From owner-chemistry@ccl.net Fri Mar 17 09:39:00 2017
From: "George Fitzgerald gfitzgerald .. oled.com" <owner-chemistry_-_server.ccl.net>
To: CCL
Subject: CCL:G: SCF Energy Significant Figures, G09
Message-Id: <-52708-170317092325-1264-rP9kos/DyuhQO1qlPoMYAA_-_server.ccl.net>
X-Original-From: George Fitzgerald <gfitzgerald^-^oled.com>
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Content-Type: text/plain; charset="us-ascii"
Date: Fri, 17 Mar 2017 13:24:28 +0000
MIME-Version: 1.0


Sent to CCL by: George Fitzgerald [gfitzgerald]_[oled.com]
Does Convg=0.5D-10 even refer to the energy difference? Documentation on the Gaussian website says that convergence is determined by RMD change in the charge density. If you insist on it, you can see more decimal places if you use formchk convert your chk file to fchk and look for SCF Energy.

I agree, though, with Susi Lehtola's comments that there is little value in looking beyond 8 decimal places.


George Fitzgerald, Ph.D.


-----Original Message-----
> From: owner-chemistry+gfitzgerald==udcoled.com-,-ccl.net [mailto:owner-chemistry+gfitzgerald==udcoled.com-,-ccl.net] On Behalf Of Susi Lehtola susi.lehtola]=[alumni.helsinki.fi
Sent: Thursday, March 16, 2017 3:17 AM
To: George Fitzgerald
Subject: CCL:G: SCF Energy Significant Figures, G09


Sent to CCL by: Susi Lehtola [susi.lehtola!^!alumni.helsinki.fi]
On 03/15/2017 06:27 PM, Elise Kenny e.kenny2#,#uq.edu.au wrote:
> Dear CCL Users,
>
> I have been using Gaussian09 to do high precision SCF energy convergences.
>
> I would like to know if there is a way to increase the number of 
> significant figures printed in the output.
>
> For example:
>
> SCF Done:  E(UB3LYP) =  -1772.85943393     A.U. after   57 cycles
>              Convg  =    0.5502D-10             -V/T =  2.0853
>
> This result has a convergence of 10^-10, but the energy is only 
> printed to eight decimal places.
>
> Can I get G09 to print more numbers or somehow scale the energy to 
> allow for 12 sig. figs. to capture the information I want?


Eight decimal places is more than enough for almost all purposes; the only case where you might need more is numerical derivatives for basis set or geometry optimization.

Note that if you really want to look at energy differences at sub-microHartree level, you'll need to crank up your DFT grid to something pretty large (likely over 300 radial points for meta-GGAs). 
You'll also need to toy with the two-electron integral screening options. IIRC a common rule of thumb is that to get an energy accurate to 10^-n, you need a cutoff of 10^-(n+2) for the integrals.

You might also want to check the stability of your SCF solution, in case you are at a saddle point which will lead to spuriously high energy.

As pointed out by Joseph Leonard, formatted checkpoint files contain energies in exponential format i.e. with full floating point (16 digits) accuracy.
--
-----------------------------------------------------------------------
Mr. Susi Lehtola, PhD             Chemist Postdoctoral Fellow
susi.lehtola##alumni.helsinki.fi   Lawrence Berkeley National Laboratory
http://www.helsinki.fi/~jzlehtol  USA
-----------------------------------------------------------------------http://www.ccl.net/cgi-bin/ccl/send_ccl_messagehttp://www.ccl.net/chemistry/sub_unsub.shtmlhttp://www.ccl.net/spammers.txt