mopac/fortran on different platforms

From: cml : at : uni-muenster.de
Subject: mopac/fortran on different platforms
Date: Tue, 12 Jan 1999 18:19:10 +0100
 Dear list members,
 I have been working with Mopac93 for quiet a long time and have
 obeserved the following problem on several occasions: Depending on
 the platform on which it has been compiled, different results are
 obtained when an optimization runs into numerical difficulties.
 I have compared the compiled programs which have been generated on
 the following systems:
 RS/6000 (AIX,xlf) - PentiumII (Linux,g77) - AMD K6/2 (Linux,g77)
 In those cases, where the geometry optimization (EF) runs straight
 into the energetic minimum, all three versions give nearly the same
 energy and geometry.
 In cases where the optimization demands a large number of steps,
 or when transition states are searched, the three versions tend to
 behave much more diverse. To give an example, in one case (TS with
 LET DDMIN=0.00) the two Linux binaries give an optimized geometry
 while the AIX version stops with a geometry 10 kcal lower in energy
 and repeating the warning "NUMERICAL PROBLEMS..".
 In another case (reaction path with three points for one dihedral
 angle) the AIX version ends without complaint while the Linux
 versions give the error message:
         CARTESIAN COORDINATES READ IN, AND CALCULATION
        TO BE RUN IN INTERNAL COORDINATES,
        BUT NOT ALL COORDINATES MARKED FOR OPTIMISATION
 although the input geometry IS a z-matrix in INTERNAL coordinates
 with some parameters marked to be held constant. This can only be
 overcome by adding XYZ to the input line, but then the two Linux
 version behave different: On the AMD K6/2, the calculation stops after
 500 cycles without finding a minimum, on the PentiumII it works, but
 gives energies that differ from the AIX-results. This of course might
 be the result of the use of cartesian coordinates for optimization.
 To conclude this story, I would like to ask the following questions:
 1) Has somebody of you already observed a similar behaviour of MOPAC-
 versions (or any other program) that have been ported to different plat-
 forms, and (if so)
 2) is this due to the different CPU architectures or to the use of
 different compilers ?
 3) What might be the problem with the internal coordinates that are not
 recognized as an internal z-matrix? Is this connected with the format in
 which the z-matrix is read in ?
 I am looking forward to helpful comments,
 sincerely
 Christian Mueck-Lichtenfeld
 ------------------------------------------------------------
 Dr. Christian M?ck-Lichtenfeld
 Organisch-Chemisches Institut
 Westf?lische Wilhelms-Universit?t
 Corrensstraáe 40                         cml : at : uni-muenster.de
 D-48149 M?nster (Germany)             phone +49 251 83-33239
 ------------------------------------------------------------