next up previous contents
Next: 8.4 LscaLAPACK Up: 8 Special switches for Previous: 8.2 Not enough memory

8.3 Elastic band method


If the elastic band method is used on the T3D scaLAPACK has to be switched of (see 8.4).

Vasp.4.2 supports the elastic band method to calculate energy barriers. The INCAR, KPOINTS, and POTCAR files must be located in the directory where VASP is started. In addition a set of subdirectories (numbered 00,01,02...) must be created, and each subdirectory must contain one POSCAR file. The tag

  IMAGES= number of images
in the INCAR file, tells VASP to run the elastic band method. The number of nodes must be dividable by the number of images (the NPAR switch can still be used as described above). VASP divides the nodes in groups, and each group then works on one ``image''. Mind, that IMAGES can only take values tex2html_wrap_inline4731 . The first group works on the directory 01, second group on 02 etc. In the elastic band method, the endpoints are kept fixed, the positions for the end points must be given in 00/POSCAR and XX/POSCAR, where XX is
  XX=number of images+1.
All output goes to the subdirectories (but because no node is running for 00 and XX, no files will be created in these sub directories). The usual stdout of the images 2,3,...,number of images is redirected to the files 02/stdout, 03/stdout etc. (only image 01 writes to the usual stdout). In addition to the IMAGES tag the spring constant can be supplied in the SPRING tag, the default is
For SPRING=0, the nudged elastic band method is used[52, 53]. In the nudged elastic band method each image is only allowed to move into the direction perpendicular to the current hyper-tangent, which is calculated as the normal vector between the neighboring two images. This algorithm keeps the distance between the images to first order constant. It is therefore possible to start with a dense image spacing around the saddle point to obtain a finer resolution around this point. One problem of this method is that the constraint (i.e movements only in the hyper-plane perpendicular to the current tangent) is non linear. Therefore the CG algorithm usually fails to converge, and it is recommended to use the RMM-DIIS algorithm (IBRION=2). However, the non-linear constraint (equidistant images) is strongly violated in the RMM-DIIS algorithm in the first view steps. We therefore recommend to use the RMM-DIIS algorithm in the first 10 steps or so with a very low dimensionality parameter (NFREE=2). This avoids that the constraint is significantly violated, and preconverges the images. An alternative is to use a steppest descent minimization without line optimization (IBRION=3, SMASS=2).

If SPRING is negative value, additional tangential springs are introduced so that the images are kept equidistant during the relaxation. Do not use too large values, because this can slow down convergence. The default values works usually quite reliable.

If all degrees of freedom are allowed to relax (isolated molecules, no surface, etc.), please make sure that the sum of all positions is the same for each cell. In other words,

must be equal for all images. Otherwise ``fake'' forces are introduced, and the images ``drift'' against each other (this will not introduce any problems in the calculation, but it is awkward if one tries to visualize the final results). Often an initial linearly interpolated starting guess is appropriated, this can be done with a small script called

found in  vamp/scripts/. The script also removes as an option the center of ``mass motion''.

next up previous contents
Next: 8.4 LscaLAPACK Up: 8 Special switches for Previous: 8.2 Not enough memory

Mon Mar 29 10:38:29 MEST 1999