Page 1 of 1

Dimer method - convergence criteria

Posted: Tue Mar 04, 2008 4:24 pm
by jhawkins
I am currently investigating Cu on TiN and am interested in determining the transition states associated with diffusion on the surface. I have identified the relevant stable adsorption sites, and after unsuccessful attempts of using the NEB method (it is likely that the transition states are not along linear paths connecting these sites), I am using the dimer method to attempt to find the transition states. The energy differences between various adsorption sites are quite small (<0.2 eV), and I anticipate that the PES is relatively flat. I have found that for some of my runs, they appear to converge prematurely - that is, the force is below the EDIFFG specified in the INCAR file (-5E-4) but the curvature is still positive. Shouldn't it continue running (since the curvature is still positive), or is the force convergence enough to end the run? A copy of the DIMCAR and a portion of the INCAR are appended below.

Thanks for your help!
~ Jeff

Code: Select all

 Step         Force        Torque        Energy     Curvature         Angle
    1       0.00470       1.73165    -195.76264       0.40234      27.99709
    1       0.00470       0.17559    -195.76264      -0.04678      14.12131
    2       0.00166       1.36137    -195.82663       0.67222      20.73637
    2       0.00166       0.92982    -195.82663       0.09024      15.69645
    3       0.00009       ---        -195.84147       ---           ---

ICHAIN = 2
IOPT = 2
DdR = 0.005
DRotMax = 4
DFNMin = 0.01
DFNMax = 1.0

Re: Dimer method - convergence criteria

Posted: Tue Mar 04, 2008 9:16 pm
by graeme
This is a good point. The Dimer method uses a force criteria to quit, and does not demand that the minimum curvature be negative. To search for a saddle starting from a minimum, it is essential to perturb the system away from the minimum so that the forces are above the convergence criteria. If this output is from the start of a run, it looks very close to an initial minimum, and should be given a kick to get started. Then, the method should converge upon a first order saddle. It is possible for it to converge upon zero modes (translation, for example) but the I feel that it's better for the method to quit instead of following the zero mode looking for a negative one. We are using the method fairly extensively now to simulate dynamics, and with a significant initial displacement we are not finding that convergence at positive modes is a problem.