1

European Summer School “Ab initio modelling in solid-state chemistry”

Torino, September 2000

An introductory guide to Gaussian basis sets in solid-state

electronic structure calculations†

Mike Towler
Theory of Condensed Matter Group
Cavendish Laboratory, University of Cambridge, U.K.

The purpose of these notes is to provide some insight into Gaussian basis set
technology as implemented in the CRYSTAL Hartree-Fock/density functional theory
program for periodic systems1. Essential differences between basis sets appropriate
for use in solids and those used in purely molecular codes are explained. Examples of
how to choose appropriate basis sets for particular problems, hints on basis set
development, and some simple exercises are also included.

1. Basis set expansions

In order to calculate molecular or crystalline structures and properties, it is
necessary to determine the eigenfunctions ^ and eigenvalues E of the Schrödinger
equation

H^ = E^ . (1.1)

where H is the electronic Hamiltonian. For systems of interest in chemistry, analytic

solutions to this problem are impossible to find, and so one normally resorts to the
variational approach involving the introduction of a trial wave function ^ (_ ) that
depends on a set of variable parameters { _ }. If the functional

^ H! ^
E (_ ) = (1.2)
^^

is minimized with respect to variations in {_ }, the energy converges from above on

the true energy in (1.1), and the wave function converges in the mean on the true wave
function as the parameter set {_ } is expanded to completeness.

†
Revised and updated Sep. 2000 from earlier version by MT, University of Torino, Sep. 1995
 2

The most obvious way of implementing this approach is to make the trial wave
function depend linearly on the parameters {_ }. The resulting linear expansion may
be written most generally as

^= -\ µ c µ . (1.3)
µ

A set of equations for the linear coefficients c µ in this expansion may be derived on
substitution of (1.3) into (1.2) by making the energy stationary with respect to
variations in the coefficients. The N-particle basis functions {\} are a set of fixed
analytic functions that depend on the co-ordinates of all electrons in the system. They
can conveniently be taken to be orthonormal, in which case the variational equations
correspond to the eigenvalue problem

Hc = Ec . (1.4)

If the N-particle basis were a complete set of N-electron functions, the use of the
variational approach would introduce no error, because the true wave function could
be expanded exactly in such a basis. However, the basis would then be of infinite
dimension, and in practice, the fact that we must work with an incomplete set of N-
particle functions is one of our major practical approximations. Furthermore, the
question arises of how to construct the N-particle basis itself. There are many possible
physically motivated ways of doing this, but in most practical quantum chemistry
techniques it is constructed using linear combinations of products of one-electron
wave functions or orbitals. These are usually antisymmetrized to account for the
permutational symmetry of the wave function (and may also be spin and symmetry
adapted):

N
\ µ = A! . q µi ( xi ) (1.5)
i

Here A is an antisymmmetrization operator and the xi are the space and spin co-
ordinates of a single electron. In this form the N-particle basis functions are called
Slater determinants. The unknown one-electron functions {q } in (1.5) are referred to
as atomic, molecular or crystalline orbitals depending on the physical nature of the
problem. To find the unknown orbitals, one generally expands them as an orthonormal
linear combination of known one-electron basis functions r a :

q µi = - r a Ca , µi (1.6)
a

One then obtains a set of algebraic equations for the optimum orbitals which may be
solved by standard matrix techniques. The set of functions {r } in (1.6) constitute the
one-particle basis set given as input to most quantum chemistry calculations. This is
what these notes are about.
So evidently the simplest truncation of the N-particle space is that in which
only one N-electron basis function is used - a single configuration which is the best
variational approximation to the exact ground-state wave function. In this case all
 3

coefficients in the expansion (1.3) are zero with the exception of that of the ground-
state configuration. Minimizing the expectation value of the Hamiltonian with respect
to the one-electron orbitals in a single determinant trial function allows us to derive
the self-consistent Hartree-Fock (HF) equations, which may be solved to find the
optimum orbitals (i.e. the best coefficients of equation (1.6)). To develop ‘correlated
wave function methods’ that give a better description of electron-electron correlation
than HF one usually uses more complex techniques involving combinations of
explicitly many-particle functions such as, for example, configuration interaction
(CI)2, MC-SCF3 or coupled-cluster4 methods (although stochastic techniques such as
quantum Monte Carlo are becoming an increasingly poplar alternative – see later).
The HF special case is sufficiently important that the difference E exact < E HF is
defined as the correlation energy. Here E HF is the Hartree-Fock energy, that is, the
SCF result in a complete one-particle basis set. The Hartree-Fock limit is thus defined
as the complete one-particle space single-configuration result.
Of course, an excellent and generally cheaper alternative to wave function
methods is Kohn-Sham density functional theory (KS-DFT). In this scheme the single
determinant of orbitals is not meant to represent the true many-electron wave function
as in HF theory. The basic idea of the Kohn-Sham scheme is to replace the calculation
of the true wave function ^ by that of a single Slater determinant that represents a
non-interacting model system and yields the same ground state density as ^. However
the expectation value of the Hamiltonian with this determinant only gives part of the
total energy. The remaining (exchange-correlation) contribution to the total energy is
not directly accessible by the determinant and is given by an (approximate) functional
of the total density. Pretty much all statements here concerning the one-particle basis
set are applicable both to HF and Kohn-Sham orbitals, but where there are distinctions
these will be noted.
The nature of the one-particle basis functions used in in the expansion of the
orbitals depends on the periodicity of the system. In the molecular case, functions
localized at the nuclear centres are generally used, consisting of products of a radial
function R(r) (such as a Gaussian) and an angular function (such as a spherical
harmonic Ylm (e , q ) ). In the periodic case by contrast, the one-particle basis must be
made up of Bloch functions r~ik (r ) i.e. products of a function periodic in the primitive
lattice and a phase factor whose frequency and direction of oscillation is dependent on
the wave vector k. These Bloch functions might be, for example. simple plane-waves
exp(i (k + G ).r ), where G is a vector in the reciprocal lattice, or as in CRYSTAL a
combination of a localized function r a and all of its periodic images with the whole
modulated by a phase factor:

