[QE-users] differently complied QE

Stefano Baroni baroni at sissa.it
Tue Jan 8 04:54:24 CET 2019


you are very right and I commend you for your thoroughness. Only figures that can be reproduced should be reported. Usually, lattice constants can be computed and reported with a higher accuracy than you observe. This being said, as Nicola says, the *numerical* accuracy that can be achieved is usually higher than needed for practical purposes and higher than the accuracy due to “unavoidable” physical approximations (first and foremost the XC functional, but also more fundamental ones, such as Born-Oppenheimer’s).

Coming to the point, that numerical results may depend on such seemingly irrelevant “details" as hardware, mathematical libraries, optimisation level, etc. is not uncommon, as surprising as it may seem, particularly on parallel machines. The result of many “commutative” operations (sums, products) depends in practice on the order in which these operations are performed, due to the non-commutativity fo rounding errors, and this order does depend on all those “details”. Some properties may be more sensitive to these details than others in some systems more than in others. For instance, structural parameters (such as lattice parameters or atomic positions) are more sensitive to the accuracy in stresses and forces in soft systems than in hard ones. Can you tell why?

Bottom line. Please, do the tests that Nicola suggests and make sure you understand the results you get (I am sure you are able to appreciate if and when you understand, which is not so common). In case of need, revert to us.

Good luck!
Stefano Baroni

> On 8 Jan 2019, at 00:49, Aleksandra Oranskaia <aleksandra.oranskaia at kaust.edu.sa> wrote:
> Thank you for the respond!
> I was confused with the difference, because if the 3rd digit is “arbitrary” (depends on libraries, compilers, hardware, etc.) than there is no sense in reporting lattice constants in papers with this accuracy; however, it is a common practice.
> I used:
> conv_thr=1.d-8
> tot_conv_thr=1.d-6
> forc_conv_thr=1.d-6
> cell_dofree=‘volume’ (for the cubic cell)
> I will try to increase conv_thr and press_conv_thr, as well as to compare gnu and intel compillations, and reply back.
> Best,
> Alex.
> ___
> Aleksandra Oranskaia (M.Sc.)
> ChemS PhD student, KAUST
> Phone: +966 50 1335254

Stefano Baroni -  SISSA, Trieste - http://stefano.baroni.me <http://stefano.baroni.me/>, stefanobaroni (Skype) 

And aren’t we all tired of those who claim to know the answer to life, death and the creation being so […] sensitive about their knowledge? [Stephen Fry]

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20190108/3a81a6b1/attachment.html>

More information about the users mailing list