CCL:G: FW: G: Different single-point DFT energy between GAMESS and Gaussian (UNCLASSIFIED)

 Sent to CCL by: "Rinderspacher, Berend (Cont, ARL/WMRD)"
 Classification: UNCLASSIFIED
 Caveats: NONE
 Dear Piotr,
 The errors between Gaussian and Gamess are on the same level as the
 errors within Gaussian or Gamess, ~10^-4 a.u. Unless you believe that
 M05 produces subchemical accuracy these results are the same for all
 chemical intents and purposes. Also M05 contains a slew of parameters.
 Do you  find the same discrepancies with "simpler" functionals (BLYP,
 -----Original Message-----
 Subject: CCL:G: Different single-point DFT energy between GAMESS and
 Several new plausible solutions to the problems have been made. Because
 the problem has been more difficult than I had thought, I've checked
 those suggestions on a simpler test case, namely: 1-fluoroethanol, using
 the geometry below:
     C  6    -1.371551   -0.189402   -0.078069
     C  6     0.055735   -0.002315    0.345325
     H  1    -1.765264   -1.132234    0.321003
     H  1    -1.422818   -0.219081   -1.174187
     H  1    -1.987781    0.640525    0.287358
     O  8     0.815913   -1.025112   -0.171888
     H  1     0.154695    0.068673    1.446323
     F  9     0.519096    1.212587   -0.140203
     H  1     1.716903   -0.919975    0.152892
 - Christopher Cramer suggested that different spatial orientation may
 cause different results. The "nosymm" Gaussian keyword indeed
 any orientation changes, so input orientation is used throughout the
 calculations. GAMESS doesn't change the orientation by default, so
 calculations performed using "nosymm" are made on the same geometry. I
 was curious if changing the initial orientation (called "InitOr in the
 table below - different than the orientation above) to the standard one,
 can change the results significantly. If "nosymm" is not present,
 Gaussian rotates the molecule to the standard orientation (StdOr). In
 order to do the same using GAMESS, "COORD=PRINAXIS" keyword is
 I've found out that when using standard orientation from the Gaussian
 output and pasting it to the GAMESS input (now without "coord=prinaxis),
 GAMESS gives different result than by rotating the molecule to the
 principal axis by itself! Therefore in all subsequent GAMESS
 calculations I was using standard orientation from Gaussian output (the
 one above). The results are presented below (using M05/cc-pVDZ and
 grid(75,302)), together with HF energies, which are pretty much
 reproducible, and M05 calculations using much better than standard grid
 -254.175261128  (InitOr,nosymm)
 -254.175313014  (StdOr transformed from InitOr by Gaussian)
 -254.1753800399 (InitOr)
 -254.1753738096 (InitOr with "COORD=PRINAXIS" keyword)
 -254.1753375023 (StdOr)
 -254.175301121  (Gaussian, StdOr)
 -254.1753365149 (GAMESS, InitOr, "PRINAXIS")
 -254.1753372591 (GAMESS, StdOr)
 -252.967151367  (Gaussian)
 -252.9671513780 (GAMESS)
 - Many people said that linear dependencies could be the cause. That's
 very unlikely with cc-pVDZ, and there are no linearly dependent MOs
 present in my GAMESS outputs.
 - The symmetry is C1 in all cases.
 - There are no transition metals present.
 - I know that GAMESS interprets CCD keyword as cc-pV(D+d)Z basis set
 instead of cc-pVDZ. That's why I was using explicitely defined cc-pVDZ
 atomic basis set in GAMESS, exactly the one used by Gaussian
 switches on printing of basis set info in Gaussian).
 - Another source of error is grid pruning in Gaussian. To avoid this,
 grid should be requested using e.g. "int(grid=75302)" instead of
 "int(grid=finegrid)". There are two weighing schemes available in
 Gaussian. The scheme of Scuseria and Stratman ("ssweights) is the
 default, and Becke scheme can be requested using "bweights". I don't
 have any ideas how weighing is performed by GAMESS. Unfortunatly
 changing those options still does not give similar results:
 -254.175313014  (Gaussian - int(grid=finegrid), ssweights)
 -254.175298048  (Gaussian - int(grid=75302), ssweights)
 -254.175297910  (Gaussian - int(grid=75302), bweights)
 -254.1753375023 (GAMESS)
 I'm afraid hat the whole idea of reproducibility of results is going to
 fail in case of DFT, but I still hope that it can see the light in the
 darkness. Somwhere...