1
r~ik (r ) = - r (r < r
t
a a < t )exp(ik u t ) (1.7)
N t

Here r at refers to the ath localized atomic function (lying at position ra in the zero
cell) in the unit cell of the crystal described by the lattice translation vector t. Note that
because the wave vector k is a continuous variable, the basis set of Bloch functions is
in principle infinite; in practice however, the problem is solved at a finite set of k
points, and the results interpolated.
Thus the one-particle basis of atomic functions/Bloch functions determines the
one-particle orbitals, which in turn determine the N-particle basis. If the one-particle
 4

basis were complete, it would in principle be possible to form a complete N-particle

basis, and hence to obtain an exact wave function variationally. Again however, such
a complete one-particle basis would be of infinite dimension, and thus the basis must
be truncated in practical applications. We must therefore use truncated N-particle
spaces that are constructed from truncated one-particle bases. These two truncations
(i.e. correlation and basis set error) are the most important sources of uncertainty in
quantum chemical calculations. It is interesting to note at this point that if we remove
the restriction of analytic integrability of basis functions in correlated wave function
methods, it is possible to construct much more efficient N-particle functions by
including terms which depend explicitly on the interparticle distances. A good
example is the Slater-Jastrow form used in quantum Monte Carlo calculations where
a single N-electron function is usually an excellent approximation to the true wave
function and gives a very accurate description of the electron correlation. Such
calculations will be the subject of my second lecture at this school but will not be
mentioned further here.
It should be noted that the ultimate accuracy of any calculation, in correlated
calculations over many N-particle basis functions as well as at the SCF level, is
determined by the one-particle basis set. This is one of the most obvious observations
about quantum chemical calculations, but worth emphasizing nonetheless. It is in
general just not possible to get the right answer for the right reason using, for
example, an STO-3G Gaussian basis set (see later). This is not necessarily an
argument against using such sets (people make bigger approximations every day) but
their limitations must be kept constantly in mind. Probably one of the most important
lessons of this school is that the choice of one-particle space is the most important
decision in setting up any calculation, since ultimately this choice determines the
reliability of the result. Nothing can overcome limitations in the one-particle basis†.
In the next section we shall examine particular properties of the one-electron
basis functions most commonly used in modern molecular quantum chemistry codes,
namely contracted Gaussian-type functions, and how these are modified in a periodic
code such as CRYSTAL.

2. One-electron basis sets constructed from Gaussian-type functions

General considerations
Having worked for the last four years in the spiritual home of plane-wave
pseudopotential calculations (the Theory of Condensed Matter Group in the
Cavendish Laboratory at the University of Cambridge), it has been rather obvious to
me that the traditional basis for construction of the one-electron Bloch functions in
solid-state electronic structure calculations has not been Gaussians. Indeed, you may
as well learn now that in certain circles the derision invited by doing so can be
tiresome. Chemists are no problem. But if you plan to interact with physicists and to
use Gaussian basis sets, it is as well to have rehearsed the arguments for and against
beforehand. Referring to what CRYSTAL does as an ‘all-electron full potential
LCGTF method’ rather than ‘Gaussian basis set calculations’ will help at first. Do not
use the phrase ‘linear combination of atomic orbitals’ under any cirumstances. When
pressed, follow the example of the director of this school at a conference some years
ago in reference to a question about why he used Hartree-Fock theory (which is
quoted to me every time I mention in conversation who I used to work for) and claim

†
Except, magically, diffusion Monte Carlo!
 5

that ‘it is some strange kind of sexual perversion’. Of course all different basis sets
have their advantages and disadvantages. The choice of one set over another is one of
personal taste (what codes do I have on my computer that I know how to use?) and a
careful consideration of the pros and cons of the various methods (do I want speed,
accuracy, forces etc.?). A short discussion along these lines will appear at the end of
this section.
So anyway in periodic systems as in molecules the fundamental idea behind
the use of localized Gaussian-type functions is the ‘atoms in molecules’ concept (i.e.
molecules are an assemblage of slightly perturbed atoms), and our consequent
expectation that localized atomic functions will prove suitable as an expansion set in
molecules is generally well-founded. The historical phrase linear combination of
atomic orbitals is often used to describe this procedure but this is both archaic and
inaccurate and should be avoided. Strictly speaking, ‘atomic orbitals’ are solutions of
the Hartree-Fock equations for the atom i.e. a wave function for a single atomic
electron in a self-consistent field. Localized basis functions are thus not atomic
orbitals, but can in spherical polar coordinates be any function of the form
R(r )O(ž )\ (q ) with properties chosen for computational convenience. The equations
determining the form of the radial functions R(r) can be solved only for one-electron
‘hydrogen-like’ atoms, but some general conclusions about the nature of the solutions
can still be drawn. In particular, due to the singularity of the potential at a point
nucleus with a charge of +Z, the wave function must have a ‘cusp’ at the nucleus. In
fact, for such a one-electron atom, it is required that

dR
= <Z (2.2)
dr r=0

At the other end of the range, an electron far away from any molecule would see the
remainder of the molecule as a positive charge without any particular structure. Like
any one-electron atom, the wave function would therefore decay exponentially. It
would thus seem reasonable to use exponential functions as basis functions, especially
since they are known to be exact solutions for any one-electron system. Historically
therefore, basis functions with exponential asymptotic behaviour - Slater-type orbitals
(STOs) - were the first to be used5. These are characterized by an exponential factor in
the radial part:

r STO = r n <1 exp(<cr )Ylm (e , q ) (2.3)

where c (‘zeta’) is called the exponent, the Ylm (e , q ) is the spherical harmonic or
angular momentum part (the function describing the ‘shape’ of the STO), and the
n,l,m are quantum numbers.
Unfortunately such functions are not suitable for fast calculations of multi-
centre integrals, so Gaussian-type functions (GTFs) were introduced to remedy the
difficulties Transforming from a local polar coordinate system to a Cartesian one,
these can be written

r GTF = exp(<_r 2 ) x l y m z n (2.4)

 6

where _ is again the exponent, and the l,m,n are not quantum numbers but simply
integral exponents of Cartesian coordinates. In this form (now called Gaussian
primitives) they be factorized into their Cartesian components i.e.

r GTF = r xGTF r GTF

y r zGTF (2.5)

where each Cartesian component has the form (introducing an origin such that the
Gaussian is located at position xa),

(
r xGTF = (x < x a )l exp < _ (x < x a )
2
). (2.6)

This simplifies considerably the calculation of integrals. (It should be clear that if
write the exponential part of an STO, exp(< _ r ), in Cartesian components we get
( )
exp < _ x 2 + y 2 + z 2 which is not so separable). Note that the absence of the STO
pre-exponential factor r n <1 restricts single Gaussian primitives to approximating only
1s, 2p, 3d etc. orbitals and not e.g. 2s, 3p, 4d etc... However, combinations of
Gaussians are able to approximate correct nodal properties of atomic orbitals if the
primitives are included with different signs. The sum of exponents of Cartesian
coordinates L = l + m + n is used analagously to the angular momentum quantum
number for atoms to mark Gaussian primitives as s-type (L=0), p-type (L=1), d-type
(L=2), f-type (L=3) etc.†
The present success of GTFs as the basis set of choice in virtually all
molecular quantum chemistry calculations was far from obvious originally. In
particular, it is clear that the behaviour of a Gaussian is qualitatively wrong both at the
nuclei and in the long-distance limit for a Hamiltonian with point-charge nuclei and
Coulomb interaction. It has therefore been a commonly held belief that STOs would
be the preferred basis if only the integral evaluation problem could be solved. It has
been claimed6 that this is not necessarily the case and that the ‘cusp’ behaviour
represents an idealized point nucleus, and for more realistic nuclei of finite extension
the Gaussian shape may actually be more realistic. If accurate solutions for a point-
charge model Hamiltonian are desired, they can be obtained to any desired accuracy in
practice by expanding the ‘core’ basis functions in a sufficiently large number of
Gaussians to ensure their correct behaviour. Furthermore, properties related to the
behaviour of the wave function at or near nuclei can often be predicted correctly, even
without an accurately ‘cusped’ wave function7. In most molecular applications the
asymptotic behaviour of the density far from the nuclei is considered much more
important than the nuclear cusp. As mentioned above, the wave function for a bound
state must fall off exponentially with distance, whenever the Hamiltonian contains
Coulomb electrostatic interaction between particles. However, even though an STO
basis would in principle be capable of providing such a correct exponential decay, this
occurs in practice only when the smallest exponent in the basis set is c min = 2I min ,
where Imin is the first ionization potential. Such a restriction on the range of exponent
values, while acceptable for atomic SCF calculations, is far too restrictive for
†
Note CRYSTAL specifics: CRYSTAL98 allows Gaussian primitives of s,p or d types. In principle,
the following six angular functions are possible for Cartesian Gaussians: 1s (1), 2p (x,y,z), 3d (x2, xy, xz,
yz, y2, z2). Note that there are only five linearly independent and orthogonal atomic d orbitals so
internally CRYSTAL uses appropriate linear combinations of Cartesian Gaussians to give the five real
spherical harmonic d basis functions: 3z2-r2 , xz, yz, x2-y2, xy (stored internally in that order).
 7

molecular and solid-state work. Some of these formal limitations have thus turned out
to be of relatively little importance in practice. Indeed the universality of Gaussian
functions in molecular quantum chemistry is virtually complete - Slater-type functions
are hardly used today and we shall not discuss them here any further. Much more
important are limitations arising from the convergence of results with the size of the
basis set.

