[ADF-LIST] Best restart practices for SCF convergence

Alexei Yakovlev yakovlev at scm.com
Wed Jan 17 09:15:56 CET 2018

Dear Tim,

Which ADF version are you using? How many iterations is it able to do in 
6 days? Can you send me your input file for the difficult SCF case 
(remember to remove adflist from the To: field when doing so)?

Regarding restarting unconverged SCF, I believe it's not very useful due 
to the reason you cited: the data from previous iterations is very 
important for accelerating the whole thing. So the rule of thumb that I 
usually follow myself: try to get the SCF converged in a single run, 
possibly using 1 smearing parameter instead of five but trying to find 
the value that will work in one go.

Kind regards,


P.S. I'm assuming you're familiar with the SCF Troubleshooting section 
of the ADF manual.

On 16/01/2018 21:49, Tim Wilson wrote:
> I have two use cases in which SCF convergence is inadequate and I’d like some advice.
> First, I have some large metal cluster single point calculations that are difficult to converge. These are running on a cluster with a runtime limit (6 days) that is too low for me to finish the single point job in one run. If I employ smearing I can get a cluster to converge at, e.g., 1Ry. In a single job I can achieve convergence through about 5 smearing steps (from 1Ry to 0.1Ry) before I approach the runtime limit. I then use the TAPE21 from the previous job converged at 0.1Ry smearing, and start a new job that lowers smearing through another 5 steps starting at 0.1Ry (the smearing level at which the previous converged job finished). However, the restart job fails to converge at 0.1Ry smearing, even though the previous job converged.
> Second, when restarting using a converged TAPE21 for the sole purpose of performing Bader analysis (on the same metal cluster, assuming it’s fully converged at ground state), the SCF might fail to converge, even though the job was provided with a converged TAPE21 for the same system. In this case, it seems ADF should simply see that the restart TAPE21 contains a successful converged termination status and just bypass the SCF procedure.
> I see in the documentation that:
> "The SCF and optimization procedures use history to improve convergence behavior. Most of such history information is not stored on a restart file. As a consequence, a restart may not continue exactly as the original run would have done if it hadn’t terminated.”
> With that in mind, I imagine that there are guidelines for utilizing a converged restart file for the use cases described above that will best accelerate SCF convergence upon restart?
> Thanks,
> Tim
> Tim Wilson
> Colorado School of Mines — Applied Chemistry
> Molecular Theory Group — Research Associate
> _______________________________________________
> ADFlist mailing list
> ADFlist at scm.com
> https://lists.scm.com/mailman/listinfo/adflist

More information about the ADFlist mailing list