I have been trying to calculate thermal conductivity of above mentioned rigid water models at 300K through LAMMPS in Gree-Kubo approach and for all the models, it over predict the results to the unacceptable range (More than 100% over prediction). I see some literature[1] [2] where they get the results within 20-50% over prediction. My results are as follows,
SPC/E - around 1 W/mK
TIP4P - around 1.2 W/mK
TIP4P/2005 - around 1.3 W/mK
Where the experimental value is around 0.61 [2]
My simulation details are as follows,
Units: real
Time_step: 1fs
pair_style: lj/cut/tip4p/long (for TIP4P) and lj/cut/coul/long (for SPC/E)
Long range Coulomb solver: pppm/tip4p (for TIP4P models) and pppm (for SPC/E)
GK Correlation length: 200ns
Angle and Bond lengths were fixed via fix shake.
Does anyone faced the same problem or wish someone can help me out. I have attached my input script (water.in) and the read_data file (data.lmp) containing water molecules.
[1] - DOI: 10.1063/1.4789961
[2] - DOI: 10.1080/00268976.2018.1562123
Hello Chamara Somarathna,
I did not work on thermal conductivity but worked on viscosity calculation using Green-Kubo method, so I hope my suggestions will be useful for your case. In order to justify your thermal conductivity measurement using Green-Kubo, you need to make sure that the system really reaches its stable state, in other words, the simulation length is enough for your calculation.
As far as I see from your input and data files, you have around 6000 atoms and have a run command for 2000000 femtoseconds. While I was calculating the viscosity, I had around 2000 atoms and it took around 5000000 femtoseconds for my viscosity calculation to converge and have a stable value. I am suggesting you that increase your runtime by one order and check out your values again. Also, you can try to calculate the thermal conductivity more than once by simply constructing a loop, and then you can graph your thermal conductivity values and figure out what is going on.
Regards
Dear Ahmet Burak Yıldırım
Thank you very much for your valuable answer.
As my understanding, velocity auto-correlation needs much more statistical data (longer run time) to be converged since water is in the liquid state ( VACF of solids converges rapidly, liquids takes time and in gases it hardly converges). And also it require a longer correlation time. Again, longer correlation time means we need longer run time. But heat current auto-correlation is the other way around. For a liquid it converges quickly (For rigid modes) since there can't be any long phonon modes. That's why I used 200fs to be the correlation length and 2M steps for the run time. My results were converged finely and have attached here with. Anyway I'll try your suggestion and again thank you very much for your time and opinion on my problem.
Cheers.
Dear Chamara Somarathna ,
I see that the methods differ between viscosity and thermal conductivity calculations. While I was going over the literature, I observed that previously 0.1 fs was specified for the simulation and also the Green-Kubo integral whereas you specified your timestep as 1 fs. In the literature, it was said that the timestep was "reduced" to 0.1 fs for a better resolution of the auto-correlation integrals and etc. If you did not particularly specify the timestep as 1 fs, your result might be different than your expectation due to this difference, in my opinion.
Regards
Dear Ahmet Burak Yıldırım
Yeah, I also came across on that too and they are using 0.1fs of time step for flexible water models like SPC/E-Fw and etc. But for rigid water models, it said 1fs of time step is enough for resolve the motion of the molecules with pretty much accuracy. Anyway did you manged to get the viscosity values of water close to experimental values?
- Badges
- Science topic