Tuesday, March 28, 2017

Computational quantum chemistry in a nutshell

To the uninitiated (and particularly physicists) computational quantum chemistry can just seem to be a bewildering zoo of multiple letter acronyms (CCSD(T), MP4, aug-CC-pVZ, ...).

However, the basic ingredients and key assumptions can be simply explained.

First, one makes the Born-Oppenheimer approximation, i.e. one assumes that the positions of the N_n nuclei in a particular molecule are a classical variable [R is a 3N_n dimensional vector] and the electrons are quantum. One wants to find the eigenenergy of the N electrons. The corresponding Hamiltonian and Schrodinger equation is


The electronic energy eigenvalues E_n(R) define the potential energy surfaces associated with the ground and excited states. From the ground state surface one can understand most of chemistry! (e.g., molecular geometries, reaction mechanisms, transition states, heats of reaction, activation energies, ....)
As Laughlin and Pines say, the equation above is the Theory of Everything!
The problem is that one can't solve it exactly.

Second, one chooses whether one wants to calculate the complete wave function for the electrons or just the local charge density (one-particle density matrix). The latter is what one does in density functional theory (DFT). I will just discuss the former.

Now we want to solve this eigenvalue problem on a computer and the Hilbert space is huge, even for a simple molecule such as water. We want to reduce the problem to a discrete matrix problem. The Hilbert space for a single electron involves a wavefunction in real space and so we want a finite basis set of L spatial wave functions, "orbitals". Then there is the many-particle Hilbert space for N-electrons, which has dimensions of order L^N. We need a judicious way to truncate this and find the best possible orbitals.

The single particle orbitals can be introduced
where the a's are annihilation operators to give the Hamiltonian

These are known as Coulomb and exchange integrals. Sometimes they are denoted (ij|kl).
Computing them efficiently is a big deal.
In semi-empirical theories one neglects many of these integrals and treats the others as parameters that are determined from experiment.
For example, if one only keeps a single term (ii|ii) one is left with the Hubbard model!

Equivalently, the many-particle wave function can be written in this form.

Now one makes two important choices of approximations.

1. atomic basis set
One picks a small set of orbitals centered on each of the atoms in the molecule. Often these have the traditional s-p-d-f rotational symmetry and a Gaussian dependence on distance.

2. "level of theory"
This concerns how one solves the many-body problem or equivalently how one truncates the Hilbert space (electronic configurations) or equivalently uses an approximate variational wavefunction. Examples include Hartree-Fock (HF), second-order perturbation theory (MP2),  a Gutzwiller-type wavefunction (CC = Coupled Cluster), or Complete Active Space (CAS(K,L)) (one uses HF for higher and low energies and exact diagonalisation for a small subset of K electrons in L orbitals.
Full-CI (configuration interaction) is exact diagonalisation. This only possible for very small systems.

The many-body wavefunction contains many variational parameters, both the coefficients in from of the atomic orbitals that define the molecular orbitals and the coefficients in front of the Slater determinants that define the electronic configurations.

Obviously, one expects that the larger the atomic basis set and the "higher" the level of theory  (i.e. treatment of electron correlation) one hopes to move closer to reality (experiment). I think Pople first drew a diagram such as the one below (taken from this paper).


However, I stress some basic points.

1. Given how severe the truncation of Hilbert space from the original problem one would not necessarily to expect to get anywhere near reality. The pleasant surprise for the founders of the field was that even with 1950s computers one could get interesting results. Although the electrons are strongly correlated (in some sense), Hartree-Fock can sometimes be useful. It is far from obvious that one would expect such success.

2. The convergence to reality is not necessarily uniform.
This gives rise to Pauling points: "improving" the approximation may give worse answers.

3. The relative trade-off between the horizontal and vertical axes is not clear and may be context dependent.

4. Any computational study should have some "convergence" tests. i.e. use a range of approximations and compare the results to see how robust any conclusions are.

No comments:

Post a Comment