Carbon Surface
The system is a hydrogen terminated Carbon surface. The surface is made up of 20 layers (atoms) in the (001) direction. We used a k-point mesh of 8x8x2 with an offset of (0,0,0.5).
Problem with the old direct minimization method
Comparisons of minimization techniques
Problem with the old direct minimization method
“Optimize insulator option”
This system is a prime example of a system where the old “optimize insulator system” breaks down. Fig. 1 shows the difference of the total energy from the final energy vs. the number of HY operations. There are 2 operations per k-point per conjugate gradient iteration. The plot readily shows instability and problems in the original algorithm “dir.mg.orig”. The new algorithm with the k-point weights incorporated differently (see write-up with GaAs) shows marked improvement in “dir.mg” and “dir.gcg”. As with the results with GaAs the GCG algorithm shows slight improvement over the MG algorithm.
Fig. 1 comparison of different direct minimization methods
Comparisons of direct minimization and self-consistent charge mixing methods
For this system, the direct minimization method exemplified by the “dir.gcg” algorithm is obviously superior to the charge mixing methods. This also has been seen in another system C system (see C881.doc) of a long skinny unit cell with one atom moved for the purposes of a frozen phonon calculation. Fig. 2 shows the significant savings of the direct method over the charge mixing schemes (see thom_fer.doc for discussion on the new charge mixing schemes). A maximum of 10 conjugate iterations per SCF cycle is used. This value usually gives the most efficient setting. Of the charge mixing schemes, Pulay-Kerker does the best. The Pulay-Thomas-Fermi (PTF) methods have instability as evidenced by the noticeable bump around 1500 seconds. The Broyden method (method in Paratec5.0 and before) shows even worse convergence.
Fig. 2 comparison of the direct “dir.gcg” method and the self-consistent charge mixing methods.
SETTINGS
