From chemistry-request@ccl.net Sat Feb 12 00:01:52 2005 Received: from smtp4.dti.ne.jp (smtp4.dti.ne.jp [202.216.228.39]) by server.ccl.net (8.12.8/8.12.8) with ESMTP id j1C51oDf003434 for ; Sat, 12 Feb 2005 00:01:51 -0500 Received: from oemcomputer (PPP132.fukuoka-ip.dti.ne.jp [211.132.92.132]) by smtp4.dti.ne.jp (3.10s) with SMTP id j1C51l0l020158 for ; Sat, 12 Feb 2005 14:01:48 +0900 (JST) Message-ID: <003c01c510bf$71eecac0$845c84d3@oemcomputer> From: "Telkuni Tsuru" To: Subject: Summary) Movement unit for geometry optimization Date: Sat, 12 Feb 2005 13:58:07 +0900 MIME-Version: 1.0 Content-Type: text/plain; charset="iso-2022-jp" Content-Transfer-Encoding: 7bit X-Priority: 3 X-MSMail-Priority: Normal X-Mailer: Microsoft Outlook Express 6.00.2800.1106 X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2800.1106 X-Spam-Status: No, hits=3.6 required=7.5 tests=IMPRONONCABLE_1, RCVD_IN_DYNABLOCK,RCVD_IN_SORBS autolearn=no version=2.61 X-Spam-Checker-Version: SpamAssassin 2.61 (1.212.2.1-2003-12-09-exp) on servernd.ccl.net Hello, CCLers. I had sent following question last week, then got 3 replies. Some of them are difficult to understand for me, but I think that I have to summarize them for CCLnet. Thanks to Dr.Campen, Dr.Pye, and Mr.Fox. --- Original Question --- > I like to ask the movement unit for geometry optimization. > > The books which I have describe the optimization method as follows: > ... when one can't find optimized structure, one should calculate > force constant of atoms as the differentiation of energy, and move > each atoms obeying the force direction. ... > > Exactly, there are many MO calculating softwares: Gaussian, Mopac, ... > and so on. I wonder what units they use. 0.1[A], 0.05[A] or 0.01[A] ? > * [A] is angstrom. > > If you know the information which concerns with this question, please > let me know. > When I receive replies, I will summarize them. --- Replies --- 1) from Dr.Campen My impression is that Gaussian (at least in 03) uses an adaptive criterion (i.e. they change the step size based on the local slope). The specifics behind this are referenced on http://www.gaussian.com/g_ur/k_opt.htm if you scroll down to the section titled THE BERNY OPTIMIZATION ALGORITHM Sorry not to be more useful. R. Kramer Campen 2) from Dr.Pye What you are referring to is an energy-only optimization with numerical differentiation. Very few programs do this by default these days. One is forced to do this if the gradients are very difficult to calculate analytically (full CI calculations might fall into this category, as might some of the more novel CC plus perturbation theories. A more practical reason is that in some codes, the calculation of the energy may be implemented, but not the calculation of the gradient (yet!). For standard Hartree-Fock, Density functional calculations, and MP2, efficient routines for calculating derivatives have generally been implemented without resorting to numerical differentiation, at least for Kohn-Sham type theories. It is a little more difficult for grid-based methods, I believe (comments?). As long as the time to calculate a derivative doesnt take N* the time to calculate the energy, it is usually worth it. N is the number of parameters (about 3 times the number of atoms). The "units" that you refer to is actually more precisely called the finite differentiation step size. You want it to be small enough so that second and higher order terms of the Taylor series about the point of interest are small enough so that they are negligible, but you also want it so that it is bigger than the machine epsilon. In practice, you actually want it bigger than the noise in the energy (which is a combination of integral accuracy and SCF convergence criteria). There are a number of energy-only optimization techniques that can be used. A good text on numerical analysis would prove handy. -Cory 3) from Douglas I am not quite sure I understand your question. When computing gradients and force constants you can use analytic derivatives, so no displacement, or numerical means and compute either first or second derivatives. For most methods in Gaussian analytic derivatives are used. For cases where no analytic derivatives are available then it uses 0.01 as the displacment for each atom. When analytic first derivatives are available then it is 0.001 angstrom displacments on each atom. See the G03 IOps Reference, page 11 for IOP(1/39). ---------------------------------------------------- Telkuni Tsuru telkuni=at=venus.dti.ne.jp