Moderators: Global Moderator, Moderator
I’m trying to calculate the thermal conductivity of a crystal structure containing light atoms like lithium. I have a well-trained force field and plan to use it for large-scale simulations. Initially, I’d like to confirm—can I reliably obtain the correct thermal conductivity for such a system using this approach?
Also, I need the thermal conductivity in units of W/(m·K). Could someone clarify the units of the heat flux (in all three directions) provided in the ML_HEAT file?
Dear Paul Tanmoy,
The units of the flux in the ML_HEAT file are (eV A)/fs. I have updated the wiki page to reflect that.
Often, the convergence of Green-Kubo is slow with respect to simulation length, box size, etc. Some tips and tricks are shared in this forum post:
heat flow and thermal conductivity.
Light elements in your system will also make a small timestep (POTIM) necessary. For Hydrogen, that would be around 0.25 - 0.5, for Lithium, you have to do your own testing.
It might be better to use the Müller-Plathe method, which has been available in VASP since version 6.5.0. However, the Green-Kubo method should, in principle, work if your force field is accurate enough and you carefully converge.
I hope you understand that I cannot guarantee that you will achieve the "correct results", especially without knowing more about your material and simulation setup.
All the best, Michael
Respected Michael,
Thank you for your insightful response. I have figured out the unit conversion part for thermal conductivity calculation. Currently I am using Green-Kubo formalism and performing long NVE simulation with Noose-Hover thermostat of a snapshot picking randomly from NVT simulation of my structure. I have attached the plots of heat flux autocorrelation and thermal conductivity, from this thermal conductivity plot can I say that it's heading to a converged value? and to finalize the value of my system's thermal conductivity how many snapshots should I consider for NVE simulation?
Sincerely,
Tanmoy
Dear Tanmoy,
I showed your graphs to a colleague who is more experienced with heat transport calculations, and he thinks your data looks good and your Green-Kubo integral should converge. He also gave a few helpful pointers that I am relaying here:
You should average your results over several MD NVE trajectories.
Compute both the Green-Kubo integral average and its variance. If the variance is not sufficiently small compared to the average value, add more trajectories until it is.
Stefano Baroni has thought deeply about Green-Kubo integrals, their convergence, and heat transport. He also wrote two software packages you could check out: QHeat and Sportran.
Please let me know if you need further support, or if this answers your questions, so I know if I should lock the thread.
Cheers, Michael
Dear Michael,
Thank you again for your confirmation regarding my result. I am planning to average 5-10 independent NVE trajectory results (snapshots of NVT 300K MD run) and for each trajectory ,suppose my simulation runs upto 250 ps, I am averaging last 1000 steps thermal conductivity value of each NVE simulation. And I am aware about Stefano Baroni's works, in future I have plan to use the SPORTRAN code.
Sincerely,
Tanmoy
Dear Tanmoy,
Great that you are familiar with Baroni's work. I cannot really see another question in your last post. Are there any open technical issues you would like to discuss?
Cheers, Michael
Dear Michael,
No, thank you I don't have any query based on this topic. I have a separate technical issue regarding Wannier90, for that I will open a separate question. For now, you can lock this conversation.
Thanks again,
Tanmoy