5
$\begingroup$

Does anyone know of a paper, textbook, or other resource which outlines the expression for the analytic gradient of the HF energy (restricted or otherwise) in terms of Dunning's contracted Gaussian-type orbitals $\chi^{CGTO}_\mu(r)$? I'm talking about an expression that someone could simply sit down and program as a free-standing code using only a converged HF calculation as input. I.e. has someone written this down in a form where the only arguments of the expression would be the primitive GTO exponents, the contraction coefficients, the atom centers, the nuclear charges, and the converged eigenvalues and eigenvectors of the Hartree Fock SCF?

Plenty of textbooks (Szabo and Ostlund, Yamaguchi et al, etc) describe the theory of analytic gradients using equations independent of basis set type, i.e. in an abstract way. Likewise, plenty of standard codes (open source or otherwise) perform these calculations with Dunning CGTOs and/or other basis set types. But because of how huge and modular (i.e. calling across many routines to read) these codes are, interpreting them takes a great deal of time.

The form of the derivative of primitive GTOs is given by Szabo and Ostlund page 442 between equations C.12 and C.13. Since the contraction coefficients are not spatially dependent, this form should be (hopefully) simple to extend to a Dunning basis. From here, the form of the various derivative integrals (such as the matrix elements $\frac{\partial V_{N N}}{\partial X_A}_{\mu \nu}$ and $\frac{\partial V_{N e}}{\partial X_A}_{\mu \nu}$ given in the middle of page 442) to obtain the HF gradient should be doable analytically in Mathematica etc. Thus, the explicit form of all these integrals and the overall gradient expression should be possible to write down in a way someone (maybe even someone without electronic structure knowledge) could just sit down and code. The expression would be long, but easily interpreted.

I'm just wondering if this full expression exists anywhere, as I haven't been able to find it anywhere online and would like to before "recreating the wheel." In particular, I will be using output from a density-fitted Hartree Fock code for my analytic gradient implementation, but any expression will do, as extending to density-fit Coulomb and exchange is probably the least tedious part of this process. Thank you!

$\endgroup$

2 Answers 2

4
$\begingroup$

Molecular Electronic-Structure Theory, by Helgaker et al., is an extremely comprehensive textbook reference for electronic structure on finite systems. I take the following from Section 10.8.3 (pp. 481–482).

This section provides an expression for the gradient in terms of the generalized Fock matrix

$$ F_{mn} = \sum_\sigma \langle \text{CSF} \vert a_{m\sigma}^\dagger [ a_{n\sigma}, H] \vert \text{CSF} \rangle; $$

for real orbitals, at least, the gradient can be written as $$ E_{mn}^{(1)} = 2(F_{mn} - F_{nm}). $$

The following equations, (10.8.19–10.8.23), provide expressions for the generalized Fock matrix in terms of one- and two-electron integrals. For the Hamiltonian $$ H = \sum_{pq} h_{pq} E_{pq} + \frac12 \sum_{pqrs} g_{pqrs} e_{pqrs} + h_{nuc}, $$ recall that the Fock matrix is defined in terms of $a_{m\sigma}^\dagger[a_{n\sigma}, H]$, so that $$ F_{mn} = \langle \text{CSF} \vert \sum_\sigma a_{m\sigma}^\dagger[a_{n\sigma}, H] \vert \text{CSF} \rangle = \langle \text{CSF} \vert \sum_q h_{nq} E_{mq} + \sum_{qrs} g_{nqrs} e_{mqrs} \vert \text{CSF} \rangle, $$ then $$ F_{mn} = \sum_q D_{mq} h_{nq} + \sum_{qrs} d_{mqrs} g_{nqrs}, $$ where by definition $$ D_{pq} = \langle \text{CSF} \vert E_{pq} \vert \text{CSF} \rangle, \\ d_{pqrs} = \langle \text{CSF} e_{pqrs} \vert \text{CSF} \rangle.. $$

If you prefer, you can also get the gradient in terms of Roothan–Hall-style equations (Eq. 10.7.80).

Note (Section 2.5 of 1, pp. 51–53) that $\lvert \text{CSF} \rangle$ refers to a configuration state function, which is a spin-adapted modification of the Slater determinant. Construction of these is detailed in Section 2.6.

$\endgroup$
3
  • $\begingroup$ Thank you! From what I can tell, this is the gradient of SCF convergence (i.e. eq. 10.7.80 is what's used to converge Hartree Fock via the DIIS scheme of Pulay) rather than the analytic HF gradient with respect to nuclear geometry or an external field. Am I mistaken? $\endgroup$
    – Rob L
    Sep 15 at 19:28
  • $\begingroup$ I think you're right, see Eq. (4.2.13) and following: it's the gradient of the energy with respect to the variational parameters in the wavefunction (here, CSF). It looks like the entire textbook assumes fixed nuclei. Sorry if it doesn't solve your problem! $\endgroup$
    – elutionary
    Sep 15 at 20:59
  • 1
    $\begingroup$ Note that this is in second quantized notation and requires extensive manipulation to end up with the explicit spatial-only quantities you need to implement. $\endgroup$ Sep 19 at 11:58
0
$\begingroup$

If you want to use Mathematica, simply write down the expression for the total energy and take the derivative with respect to the nuclear coordinates.

The explicit form of the expressions can be found in standard textbooks and the literature; see e.g. the seminal work of Pulay. All you need is density matrix elements, and the molecular (gradient) integrals. Automated differentiation is another option, and you can also find these works easily in the literature.

$\endgroup$
2
  • $\begingroup$ Hi Susi, thank you for your response. Maybe I wasn't clear in my question: I'm looking for an expression of the gradient in terms of the actual basis functions. I.e. the the only arguments of the expression would be the primitive GTO exponents, the contraction coefficients, the atom centers, the nuclear charges, and the converged eigenvalues and eigenvectors of the Hartree Fock SCF. The expressions in standard textbooks etc are helpful in understanding the theory of analytic gradients but don't translate directly to a working code. $\endgroup$
    – Rob L
    Sep 12 at 16:00
  • $\begingroup$ So you don't know how to evaluate integrals? $\endgroup$ Sep 13 at 9:48

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .