[Wien] QTL-B and Eigenvalues Below...
Dear wien2k community, I want to come back to this thread, as I did not yet succeed to identify what is really going on. To summarize the original problem: a crash in lapw2 for impurities in Ge, with very large (actually **) QTL-B warnings and (sometimes) Eigenvalues below warnings in the preceding case.output1. Earlier suggestions from the mailing list: keep Ge-3p in the core, take larger sphere sizes for Ge. This helps in some cases, but then the problem reappears in very similar cases that were before running without problem. I have now 2 test cases that show the problem, and that are as simple and clean as possible. Both are a 64-atom supercell of pure bulk Ge (no impurities), with the symmetry artificially lowered such that there are 10 inequivalent positions. This is the diamond structure, with 4.71 au between nearest neighbours. What I call case1 has Rmt=2.29 and RKM=8.01 for all Ge atoms (Kmax=3.5), case2 has Rmt=2.34 and RKM=8.19 (again Kmax=3.5). Ge-3p states are in the core, sphere sizes are larger than in my previous post, and there is now only one element involved (Ge). In case1, -in1new is used, in case2 not. I discuss both these cases independently. They give more or less the same errors, although some details of the errors are different. I do not succeed to reproduce this error in a smaller cell (e.g. the 2-atom unit cell for the diamond structure). And it looks like the problem is connected to the basis set size: often a small basis set works (Kmax=2.5), but the problem appears when going to larger basis sets as the Kmax=3.5 from the following two examples. Any hints on what is going wrong? Thanks, Stefaan Case1 This case (Rmt=2.29, RKM=8.01) was restored from a previous run, that successfully finished at RKM=8.01 and without in1new. The only thing that was changed for the present run, was to add -in1new 1. It crashes right in the first iteration, with this error in case1.output2_4 (no errors in the other 3 parts of this k-point parallel run): QTL-B VALUE .EQ. 1844.97972 in Band of energy -9.34906 ATOM= 5 L= 2 Check for ghostbands or EIGENVALUES BELOW XX messages Adjust your Energy-parameters or use -in1new switch, check RMTs !!! There are no eigenvalues around -9.3 Ry, however, as can be seen in the last iteration of the successful preceding run without -in1new: K= 0.12500 0.12500 0.125001 :RKM : MATRIX SIZE 8335LOs: 896 RKM= 8.01 WEIGHT= 8.00 PGR: EIGENVALUES ARE: :EIG1: -1.4769260 -1.4769260 -1.4769260 -1.4769258 -1.4769257 :EIG6: -1.4769257 -1.4767479 -1.4767478 -1.4767478 -1.4766416 :EIG00011: -1.4766416 -1.4766416 -1.4766414 -1.4766414 -1.4766414 (--snip--) This -9.3 Ry rather lies in the region of the core state energies: 5.ATOM Ge4 7 CORE STATES CORE-FORCE in mRy/a.u. = |F| Fx Fy Fz :FCO005: 5.ATOM 0.004 0.000 -0.004 0.000 :1S 005: 1S-804.696867125 Ry :2S 005: 2S-100.093011168 Ry :2PP005: 2P*-88.951919057 Ry :2P 005: 2P -86.609329514 Ry :3S 005: 3S -11.689696508 Ry :3PP005: 3P* -8.017205664 Ry :3P 005: 3P -7.702933865 Ry In the crashed iteration, there is an 'eigenvalues below' warning in case1.output1_4 (not in the other 3 output1_* files): 1.96690051.96771561.97307281.98089751.9813078 1.98156181.99719701.99733381.9984643 1 EIGENVALUES BELOW THE ENERGY -10.0 The automatically produced case1.in1c looks like this (only one atom shown, all other are identical to this): .13832 5 0 global e-param with N other choices, napw 0 -0.124 0.000 CONT 1 0 -0.421 0.000 CONT 1 10.129 0.000 CONT 1 20.106 0.000 CONT 1 2 -1.464 0.000 CONT 1 Case2 This case (Rmt=2.34, RKM=8.19) was started from scratch and fails in the 20th iteration. In contrast to case1, no -in1new was ever used. These are the first few eigenvalues in the last successful iteration (we will need them later): K= 0.12500 0.12500 0.125001 :RKM : MATRIX SIZE 8335LOs: 896 RKM= 8.18 WEIGHT= 8.00 PGR: EIGENVALUES ARE: :EIG1: -1.4742860 -1.4742754 -1.4742754 -1.4742486 -1.4742368 :EIG6: -1.4742368 -1.4740887 -1.4740887 -1.4740748 -1.4739971 :EIG00011: -1.4739788 -1.4739788 -1.4739514 -1.4739514 -1.4739368 :EIG00016: -1.4737167 -1.4736918 -1.4736918 -1.4736907 -1.4736581 (--snip--) All Ge atoms show the expected linearization energies in case.scf: ATOMIC SPHERE DEPENDENT PARAMETERS FOR ATOM Ge8 :e__0009: OVERALL ENERGY
[Wien] QTL-B and Eigenvalues Below...
