Second-order Møller-Plesset theory^{665} (MP2) is probably the simplest useful wave function-based electron correlation method. Revived in the mid-1970s, it remains highly popular today, because it offers systematic improvement in optimized geometries and other molecular properties relative to Hartree-Fock (HF) theory.^{363} Indeed, in a recent comparative study of small closed-shell molecules,^{364} MP2 outperformed much more expensive singles and doubles coupled-cluster theory for such properties! Relative to state-of-the-art Kohn-Sham density functional theory (DFT) methods, which are the most economical methods to account for electron correlation effects, MP2 has the advantage of properly incorporating long-range dispersion forces. The principal weaknesses of MP2 theory are for open shell systems, and other cases where the HF determinant is a poor starting point.
Q-Chem contains an efficient conventional semi-direct method to evaluate the MP2 energy and gradient.^{361} These methods require $OVN$ memory ($O$, $V$, $N$ are the numbers of occupied, virtual and total orbitals, respectively), and disk space which is bounded from above by $OV{N}^{2}/2$. The latter can be reduced to $IV{N}^{2}/2$ by treating the occupied orbitals in batches of size $I$, and re-evaluating the two-electron integrals $O/I$ times. This approach is tractable on modern workstations for energy and gradient calculations of at least 500 basis functions or so, or molecules of between 15 and 30 first row atoms, depending on the basis set size. The computational cost increases between the 3rd and 5th power of the size of the molecule, depending on which part of the calculation is time-dominant.
The algorithm and implementation in Q-Chem is improved over earlier methods,^{358, 268} particularly in the following areas:
Uses pure functions, as opposed to Cartesians, for all fifth-order steps. This leads to large computational savings for basis sets containing pure functions.
Customized loop unrolling for improved efficiency.
The sort-less semi-direct method avoids a read and write operation resulting in a large I/O savings.
Reduction in disk and memory usage.
No extra integral evaluation for gradient calculations.
Full exploitation of frozen core approximation.
The implementation offers the user the following alternatives:
Direct algorithm (energies only).
Disk-based sort-less semi-direct algorithm (energies and gradients).
Local occupied orbital method (energies only).
The semi-direct algorithm is the only choice for gradient calculations. It is also normally the most efficient choice for energy calculations. There are two classes of exceptions:
If the amount of disk space available is not significantly larger than the amount of memory available, then the direct algorithm is preferred.
If the calculation involves a very large basis set, then the local orbital method may be faster, because it performs the transformation in a different order. It does not have the large memory requirement (no $OVN$ array needed), and always evaluates the integrals four times.
There are three important options that should be wisely chosen by the user in order to exploit the full efficiency of Q-Chem’s direct and semi-direct MP2 methods (as discussed above, the LOCAL_OCCUPIED method has different requirements).
MEM_STATIC: The value specified for this $rem variable must be sufficient to permit efficient integral evaluation (10-80MB) and to hold a large temporary array whose size is $OVN$, the product of the number of occupied, virtual and total numbers of orbitals.
N_FROZEN_CORE: The computational requirements for MP2 are proportional to the number of occupied orbitals for some steps, and the square of that number for other steps. Therefore the CPU time can be significantly reduced if your job employs the frozen core approximation. Additionally the memory and disk requirements are reduced when the frozen core approximation is employed.