[gpaw-users] Convergance Issues with B3LYP and PBE0 functionals
Elsebeth Schröder
elsebeth.schroder at chalmers.se
Mon Mar 21 22:05:52 CET 2011
Dear Andy,
Without going into too much detail (since I am still half-heartedly testing) your problem looks somewhat like mine.
Background: some 10-12 years ago I did GGA DFT calculations in the PW91-version with dacapo of the surface of Mg, 12 atoms per unit cell. I did my own (ultrasoft) 2-valence-electron pseudopotential then, since Mg was not available. I never had the slightest convergency problem, never saw any abnormal number of electronic iterations.
Fast-forward to 2011: I want to repeat some of those calculation, but now with gpaw. Using PBE the convergence is difficult (several hundred electronic iterations if in fact it converges), no matter how I try to trick the convergency. Like you I have tested various eigensolvers and many other settings. I have created alternative pseudopotentials, also some with 8 valence electrons. Nothing helps (sufficiently). Using the selfconsistent vdW solver in the Soler implementation the system plainly refuses to converge, no matter what. It bugs me that what really was a simple system a decade ago can so persistently cause me problems now. I might have introduced some simple obvious error (as I said: I am still testing), but it seems less likely the more I test.
I am sorry, I have no suggestions for solving your (or my) problem at the moment.
/Elsebeth
________________________________________
From: gpaw-users-bounces at listserv.fysik.dtu.dk [gpaw-users-bounces at listserv.fysik.dtu.dk] On Behalf Of Andrew Long [andrew.wilson.long at gmail.com]
Sent: 21 March 2011 18:36
To: gpaw-users at listserv.fysik.dtu.dk
Subject: [gpaw-users] Convergance Issues with B3LYP and PBE0 functionals
I've been running into the "gpaw.KohnShamConvergenceError: Did not converge!" issue trying to use the PBE0 and B3LYP functionals on a 1x1x1 unit cell of LaMnO3 and I was wondering if anyone had any tips for me, if you've had similar issues, or if people are seeing convergence faster than I am (well over 300 iterations). Also, I apologize that this is long and has a lot of raw data to parse through, but I figured it would be best to be complete in giving examples of the issues.
Note before starting: I've run the exact same structure in GPAW with xc='PBE' and been able to find an appropriate lattice constant for LaMnO3, so my structure should be fine (using default eigensolver, gpts = (32,32,32), kpts=(12,12,12), and nbands = 38), and that converged in 65 iterations.
Setup generation for BLYP files:
I created the BLYP files using the gpaw-setup string `gpaw-setup --exact-exchange -f BLYP La Mn O`, and I've added the directory for these files to my GPAW_SETUP_PATH.
Calculation Parameters:
I've tried varying several different parameters to just converge the a=4 Angstrom cubic LaMnO3 unit cell. I've tried using both the rmm-diis and cg eigensolvers, I'm now using nbands = -15 (to ensure an appropriate number of empty bands, this might need to be higher/I'm running calculations with more empty bands to check this now), and gpts in [32,40,48,56,64] (kpt sampling is unsupported for hybrid functionals). It seems that the rmm-diis solver is converging the PBE0 calculations better than the cg, as in 300 iterations at 32 gpts,
xc='PBE0', Eigensolver = rmm-diis, gpts=(32,32,32) grep on 'iter:' from output txt file:
iter: 294 17:18:17 -4.0 -3.3 -268670.68734 0 4
iter: 295 17:35:51 -4.1 -3.3 -268670.68991 0 4
iter: 296 17:53:06 -4.2 -3.4 -268670.69695 0 4
iter: 297 18:07:30 -4.3 -3.5 -268670.69772 0 4
iter: 298 18:21:42 -4.3 -3.6 -268670.69917 0 4
iter: 299 18:37:00 -4.1 -3.7 -268670.70235 0 4
iter: 300 18:51:06 -4.7 -3.9 -268670.70431 0 3
xc='PBE0', Eigensolver = cg, gpts=(32,32,32) grep on 'iter:' from output txt file:
iter: 294 23:33:26 +0.7 -1.1 -268668.31406 0 7
iter: 295 23:45:39 +0.8 -1.1 -268674.86126 0 8
iter: 296 23:57:30 +1.3 -0.8 -268693.73136 0 6
iter: 297 00:09:26 +1.4 -0.8 -268673.48216 0 7
iter: 298 00:21:28 +0.7 -0.8 -268663.46448 0 8
iter: 299 00:32:53 +1.1 -0.9 -268665.24148 0 8
iter: 300 00:44:17 +0.8 -1.2 -268667.46744 0 7
Even still, the PBE0 with rmm-diis has not reached the convergence criteria in 300 iterations, although it is certainly approaching it better than the cg solver. If others using hybrid functionals have noticed this, then I guess I'm just stuck running extremely long number of iteration calculations, but this seemed abnormal.
B3LYP has some much bigger issues, namely that it occasionally diverges, and the conjugate gradient solver seems to cause this divergence to happen almost immediately with almost no chance of returning to stable values.
xc='B3LYP', Eigensolver = rmm-diis, gpts=(32,32,32) grep on 'iter:' from output txt file
iter: 294 20:31:19 +1.3 -1.0 -268766.37986 0 8
iter: 295 20:43:42 +0.6 -0.5 -268766.01134 0 7
iter: 296 20:55:58 +0.3 -0.5 -268729.09616 0 7
iter: 297 21:08:27 +11.9 -0.5 -22638322.11476 0 9
iter: 298 21:23:06 +11.9 -0.1 -15204779.31868 0 9
iter: 299 21:36:55 +12.0 -0.2 -11321250.78445 0 9
iter: 300 21:50:42 +5.9 -0.4 -269415.17823 0 11
xc='B3LYP', Eigensolver = cg, gpts=(32,32,32) grep on 'iter:' from output txt file
iter: 26 17:10:17 +1.5 -0.5 -268972.15494 0 7
iter: 27 17:23:58 +1.5 -0.4 -268769.43458 0 8
iter: 28 17:38:25 +10.0 -0.5 -12946072.36830 0 9
iter: 29 17:59:36 +10.2 +0.2 -29325678.97489 0 9
iter: 30 18:24:48 +3.5 +0.3 -267327.36211 0 9
iter: 31 18:43:15 +12.1 -0.1 -79931041.86473 0 9
iter: 32 19:08:54 +11.6 +0.1 -41043447.56353 0 9
iter: 33 19:32:13 +12.7 +0.1 -185879403.34907 0 9
The xc=B3LYP with rmm-diis looks like it might fully converge at gpts=40, though this calculation has not finished yet so I'm unsure of whether it will (been running for almost 100 hours and at 209 iterations):
iter: 205 11:26:21 -4.9 -3.8 -268696.15484 0 3
iter: 206 11:52:16 -4.7 -3.8 -268696.15383 0 3
iter: 207 12:17:03 -4.9 -3.8 -268696.15325 0 3
iter: 208 12:39:00 -5.0 -3.9 -268696.15316 0 4
iter: 209 13:01:48 -4.9 -4.2 -268696.15376 0 3
I'm also unsure of what happened but for gpts=[40,48,56] with rmm-diis, I see energy spikes that come back and stabilize. I don't know if these trends exist in the gpts=64 case, as 100 hours has only resulted in ~60 iterations. This seemed extremely abnormal, as I don't see any of these peaks in the PBE0, meaning something with B3LYP is specifically causing this.
xc='B3LYP', Eigensolver = rmm-diis, gpts=(40,40,40) grep on 'iter:' from output txt file
iter: 36 03:07:54 +1.0 -0.6 -268748.34025 0 8
iter: 37 03:31:07 +9.6 -0.7 -1316167.90156 0 10
iter: 38 03:57:22 +7.6 -0.3 -313190.14697 0 7
iter: 39 04:25:28 +1.9 -0.3 -268634.10803 0 9
xc='B3LYP', Eigensolver = rmm-diis, gpts=(48,48,48) grep on 'iter:' from output txt file
iter: 54 12:31:59 +0.4 -0.8 -268709.67470 0 7
iter: 55 13:22:54 +8.4 -0.9 -312728.04233 0 8
iter: 56 14:30:57 +5.6 -0.3 -268946.26055 0 6
...
iter: 67 01:31:01 +1.8 -0.3 -268679.04502 0 9
iter: 68 02:15:58 +9.1 -0.3 -447263.27650 0 9
iter: 69 03:10:37 +8.8 -0.4 -362153.71279 0 8
iter: 70 04:19:19 +7.0 -0.3 -274627.59113 0 8
xc='B3LYP', Eigensolver = rmm-diis, gpts=(56,56,56) grep on 'iter:' from output txt file
iter: 76 18:30:07 +0.7 -0.8 -268682.77118 0 6
iter: 77 19:30:04 +9.9 -0.8 -1090384.83029 0 10
iter: 78 20:35:55 +2.0 -0.2 -268620.74846 0 10
iter: 79 21:41:14 +1.4 -0.7 -268718.26284 0 7
If anyone has some ideas or sees some issues I'd be extremely grateful. I'm wondering if others have noticed that hybrid functionals take a much higher number of iterations to converge, or if my results are seemingly abnormal. Also, I think I've setup the gpaw setup files properly for B3LYP/PBE0 (don't need anything except for defaults from the setup package for PBE0 and need to generate BLYP files with exact exchange for B3LYP), but if I could get confirmation that this is correct, that'll also narrow down my range of possible issues.
Thank you so much for taking the time to read through this,
Andy Long
More information about the gpaw-users
mailing list