I am trying to calculate the vibrational frequencies of the excited state of 9methylantracene. A problem rises when the program is trying to calculate the MO derivatives via CPSCF. The system didn’t converge after more than 100 iterations and ‘Collapsing Subspace’ message is printed after 5-8 iterations
With RPA=1, I have this error message:
Warning: Singularity detected in AXeqB0 169 130 1
Q-Chem fatal error occurred in module 0, line 118:
Disabling the RPA didn’t help.
Any suggestion please?
BASIS = aug-cc-pVDZ
CIS_N_ROOTS = 1
CIS_STATE_DERIVATIVE = 1
CIS_TRIPLETS = 0
EXCHANGE = B3LYP
GUI = 2
JOB_TYPE = Frequency
MEM_STATIC = 3500
METHOD = HF
RPA = 1
SCF_CONVERGENCE = 8
SCF_MAX_CYCLES = 100
CIS_MAX_CYCLES = 150
SYM_IGNORE = 1
XC_GRID = 2
S2THRESH = 17
THRESH = 14
Couple of things here. First is that your input file doesn’t even run for me unless I set MEM_TOTAL to something larger than the default. Anything < 13000 and you wind up solving CPSCF equations in segments, which you don’t want if you have the memory to avoid so I just set MEM_TOTAL to be all the memory on my compute node.
That aside, I can reproduce your problem with the CPSCF convergence if I use your $rem just with MEM_TOTAL specified. I notice that you have set both METHOD = HF and EXCHANGE = B3LYP, which is ambiguous. (We should probably trap that, but we don’t apparently. I will submit that as a bug ticket.) If I comment out METHOD = HF, this job runs successfully, CPSCF converges in 6 cycles. In your original input, the SCF is run at the HF level but it’s possible that the CPSCF solver is trying to use B3LYP and that may be what’s causing the convergence problem.
Last thing that I notice from your sample is XC_GRID=2 (SG-2 grid), which is probably overkill for B3LYP, for which the default is SG-1. Unless you’ve had issues with SG-1, I would leave the default in place… saves about 15% for this job, according to my tests.
Thank you so much for your feedback. I think when the Mem_total is not specified it takes all the memory of the node and that’s why I didn’t have a problem at this point. The CPSCF is indeed converged quickly by disabling the HF method.
I am using the xc_grid =2 because if helps eliminating the imaginary frequencies. Tightening the convergence or twisting the geometry didn’t help much for these types of molecules.
I see. Probably landing on the TS for methyl rotation with the looser grid, I’ve seen that before many times. (And as the designer of SG-2 I am happy to have you use those grids, hope you will cite the paper: DOI 10.1002/jcc.24761). I have submitted a bug ticket about the specification of two different functionals so I hope that a future release will prevent you from doing this.
Thank you for the paper.
Do you think that the unpruned SG-2 would be more efficient to avoid the imaginary frequencies?
Are they still present with SG-2? I would try SG-3 before trying the unpruned grids.