Orbital gradient RMS convergence issue

Rulin · November 19, 2016, 12:29am

Dear developers:

I’m now trying to perform a ROHF-BCCD(T)-DK3/Userdefined Basis calculation, and I’m now having a SCF convergence issue.
The problem really is just about the density rms convergence because the energy is converged to E-10.
The question is how do I improve the density rms convergence?
I have looked into the manual but maybe I was negligent that I did not find anything quite useful.
Here are the input and part of the output:

#! Sample

memory 450 mb

molecule Pu {
0 7
Pu
}

set {
basis puo_psi4
BASIS_RELATIVISTIC puo_psi4_decon
basis_guess guess_psi4
reference rohf
maxiter 300
soscf true
docc [13,3,3,3,1,7,7,7]
socc [0,0,0,0,1,1,2,2]
relativistic dkh
dkh_order 3
frozen_docc [8,2,2,2,1,5,5,5]
units angstrom
print_MOs true
print 2
SCF_TYPE PK
}

energy(‘ccsd(t)’)
energy(‘bccd(t)’)

                    Total Energy        Delta E     RMS |[F,P]|

@ROHF iter 0: -59318.02481777605135 -5.93180e+04 0.00000e+00
@ROHF iter 1: -58615.82814712078834 7.02197e+02 1.83322e+00
@ROHF iter 2: -59150.80897770584852 -5.34981e+02 6.93622e-01 DIIS
@ROHF iter 3: -59136.86862110742368 1.39404e+01 5.50682e-01 DIIS
@ROHF iter 4: -59313.92296621794230 -1.77054e+02 1.64385e-01 DIIS
@ROHF iter 5: -59320.11249645098724 -6.18953e+00 6.10149e-02 DIIS
@ROHF iter 6: -59324.79973564973625 -4.68724e+00 5.55518e-02 DIIS
@ROHF iter 7: -59326.42060920816584 -1.62087e+00 2.18728e-02 DIIS
@ROHF iter 8: -59325.43127388770517 9.89335e-01 3.00068e-02 DIIS
@ROHF iter 9: -59325.82774186931783 -3.96468e-01 2.32858e-02 DIIS
@ROHF iter 10: -59326.03309355521924 -2.05352e-01 1.93570e-02 DIIS
@ROHF iter 11: -59326.23323939429974 -2.00146e-01 1.49007e-02 DIIS
@ROHF iter 12: -59326.29212223774084 -5.88828e-02 1.37931e-02 DIIS
@ROHF iter 13: -59326.30924348635017 -1.71212e-02 1.30468e-02 DIIS
@ROHF iter 14: -59326.70388870295574 -3.94645e-01 3.70258e-03 DIIS
Gradient element too large for SOSCF, using DIIS.
Did not take a SOSCF step, using normal convergence methods
@ROHF iter 15: -59326.82042048194853 -1.16532e-01 9.49905e-03 DIIS
Gradient element too large for SOSCF, using DIIS.
Did not take a SOSCF step, using normal convergence methods
@ROHF iter 16: -59326.85632137212087 -3.59009e-02 8.97002e-03 DIIS
@ROHF iter 17: -59326.85688851574378 -5.67144e-04 6.90996e-03 SOSCF, nmicro = 5
Gradient element too large for SOSCF, using DIIS.
Did not take a SOSCF step, using normal convergence methods
@ROHF iter 18: -59320.78262491374335 6.07426e+00 8.81235e-02 DIIS
@ROHF iter 19: -59326.85265073690971 -6.07003e+00 6.62735e-03 DIIS
@ROHF iter 20: -59326.88470410564332 -3.20534e-02 6.89174e-03 SOSCF, nmicro = 5
Gradient element too large for SOSCF, using DIIS.
Did not take a SOSCF step, using normal convergence methods
@ROHF iter 21: -59326.88154678140563 3.15732e-03 9.50459e-03 DIIS
@ROHF iter 22: -59326.91504712727328 -3.35003e-02 3.66728e-03 SOSCF, nmicro = 5
Gradient element too large for SOSCF, using DIIS.
Did not take a SOSCF step, using normal convergence methods
@ROHF iter 23: -59326.49063838031725 4.24409e-01 1.30042e-02 DIIS
@ROHF iter 24: -59326.91210847272305 -4.21470e-01 3.52125e-03 DIIS
Gradient element too large for SOSCF, using DIIS.
Did not take a SOSCF step, using normal convergence methods
@ROHF iter 25: -59326.02386211624980 8.88246e-01 1.97166e-02 DIIS
@ROHF iter 26: -59326.02991803424084 -6.05592e-03 1.96624e-02 DIIS
@ROHF iter 27: -59326.03237606327457 -2.45803e-03 1.95798e-02 DIIS
@ROHF iter 28: -59326.03205584034731 3.20223e-04 1.95716e-02 DIIS
@ROHF iter 29: -59326.03200173590449 5.41044e-05 1.95747e-02 DIIS
@ROHF iter 30: -59326.03202121431968 -1.94784e-05 1.95748e-02 DIIS
@ROHF iter 31: -59326.03201963027823 1.58404e-06 1.95747e-02 DIIS
@ROHF iter 32: -59326.03201918516424 4.45114e-07 1.95747e-02 DIIS
@ROHF iter 33: -59326.03201926116890 -7.60047e-08 1.95747e-02 DIIS
@ROHF iter 34: -59326.03201927398914 -1.28202e-08 1.95747e-02 DIIS
@ROHF iter 35: -59326.03201926928887 4.70027e-09 1.95747e-02 DIIS
@ROHF iter 36: -59326.03201926925249 3.63798e-11 1.95747e-02 DIIS
@ROHF iter 37: -59326.03201926944166 -1.89175e-10 1.95747e-02 DIIS
@ROHF iter 38: -59326.03201926944894 -7.27596e-12 1.95747e-02 DIIS
@ROHF iter 39: -59326.03201926945621 -7.27596e-12 1.95747e-02 DIIS
@ROHF iter 40: -59326.03201926946349 -7.27596e-12 1.95747e-02 DIIS
@ROHF iter 41: -59326.03201926944166 2.18279e-11 1.95747e-02 DIIS
@ROHF iter 42: -59326.03201926943439 7.27596e-12 1.95747e-02 DIIS
@ROHF iter 43: -59326.03201926944894 -1.45519e-11 1.95747e-02 DIIS
@ROHF iter 44: -59326.03201926944894 0.00000e+00 1.95747e-02 DIIS
@ROHF iter 45: -59326.03201926941256 3.63798e-11 1.95747e-02 DIIS

