Green-Kubo thermal conductivity for graphene

Dear all,

On this forum (thermal conductivity of graphene with green-kubo) I read that

Green-Kubo method is not appropriate for dealing with your graphene nanoribbons.

There is something in the literature (Redirecting) reporting calculations done with LAMMPS on monolayer 2D graphene employing the Green Kubo method and achieving values close to the experiments for pristine graphene.

Can someone with experience in merit comment if the first affirmation is true, and if so, why? Also, this is due to the nature of graphene or nanoribbon geometry?


Hi @lcaputo,

The system in the original post was probably joined in the original mail. This is a mirror of the archived of the now deactivated LAMMPS mailing-list.

The reason the method zas inappropriate is stated in the end of the message:

Your system is way small to calculate thermal conductivity because of the thermal noise.

Without the initial system it is impossible to tell how accurate this statement is but it is correct that for system with too few atoms, thermal noise will make GK method very slow to converge. See how the paper also uses 20 different simulation to determine the thermal conductivity of each graphene system.

Thank you for your reply.

I have noticed the average results over 20 calculation (for each different model). However, one of my doubt here is, the calculations are exactly the same? But because of the noise one has different results, and thus the average?

Of course you need decorrelated initial states for the trajectories to average over. So either you equilibrate the same system with different starting velocities (using a different random number seed) or collect snapshots from the equilibration trajectory toward the end that are at least a few picoseconds apart and then use those to seed multiple calculations. Due to the Lypunov instability, those trajectories will diverge exponentially and thus soon be independent and thus can be used to average over. Check out your favorite statistical mechanics text book for more insight into this.

Thanks. So, if I understand correctly, it would enough to do n calculations with different velocities in the velocity create option. Would it be necessary to start with velocities still close to the ones used in the npt and nvt step (±10K)?

Please understand that this forum is no replacement for talking to your adviser/mentor/tutor about the technical details if you have trouble understanding suggestions that were made.

I thought I made it clear enough that you need to change the random number seed.
Please note that you can have a different distribution of velocities and still the same temperature.