Orbitals should be multiplied by sqrt(occ) not occ #3138
Conversation
|
Yes, by testing the guess energy for H atom. |
|
Looks like there are some unrelated test failures, but this one is ready to go. As shown by the changes in the pytest guess energies, the new code reaches a considerably lower SAD guess energy for the HF molecule: -100.02909951427 vs -99.63941801281894. Large improvements are expected especially for systems containing hydrogen atoms, since the existing code only includes 0.5 electrons on those atoms. |
tests/pytests/test_scf_options.py
Outdated
| "sap": -99.50144270485751, | ||
| "core": -88.99202896495, | ||
| "gwh": -94.35585144645, | ||
| "sad": -99.34836107817, |
There was a problem hiding this comment.
Is it a concern that with the change SAD UHF & ROHF (and often sadno, *huckel) guess energies are less deep?
There was a problem hiding this comment.
No, the issue is that the test here is FH+, which has orbitals that are more compact than the neutral FH. It is logical that the proper guess for the neutral molecule yields a slightly higher energy for the cation.
Like I mention above, right now the SAD code is placing 0.5 electrons for the hydrogen atom, whereas fluorine probably has 7 electrons distributed evenly on 4 spatial orbitals [no symmetry, this is the key feature in the OpenOrbitalOptimizer guess in #3136] which means that its number of electrons is also wrong by -0.875. Therefore, the current guess is spuriously good for FH+ as it corresponds to a cation with charge 1.375.
susilehtola
force-pushed
the
fix_sad_occ
branch
2 times, most recently
from
14c9a6f
to
dc01cc9
Compare
* Orbitals should be multiplied by sqrt(occ) not occ, since otherwise the density matrix becomes C occ^2 C^T. * Update pytest reference values * Update number of screened shells based on SAD guess * Specify occupation in large atoms test * Increase gradient test tolerance due to very small mismatch * clang-format * Relax test criteria due to small failures * Update test variable to remove deprecation warning * SAD guess energy check is for SAD energy not first variational energy * There is a 1e-7 change to the orbital energy in the HF/DZ test on SH2. * Wave function needs to be converged more tightly for test to pass * Increase test limit * Bump tolerance also for other gradient tests
…he density matrix becomes C occ^2 C^T.
Description
Fixes a bug in
sad.cc: pseudo-orbitals should be obtained by multiplying the orbital with the square root of their occupation to reproduce the correct density matrix C occ C^T.Closes #3137.
User API & Changelog headlines
Dev notes & details
Questions
Checklist
Status