Hi there, I am trying to optimize this system with OO-RI-MP2/aug-cc-pVTZ. After few successful iterations, the following error occurs.
Fatal Error: DF-OCC iterations are diverging
Error occurred in file: /scratch/psilocaluser/conda-builds/psi4-multiout_1563933640427/work/psi4/src/psi4/dfocc/occ_iterations.cc on line: 284
The input is as follows, Output file I am attaching.
memory 5 gb
Cl 0.20345 -1.85444 0.00007
Br -0.01639 0.28688 -0.00013
O -0.31233 2.88500 0.00006
H -1.26681 3.06549 -0.00000
4e2-dfomp2.txt (106.4 KB)
You don’t need to mess with damping or SOSCF for this system.
Once again, pinging @bozkaya… I’m out of recommendations.
Okay, I will try without damping and SOSCF. Lets see what happens. @jmisiewicz
However, I succeeded to optimize at DF-OMP2 level for a few systems. The change in geometry was very negligible, although I am yet to analyze all the systems under consideration.
Regarding this question, How expensive/accurate is OMP2?
Damping and SOSCF will only affect the Hartree-Fock part, which isn’t the issue here.
I don’t understand your second question. Are you asking about changing from DF-OMP2 to OMP2 with conventional integrals?
For few of the systems, SCF convergence was the issue, that is why, I have used SOSCF and damping.
I tagged the second question because there I asked previously about the accuracy of OOMP2, albeit with RI approximation. It doesn’t concern me now.
Hi Prasanta, another strange molecule? Why OO-MP2?
I optimized your molecule with BLYP/def2-tzvp in 14 steps:
CL 0.2041036785 -2.0069077791 -0.0002349663
BR -0.0359520630 0.3799283765 0.0001541295
O -0.1955161160 2.3541862951 -0.0002393798
H -1.1636254352 2.5212092636 -0.0001173633
From this point further optimization with B3LYP leads to:
CL 0.0344598886 -1.9414802018 -0.2591692457
BR 0.0594165198 0.3860394632 0.1788781019
O -0.2910374580 2.1929722581 -0.3094792407
H -1.2293368467 2.3310975351 -0.1030111628
But it took 46 steps!
In your attached file I found
So you want to combine dft with omp2?!?
In my opinion OO-MP2 makes no sense if you need 50 or 100 OO steps. This is really very expensive.
B2PLYP/def2-tzvp optimization leads to
CL 0.016535060839 -1.913565018007 -0.275977224241
BR 0.067582392820 0.382674851349 0.189829687098
O -0.292152854836 2.151232965816 -0.326188970663
H -1.229124292780 2.288415561779 -0.112193779305
I tested energy(‘OMP2’, set_dft_functional=‘bhandhlyp’). DFT was fast, the OO was slow: 50 iterations.
Hence I checked the anion: OO convergence in 7 iterations!
Just have a look on your orbital energies. Maybe the single UHF reference for the neutral molecule is bad?
The idea is that instead of UHF orbitals you have UKS orbitals that include correlation effects as a better start for the OO procedure.
Wanting OO-MP2 results here, even if convergence is slow, is not that outlandish in my opinion.
@OliP , the reason for using DFT is what @hokru just quoted.
I basically understand that it makes sense to combine DFT and OO-MP2, even if convergence is slow, but in this case the OO procedure fails (after more than 90 iterations) while optimizing the geometry of an open shell molecule.
Do we need a multi reference approach?
In this case OO-MP2 would not be my preferred solution…
Have you used Psi4’s DF-OMP2 code before? The orbital convergence is very poor due to the details of how the DIIS extrapolation is performed. Multireference systems may cause problems converging orbitals in DF-OMP2, but problems converging orbitals in DF-OMP2 don’t suggest the system is multireference.
Sorry for the misunderstanding. I think DF-OMP2 will not be helpful for the geometry optimization of ClBrOH. And I suggested to check a multireference approach instead.
@OliP I have used DFOMP2 because the interaction energies in open-shell systems can be quite flawed for MP2 and OO-MP2 is better, as suggested in analyses based on the O23 set by Bozkaya in J. Chem. Theory Comput. 2013, 9, 4679−4683 DF approximations are used to decrease the computational resources.
This is also my first encounter with DFOMP2.
Okay, DFOMP2 is better than MP2, analytical gradients are faster than CCSD.
As mentioned above, I optimized the geometry on BLYP/def2-tzvp and the B3LYP/def2-tzvp. And then I started an optimization with CCSD/def2-tzvp. It took 12 steps. In the first steps CCSD neded about 30 iterations to converge, in the last only 10 (starting with previous t2 amplitudes). Geometry:
CL -0.0337669722 -1.9479303270 -0.3097749519
BR 0.0947484159 0.4064204619 0.2096183867
O -0.3149431750 2.1130396587 -0.3492392175
H -1.2493299557 2.2275419037 -0.1232440994
With this geometry the orbital optimization for DFOMP2 (set_dft_functional=‘bhandhlyp’) converged after 64 iterations…