DFT calculations provide a means to calculate material properties rather quickly, and quite reliably. As seen above, they are also capable to describe hydrogen bonded ice. However, there is no systematic scheme to improve DFT results towards experimental accuracy. Wave function based methods offer this systematic improvement, be it configuration interaction (see section 2.3.2) or coupled cluster theory (see section 2.3.3). For extended systems, pe- riodic Hartree-Fock calculations are well established but neglect electron cor- relation which often proves crucial when calculating material properties. It is highly desirable to go beyond the periodic HF scheme and include electron correlation for extended systems. However, the treatment of electron corre- lation from first principlesmethods constitutes one of the most fundamental problems in solid state physics. It is currently not known how to treat sys-
tems with small band gaps or even metals using accurate electron correlation methods such as the above mentioned CI or CC theories [112, 113].
A major improvement in this direction consists of the introduction of the incremental correlation method by Stoll and co-workers [114, 115], the suc- cessful application of which was reported even for metallic systems like mer- cury [116]. Note, however, that while this method is a way to systematically obtain correlation energy corrections, it is not a true two-particle theory that provides correlated wave functions in a periodic system. Instead, it combines an effective single-particle theory (Hartree-Fock) used in a periodic system with local correlation methods.
It was shown that the electron correlation energy of small water clusters converges rapidly with the order of the many-body decomposition in the in- teraction energy, much faster than the total interaction energy [117]. While this illustrates the local nature of electron correlation, it also raises the ques- tion whether solid water could be described by combining periodic HF cal- culations with localized correlation energy calculations truncated at two- or three-body level (see below). If the former is sufficient, for instance liquid water could be simulated from ab initio wave function based methods by combining HF calculations and a parametrized dimer correlation potential, in the spirit of a recently presentedab initiowater pair potential [75].
In the incremental approach, the total interaction energy of ice is separated into the HF energyEHFand the electron correlation energyEc:
E= EHF+Ec. (3.13)
EHFis obtained from periodic Hartree-Fock calculations, whereasEc(normal-
ized to one unit cell) is expanded in a many-body decomposition [118]:
Ec =
∑
n Ec(n) =∑
i E(c1)(i) +∑
i,j E(c2)(ij) +∑
i,jk Ec(3)(ijk) +. . . (3.14)The indexiin each sum runs over all correlation units in the crystalline unit cell, whereas the higher indices j,k, . . . run over the whole crystal. There is
an infinite number of many-body sums. The individual terms are defined as
Ec(1)(i) =ϵc(i) (3.15)
E(c2)(ij) =ϵc(ij)−ϵc(i)−ϵc(j) (3.16)
Ec(3)(ijk) =ϵc(ijk)−ϵc(ij)−ϵc(jk)−ϵc(ki) (3.17)
+ϵc(i) +ϵc(j) +ϵc(k)
where ϵc(x) denotes the total correlation energy of correlation unit x. The
incremental scheme proves a valuable computational tool if the series (3.14) converges sufficiently fast (i) in terms of number of sums that have to be considered, and (ii) in terms of number of contributions within each sum that have to be included. The first condition asks for the physical or chemical nature of the interaction in the system studied: how good is it described with a pure pair interaction, or how many higher-order interactions are important. The second condition asks for the range, or the screening, of the interaction: up to which distance are for instance pair interactions important. If the series (3.14) converges sufficiently fast, it can be truncated after the two- or three- body terms E(c2)(ij) or E
(3)
c (ijk). Another choice has to be made regarding
which subsets of the atomic basis form the primitive correlated units. In case of ice, the natural choice is to treat every water molecule in the unit cell as smallest independent correlated unit.
The FORTRAN90 programcorrpbcwas written to perform the incremen- tal scheme calculations. It interfaces with the CRYSTAL06 [119] program package for the periodic boundary condition Hartree-Fock calculations, and with theMOLPROprogram package (2006 release version) for the local corre- lation energy calculations. Input data forcorrpbcincludes a completeCRYS- TAL06 input file (including unit cell and atomic basis parameters, basis sets for the periodic HF calculations, k-point sampling, and all other computa- tional parameters needed); basis sets for the correlation calculations; an en- ergy threshold for the geometry optimization. Supported correlation meth- ods are MBPT2 (see section 2.3.1) and CCSD(T) (see section 2.3.3). corrpbc
can perform single point calculations and geometry optimizations; for the latter, it can optimize internal coordinates and/or unit cell parameters (ei- ther globally or at fixed unit cell volume). Gradients for all internal coordi-
nates are calculated analytically – for the HF energyEHFfrom the periodic HF
calculations, and for the correlation energyEc from the localized correlation
calculations, by∇Ec = ∇Eloctot − ∇ElocHF. Gradients for the unit cell parame-
ters are calculated numerically. Update of atomic coordinates and unit cell parameters (possibly constrained by the constant volume condition) is done by using either a conjugate gradient or a quasi-Newton optimization algo- rithm [120, 121]. Structural optimizations run until the change in the total energyE= EHF+Ec is smaller than some energy criterionEmin. Implement-
ing a gradient based criterion should also be possible. Figure 3.6 shows a flowchart diagram ofcorrpbc’s operation mode.
The program can be used for atomic and molecular crystals, with arbitrary choice of primitive correlation units. It can handle many-body expansions up to the trimer termEc(3)(ijk). The monomer termE
(1)
c is of interest in molecu-
lar crystals; although it is not expected to influence binding energies, it could contribute to changes of the equilibrium geometry. Note that no “embedding procedure” of any kind has been implemented; all localized units are calcu- lated in the gas phase. Surrounding them by appropriate basis set centres, polarizable fields, or simply a dielectric background would make the gen- eration of the local units more demanding, but take into account a screening effect of the interaction that may lead to faster convergence of the many-body expansion.
The selection of monomer, dimer, and trimer units is not automatised. Thus, all monomers, dimers, and trimers have to be handed tocorrpbcex- plicitly. It should, however, be possible to extend the program to automati- cally generate all sub-units based e.g. on a distance cutoff criterion. Another optimization route would be parallelizing the correlation calculations, which can be performed independently.