Using restart file for resuming simulation

LAMMPS LAMMPS Mailing List Mirror
1

Dear All,

It will be great if someone answers my question.

Suppose we have 2 runs for evaluating surface tension of water vapor/liquid equilibrium system.

First one : is for 5 ns . Surface tension is evaluated over the last 2 ns.

Second one: Same system with continuous run for 3 ns from the start , followed by another 2 ns run using the restart file of the previous 3 ns run. For this system as well, the surface tension is evaluated over the last 2 ns

Should the results be the same in both the scenarios? Or does using the restart file to resume the simulation give different results?Being new to Lammps, I am confused.

ns stands for nanoseconds.

Thank you.

Sincerely,

Neha, India

2

They should give statistically identical results.

They may not be exactly the same for various

reasons, many of which are discussed on the read_restart

doc page.

Steve

3

Dear All,

It will be great if someone answers my question.

Suppose we have 2 runs for evaluating surface tension of water
vapor/liquid equilibrium system.
*First one :* is for 5 ns . Surface tension is evaluated over the last 2
ns.

*Second one:* Same system with continuous run for 3 ns from the start ,
followed by another 2 ns run using the restart file of the previous 3 ns
run. For this system as well, the surface tension is evaluated over the
last 2 ns

Should the results be the same in both the scenarios? Or does using the
restart file to resume the simulation give different results?Being new to
Lammps, I am confused.

​this is less a question about LAMMPS, but on MD in general. i suggest you
grab an MD text book and read up about statistical errors ​in MD,
statistical efficiency, statistical relevance and convergence of results.
you should also check out lyapunov stability (or instability), more
popularly known as "the butterfly effect".

a simple strategy to get some sense of convergence and statistical errors
of your computed properties is the following:

suppose you have 2ns worth of production simulation data. then you do your
analysis for chunks of 250ps each, then 500ps each, then 1ns each and 2ns.
statistically, you should expect a larger variance of the small chunk
results and those should converge to the 2ns result. if the variance is
"too large", you may need a longer production simulation, if the results
are very close, you have already collected sufficient simulation data
(assuming there are no "rare events" that trigger a significant change).

​axel.​

4

Hi Axel,

I did my analysis for 2 ns worth of production run, as you explained above, and I got something like below,


 Avg Surface tension (mN/m) Standard deviation variance
250 ps each
 64.3865 1.7753 3.1518
500 ps each 64.3864 0.9844 0.969
1 ns each
 64.3865 0.784 0.6147
2 ns each
 64.3862

So, I believe that I have collected sufficient simulation data.

Thank you so much.

Thanks Steve.

Best Wishes,

Neha