Re: [Opium-talk] DNL boxes and ghost states

[Ben T. <b.s...@br...>]

Dear Eric

Many thanks for your work fixing the bug with ghost diagnosis for version 
3.6. Now that we can properly detect and avoid ghosts during the 
optimization, the quality of the pseudopotentials we produce seems to be much 
more consistent.

As an example of a DNL pseudopotential optimized (apparently correctly) using 
differential evolution, I attach zinc augmented with 10 boxes that have been 
chosen to minimize the eigenvalue and tail norm errors (per your helpful 
suggestion a little while ago) of each of the listed configurations. Although 
we haven't tested this pseudopotential extensively, the initial results 
appear promising; a CASTEP calculation using the PBE functional produces 
lattice parameters for B4 ZnO of a=3.2537 and c=5.2437 Angstroms (+0.1% and 
+0.7% versus experiment, respectively; about what one expects from PBE). For 
comparison, the best I was able to achieve by manually placing the boxes was 
a=3.267/c=5.264, which is obviously somewhat worse.

Although we're also interested to see how the transferability and/or accuracy 
of the optimized pseudopotentials vary with the choice of optimization 
criterion (eigenvalue and tail norm errors, logarithmic derivative errors, or 
both), the logarithmic derivative optimization is slower because of the need 
to loop over configurations and radii, and so hasn't yet produced any 
definitive results. We'll release the code to the community as soon as we 
have reasonable evidence of its reliability and some coherent guidance as how 
it can best be used.

As an aside, with relevance to CASTEP specifically, the .recpot output from 
version 3.6 seems to be invalid due to a missing file format version line 
immediately after the comment. Since this was present in .recpot files 
produced by OPIUM 3.5, the relevant line in do_recpot.c may have been removed 
by mistake. A patch is attached to add it back into version 3.6.

Best regards

Ben Truscott