Restarting SCF starts SCF all over again

Questions
#1

I run a job using this command line:

qchem -nt 60 jobname.in jobname.out jobname.scr

It fails to converge with DIIS and the DIIS error
is 3.24e-05

I add SCF_GUESS=READ to jobname.in, delete jobname.out and resubmit using the same command line:

qchem -nt 60 jobname.in jobname.out jobname.scr

The SCF DIIS error starts at 1.09e-03

Why does the SCF DIIS error not start at 3.24e-05?

jobname.in is unaltered with the exception of the addition of SCF_GUESS. The jobname.scr directory and files contained within are unaltered.

I’m running qchem 6.0.2 for Intel X86 EM64T Linux

#2

DIIS is an extrapolation technique based on previous few Fock matrices and with SCF_GUESS = READ you have lost that subspace (only the MOs at the last cycle are retained and thus only the Fock matrix at the last cycle can be reconstructed). The error is different because the extrapolation is starting over.

#3

Argh.

Thanks for the reply. Are there any input parameters that would allow storage of previous Fock matrices? This behavior by QChem rather eliminates the utility of doing restarts.

If there is no way to control this in the current incarnation of QChem, have any feature requests been filed to modify QChem to store Fock matrices in the scratch folder for reuse?

#4

It’s not a restart, which we don’t have, it’s a reading of the density matrix from the last SCF cycle. There is no simple way to preserve the DIIS subspace.