Many thanks, and happy holiday!
-Rulin

dgasmith · November 19, 2016, 3:12pm

Atom this heavy are usually outside our experience, but ill see what we can do.

First up, SOSCF is probably not a good choice for heavy atoms. Since SOSCF uses rotations its much more likely to find a local minimum and stay there. In addition, the shape of your local minima are not likely to be quadratic which can lead to some odd behavior. This can be seen in some of the SOSCF steps where the energy increases.

I have seen ROHF get “stuck” before which was due to the gradient being constructed improperly. I rewrote the gradient this year so that the gradient the Fock matrix now contain the same rotations, so it shouldnt happen anymore. However, clearly it does. Ill have to check back into it. You can always simply override the convergence issue if you are happy with the state Psi found with die_if_not_converged False.

Stability analysis would probably be useful in getting out of local minima, option: stability_analysis follow. Again, im not sure how well this will work for these systems.

Playing around with diis_start and diis_max_vecs may help as well.

Finally, im not sure how well DKH will perform for plutonium. It would seem Breit terms would play a large roll for transuranic compounds. These computations are usually performed with ECP’s to offset such effects (which Psi unfortunately does not have at the moment).

Good luck!

Rulin · November 21, 2016, 5:20pm

Thanks dgasmith, I am trying to run some heavy element containing molecules, Pu atom is partly for the test calculations. The cases I tried are not likely to converge due to the same density RMS issue.
I will be very happy to provide additional test information if you need any.

loriab · November 21, 2016, 5:35pm

I’m not familiar with these heavy-atom calcs either, but damping might be another knob to tweak.

Rulin · November 21, 2016, 9:32pm

Thanks loriab, I did not know damping is implemented for version 1.1 for rohf, I’ll trying that too.