Contraction schemes
Both the number of integrals over basis functions to be stored on disk and the
total CPU time nominally scale rather unpleasantly with the number of functions in
the basis set. Thus it usually pays to consider the issue of basis set compactness, that
is, the ability to expand the orbitals as accurately as possible using the minimum
number of basis functions. Furthermore, although we wish to perform this expansion
in terms of localized ‘atomic-orbital’-like functions, we have seen that our basis
functions of choice (Gaussians) do not in themselves resemble exact HF atomic
orbitals particularly closely. In most applications therefore, Gaussian-type basis
functions are expanded as a linear combination (or ‘contraction’) of individually
normalized Gaussian primitives gj(r) characterized by the same centre and angular
quantum numbers, but with different exponents,

L
r i (r ) = - d j g j (r )
j =1
where (2.7)
g j (r ) > g (r; _ , l , m) = N lm (_ )r l Ylm (e , q ) exp(<_ j r 2 )

where L is the length of the contraction, the _j are the contraction exponents, the dj
contraction coefficients and I have now written the Gaussian primitives in terms of
real spherical harmonics including a normalization constant. By proper choice of these
quantities, the ‘contracted Gaussians’ may be made to assume any functional form
consistent with the primitive functions used. One may therefore choose the exponents
of the primitives and the contraction coefficients so as to lead to basis functions with
desired properties, such as reasonable cusp-like behaviour at the nucleus (e.g.
approximate Slater functions or HF atomic orbitals). Integrals involving such basis
functions reduce to sums of integrals involving the Gaussian primitives. Even though
many primitive integrals may need to be calculated for each basis function integral,
the basis function integrals will be rapidly calculated provided the method of
calculating primitive integrals is fast, and the number of orbital coefficients in the
wavefunction will have been considerably reduced.
The exponents and contraction coefficients are normally chosen on the basis of
relatively cheap atomic SCF calculations so as to give basis functions suitable for
describing exact Hartree-Fock atomic orbitals. An approximate atomic basis function,
whose shape is suitable for physical and chemical reasons, is thus expanded in a set of
primitive Gaussians, whose mathematical properties are attractive from a
computational point of view. Note that the physical motivation for this procedure is
that, while many primitive Gaussian functions may be required to provide an
acceptable representation of an atomic orbital, the relative weights of many of these
primitives are almost unchanged when the atoms are formed into molecules or
crystals. The relative weights of the primitives can therefore be fixed from a previous
 8

calculation and only the overall scale factor for this contracted Gaussian function need
be determined in the extended calculation. It is clear that contraction will in general
significantly reduce the number of basis functions. For example a so-called STO-3G
basis, where three Gaussian primitives are used to form a contracted function which
resembles a Slater-type orbital, the reduction in size from the primitive basis is a
factor of 3, corresponding to a nominal reduction factor of 81 (N4) on the number of
two-electron integrals - clearly a significant reduction.
How does this scheme transfer to the solid state? A finite number p of GTFs,
constructed from contractions of Gaussian primitives, are attributed to each of the
non-equivalent atoms in the reference zero cell. The same GTFs are then formally
associated with all N translationally equivalent atoms in the crystal by direct
translations of the lattice vectors t. This gives a total of Np GTFs from which Np
Gaussian-type Bloch functions (GTBF) are then constructed according to

r~ak (r ) = - r a (r < ra < t ) exp(ik u t ) . (2.8)

t

where the ra are the co-ordinates of the basis atom in the reference zero cell with
which ra is associated. In fact, for solid-state calculations, there are no practical
differences in the form of the basis set input compared to the molecular case, as the
transformation of the one-electron basis functions to their Bloch form is done
internally after the definition of the localized atomic functions. However, the
exponents and contraction coefficients in the two cases will generally be rather
different, and with some exceptions such as molecular crystals and certain covalent
systems, molecular basis sets are not directly transferable to the study of crystalline
solids. We shall return to this point in the next section.
Two further points may be made with respect to basis set contraction schemes.
The first concerns the way AOs belonging to a given atom are grouped into shells. In
general, a shell contains all functions characterized by the same n and l quantum
numbers (e.g. all the different d functions in a 3d shell); this allows the partitioning of
the total charge density into ‘shell charge distributions’ and is useful in the selection
of bielectronic integrals and in the evaluation of long-range interactions. A feature of
the contraction schemes originally used in basis sets of the Pople type (and often
useful in calculations with CRYSTAL) is the additional grouping of AOs with only
the same principle quantum number into shells; e.g. a 2sp shell, in which both 2s and
2p functions have the same set of exponents _j but different contraction coefficients
dj. This procedure reduces the number of auxiliary functions to be calculated in the
evaluation of electron integrals. In fact basis sets with sp shells can give a saving
factor as large as four in the CPU time, compared with the case where s and p have
different exponents. Note that CRYSTAL is restricted to s, p and d basis functions and
that only sp shells may be formed in this way. One should note however, that while
the use of sp shells may reduce the computational effort somewhat, in certain
circumstances it may actually represent an important constraint on the form of the
basis functions. For relatively small calculations where the time and storage
limitations are not an important factor, some consideration should be given to
describing the s and p functions with separate sets of exponents.
The second point is not directly relevant to the CRYSTAL program but may be
encountered in the literature and as such one should be aware of it, namely the so-
called general contraction scheme8. Most standard codes such as CRYSTAL use what
is known as a segmented contraction, in which the transformation from the larger
 9

