About structural optimization of atomic cluster

[chp]

Recently I have written a small python script based on ATK for structural optimization of randomly generated atomic clusters. It works well for small molecular, however, when it is used for medium-sized cluster, it cannot converge after thousands of steps. I have tried to adjust some parameters, but no obvious improvement has been observed.

Please see the following script for structural optimization, an randomly generated cluster compose of 34 oxygen atoms and 34 zinc atoms, and output file.

What is the matter ? How to deal with it? Thanks everyone !!!

[Anders Blom]

The "standard" tricks for better convergence would be to lower the diagonal mixing parameter, perhaps to 0.05 or lower. Sometimes it helps to have more history steps (12-20), sometimes fewer. I would probably increase the mesh cut-off, provided you have enough memory for it.

Hope these simple suggestions help.

[Nordland]

Based on the output I would guess that two atoms are sitting really really close to one another, if not on top of one another.
This is the only case where I have seen ATK having problems with mocules when the other parameters is decent.

If anything else I would try increasing the mesh-cutoff slightly any way, as zinc has some really nasty d-orbitals in the pseudo-potentials.

[chp]

Thank you for your nice advice!

Now it can work. But I have another question. The main aim of my script is to optimize any randomly generated molecular roughly and quickly, however, the optimization steps are often very long due to the randomly initial geometry. Is there any “tricks” to improve this situation? I have tried to adjust some parameters such as the “optimizer” of “geometricOptimizationParameters”, But no clear hints are obtained. For the optimization of molecular, which of the “optimizer algorithms” (QuasiNewton and SteepestDescent) is better? In the case of “SteepestDescent”, how to set the best value of the “time_step” parameter to finish quickly the optimization?

Thanks a lot !

[zh]

For the optimization of molecular, which of the “optimizer algorithms” (QuasiNewton and SteepestDescent) is better?

Usually the QuasiNewton optimizer is more efficient than the steepest descent one. You could first use the steep descent method for  a couple of optimization steps and then switch into the QuasiNewton optimizer.