Nudged Elastic Band (NEB, BAND): understanding and improving efficiency

[Daniele Ongari]

Dear cp2k developers, 
I'm working with BAND algorithm in cp2k, to explore phase transition in 
metal organic frameworks.

I'm using a starting configuration and a final configuration (sometimes I 
also add an intermediate configuration to avoid strange geometry 
interpolations) to generate 20 replicas.

Using a DFT/PBE+D3 calculation on ~300 atoms efficiency is very important 
to me and I'm asking your help to understand better how to chose the best 
input:

1) If i use a total of 400 cpus to run 20 replicas, I know that with 
"NPROC_REP 100" the SCF of 4 replicas are done in parallel and is better 
for scalability issues. But I found out that BAND starts with replica 1 2 3 
4 and move to 5 6 7 8 only when the last SCF is terminated, making lots of 
cpus waiting for nothing. Am I right or I misunderstood something? Is there 
an option to avoid this nonsense?

2) Why starting and final configuration need a SCF for each BAND cycle? The 
configuration and the energy appear to be always the same, so why run a SCF 
more than once?

3) In an old post 
(https://groups.google.com/forum/#!topic/cp2k/_icaeFBVnZ4) I found that for 
each BAND cycle there are more than one SCF running without geometry 
change, and this is related to NO_LS and NP_ LS (=2 by default). What is 
the scope of this? Can NO_LS be safely used? Does NO_LS=T correspond to 
NP_LS=1?

4) Fixed cell for NEB is a serious constraint: are you developing a 
variable-cell-neb in the next version? An alternative seems to be USPEX 
which is interfaced with cp2k but not properly documented for it (in my 
opinion). Do you have any input example to share?

Thanks for your help and for your continuous effort to improve this 
wonderful software.

Daniele Ongari
PhD candidate, LSMO, EPFL Switzerland
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20160727/13fbc6a7/attachment.htm>