Hello again, I have been too optimistic, yesterday. The reported problem (eigenvalues below...) did indeed disappear when I modified the sphere radii and/or put deep semi-core states in the core, as reported yesterday. But when I applied the same modification to a Cu atom at a different site in Ge (a case that was working well with the old settings), exactly the same problem now appeared there... This happens, in the very first iteration, in a directory that was freshly initialized with not old files lingering around: an error message in lapw2_3.error: 'l2main' - QTL-B.GT.15., Ghostbands, check scf files end of case.scf2_3 : QTL-B VALUE .EQ. ** in Band of energy -5.71041 ATOM= 5 L= 2 Check for ghostbands or EIGENVALUES BELOW XX messages Adjust your Energy-parameters or use -in1new switch, check RMTs !!! But atom 5 (which is Ge) does not have states at that energy: 0.304 0 (GLOBAL E-PARAMETER WITH n OTHER CHOICES, global APW/LAPW) 2 -1.82 0.010 CONT 1 20.30 0.000 CONT 1 00.30 0.000 CONT 1 10.30 0.000 CONT 1 The only other element in the cell is Cu (ATOM 1), which has these linearization energies: 0.304 0 (GLOBAL E-PARAMETER WITH n OTHER CHOICES, global APW/LAPW) 10.30 0.000 CONT 1 1 -5.36 0.005 STOP 1 20.30 0.010 CONT 1 00.30 0.000 CONT 1 It looks like eigenstates for Cu and Ge are somehow interchanged by the error message...? The end of case.scf1_3 : :EIG00936: 1.95876521.96050221.96203001.96516811.9677904 :EIG00941: 1.97092361.97591201.97839751.98225651.9843245 :EIG00946: 1.98581091.98808901.99031351.99273091.9944311 :EIG00951: 1.99639981.99680781.9992533 2 EIGENVALUES BELOW THE ENERGY -9.0 :KPT : NUMBER OF K-POINTS: 1 All other case.scf1_* files have 0 eigenvalues below. And these are the deepest eigenvalues in case.scf1: :RKM : MATRIX SIZE 5615LOs: 908 RKM= 6.53 WEIGHT= 8.00 PGR: EIGENVALUES ARE: :EIG1: -4.9618775 -4.9618775 -4.9618749 -1.4433134 -1.4425104 :EIG6: -1.4425104 -1.4424955 -1.4415851 -1.4415851 -1.4414788 :EIG00011: -1.4414788 -1.4414291 -1.4414284 -1.4412709 -1.4412709 :EIG00016: -1.4411669 -1.4411669 -1.4410760 -1.4410224 -1.4409746 Here are the linearization energies that lapw1 determined (no in1new): For Cu: ATOMIC SPHERE DEPENDENT PARAMETERS FOR ATOM Cu1 :e__0001: OVERALL ENERGY PARAMETER IS0.3000 OVERALL BASIS SET ON ATOM IS LAPW :E1_0001: E( 1)=0.3000 APW+lo :E1_0001: E( 1)= -4.9650 E(BOTTOM)= -4.995 E(TOP)= -4.935 LOCAL ORBITAL :E2_0001: E( 2)= -0.0450 E(BOTTOM)= -0.270 E(TOP)=0.180 APW+lo :E0_0001: E( 0)=0.3000 APW+lo For Ge: ATOMIC SPHERE DEPENDENT PARAMETERS FOR ATOM Ge3 :e__0004: OVERALL ENERGY PARAMETER IS0.3000 OVERALL BASIS SET ON ATOM IS LAPW :E2_0004: E( 2)= -1.4450 E(BOTTOM)= -1.530 E(TOP)= -1.360 APW+lo :E2_0004: E( 2)=0.3000 LOCAL ORBITAL :E0_0004: E( 0)=0.3000 APW+lo :E1_0004: E( 1)=0.3000 APW+lo I used RKMax=6.53 (Kmax=3.00), and 4 k-points -- values that should be more than good enough for reasonable results. Changing the initial linearization energies did not help to the values reported case.scf1 did not help. The real reason of this behaviour must lie elsewhere...? Stefaan Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm
[Wien] QTL-B and Eigenvalues Below...
