Question on temperature

[ipsecog]

I noticed that the default electron temperature is 300 K. In the example on optimizing a two-probe (manual chapter "Optimizing a two-probe system") the temperature is set to 1300 K, however I didn't see any comment on that choice. Is that a reasonable choice (1300 K)? Would a lower temperature hinder the optimization, and are there any drawbacks to choosing a higher temperature such as 1300 K? I was just trying to get a feel for that parameter, I'm using 1300 K at the moment since I'm not sure what else to use.

[Anders Blom]

The higher temperature is something to be careful with. Unless there is any particular reason to increase it, one should stick to the default (or lower).

In some cases, the default temperature does however cause problems for the convergence, because it gives a relatively sharp Fermi function. This can make the SCF loop take a lot of steps, or it doesn't converge at all. In this case, you can smooth the Fermi function by increasing the temperature to 1000-15000 K, and often this will mean convergence is much better.

However, doing so actually influences the results, because you smear out the transport properties over a much larger energy interval. Therefore, the proper approach in this case is to take the results (i.e., the converged density matrix, contained in the NetCDF checkpoint file) obtained from the high-temperature calculation, and use those as a starting point for a low-temperature calculation, and in this way "anneal" the system. Usually the low-temperature calculation is able to converge once it has a reasonable starting guess, as provided this way.

So, why the higher temperature in the example in the manual? Well, basically, it was chosen to make the calculation converge fast and easily. It's not a really physically proper system anyway, and the focus of that chapter is how to relax a two-probe system, so the parameters were chosen as conveniently as possible...

[zhangguangping]

The higher temperature is something to be careful with. Unless there is any particular reason to increase it, one should stick to the default (or lower).

In some cases, the default temperature does however cause problems for the convergence, because it gives a relatively sharp Fermi function. This can make the SCF loop take a lot of steps, or it doesn't converge at all. In this case, you can smooth the Fermi function by increasing the temperature to 1000-15000 K, and often this will mean convergence is much better.

However, doing so actually influences the results, because you smear out the transport properties over a much larger energy interval. Therefore, the proper approach in this case is to take the results (i.e., the converged density matrix, contained in the NetCDF checkpoint file) obtained from the high-temperature calculation, and use those as a starting point for a low-temperature calculation, and in this way "anneal" the system. Usually the low-temperature calculation is able to converge once it has a reasonable starting guess, as provided this way.

So, why the higher temperature in the example in the manual? Well, basically, it was chosen to make the calculation converge fast and easily. It's not a really physically proper system anyway, and the focus of that chapter is how to relax a two-probe system, so the parameters were chosen as conveniently as possible...

Hi,Anders Blom,
Since the temperature does not corresponding the physical temperature of the system,it is only a convergence parameter,how can we know 300K is reasonable in fact we should have the result of 0 K.
I now have a system does not converge at negative bias,for there is transmission peak at the Fermi level.Only I increase the temperature to 600K can solve this problem,and use the converged result to initialize the 300K calculation,but it seems not to converge at all.So can I safely use 600K for my calculation.Interestingly,the convergence is good under positive bias for 300K.
Any suggestions?

[Anders Blom]

Is this for 2008.10 or 10.8? We are quite interested in collecting systems that don't converge in 10.8, so we can improve the convergence. We have some ideas about pre-conditioning that might help, but we need good test systems. So, please send it to me, if you don't mind, either by mail or post it here.

Otherwise and in general, you have the right idea about the temperature. The relevant comparison is between the temperature broadening (25 meV @ 300 K) and the spacing of the molecular energy levels or the distance between or sharpness of features in the density of states in the electrode.

As you have noted in another post, T(E) may depend on this temperature, although it is not expected to have a major influence on the peak structure and total current. But in the end, you have to state the used parameters and motivate the choice (by checking their influence at least in a few cases); it's hard to give any general advice.

[zhangguangping]

Is this for 2008.10 or 10.8? We are quite interested in collecting systems that don't converge in 10.8, so we can improve the convergence. We have some ideas about pre-conditioning that might help, but we need good test systems. So, please send it to me, if you don't mind, either by mail or post it here.

Otherwise and in general, you have the right idea about the temperature. The relevant comparison is between the temperature broadening (25 meV @ 300 K) and the spacing of the molecular energy levels or the distance between or sharpness of features in the density of states in the electrode.

As you have noted in another post, T(E) may depend on this temperature, although it is not expected to have a major influence on the peak structure and total current. But in the end, you have to state the used parameters and motivate the choice (by checking their influence at least in a few cases); it's hard to give any general advice.

I have use 08.10. Yes,after I have post the information I saw your reply here.http://quantumwise.com/forum/index.php?topic=315.msg1811#msg1811.There are two influences for temperature.One is for the SCF which we discussed many times for non-converged system. Another influence is on the T(E),and then on the current.
For my situation,there is T(E) at Fermi level,and it is that cause the non-convergence.You have also mentioned lower the temperature may also solve the non-converged problem.But for my situation where there is Dos and T(E) at Fermi level,it is useful?
You have also mentioned we could first increase the system and then anneal it (I am testing this method,but is seems useless).But I can not affort the time.Even if I increase the temperature, the convergence can be achived,but it also needs a lot of SCFs.
If I use 600K for calculation instead 300K,does it make sense for the calculation? I think as long as I use 600K for all may calculation, the results can be compared among them.And the current from different bias can be compared too.

Am I right? By the way, thanks for your quick answers.

[Anders Blom]

I think that should be fine. I hope you will have opportunity to try 10.8 and its better convergence later on.