primitive set to the smaller contracted set is restricted in such a way that each
Gaussian primitive gj contributes to exactly one contracted GTF. The algorithms
involved are relatively simple if the transformation is reduced to a series of small,
independent summations within mutually exclusive sets. In contrast, the general
contraction scheme makes no such assumptions, and allows each Gaussian primitive
to contribute to several contracted GTFs. A considerable advantage of the general
scheme is that the contracted GTFs reproduce exactly the desired combinations of
primitive functions. For example, if an atomic SCF calculation is used to define the
contraction coefficients in a general contraction, the resulting minimal basis will
reproduce the SCF energy obtained in the primitive basis. This is not the case with
segmented contractions. There are other advantages with a general contraction: for
example, it is possible to contract inner-shell orbitals to single functions with no error
in the atomic energy, making calculations on heavy elements much easier. Another
advantage is a conceptual one: using a general contraction, it is possible to perform
calculations in which the one-particle space is a set of atomic orbitals, a true LCAO
scheme, rather than being a segmented grouping of a somewhat arbitrary expansion
basis. The MOs can then be analysed very simply, just as for the original qualitative
LCAO MO approach, but in terms of ‘exact AOs’ rather than relatively crude
approximations to them.

Computational aspects
One of the most attractive features of Gaussian basis functions is their
separability into Cartesian components, as in (2.5). This allows a computationally
efficient transition from the spherical symmetry of the atom, naturally represented in a
polar co-ordinate system, to a more general Cartesian representation which is useful
for describing molecular/crystalline geometries. Another equally important reason for
the usefulness of a Gaussian basis set is embodied in the Gaussian product theorem
(GPT), which in its simplest form states that the product of two simple Gaussian
functions with exponents _ and `, located at centres A and B, is itself a simple
Gaussian with exponent a, multiplied by a constant factor F, located at a point C along
the line segment A–B, where

a =_ + ` (2.9a)

£ _A + `B ¥
C=² ´ (2.9b)
¤ a ¦

£ _` ¥
F = exp² < (A < B) 2 ´ (2.9c)
¤ a ¦

The product of two polynomial GTF, of degree µ and i and located at points A and B
is therefore another polynomial GTF located at C of degree µ+i in xC, yC and zC, which
can be expressed as a short expansion of one-centre Gaussians:

µ +i
r ax ( x ) r bx ( x ) = - Ciµ +i   ci ( x< x )
c
(2.10)
i =0
 10

2 _xa + `xb
where   ci ( x ) = x i e <_p ( x< xc ) and x c = .
_ +`

The product of two Gaussians which are functions of the co-ordinates of the
same electron is referred to as an overlap distribution, and all the integrals which must
be calculated involve at least one such overlap distribution. The most important
consequence of the GPT is that all four-centre two-electron integrals can be expressed
in terms of two-centre quantities. Some properties of the GPT and overlap
distributions are explored in Exercise 1.
CRYSTAL actually uses a common and more efficient approach for the
evaluation of integrals over Gaussian basis functions9, in which Hermite Gaussian
functions (HGFs) are used instead of the usual Cartesian Gaussians in the re-
expansion (2.10). Hermite Gaussians are defined as derivatives of an s type Gaussian:

£ 2¥ di
¤ (
R i (j ) = Hi (j ) exp² <_ p x < x p ) i
´ = (<1)
¦ dj i
exp(<j 2 ) (2.11)

where Hi (j ) is a polynomial of order i, and

1
j = _ p2 ( x < x p ) (2.12)

The set of HGFs spans the same space as the expansion functions in (2.10) and as a
consequence they can be used for expanding the basis function products:

la + lb
r ax ( x ) r bx ( x ) = - Cila +lb R i (j ) (2.13)
i =0

where the expansion coefficients must now be redefined. Because of the natural
relations between Hermite polynomials and Gaussians, the necessary two-centre
integrals can be evaluated with very high efficiency (see Ref. 10 for a useful
discussion). Basis functions with higher quantum numbers can be generated through
repeated differentiation of an s-type Gaussian.
Even though the four-centre bielectronic integrals can be written in terms of
two-centre quantities, the cost of evaluating them still scales nominally as N4, where N
is the number of functions in the expansion. This scaling is far from satisfactory and
this must be reduced in order to treat large systems. One way of doing this which is
used in CRYSTAL is the method of pre-screening where, rather than attempting to
calculate the integrals more efficiently, one seeks where possible to avoid their
evaluation altogether. Since the expression for an integral over primitive Gaussians
can be formally written as

ab cd = S ab S cd Tabcd (2.14)

