Phonon transmission calculation error with DFT

[su zhongbo]

Dear Prof.,
I get these errors when I calculate the phonon transmission with DFT.  What's the problem? Looing forward to your help!

** Back Engine Exception : The CSR matrix has no such entry
** Location of Exception : sparsecsr.h:800



** Back Engine Exception : The CSR matrix has no such entry
** Location of Exception : sparsecsr.h:800



** Back Engine Exception : The CSR matrix has no such entry
** Location of Exception : sparsecsr.h:800



** Back Engine Exception : The CSR matrix has no such entry
** Location of Exception : sparsecsr.h:800


rank 5 in job 3  node049_48674   caused collective abort of all ranks
  exit status of rank 5: killed by signal 9
rank 2 in job 3  node049_48674   caused collective abort of all ranks
  exit status of rank 2: killed by signal 9

[Anders Blom]

Strange error - will need to see input script.

[yqxie]

I got the same errors.  I used ATK 2014.2. Did you find the answers?

[Anders Blom]

Still no script provided so we can't test it

[yqxie]

Hello, Dr. Anders Blom,

  attached is the input file and the last few lines of the outputfile. The displacements of atoms for forces calculations  were finished, and then the job terminated accidently with the errors.

And in this case, I want to restart my calculation becasue it takes several days to complete all the forces calculations.   So could you please give me some suggestions to my questions raised in another post "   How to restart a phonon calculation?, and about the parrallelization strategies"

Thanks

[Anders Blom]

A calculation like this should be broken down into several steps, instead of making one huge script, since it's very difficult to follow. Now, in ATK 2014 it's not possible to save the dynamical matrix, so this is hard, but it will be possible in ATK 2015. So, until that is done there is no way to restart a broken calculation (and indeed, even in 2015 there is no checkpoint file for the dynamical matrix).

There are a few problems in your calculation: you are using 1 k-point for the electrodes in C, that is probably a bad idea, and also the central region is really too short, only the 2 middle layers can actually move, and that will not give any sensible results. I suspect that is causing the error you see.

Realistically you probably need to double the central region, which of course means a lot more atoms, and more calculation time.

Therefore, it may not be a good idea to try this in DFT, but on the other hand I  suppose there may be no good classical potentials for "2D arsenic" (if indeed that material even exists in nature...). You could search for one, or try the InAs or GaAs potentials and compare their phonon spectra to DFT pronon spectra for bulk Arsenic.

[yqxie]

Thanks very much for your timely  reply!
I thought that for a periodical system, such as the two-dimensional sheet in my script and any 2D nanoribbons, there maybe an easy way to calculate the transmission spectrum, as the leads and the center region are all the same. So we can just calculate the Hamiltonian, dynamic matrix of one lead, etc., and then construct the green functions and the self energy.  In this way, the workload will be largely reduced for phonon transport calculations.  In the preseent way,  we have to construct a center region which is at least two times larger than the leads, and to rpeat it by at least 3 times in order to calculate dynamic matrix, so the workload is tremendous for DFT calculations.

[Anders Blom]

Well, yes, similarly that you for a periodic system can calculate the electronic transmission directly from the bulk system basically from the band structure, one could imagine doing the same from the phonon band structure. But this functionality is not defined yet. It can probably be done by hand though. If you follow http://quantumwise.com/documents/tutorials/latest/LowLevelEntities/index.html/chap.electrontransmission.html and replace the electronic part with the phononSelfEnergies etc (see the reference manual) you could make your own NEGF method and obtain the phonon transmission this way.

[yqxie]

Hello, according to the reference manual,
calculatePhononSelfEnergy(device_configuration, energy, contribution=Right),
to calculate the selfenergy we need a device configuration. It seems that we can't   provide the infromations from a bulk  calculation since the   "device_configuration" is a parameter must be provided.   I guess that only the informations from a lead (like lead's Hamiltonian, etc.) are required to calculate selfenergy. So if using  "Hamiltonian" and other informatin as augments for " calculateSelfEnergy()",  it will be possible for us to obtain selfenergy and green functions from a bulk calculation by hand. 
Thanks.

[yqxie]

  Hello Dr. Anders Blom,

      I got the same erros in my calculation once again.  I tried to calculate the phonotransmission of two graphene nanoribbons with DFT. The phonobandstructure calculations were properly fininshed, but when doing phonontransmission calculations the jobs terminated  with erros  like "Back Engine Exception : The CSR matrix has no such entry ...".   As the same errors happened for three different calculation exmaples, I was wondering whether there are some problems in my input script or bugs in ATK? 
    Attached are the input scripts for my calculations ,  the output file "out2.log" and the phononbandstructures for two nanoribbons.
    By the way, for the small nanoribbons, it takes 8.5 hours to complete all the dymnamic matrix calculation, using 3nodes with 16 cores, and then the job terminated with errors. I thought you might be able to repeat my calcuation with this smaller system.
     Or,have you ever tested the phonontransmission calculations with DFT ?  If yes, could you please  give me the script so that I can repeat it.
   
Thanks,

xie

[Umberto Martinez]

We are investigating the issue but we are not able to reproduce the error.
It would be helpful to know how you run exactly your script (mpiexec settings) and please attach the whole output log file.

[yqxie]

  I use 3 nodes ( each has 16 cores ), and assign 2 MPI processes on each node (, using OpenMP, multithreads). CPU is Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GH. The output file is too large, so I keep the only the first and last hundreds of lines.  Thanks.  I use mvapich2 with pbs.  Thanks