My Community

Hi all,

Due to the wallclock time limit on the cluster we have access to, for a rather large periodic system we can only finish one iteration (NELM=1, i.e. G0W0) but cannot finish all 4 iterations (NELM=4) of a GW0 calculation . So I try to update the WAVECAR after a G0W0 single-shot calculation. Then I continue from that to run another G0W0 (NELM=1), hoping that it gives me a G1w0 result. This is repeated two more times, hoping that G2W0 and G3W0 are obtained. The last step would essentially give me GW0 result if the calculations are done as expected. Note that at the end of each step, the WAVECAR is updated for the next iteration. However, I compared the final result of such a step-by-step method with the GW0 result obtained using NELM=4 for a small system as the test case and found that the results of the step-by-step method are incorrect. So here I'd like to ask vasp experts to give me a suggestion how to break down the 4 iterations into a step-by-step method so that I can obtain the GW0 results within the wallclock time limit. Anything to change in the INCAR or the source code?

Thanks a lot.

huangj3

As you checked this way it is not working properly. However, using LUSEW=.TRUE. it should go.
Start the first calcualtion with INCAR as you did, then in next three steps use LUSEW=.TRUE.

Hi Admin,

Thank you for your reply. I tried the 2nd step with that and saw that in the OUTCAR LUSEW=T (use screened WXXXX.tmp files instead of (screened) HF kernel).

Unfortunately, this still does not work, so I did not try the 3rd and the 4th steps. Note that I started from the WAVECAR, WAVEDER, and WXXXX.tmp files from the first step of G0W0 calculation. Any other suggestions?

Thanks a lot.

huangj3

The step by step calculation requires setting of additional parameters which are not documented.

Dear Admin,

Do you mean that in addition to the LUSEW=T you said earlier, I need some other tags in order to run the step-by-step GW0 calculations? Is it possible you please tell me what the additional tags are?

Thanks a lot for your kind help.

huangj3

Happy holidays!

So, does any one, esp. the Admin, have any further suggestions on how to break down the 4 iterations of GW0 to four G0W0 calculations? Knowing how to do this would be very helpful. I believe other users may also need this.

Best,

huangj3

I tried the same thing with LUSEW=.TRUE. . When I use it on G0W0 (.e. ALGO = GW0), the new results are all jumbled up and wrong. I wonder if this works on the partially self consustent ALGO=SCGW0. Could the admin please clarify, if there have been any advancements on these front and if these is any possibility to break down the iterations ? I want to use the new WAVECAR and WAVEDER, without any LUSEW obtained from the G0W0 and rerun the G0W0. Will that given me better results as the starting point for iterations is now from a better WAVECAR and WAVEDER file ?

It looks like I am not alone in this regard. It would be helpful to come up with a way to checkpoint the intermediate results.

the "logistical problem" i am facing that the smaller clusters, where i can run the problem for more than a couple of days, have massive memory issues. The larger scale clusters where memory problems can still be overcome using mkl_threads do not allow for more than 1 day of simulation time. the third option would be to just ease the parameters - but then why even do GW in the first place.