where Sab is a radial overlap between functions ra and rb , and Tabcd is a slowly varying
angular factor. In many situations the product SaSb thus constitutes a good estimate of
the magnitude of the integral, and it may seem attractive to use that product as an
estimate in screening out small integrals. In order to estimate these overlaps quickly, a
single, normalized s-type Gaussian (called an ‘adjoined Gaussian’ in the CRYSTAL
 11

literature) is associated with each shell, whose exponent _ is the smallest of the
exponents in the shell contraction. This function thus reproduces approximately the
absolute value of the corresponding AOs at intermediate and long range. The adjoined
Gaussian is used in fast algorithms for estimating overlaps on the basis of which
integrals are either evaluated exactly, approximately, or not at all. The level of
approximation is user-definable through a set of tolerances given in the input. Such
algorithms, and a consideration of the crystalline symmetry, mean that the CRYSTAL
integrals part currently scales at between N and N2, depending on the size of the
system. The most unpleasant scaling in CRYSTAL is thus the SCF part which, since it
involves diagonalization of the Fock matrix, scales as approximately N3.

Plane waves vs. Gaussians

Finally I shall briefly compare the advantages and disadvantage of the simplest
alternative to Gaussians in solid-state calculations i.e. plane waves. Plane waves are
an orthonormal complete set; any function belonging to the class of continuous
normalizable functions (which are those of interest in quantum mechanics) can be
expanded with arbitrary precision in such a basis set. The set is universal, in the sense
that it does not depend on the positions of the atoms in the unit cell, nor on their
nature. We thus do not have to invent a new basis set for every atom in the periodic
table nor modify them in different materials as is the case with Gaussian functions,
and the basis can be made better (and more expensive) or worse (and cheaper) by
varying a single parameter. This characteristic is particularly valuable in ab initio
molecular dynamics calculations, where nuclear positions are constantly changing.
The algorithms mainly (involving fast Fourier transforms) are easier to program since
the algebraic manipulation of plane-waves is very simple. It is relatively easy to
compute forces on atoms, with all the wealth of new physics and chemistry which that
implies. Finally, plane-wave calculations do not suffer from basis set superposition
error (see later). In practice, one must use a finite set of plane waves, and this restricts
the detail that can be revealed in real space to such an extent that core electrons cannot
be described in this manner. One must either augment the basis set with additional
functions (as in e.g. the (F)LAPW scheme), or use pseudopotentials to describe the
core states.
In comparison with plane waves, the use of all-electron Gaussian calculations
allows us to describe accurately electronic distributions both in the valence and the
core region with a limited number of basis functions. The local nature of the basis
allows a treatment both of finite systems and of systems with periodic boundary
conditions in one, two or three dimensions. This has advantages over plane wave
calculations of molecules, polymers or surfaces which work by imposing artificial
periodicity: the calculation must be done on e.g. a three-dimensional array of
molecules with a sufficiently large distance between them. Gaussian total energies can
be made very precise (i.e. reliable to many places of decimals) since all integrals can
be done analytically (in practice, this is only true for Hartree-Fock calculations;
density functional theory calculations with CRYSTAL require a numerical integration
of the exchange-correlation potential which reduces the attainable precision, although
this was improved with the release of CRYSTAL98). Having an ‘atomic-like’ basis
facilitates population analyses, the computation of properties such as projected
densities of states, and ‘pre-SCF alteration of orbital occupation’ (the CRYSTAL
 12

‘EIGSHIFT’ option)†. Many plane-wave programs cannot compute exact non-local

exchange which is required not only for Hartree-Fock calculations, but also in the
‘hybrid’ DFT exchange-correlation functionals and in overcoming self-interaction
problems in DFT calculations of Mott insulators11. It is now well-understood that
localized basis functions are essential for the various new linear scaling DFT
algorithms that have been the subject of a great deal of recent research12 and will in all
probability be incorporated in a future version of CRYSTAL. The cost that is paid for
using Gaussians is the loss of orthogonality, of universality, the need for more
sophisticated algorithms for the calculations of the integrals, the difficulty of
computing forces (although the next release of CRYSTAL scheduled for early 2001
will include forces for the first time), and an overly heavy reliance on the presence of
lots of space group symmetry operators for efficient calculations. As a final note, I
have observed that Gaussians are generally more efficient than plane-waves in highly
accurate quantum Monte Carlo calculations where symmetry is irrelevant. CRYSTAL
may be used to provide initial trial wave functions for the Cambridge QMC code,
CASINO13.

