Inconsistent handling of guess vectors between escf and egrad

[marin.sap]

Dear developers,

when running escf in a directory where a previous calculation has been performed, it will use the values found in the sing_a/ciss_a file:

Code: [Select]

  logfile sing_a    exists already

 found      4 converged vectors on sing_a
 read       4 vectors from logfile sing_a
while egrad will always construct new guess vectors even when these files exist:

Code: [Select]

logfile sing_a    will be constructed
and consequently when running calculations along paths with small displacements between subsequent geometries, escf will need significantly fewer iterations than egrad. Is there a way to get egrad to behave in the same way as escf in this case?

This would be a useful option to have since it could significantly speed up calculations in some cases. We have also been testing a recently published method for generating guess vectors (https://pubs.aip.org/aip/jcp/article/159/12/121101/2912104), but can't use it with Turbomole when running our trajectories because egrad just ignores the guess vectors.

[uwe]

Hello,

the geometry does change from one egrad call to the next in a geometry optimization. Using the previous solution of the excitation vectors is in general not faster than generating them from scratch (that was a finding during the implementation and later usage), but has the benefit of having a cleaner initial guess which is not biased from the step before.

But you wrote that you are running calculations with small displacements only, and that there is hope to safe a few CPKS iterations. Well, in that case it might work.

So is there a way to let egrad read in the restart files? Yes, there is an undocumented keyword which allows exactly this.

But if I were to write this keyword here in the forum, it would no longer be undocumented... and I really think it could be dangerous to use it mindlessly.

I will write you a personal message, hope that's OK.

Cheers