From owner-chemistry-: at :-ccl.net Tue Nov 22 14:11:00 2016
From: "Horkel Ernst ernst.horkel**tuwien.ac.at"
To: CCL
Subject: CCL:G: AW: Gaussian 09 Frequency Jobs
Message-Id: <-52507-161122140657-18083-D5Xbr2dKxWzVT9JZE8IaRg,,server.ccl.net>
X-Original-From: Horkel Ernst
Content-Language: de-DE
Content-Transfer-Encoding: 8bit
Content-Type: text/plain; charset="iso-8859-1"
Date: Tue, 22 Nov 2016 19:06:46 +0000
MIME-Version: 1.0
Sent to CCL by: Horkel Ernst [ernst.horkel(a)tuwien.ac.at]
Hi,
I had a similar problem some time ago and got the following response from the Gaussian support:
**************************
Hello Ernst,
This is symptomatic of numerical accuracy issues. For this job you were running into poor convergence in the iterative solution of the CPHF equations when doing the differentiation with respect to nuclear coordinates. See this part of the output file:
Differentiating once with respect to nuclear coordinates.
...etc...
11 vectors produced by pass 79 Test12= 2.02D-13 1.00D-09 XBig12= 3.36D-14 3.80D-09.
11 vectors produced by pass 80 Test12= 2.02D-13 1.00D-09 XBig12= 5.49D-14 4.96D-09.
11 vectors produced by pass 81 Test12= 2.02D-13 1.00D-09 XBig12= 2.65D-14 3.22D-09.
11 vectors produced by pass 82 Test12= 2.02D-13 1.00D-09 XBig12= 4.36D-14 4.36D-09.
11 vectors produced by pass 83 Test12= 2.02D-13 1.00D-09 XBig12= 4.65D-14 4.42D-09.
11 vectors produced by pass 84 Test12= 2.02D-13 1.00D-09 XBig12= 5.02D-14 5.49D-09.
11 vectors produced by pass 85 Test12= 2.02D-13 1.00D-09 XBig12= 2.14D-14 3.26D-09.
11 vectors produced by pass 86 Test12= 2.02D-13 1.00D-09 XBig12= 5.85D-14 5.20D-09.
11 vectors produced by pass 87 Test12= 2.02D-13 1.00D-09 XBig12= 3.95D-14 4.24D-09.
11 vectors produced by pass 88 Test12= 2.02D-13 1.00D-09 XBig12= 4.44D-14 5.09D-09.
11 vectors produced by pass 89 Test12= 2.02D-13 1.00D-09 XBig12= 4.34D-14 4.30D-09.
The values printed after "XBig12" do not steadily approach the values printed before that (the ones between "Test12" and "XBig12"), and that these values have been oscillating near convergence for a large number of iterations ("passes"), in this case more than 89. For most frequency calculations, convergence here is achieved in less than 20 "passes". A larger number of passes and oscillations in the "XBig12" values as it approaches convergence are indicative of numerical accuracy issues. For certain molecular systems, certain combinations of computational methods and basis sets might need larger integration grids (in the case of DFT methods) and/or increased integral accuracy (especially calculations with very diffuse basis functions).
The most general procedure to reduce the numerical noise in the calculation is to increase the integral accuracy. By default, the integral accuracy cutoff is 10^-10, so increasing the accuracy to 10^-11 or 10^-12 should have a positive impact in reducing numerical noise. In G09, there is a keyword to control integral accuracy, "Integral=(Acc2e=N)", which will set the accuracy to 10^-N. So, for instance, "Integral=(Acc2e=12)" sets the integral accuracy cutoff to 10^-12.
Since this is a DFT calculation, the numerical integration grid is also another parameter that can impact the numerical stability of the calculation. By default, DFT energy and gradient calculations use the "Fine" grid (a pruned grid with 75 radial shells and 302 angular points, or 75,302), while the default grid for the CPHF equations in DFT frequency calculations is the "Coarse" grid (a pruned 35,110 grid). Increasing the number of points in the DFT integration grid also helps when problems due to numerical inaccuracies appear. Usually the natural increase in the number of points would be to use the "Ultrafine" grid (a pruned 99,590 grid) for the DFT energies and gradients, which would automatically set the use of the "SG1" grid (a pruned 50,194 grid) for the CPHF equations. If required (in special cases, such as using uncontracted basis sets), one can specify directly the number of radial shells and angular points.
I would try first increasing integral accuracy by one or two orders of magnitude over the default ("Int=(Acc2e=11)" or "Int=(Acc2e=12)"). If you still have this issue, you may also need to increase the size of the numerical integration grid, but try first something like the following:
%nprocshared=8
%mem=16GB
%chk=o-PCzPOXD_DFT_b3lyp_6-311+d_freq_2ndtry.chk
# freq b3lyp/6-311+g(d) geom=connectivity scf=noincfock integral=(acc2e=11)
[No Title]
0 1
C 0.34900900 2.46465100 0.10154300
N -0.73346000 2.23368700 -0.57575100
N -1.08971300 0.91579800 -0.32567600
...etc...
*********************
Indeed, setting "integral=(acc2e=11)" solved the problem. So you might give this a try...
Good luck an keep computing,
Ernst
Senior Lecturer Dipl.-Ing. Dr.techn. Ernst Horkel
Institute of Applied Synthetic Chemistry,
Vienna University of Technology Tel.: +43-1-58801-163609
Getreidemarkt 9/163OC, +43-664-60588-7122
A-1060 Vienna, Austria Fax: +43-1-58801-15499
email: ernst.horkel a tuwien.ac.at
-----Ursprüngliche Nachricht-----
Von: owner-chemistry+ehorkel==ioc.tuwien.ac.at a ccl.net [mailto:owner-chemistry+ehorkel==ioc.tuwien.ac.at a ccl.net] Im Auftrag von Alan Wilfred Humason ahumason*_*smu.edu
Gesendet: Montag, 21. November 2016 04:56
An: Horkel, Ernst
Betreff: CCL: Gaussian 09 Frequency Jobs
Sent to CCL by: "Alan Wilfred Humason" [ahumason]~[smu.edu] I am running a large molecule with a large basis set. (It's an expensive calculation, but I've been given the
resources.) After optimization, the program runs a set of iterations, giving the message:
3 vectors produced by pass897 Test12= 6.42D-13 1.00D-09 XBig12= 5.97D-13 1.11D-08.
The calculation gets 'stuck' at 3 vectors, even after 3000 iterations, and several weeks on 24 cores.
1) What do the vectors represent?
2) What to the passes represent?
3) How can I get my job to finish?
Alan Humason
ahumason||smu.eduhttp://www.ccl.net/cgi-bin/ccl/send_ccl_messagehttp://www.ccl.net/chemistry/sub_unsub.shtmlhttp://www.ccl.net/spammers.txt