The only times I have seen things like this is when, for whatever reason, the case.clmsum was messed up. I would check that lstart and dstart worked OK, and that the output from lapw0 looks reasonable. If all else fails, perhaps move to an LAPW basis set until you have at least partially converged the calculation. Thanks, Laurence. Nothing suspicious in case.outputst, for sure. In case.outputd and case.output0, I see at least no error messages or anything particularly unfamiliar, but I admit I do not know about every number printed there what exactly it means. Changing completely to LAPW got it through the first iteration, although there was still the message that one eigenvalue was below -12 Ry in case.output1 (only 1, not 8). And eventually, the scf cycle never converged with LAPW, obviously due to that 'eigenvalue below' problem. Meanwhile, I got the same structure converged after artificially lowering the symmetry. But that is not how things should be done. This case isn't something very exotic (a substitutional Cu impurity in a 64-atom Ge supercell), it should be possible to run this in a normal way, isn't it? Any other hints...? Stefaan On Fri, Jul 25, 2008 at 10:18 AM, Stefaan Cottenier stefaan.cottenier at fys.kuleuven.be wrote: Dear wien2k community, I ran into a case that gave a QTL-B error in the first iteration. Following the FAQ, I therefore set the default 0.30 linearization energy to about 0.2 Ry below the :FER that is found in case.scf2, and adjust the linearization energies of the deeper states to the eigenvalues found in case.scf1. I have then the following case.in1c (the first atom is Cu, the second is Ge): WFFIL(WFPRI, SUPWF) 6.15 104 (R-MT*K-MAX; MAX L IN WF, V-NMT 0.105 0 (GLOBAL E-PARAMETER WITH n OTHER CHOICES, global APW/LAPW) 00.10 0.000 CONT 1 0 -8.12 0.002 STOP 1 10.10 0.000 CONT 1 1 -4.99 0.005 STOP 1 20.10 0.010 CONT 1 0.105 0 (GLOBAL E-PARAMETER WITH n OTHER CHOICES, global APW/LAPW) 10.10 0.000 CONT 1 1 -7.76 0.002 STOP 1 2 -1.42 0.005 CONT 1 20.10 0.000 CONT 1 00.10 0.000 CONT 1 (snip) Nevertheless, the same QTL-B error reappears (last lines from case.output2): QTL-B VALUE .EQ. ** in Band of energy -8.93817 ATOM= 2 L= 2 I'm a bit puzzled by this, as atom nr. 2 (Ge) should have no states near that energy, certainly no d-states.The linearization energies used for this second trial are nicely consistent with the content of the new case.scf2 and case.scf1: grep :FER case.scf2 == :FER : F E R M I - ENERGY(TETRAH.M.)= 0.33183 case.scf1: :EIG1: -8.1239527 -7.7633883 -7.7633873 -7.7633873 -7.7633588 (snip) :EIG00186: -7.7595614 -7.7595564 -7.7595506 -7.7595506 -7.7595488 :EIG00191: -4.9882067 -4.9882067 -4.9882066 -1.4337422 -1.4337422 :EIG00196: -1.4335714 -1.4333581 -1.4332873 -1.4332873 -1.4332455 A second inspection shows there is a message 8 EIGENVALUES BELOW THE ENERGY -12.0 in case.output1 (actually in case.output2_3 of a k-point parallel run, which corresponds to the case.output2_3 where I got the QTL-B message). That message is there with and without adjusting the linearization energies. There should not be any eigenvalues so deep, and changing Emin in case.in1c to a very low value of -20.0 leads to the same message (now 8 eigenvalues below -20.000). This message is probably the source