 Best Regards,
 On Tue, May 25, 2010 at 7:53 PM, Piotr Nowak
 piotrnowak[*] <owner-chemistry(!)> wrote:
 	First of all, thanks everyone for response. Some questions and
 suggestions appeared; I'll try to answer them briefly:
 	-I've been using exactly the same structures for the single
 point energy calculations;
 	-I have been using spherical harmonics in both programs.
 Gaussian uses them by default, and I have ensured their use in GAMESS
 with "ISPHER=+1" keyword. The number of cartesian basis fuinctions is
 the same;
 	-Gaussian manual states that default grid uses 75 radial shells
 and 302 angular points/shell. I have been using the same grid in GAMESS
 thanks to "NRAD=75" and  "NLEB=302" keywords. I also
 suspected that grid
 handling might be implemented differently in both programs, therefore I
 tried some "super-ultra-extra-fine" grid with 250 radial shells and
 angular points/shell (using "Int(Grid=250974)" keyword in Gaussian).
 Unluckily, the energy difference remained within the same order of
 magnitude as it was with former grid;
 	-The relative energies are still different. If you compare e.g.
 different geometries of the same molecule, or activation energies, the
 error is still 10^(-4) hartree.
 	I would agree with Soren - there must be some "hidden"
 adjustable parameters, but I have no idea which one can cause these
 differences. I still hope it is possible to get the same results using
 both programs.
 	Kind regards,
 	On Tue, May 25, 2010 at 12:50 AM, Piotr Nowak
 piotrnowak!^! <owner-chemistry]^[
 <blockedmailto:owner-chemistry]> > wrote:
 		Sent to CCL by: "Piotr  Nowak"
 		Dear CCL users,
 		I'm trying to reproduce single point energy obtained
 with Gaussian 03 using
 		GAMESS US. Hartree-Fock energy is almost exactly the
 same e.g.
 		Gaussian: -1849.26414782
 		GAMESS:   -1849.2641478646
 		Unfortunately my attempts to get the same results using
 DFT failed. The
 		energy differences between both programs are
 unreasonably huge. Here are some
 		examples of results for different functionals (the same
 case as above-
 		mentioned HF example):
 		Gaussian: -1855.79754118
 		GAMESS:   -1855.7976587495
 		Gaussian: -1845.45112047
 		GAMESS:   -1845.4510666810
 		Slater (also known as Dirac, one of the simplest LDA
 functionals, so I'm sure
 		it has the same definition in both programs)
 		Gaussian: -1833.20351470
 		GAMESS:   -1833.2034704727
 		I have done those calculations using the same grid,
 using tight convergence
 		criteria. I've found out that Gaussian uses slightly
 different cc-pVDZ basis
 		set than the one present in Basis Set Exchange, but
 using this basis set with
 		GAMESS has left the results unchanged. I have also tried
 different guesses,
 		and SCF algorithms, but without success.
 		Here are keywords used in inputs for above calculations.
 		#p m05/cc-pvdz nosymm iop(6/7=3) scf=tight
 		 $DFT NRAD=75 NLEB=302 $END
 		I would appreciate any kind of help.
 		Best regards,
 		Piotr Nowak
 		-= This is automatically added to each message by the
 mailing script =-
 		E-mail to subscribers: CHEMISTRY]^[
 <blockedmailto:CHEMISTRY]>  or use:<blocked>;
 		E-mail to administrators: CHEMISTRY-REQUEST]^[
 <blockedmailto:CHEMISTRY-REQUEST]>  or
 		<blocked>; <blocked>;
 Classification: UNCLASSIFIED
 Caveats: NONE