3. Terminology and notation connected with Gaussian basis sets

There is a vast amount of historical quantum chemistry jargon associated with
Gaussian basis sets, much of it redundant. Unfortunately it is necessary to understand
a certain proportion of it in order to read the literature, but please try to avoid using
too much of it yourself.
It has been usual to make a distinction between ‘core’ basis functions and
‘valence’ basis functions in Gaussian basis sets. Contractions consisting of primitives
with large exponents are associated with the core while more diffuse (small exponent)
functions are associated with the valence. This is actually rather arbitrary and is a
lingering echo from the past era of Slater orbitals. I stress again that basis functions
are not atomic orbitals, and in many cases, they do not even resemble orbitals of
isolated atoms. In fact, examining coefficients of true atomic/molecular/crystalline
orbitals expanded in such a basis usually reveals that these ‘core’ basis functions
contribute substantially to the highest occupied valence orbitals. This is a consequence
of the fact that basis functions on a given center are usually not orthogonal; in addition
they are often not really all that compact and overlap to some extent with ‘core’
functions on neighbouring centers – a situation not likely to occur with true atomic
core orbitals.
The early Gaussian contractions were obtained by a least squares fit to Slater
orbitals. The number of contractions (not primitives) used for representing a single
Slater orbital (i.e. zeta) was a measure of the goodness of the set. So, a single zeta (or
minimal) basis set is one that has a single basis function corresponding to each of the
atomic orbitals that are occupied in the atom. It is the smallest set one can reasonably
use in any calculation, and one should not expect any quantitative accuracy with such
a basis. The double-zeta basis set consists of two basis functions per atomic orbital,
and is thus twice as large as the minimal. In the same way, basis sets of triple-zeta,
quadruple-zeta etc. quality can be built. One often encounters the term split-valence
basis which basically means a set in which more contractions are used to describe
valence orbitals than core orbitals. The letter V denotes split valence sets e.g. DZV

†
This technique is important, for example, in driving the calculation into particular states in cases like
the various transition metal materials where d orbital degeneracies are not broken by the crystal field
but by high order effects such as spin-orbit coupling e.g. CoO.
 13

represents a basis set with only one contraction for core orbitals and two contractions
for valence orbitals. The fact that more basis functions are assigned to valence orbitals
does not mean the valence orbitals incorporate more primitives. Usually the core
orbitals are long contractions consisting of many primitives to represent well the cusp
of the s-type function at the nucleus.
One very important and still useful concept is that of polarization functions,
which are nominally functions of higher angular quantum number than the highest
occupied orbital in the system. As an example of why these may be needed, consider
an isolated hydrogen atom, the exact wave function of which is just the 1s orbital. If
the hydrogen atom is placed in a uniform electric field then the charge distribution
about the nucleus becomes asymmetric - it is polarized. The lowest order solution to
this problem is a mixture of the original 1s orbital and a p-type function i.e. the
solution can be considered to be a hybridized orbital. A hydrogen atom in a molecule
experiences a similar, but non-uniform electric field arising from its non-spherical
environment. By adding polarization functions i.e. p-type functions to a basis set for H
we directly accommodate this effect. In a similar way, d-type functions which are not
occupied in first-row atoms, play the role of polarization functions for the atoms Li to
F. Note that the exponents of polarization functions cannot be optimized in atomic
SCF calculations and must be reoptimized specifically for the molecule or solid. The
‘zeta’ terminology is often augmented with a description of the polarization functions.
Thus, DZP means double-zeta plus polarization, TZP for triple-zeta plus polarization
etc. Sometimes the number of polarization functions is given e.g. TZDP, TZ2P,
TZ+2P all stand for triple-zeta plus polarization. The creativity here is evidently
extensive.
In molecular work, systems are commonly encountered for which the charge
distribution is expected to be considerably more diffuse than in the neutral atom. This
is especially true for negatively charged species, or polar systems where a part of the
molecule can be expected to carry an excessive negative charge. In this case, it is often
advantageous to augment the basis set with diffuse functions, i.e. functions that have
smaller orbital exponents than those normally used. Diffuse functions are also helpful
in calculations when an accurate account of the outer region of the charge density
cloud is essential, such as in the calculation of higher-order moments or
polarizabilities. It is essential to realize however, that functions involving diffuse
primitives are of very little use in the solid state and may even be dangerous, for at
least the following three reasons: first the number of integrals to be explicitly
calculated increases very quickly as you decrease the exponent; secondly, the accuracy
of the calculation must be particularly high in order to avoid pseudo-linear
dependence catastrophes (especially when computing exact Fock exchange in HF or
hybrid DFT calculations); thirdly diffuse functions are not of much use in densely
packed crystals, because their tails are found in regions where there is large variational
freedom associated with functions on other atoms.
We have already seen some of the bewildering array of acronyms for the many
different kinds of basis sets available. These are essentially just cryptic shorthand for
the way the contractions from Gaussian primitives were performed with possibly
some description of how the set was modified afterwards. The way in which
contractions are derived is not easy to summarize in general, and moreover, it depends
upon the intended use for the basis functions. It is a good idea to always read the
original paper which describes the contraction procedure. Some basis sets are good for
geometry and energies, some are aimed at properties (such as polarizability), some are
optimized only with Hartree-Fock in mind, and some are tailored for correlated
 14

