code runs smoothly now.
Effectively, I compare the energy change $\Delta E$ with KT, so it's size dependent.
However, if I compare with NKT it becomes totally unstable.
I can't figure why.
Is there a way to make the metropolis algorithm independent of the system size?
Note that I'm doing NPT simulation so I consider the Gibbs energy $\Delta G$ =$\Delta E + P*\Delta V - NKT*ln(V2/V1)$
- Badges
- Science topic
- Similar topics
- Computer Communications (Networks)
- Simulators