of the error, rather than the linearization energies. But I do not understand where it comes from...? Another observation is that in case.scf2_3 there are many '' strings, for all atoms: :POS002: AT.NR. -2 POSITION = 0.37456 0.12507 0.12507 MULTIPLICITY = 12 LMMAX 28 LM= 0 0 1 1 -1 1 2 0 2 2 -2 2 3 1 -3 1 3 3 -3 3 4 0 4 2 -4 2 4 4 -4 4 5 1 -5 1 5 3 -5 3 5 5 -5 5 6 0 6 2 -6 2 6 4 -6 4 6 6 -6 6 :CHA002: TOTAL CHARGE INSIDE SPHERE 2 = :PCS002: PARTIAL CHARGES SPHERE = 2 S,P,D,F,PX,PY,PZ,D-Z2,D-X2Y2,D-XY,D-XZ,D-YZ :QTL002: Q-s-low E-s-low Q-p-low E-p-low Q-d-low E-d-low Q-f-low E-f-low :EPL002: -8.8617 -8.8746 -8.8802 -8.8790 Q-s-hi E-s-hiQ-p-hi E-p-hiQ-d-hi E-d-hiQ-f-hi E-f-hi :EPH002: 0.3394 -0.25550.3053 0.12770.0300 -0.40980.0014 -0.0723 QXX QXY QYY QZZ UP TO R :VZZ002: 2.050 It may be useful to mention the sphere radii, as they are rather small for Ge (due to atoms that lie close to each other): RMT = 2.31 for Cu, which leads to this in case.outputst: TOTAL CORE-CHARGE: 10.00 TOTAL CORE-CHARGE INSIDE
[Wien] QTL-B and Eigenvalues Below...
Thanks Peter and Laurence. The suggestion on core states/Rmt value seem to solve it. Here are the results of the tests both of you suggested: I've two suspicions: a) Some problem with the struct file. It could be identical positions of one atom, wrong symmetry operations, wrong multiplicity,... Make sure that the analysis from nn, sgroup and symmetry fit to each other. You may also remove case.vns temporarely to check if the problems may come from the Fourier coefficients (symmetry!) Definitely nothing wrong with the struct file. All structural info is consistent, and the same case.struct has worked before with several other impurity/host lattice combinations. b) The low lying Cu-3s (and maybe Ge-p) states. Definitely one does not need the Cu-3s states as valence for a large Cu sphere, and I'm not even sure that the Ge-p states are necessary. At least I'd test this and remove them temporarely to see of the states below are gone. Eventually, you may slightly change RMTs (larger Ge sphere at the cost of Cu). All tests along this line do cure the problem: a) Slightly increasing the Rmt's of both Cu and Ge until they are nearly touching (Cu from 2.31 to 2.37, Ge from 2.05 to 2.10) and keeping Cu-3s and Ge-3p as valence states, does work. b) Trading Ge size for Cu size (Cu from 2.31 to 2.26, Ge from 2.05 to 2.10) and still Cu-3s and Ge-3p as valence states, does work. c) Keep the original Rmt's (2.31 and 2.05), but put Cu-3s and Ge-3p in the core (Ecut=-6 Ry), does work. Apparently the Cu and Ge Rmt's were too different from each other, in particular for deep semi-core states (I should have known this, the FAQ on choosing Rmt's warns they should not be more than 10-20% different from each other -- the present choice arose because I wanted to do several different Cu sites with the same Rmt-settings). What I do not understand is why lowering the symmetry (and displacing one atom) did not produce the eigenvalues below error, for the same Rmt-values. Also Cu on slightly different sites in Ge ran fine. It must be quite a coincidence that this problem shows up. Anyway, I'll choose more similar Rmt values for this entire set of calculations from now on, to be safe. 1) Try x patchsymm. It's an undocumented utility I put in with pairhess. It will put into case.struct_new a set of atomic positions averaged over the symmetry with an output in case.outputpatch which will show any problems. The difference between the new/old positions and what patchsymm gives should be less that 1D-7. Sometimes there are subtle issues. The largest differences are 1D-15, no problems here (nice tool, by the way). 2) Check case.struct and ensure that the symmetry operations are what they should be, and that there are no numerical errors giving nonzero (inappropriate) translations. All look like this: 0 0-1 0. 0 1 0 0. -1 0 0 0. Nowhere nonzero translations. 3) Look at case.scf0 and see if there are any very poor fits to Vxc. Nothing suspcious here too: ATOM 1 AT RMT: SIGMA OF V-XC FIT:0.29702E-02 ATOM 2 AT RMT: SIGMA OF V-XC FIT:0.99336E-03 ATOM 3 AT RMT: SIGMA OF V-XC FIT:0.10263E-02 ATOM 4 AT RMT: SIGMA OF V-XC FIT:0.10271E-02 ATOM 5 AT RMT: SIGMA OF V-XC FIT:0.10161E-02 ATOM 6 AT RMT: SIGMA OF V-XC FIT:0.99735E-03 ATOM 7 AT RMT: SIGMA OF V-XC FIT:0.99394E-03 ATOM 8 AT RMT: SIGMA OF V-XC FIT:0.99534E-03 ATOM 9 AT RMT: SIGMA OF V-XC FIT:0.10505E-02 ATOM 10 AT RMT: SIGMA OF V-XC FIT:0.98391E-03 ATOM 1 LARGEST SIGMA AT R(781)= 2.310 and MAX DIFF: 0.29702E-02 0.91750E-02 ATOM 2 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.99336E-03 0.26818E-02 ATOM 3 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.10263E-02 0.26926E-02 ATOM 4 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.10271E-02 0.27565E-02 ATOM 5 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.10161E-02 0.28470E-02 ATOM 6 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.99735E-03 0.26202E-02 ATOM 7 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.99394E-03 0.26300E-02 ATOM 8 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.99534E-03 0.26776E-02 ATOM 9 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.10505E-02 0.34587E-02 ATOM 10 LARGEST SIGMA AT R(781)= 2.050 and MAX DIFF: 0.98391E-03 0.25929E-02 4) Stare at case.in2c and convince yourself that it is right. (not yet done, as Peter's reply intervened) 5) Maybe put more states as semicore rather than core, although it makes no sense to me that this should matter. For Cu, I had taken the 3s as semi-core: 2P* -68.81376 -68.78206 2P -67.30838 -67.27554 3S-8.85641 -8.77179 3P* -5.84815 -5.76476 3P
[Wien] QTL-B and Eigenvalues Below...
Laurence Marks wrote: Wow, to me this is a clear red flag -- you have 10 too many electrons. All looks fine here too (data taken from the diverging run with LAPW): :NEC01: NUCLEAR AND ELECTRONIC CHARGE 2063.0 2073.45604 0.99496 :NEC02: NUCLEAR AND ELECTRONIC CHARGE 2063.0 2063.0 1.0 :NEC03: NUCLEAR AND ELECTRONIC CHARGE 2063.0 2063.1 1.0 In that case I learned even more than 1 thing today ;-). As this was in the unmixed density, I wasn't alarmed. Summarized, for the archive: * Inproper Rmt's led to a crash in the first iteration, with eigenvalues below error in case.output1. * Going to full LAPW softens this error, but leads to a diverging scf-cycle (of which the above :NEC message shows the origin) * Taking proper Rmt's makes all problems fade away (in the first iteration as well as for the convergence) Stefaan Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm