Hello Evgeniy,
restarts are the default for more or less all Turbomole modules.
So if you have calculated the first three excited states, change the input to get the first six and then rerun the calculation with ricc2, it will by default read in the converged vectors of the first three states and only calculates the remaining three.
That's the short answer.
.
.
.
.
.
.
A more detailed answer is following.
In the first run you have e.g.:
$excitations
irrep=a multiplicity= 1 nexc= 3
spectrum states=all operators=diplen
After the job has finished, there will be a number of CCXXX-1--1--N files in the job directory. It is important to not delete them, otherwise the restart will fail and ricc2 will start the CC2 calculation from the start.
Next step is to change the input to:
$excitations
irrep=a multiplicity= 1 nexc= 6
spectrum states=all operators=diplen
So the number of excited states has changed from 3 to 6.
The next ricc2 run will print to the output if restart was successful. If it was, the output contains several hints about the restart:
global parameters for ricc2 program:
hard restart (reuse of interm.) : enabled
soft restart (reuse of vectors) : enabled
[...]
restarted calculation begins with wavefunction model nb. 1 which is CC2
**************************************************************************
* *
* OPTIMIZATION OF THE GROUND STATE CLUSTER AMPLITUDES *
* *
**************************************************************************
restart CC2 from CC2 amplitudes found on the R0 file.
[...]
summary of start vectors generation:
-------------------------------------------
type of solution vectors : RE0
symmetry : a
number of vectors requested : 6
number of vectors generated : 6
-- restarted from file : 3
-- guessed as rhs/(e_a-e_i) : 3
-------------------------------------------
[...]
Please note that, in order to solve the equations for S4-S6 the first three need to be taken into account. So the size of the calculation increases, the matrices are bigger, the time to calculate S1-S3 is not the same as the time for S4-S6, even if the first three vectors are already converged. Nevertheless faster than without the restart data.
Further it can happen that including S4 in the calculation will also have an effect on S3, such that S3 needs to be refined.
Example: Use benzene as input, run in C1 symmetry such that S3 and S4 are degenerated. Restarting from the converged vectors for S1, S2 and S3 and including S4 to S6 will result in:
======== LINEAR CC RESPONSE SOLVER ========
threshold for convergence: 1.00E-06
maximum number of simultaneous Jacobi matrix transformations: 6
summary of start vectors generation:
-------------------------------------------
type of solution vectors : ME0
symmetry : a
number of vectors requested : 6
number of vectors generated : 6
-- restarted from file : 3
-- guessed as rhs/(e_a-e_i) : 3
-------------------------------------------
Iter #Vectors time (min) max. residual
--------------------------------------------
1 6 0.01 0.10E+01 ( 3)
2 4 0.01 0.80E-01 ( 3)
3 4 0.01 0.55E-02 ( 3)
4 4 0.01 0.46E-03 ( 3)
5 4 0.01 0.38E-04 ( 3)
6 4 0.01 0.29E-05 ( 3)
7 1 0.00 0.86E-06 ( 6)
--------------------------------------------
converged in 7 iterations
used 27 transformations
Total time 0.05 minutes
As you can see, the procedure starts with 6 vectors to be calculated. After the first iteration, one would expect 3 to-be-solved vectors, as the first three are already converged. For systems where S3 and S4 are not close, this is indeed the case.
But here S1 and S2 are found to be converged, while the remaining 3 states, S3-S6, need to be calculated simultaneously.
A final hint: If you use TmoleX and run remote jobs from the graphical user interface, the CC restart files are deleted before copying back the files to the local machine. So restarts are not possible. Larger coupled-cluster calculations generate quite big files and TmoleX brings back the results and minimizes the network traffic by deleting large restart files in those cases.
Best Regards,
Uwe
