Thermal conductivity: Muller-Plathe vs Green Kubo

Dear all,

I have done a lot of simulations of thermal conductivity in carbon nanostructures using 3 methods using both LAMMPS (airebo potential) and my own MD simulator (Brenner Potential).

  1. Fixed-temperature reservoir with rigid boundary
  2. Muller Plathe with periodic boundaries
  3. Green Kubo (LAMMPS heat/flux), using only the z component of the conductivity vector.

My own simulator did not have Green-Kubo implemented, so I am very interested in using LAMMPS for my work.

My results using all 3 methods and the 2 potentials/simulators all show qualitatively similar results (300-400 W/m-K for a (10,10) single-walled nanotube @300K).

I was wondering if it seems reasonable that the Green Kubo method would have similar results. I had expected a more accurate result because it is using the autocorrelation of velocity over all atoms in the simulation at a constant temperature.

I had a lot of trouble choosing an appropriate correlation length because it seems to be unstable depending on the seed used for velocity calculation. I ended up getting fairly stable results with ~25 ps, but I am testing further just to make sure it is actually stable. Another email on this list from a couple years ago expressed similar problems up to about 500 ps correlation length.

Has anyone had experience relating the various methods of determining thermal conductivity using NEMD and EMD. I would like to get an idea of what I might expect because I am less familiar with the Green-Kubo formalism at this time. Of course, the complexity here may be a result of using the non-bonded interaction of Airebo or the 1-dimensionality of the conductor. If you have any thoughts I would appreciate the discussion.

Derek Thomas

I used EMD with Green Kubo for a couple of different fluids based on the thermal conductivity code in the manual. Correlation length and sample interval etc values in manual can be reliable as I could more or less achieve consistent results with experiments using them. Also I changed those parameters a bit but significant difference didn't show up and it makes sense because may be we shouldn't expect very large deviations due to statistical formulas once we set up the problem dynamics accurately. If you think you arranged the dynamics, potential and ensemble correctly may be you should be careful if you are calculating heat current vector correctly.


2013-03-08 06:57, Derek Ashley Thomas yazmış: