VTST 2.x code updates
Posted: Sun Feb 25, 2007 4:16 pm
The vtstcode.tar.gz has been updated to version 2.0. This new code has many significant changes including:
- Force based optimizers that work with the NEB/dimer/Lanczos methods. These optimizers include conjugate gradients, quick-min, lbfgs, and fire. The optimizer is selected with the IOPT variable. If one of these optimizers is selected, the INCAR variables IBRION should be set to 3 and POTIM to 0. This tells vasp to do damped dynamics with a zero time step. In this way, the built-in vasp optimizers will not move the ions. The new optimizers are designed to improve the efficiency of the transition state finding methods, but they can also be used with normal vasp minimization runs.
- The dimer method has been modified to work with a single image instead of two. This has reduced the total number of force evaluations required to find a saddle. The setup for a dimer and lanczos run are now the same. Both use a MODECAR file to store the direction along the lowest curvature mode.
- The NEB code has been simplified. We no longer support variable spring constants along the band -- this is not a popular option given that the climbing image does the job of finding the saddle point. The double-nudging force-projections have been implemented, although we have not found this to improve convergence of the NEB.
Known issues:
- The force-only conjugate gradient and lbfgs algorithms use a finite-difference / Newton's method line minimization step. In the finite-difference step, the forces can increase. Close to a minimum, every other step will have this artificial rise in the force -- it does not mean that the method is not converging. In the future, we will find a way to hide or systematically reduce the finite difference step so that convergence does not appear rough.
- These methods have only been tested in our group, and there are bound to be a few bugs. If you find any, or have problems with the methods, please let us know and we’ll fix them.
- Force based optimizers that work with the NEB/dimer/Lanczos methods. These optimizers include conjugate gradients, quick-min, lbfgs, and fire. The optimizer is selected with the IOPT variable. If one of these optimizers is selected, the INCAR variables IBRION should be set to 3 and POTIM to 0. This tells vasp to do damped dynamics with a zero time step. In this way, the built-in vasp optimizers will not move the ions. The new optimizers are designed to improve the efficiency of the transition state finding methods, but they can also be used with normal vasp minimization runs.
- The dimer method has been modified to work with a single image instead of two. This has reduced the total number of force evaluations required to find a saddle. The setup for a dimer and lanczos run are now the same. Both use a MODECAR file to store the direction along the lowest curvature mode.
- The NEB code has been simplified. We no longer support variable spring constants along the band -- this is not a popular option given that the climbing image does the job of finding the saddle point. The double-nudging force-projections have been implemented, although we have not found this to improve convergence of the NEB.
Known issues:
- The force-only conjugate gradient and lbfgs algorithms use a finite-difference / Newton's method line minimization step. In the finite-difference step, the forces can increase. Close to a minimum, every other step will have this artificial rise in the force -- it does not mean that the method is not converging. In the future, we will find a way to hide or systematically reduce the finite difference step so that convergence does not appear rough.
- These methods have only been tested in our group, and there are bound to be a few bugs. If you find any, or have problems with the methods, please let us know and we’ll fix them.