12
$\begingroup$

As I pointed out in the other question Convergence test I: same atoms and different structures, each time you start a DFT calculation you need to do a convergence test.

Normally, we use the system energy, but this energy per se, is meaningless.

The question now is: What is the adequate property to do the convergence test? Should we use another calculated property instead of the system energy?

$\endgroup$
11
$\begingroup$

The question you should be asking yourself is: what properties do I want to compute? You, naturally, want to make sure that property is converged. For instance, if you wish to compute the band gap, then it makes more sense to confirm that your band gap is converged with your given settings. The band gap could very well be more (or less) sensitive to a given parameter than the absolute energy, in which case your convergence test on the absolute energy may not translate well for the property of interest. There is no inherent reason to do a convergence test on the absolute energy -- it is just a commonly computed property and generally will indicate that other properties are well-converged, but that's certainly not a guarantee.

As an aside, there may be cases where converging the absolute energy is recommended but not required. If you are computing a reaction energy or any other energy for that matter, what you are often reporting is an energy difference. In many cases, there is significant cancellation of error that occurs, which means that the energy difference will often converge faster than the absolute energy.

Interestingly enough, there are even subtle cases with certain codes where it is impossible to converge the absolute energy. I found this to be the case in VASP when using the LASPH flag, which determines whether to include non-spherical contributions related to the gradient of the density in the projector-augmented wave spheres. When enabling LASPH, the absolute energy never converges with respect to increasing plane-wave kinetic energy cutoff because the non-spherical contribution is always changing. In this case, the only option is to check that an energy difference is converged.

$\endgroup$
6
  • $\begingroup$ I got it. It is worth to mention that using any other properties will increase substantially the computational time :( $\endgroup$ – Camps May 3 '20 at 18:14
  • $\begingroup$ Increase the computational time how? In terms of doing the convergence test or in terms of the settings to use for further DFT calculations? If the former, that's certainly not always the case -- band gap is just as easy to compute as energy, relative energies are only 2 absolute energy calculations, and so on. If the latter, well, that might mean you need to use more computationally expensive settings. Unfortunately, that is sometimes the reality. However, you should think about what properties are most important and make sure they are converged. $\endgroup$ – Andrew Rosen May 3 '20 at 18:17
  • $\begingroup$ Well, in my case, using SIESTA and a system with 120 atoms, a geometry optimizations took around 12 hours. The band calculation for the optimized geometric took 3-4 hour. When doing the convergence test for k-points, I used the optimized geometric and freeze it. If I run all the calculation (geometry optimization + band calculation) for each k-point or mesh cut-off, I have to multiply the time (16 hours) for the number of points used. So, the computational time will increase substantially. $\endgroup$ – Camps May 3 '20 at 18:38
  • $\begingroup$ @I.Camps. How did you choose the mesh cut off for the geometry optimization? $\endgroup$ – Thomas May 12 '20 at 16:05
  • $\begingroup$ @Thomas, normally, I use a bash script that, inside a for, change the value of the mesh cut-off, write the input file with the new values, run the calculation, extract the system single point energy value, and write a file with mesh cut-off vs energy. Then, plot the graph and visually inspect for a plateau. Normally, I choose a value higher than the indicating in the plateau. Then, with this value of mesh cut-off, I run the geometry optimization. $\endgroup$ – Camps May 12 '20 at 16:11
7
$\begingroup$

The hydrostatic pressure on a given unit cell is an experimentally observable property very sensitive to numerical precision. If this property is converged many other less sensitive properties will be converged too. Hence checking for convergence of pressure will be a better bet to make sure that other properties are also converged

Reference

Source

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.