Next: Comparison of SQVV and Up: Results Previous: Slow-response Quenched Velocity Verlet Contents
We find that if the force constant is too small many more iterations are needed to converge the images to the required RMS tolerance, regardless of the type of quenching. In addition, the path exhibits a more uneven image distribution. This result occurs because at the initial stage the images may have very different gradients from the true potential along the band, because they lie far from the required path, and the gradient of the true potential governs the optimisation. When the true RMS force is reduced the springs start to play a more important role. But at this stage the forces are small and so is the QVV step size. The influence of the springs is actually most important during the initial optimisation stage, for it can determine the placement of images in appropriate regions. It is less computationally expensive to guide an image into the right region at the beginning of an optimisation than to restore the distribution afterwards by dragging it between two minima through a transition state region.
Choice of the Force Constant
If is too big the NEB never converges to the required RMS gradient tolerance value. Instead, it stays in proximity to the path but develops oscillations: adjacent images start to move in opposite directions. For all types of quenching we observed similar behaviour when large values of the force constant were used. This problem is related to the step in coordinate space that the optimiser is taking: for the SQVV case simply decreasing the time step remedies this problem.
Next: Comparison of SQVV and Up: Results Previous: Slow-response Quenched Velocity Verlet Contents Semen A Trygubenko 2006-04-10