A question about optimization

[mldavidhuang]

In the tutorial, the optimization is conducted under the sing zeta basis set to save the time, in practice I have try optimization of a bilayer system consisting of 150 atoms with the single zeta polarized basis set and find the process really time consuming. So is it acceptable to conduct the optimization using simple basis set while conducting analysis using more complex basis set? Optimization under simple basis set will influence the accuracy of the analysis result?

[Anders Blom]

SingleZeta is not recommended for any serious work. It very hard to say exactly how sensitive the geometries are to the basis set, however. The best thing would be to test it for a subsystem system; for instance, if you are looking at some nanotube with a defect, just as an example, then you can compare the nanotube unit cell relaxed with different basis sets. The unit cell is small, and shouldn't take long even with DoubleZetaPolarized.

[mldavidhuang]

But how to choose the subsystem, if the system I am studying is a system with a big molecule conneted with the nanotube, should the subsystem include the whole molecule with unit cell of nanotube or , just just one atom of all the elements in the big molecule is OK？

[Anders Blom]

The molecule plus some part of the tube will under any circumstance be what you want to relax. There is rarely any need to relax the entire system, since the tube, far away from the molecule, will not be affected by it.

[mldavidhuang]

1.You mean the choice of the subsystem or the way we relax the entire system? Is it possible the keep choose the range of relaxation in a unit cell constrained in only part of the system ?
2.I don not quite understand the constrain cell choice in the optimization panel,could you explain to me? when it is necessary to uncheck it?
3.Is it necessary to relax system under each bias when calculating the I-V curve?

[Anders Blom]

1. I'm not sure I understand your comment... If it's a large molecule on a nanotube, it will be hard to relax anything except the full molecule + the region of the tube to which it connects
2. You uncheck those if you also want to relax the cell, which probably is not necessary for your case
3. Most likely the effect of the bias on the geometry is very small, and often only necessary if you are studying that effect specifically

[mldavidhuang]

Sorry for poor statement! For the first question, I mean the whole molecule+the region of the tube to which it connects is the subsystem I could use to test the basis set, after that, I use the basis set which combines accuracy and time-efficiency to relax the whole system or I just relax the whole system with the part far away from the molecule constrained?

[Anders Blom]

Yes, except I doubt optimizing the whole system is worth the effort, the main relaxation effects will occur right where the molecule binds.

If you want to have an extremely accurate basic nanotube, then just relax the single unit cell of the tube, which again is a small subsystem.

One thing to keep in mind is to use the same basis set used in the optimization for the main calculation of the transport properties, as otherwise there is a mismatch, of sorts, between the structure and the transport properties. Similarly, it's not always the best idea to relax with one software and run the calculations with another, since the force minimum is not necessarily precisely the same, and if your electronic structure is sensitive to such effects, it's best to use the minimum of the code at hand, as long as the geometry is reasonable. This is fact why we relax the structure, instead of using experimental bond lengths or lattice constants; LDA might predict a slightly different structure, and the most accurate calculation is that which is consistent within itself, rather than that which necessarily uses the experimental structure.

[mldavidhuang]

Your reply is really instructive! In the past, I have used single-zeta polarized to relax the system in order to save the time while calculating the transport properties using double-zeta polarized. So according to what you said, my result may be inaccurate in some way.

I did so because I found the relaxation costs much more time than transport calculation.Is this the usual case? What if the case is that I can afford the transport calculation using double-zeta polarized while relaxation using double-zeta polarized may be too-much time consuming? 

[Anders Blom]

Indeed the relaxation is time-consuming, but not necessarily more than the transport, if you only relax subsystems. The trick is to avoid relaxing the full structure, although it's not always possible.

Calculations of this type are time-consuming, it's the name of the game. We try to make the code faster with every release, and we have succeeded in doing so rather well historically. I believe ATK today is at least 10x faster than version 2.0, and we have more ideas to implement. Also, running in parallel, unless you are doing so already, can mean a difference of another factor 4-5. It's also worth noting that ATK is fast compared to e.g. plane-wave codes.

[mldavidhuang]

Thanks sincerely for your answer!