Page 1 of 2
Ultralong computational time in The Real Time Collisions calculations
Posted: Wed May 24, 2023 9:41 am
by Dean
Dear all,
I am calculating the collisions for a 2D material, but the calculation lasted a very long time.
The out file remains unchanged for more than 3 days.
The end of out file is shown:
Code: Select all
[WR./2D_WR_WC//ndb.COLLISIONS_HXC_header]---------------------------------------
Brillouin Zone Q/K grids (IBZ/BZ) : 112 112 112 112
RL vectors : 37831 [WF]
COLLISIONS name : HARTREE+SEX
Total number of COLLISIONS : 19152
COLLISIONS kind : 1327
Bands : 25 42
Bands mixing : 100.0000
CV only scattering : no
Collisions Cutoff : 0.000000
Hartree RL vectors : 13391
Exchange RL vectors : 13391
- S/N 006038 ---------------------------------------------- v.05.01.00 r.21422 -
I find that the "ndb.COLLISIONS_HXC_header" is wrote successfully, but yambo is writing the file "ndb.COLLISIONS_HXC" for 3 days.
So, is it reasonable? Is there any way to speed up this calculation?
My input and output files are attached.
Thanks in advance.
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Fri May 26, 2023 3:54 pm
by claudio
Dear Dean
did you parallelize the calculation?
how many k-points?
try to reduce the number of plane waves and bands.
Just for info we plain to implement some faster but a bit less accurate method in the new versions
best
Claudio
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Sat May 27, 2023 3:02 am
by Dean
Dear Claudio,
Thanks for your reply.
First, I did not parallelize calculation. Is there some tutorials?
Second, there are 112 k-points.
Third, if I reduce the number of plane waves and bands, how can I get accurate results in the following calculations such as SHG？
Thanks in advance！
Best，
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Tue May 30, 2023 8:20 am
by Davide Sangalli
Dear Dean,
the generation of collisions may take a long time, depending on the parameters.
I'd suggest to first start with low energy cutoff, few k-points and few bands, and see if the run completes fine.
Later you can increase the parameters.
The automatic parallelization should work fine (I see you are running on 44 MPI tasks)
In the log file the strategy automatically selected should be reported.
The code will first create a very large empty file (2D_WR_WC/ndb.COLLISIONS in your case), and then it will progressively fill it.
Was this file created for your run, and which was the size ?
Best,
D.
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Wed May 31, 2023 1:44 am
by Dean
Dear Davide Sangalli,
Thanks for your reply. This calculation is finished successfully and takes about 4 days.
The size for ndb.COLLISIONS_HXC is 5.2 G and 320 K for ndb.COLLISIONS_HXC_header.
I do start the calculation of collisions with low energy cutoff, few k-points and few bands.
The question is how to choose these parameters to get reasonable collision files? In other words, how to test the convergence of collision?
Best,
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Wed May 31, 2023 9:09 am
by Davide Sangalli
Dear Dean,
the TD-HSEX is equivalent to the BSE in the low field intensity limit.
Accordingly, a correct modelling of the absorption of the laser pulse to first order, e.g. in the linear regime, is obtained by using the same parameter of a converged BSE simulation.
The second suggestion is to set in your input file for the collisions generation
Code: Select all
HXC_Potential= "HARTREE+SEX+CVONLY"
This will include only the scatterings between valence (v) and conduction (c) states (e.g. in the "cv" channel needed to model the physics of excitons), while neglecting the terms in the "cc" and "vv" channel. As a consequence the size of the ndb.COLLISIONS file will be reduced.
Best,
D.
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Thu Jun 01, 2023 1:51 am
by Dean
Dear Davide Sangalli,
Thanks for your suggestions. I will do some tests.
Best,
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Tue Jun 06, 2023 3:07 am
by Dean
Davide Sangalli wrote: ↑Wed May 31, 2023 9:09 am
Dear Dean,
the TD-HSEX is equivalent to the BSE in the low field intensity limit.
Accordingly, a correct modelling of the absorption of the laser pulse to first order, e.g. in the linear regime, is obtained by using the same parameter of a converged BSE simulation.
The second suggestion is to set in your input file for the collisions generation
Code: Select all
HXC_Potential= "HARTREE+SEX+CVONLY"
This will include only the scatterings between valence (v) and conduction (c) states (e.g. in the "cv" channel needed to model the physics of excitons), while neglecting the terms in the "cc" and "vv" channel. As a consequence the size of the ndb.COLLISIONS file will be reduced.
Best,
D.
Dear Davide,
According to your suggestions, I set the parameter, HXC_Potential= "HARTREE+SEX+CVONLY". And, the generation of collisions may take a much shorter time. But, the following calculated SHG resuts show a litter different.
Please see the Figure. In the low energy (below 0.5 eV) and high energy （3.5-4.5 eV）region, the SHG coefficient shows big differences. I am confused about the results and wonder which is more resonable? Thanks in advance.
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Tue Jun 06, 2023 8:43 am
by Davide Sangalli
Dear Dean,
the "CVONLY" collisions are meant to reproduce linear response, and thus a way to get the convergence parameters for the linear response.
Beyond the linear regime, it is an open questions what happens.
First it is not sure that the convergence parameters, which are good for linear response, will also work fine beyond.
Second, it is perfectly possible that "CC" and "VV" collisions matter, and I'm not sure how to interpret this.
The effect (and the implementation) should be tested.
Best,
D.
Re: Ultralong computational time in The Real Time Collisions calculations
Posted: Mon Jun 26, 2023 2:17 am
by Dean
Dear Davide,
These days, I want to rerodce the linear response based on the RT-BSE mathods.
However, these is always a difference (0.25 eV) of the first peak energy between RT-BSE (1.27 eV) and G0W0-BSE (1.52 eV).
I wonder how to improve my calculations of RT-BSE?
Any Suggestions are greatly appreciated. The input and output files are attached.
Best,