Dear Psi4 Developers,
Many thanks for making Psi4 available to scientific community! When using the 4th order DKH correction to calculate neutral atoms at the Hartree-Fock level, SCF does not seem to converge, unlike with the 2nd or 3rd order DKH Hamiltonians. For example,
memory 8 GB
molecule Kr {
0 1
Kr 0. 0. 0. }
set {
basis cc-pvdz-dk
basis_relativistic cc-pvtz-dk
reference uhf
maxiter 100
ints_tolerance 1.e-14
scf_type pk
molden_with_virtual false
relativistic dkh
dkh_order 2 }
E, wfn = energy(‘scf’, return_wfn=True)
molden(wfn, ‘Kr_0.molden’)
converges very quickly:
Total Energy Delta E RMS |[F,P]|
@UHF iter 1: -2793.10359581754301 -2.79310e+03 2.14464e-01 ADIIS
@UHF iter 2: -2797.68286988455566 -4.57927e+00 6.82176e-02 DIIS/ADIIS
@UHF iter 3: -2797.82573285833632 -1.42863e-01 1.07805e-02 DIIS/ADIIS
@UHF iter 4: -2797.83169453878054 -5.96168e-03 9.24308e-04 DIIS/ADIIS
@UHF iter 5: -2797.83175188186442 -5.73431e-05 1.20560e-04 DIIS/ADIIS
@UHF iter 6: -2797.83175295535511 -1.07349e-06 6.52788e-07 DIIS
@UHF iter 7: -2797.83175295539877 -4.36557e-11 5.70159e-08 DIIS
Energy and wave function converged.
So does the calculation with dkh_order set to 3:
Total Energy Delta E RMS |[F,P]|
@UHF iter 1: -2793.11277575518261 -2.79311e+03 2.14462e-01 ADIIS
@UHF iter 2: -2797.69195653273755 -4.57918e+00 6.82143e-02 ADIIS/DIIS
@UHF iter 3: -2797.83480826836467 -1.42852e-01 1.07796e-02 ADIIS/DIIS
@UHF iter 4: -2797.84076907161671 -5.96080e-03 9.24270e-04 ADIIS/DIIS
@UHF iter 5: -2797.84082641155828 -5.73399e-05 1.20559e-04 ADIIS/DIIS
@UHF iter 6: -2797.84082748502942 -1.07347e-06 6.52801e-07 DIIS
@UHF iter 7: -2797.84082748507217 -4.27463e-11 5.70201e-08 DIIS
Energy and wave function converged.
However, when using “dkh_order 4”, the calculation does not converge
Total Energy Delta E RMS |[F,P]|
@UHF iter 1: -2793.11383899603334 -2.79311e+03 2.14462e-01 ADIIS
@UHF iter 2: -2797.69301678776128 -4.57918e+00 6.82142e-02 ADIIS/DIIS
@UHF iter 3: -2797.83586822539519 -1.42851e-01 1.07795e-02 ADIIS/DIIS
@UHF iter 4: -2797.84182898827476 -5.96076e-03 9.24444e-04 ADIIS/DIIS
@UHF iter 5: -2797.84188632782343 -5.73395e-05 1.21848e-04 ADIIS/DIIS
@UHF iter 6: -2797.84188740112540 -1.07330e-06 1.77465e-05 DIIS
@UHF iter 7: -2797.84188740116861 -4.32010e-11 1.77327e-05 DIIS
@UHF iter 8: -2797.84188740116588 2.72848e-12 1.77327e-05 DIIS
…
@UHF iter 97: -2797.84188740116679 4.54747e-13 1.77325e-05 DIIS
@UHF iter 98: -2797.84188740116633 4.54747e-13 1.77325e-05 DIIS
@UHF iter 99: -2797.84188740116770 -1.36424e-12 1.77325e-05 DIIS
@UHF iter 100: -2797.84188740116815 -4.54747e-13 1.77325e-05 DIIS
PsiException: Could not converge SCF iterations in 100 iterations.
Failed to converge.
It seems that while the energy converges well, RMS [of the orbital gradient] is “stuck”
I have tried using SOSCF instead of DIIS, and tried to include damping and level shifting but whaever I tried, did not improve the convergence… I have also tried different atoms (Po, Sn, Ge and even Si) and different basis sets (ultimately, I would like to do calculations with the UGBS basis sets available at the Basis Set Exchange website). For example, for Po with the cartesian UGBS basis set, DKH2 and DKRH3 converge in under 15 cycles:
DKH2:
@UHF iter 9: -22170.85171556465866 -5.04765e-04 1.01320e-04 ADIIS/DIIS
@UHF iter 10: -22170.85174134019690 -2.57755e-05 1.18231e-05 DIIS
@UHF iter 11: -22170.85174213978098 -7.99584e-07 2.59736e-06 DIIS
@UHF iter 12: -22170.85174219897453 -5.91936e-08 4.58594e-07 DIIS
Energy and wave function converged.
DKH3:
@UHF iter 9: -22199.95106083959035 -5.03529e-04 1.01268e-04 DIIS/ADIIS
@UHF iter 10: -22199.95108658786921 -2.57483e-05 1.18153e-05 DIIS
@UHF iter 11: -22199.95108738750059 -7.99631e-07 2.59613e-06 DIIS
@UHF iter 12: -22199.95108744670142 -5.92008e-08 4.55252e-07 DIIS
Energy and wave function converged.
but in the DKH4 calculation, RMS never reaches convergence (energy is Ok):
@UHF iter 97: -22194.59984812520270 4.36557e-11 1.20787e-01 ADIIS
@UHF iter 98: -22194.59984812508628 1.16415e-10 1.20787e-01 ADIIS
@UHF iter 99: -22194.59984812504263 4.36557e-11 1.20787e-01 ADIIS
@UHF iter 100: -22194.59984812500988 3.27418e-11 1.20787e-01 ADIIS
PsiException: Could not converge SCF iterations in 100 iterations.
Failed to converge.
I am using Psi4 1.9.1 binaries installed via conda, and have attached to this message my input and output files for Kr.
I would appeciate if you could help me figure out what I am doing wrong.
Thank you,
Anatoliy
Kr_0.in.txt (332 Bytes)
Kr_0_DKH2.out.txt (11.5 KB)
Kr_0_DKH3.out.txt (11.5 KB)
Kr_0_DKH4.out.txt (16.2 KB)