Performance of autocas

[LeonardoBarneschi]

Dear developers,
I am attempting an autocas calculation on a small molecule, attached.

After having compiled a working installation of OpenMolcas-QCMaquis (+NEVPT2) and scine-autocas, I was trying to
perform for the first time an active space determination.

I tried by executing the following
python3 -m scine_autocas --xyz_file mol2.xyz --basis_set 6-31G* --plot --interface Molcas

The preliminary &SCF calculation was terminated rather quickly, while the follow-up DMRGSCF calculation has been running for 12 hours already. Since it's the first time trying this, I wanted to ask you if this sounds reasonable or if there's something I should be looking into to enhance performance.

I am running with 40 OMP threads and using all the memory at disposal of my computing nodes. BTW, by looking at active processes I found out that at this stage, the calculation is using a single cpu.
I attach also the output of the completed &SCF calculation and ongoing &DMRGSCF calculation, together with the QCMaquis log file.

Finally, I know it is recommended to provide an input yml file but I come across the same issues reported in #5.

Any help is deeply appreciated.
Kindest regards,
Leonardo

My geometry:
27

C -1.08745 0.31737 -0.18798
N 0.00901 1.08097 0.03399
C 1.09740 0.30502 0.25267
C 0.72854 -1.05070 0.17543
C -0.73273 -1.04244 -0.11662
C -1.15761 -2.36303 -0.20081
C -0.01317 -3.16903 0.02491
C 1.13964 -2.37602 0.25388
C 3.59051 0.05087 0.21939
C -3.58313 0.08905 -0.15666
H -2.15229 -2.72979 -0.39441
H -0.01879 -4.25011 0.02261
H 2.13045 -2.75393 0.44587
H 3.55558 -0.23466 -0.83223
H 3.57515 -0.84742 0.83714
H 4.52016 0.58563 0.40700
H -4.50709 0.63435 -0.34205
H -3.55163 -0.20187 0.89360
H -3.57680 -0.80637 -0.77872
C -2.40098 0.97716 -0.48761
H -2.42102 1.20454 -1.56440
C 2.41780 0.94963 0.55513
H 2.44051 1.17141 1.63304
O -2.53116 2.17856 0.24389
H -1.65986 2.59023 0.28515
O 2.56028 2.15326 -0.17039
H 1.69305 2.57357 -0.21034

dmrg.log
scf.log
autocas_project.QCMaquis.log

[mmoerchen]

Hi Leonardo,

our DMRG code is currently only OpenMP parallelized, so that it cannot utilize multiple nodes (we are working on an MPI version, but this will take a few months).
Since your initial active space is rather large even for DMRG, it might take while to find the final active space.
You could try the "large active space protocol", which is accessible through the yaml input, in order to speed up the CAS search.

Best Max

[LeonardoBarneschi]

Thank you very much for your answer!
Indeed I was trying here and there and I was finally both able to use a yaml file to launch autocas as well as using the large active space protocol. I take advantage (if I may) to ask for an extra suggestion. Since this molecule I was using to test autocas is planar and rigid with an extended pi conjugation, I was expecting it to be a fairly easy case for autocas.

I tried to play with some input parameters here and there but in most cases all of the orbitals are dropped from the active space as weakly correlated.
In all cases I am using:

• weak_correlation_threshold: 0.02
• single_reference_threshold: 0.10

I found a bit of improvement by increasing the maximum number of orbitals in the "large active space protocol", meaning that I am able to recover correct CAS sub-active spaces, e.g. (4,4) or (6,6) with respect to the (8,8) which I would be using to include all the pi-pi* couples.

I am using a very small basis (STO-3G) to keep the computational time reasonable.

Is there anything I could be looking into?

Best regards,
Leonardo

[mmoerchen]

The input parameters look good, but the size of the sub-CAS is probably still to small.
You could try something around CAS(20, 20) to prevent orbitals from dropping out of the CAS.

Best,
Max