calculations. Finally, some are good for anions and others for cations and neutral
molecules. For some calculations, a good representation of the inner (core) orbitals is
necessary (e.g. for properties required to analyze NMR spectra or hyperfine
interactions (e.g. ISOTROPIC/ANISOTROPIC keywords in CRYSTAL) or for all-
electron Gaussian quantum Monte Carlo calculations), while others require the best
possible representation of the valence orbitals.
The most widely-known notation other than the ‘zeta’ system consists of
acronyms like 6-31G. This denotes a basis set where six Gaussian primitives have
been used to expand each of the ‘core atomic orbitals’, whereas the ‘valence orbitals’
are described by two functions - the inner one expanded in three Gaussians, the outer
one uncontracted. It is usual to leave the most diffuse basis functions uncontracted -
the outer part of the valence is so strongly distorted from the atomic picture that
flexibility is more important than atomic resemblance. To indicate the presence of
(any number of) polarization functions an asterisk is added to the basis set symbol. In
practice, hydrogen atoms are often treated differently from other atoms in a molecule
with regard to the choice of basis set, and polarization functions are not always added
to hydrogen atoms. Thus for a set with polarization on all atoms we add two asterisks,
6-31G**. Diffuse functions are treated similarly; a ‘+’ denotes the presence of diffuse
functions, ‘++’ denotes that such functions are used on all atoms. A symbol such as
e.g. 6-311G**+ would thus be interpreted as follows:
(1) Each atom core orbital is represented by one basis function, expanded in
six primitive Gaussians.
(2) Each atom valence orbital is represented by three basis functions, the
tightest expanded in three Gaussians, the other two uncontracted.
(3) A set of uncontracted polarization functions has been added on each atom
(p-orbitals on hydrogen, d-orbitals on all other atoms).
(4) A set of diffuse functions (with the same l-values as those occurring in the
valence orbitals) have been added on all non-hydrogen atoms.
Just for fun before we go any further, let’s take a look at the EMSL web library
of Gaussian basis set (www.emsl.pnl.gov:2080/forms/basisform.html) used by
molecular quantum chemists. I can tell you that the lucky punter is given the choice of
the following basis set types (which I am typing in only because I want to concentrate
on watching Ally McBeal for half an hour – OK?):
STO-2G, STO-3G, STO-6G, STO-3G*, 3-21G, 3-21++G, 3-21G*, 3-21GSP,
4-31G, 4-22GSP, 6-31G, 6-31G-Blaudeau, 6-31++G, 6-31G*, 6-31G**, 6-31G*-
Blaudeau, 6-31+G*, 6-31++G**, 6-31G(3df,3pd), 6-311G, 6-311G*, 6-311G**, 6-
311+G*, 6-311++G**, 6-311++G(2d,2p), 6-311G(2df,2pd), 6-311++G(3df,3pd),
MINI (Huzinaga), MINI (Scaled), MIDI (Huzinaga), MIDI!, SV (Dunning-Hay),
SVP+Diffuse (Dunning-Hay), DZ (Dunning), DZP (Dunning), DZP+Diffuse
(Dunning), TZ (Dunning), Chipman DZP+Diffuse, cc-pVDZ, cc-PVTZ, cc-pVQZ, cc-
pV5Z, cc-pV6Z, pV6Z, pV7Z, cc-pVDZ(seg-opt), cc-pVTZ(seg-opt), cc-PVQZ(seg-
opt), cc-pCVDZ, cc-pCVTZ, cc-pCVQZ, cc-pCV5Z, aug-cc-pVDZ, aug-cc-pVTZ,
aug-cc-pVQZ, aug-cc-pV5Z, aug-cc-pV6Z, aug-pV7Z, aug-cc-pCVDZ, aug-cc-
pCVTZ, aug-cc-pCVQZ, aug-cc-pCV5Z, d-aug-cc-pVDZ, d-aug-cc-pVTZ, d-aug-cc-
pVQZ, d-aug-cc-pV5Z, d-aug-cc-pV6Z, Feller Misc. CVDZ, Feller Misc cVTZ, Feller
Misc. CVQZ, NASA Ames ANO, Roos Augmented Double Zeta ANO, Roos
Augmented Triple Zeta ANO, WTBS, GAMESS VTZ, GAMESS PVTZ, Partridge
Uncontr. 1, Partridge Uncontr. 2, Partridge Uncontr. 3, Ahlrichs VDZ, Ahlrichs,
pVDZ, Ahlrichs VTZ, Ahlrichs TZV, Binning/Curtiss SV, Binning/Curtiss VTZ,
Binning/Curtiss SVP, Binning-Curtiss VTZP, Mclean/Chandler VTZ, SV+Rydberg
 15

(Dunning-Hay), SVP+Rudberg (Dunning-Hay), SVP+Diffuse+Rydberg, DZ+Rydberg

(Dunning), DZP+Rydberg (Dunning), DZ+Double Rydberg (Dunning-Hay),
SV+Double Rydberg (Dunning-Hay), Wachters+f, Bauschlicher ANO, Sadlej pVTZ,
Hay-Wadt MB(n+1)ECP, Hay-Wadt VDZ(n+1)ECP, LANL2DZ ECP, SBKJC VDZ
ECP, CRENBL ECP, CRENBS ECP, Stuttgart RLC ECP, Stuttgart RSC ECP, DZVP
(DFT Orbital), DZVP2 (DFT Orbital), TZP (DFT Orbital), DeMon Coulomb Fitting,
DGauss A1 DFT Coulomb Fitting, DGauss A1 DFT Exchange Fitting, DGauss A2
DFT Coulomb Fitting, DGauss A2 DFT Exchange Fitting, Ahlrichs Coulomb Fitting,
cc-pVDZ-fit2-1, cc-pVTZ-fit2-1, cc-pVDZ_DK, cc-pVTZ_DK, cc-pVQZ_DK, cc-
pV5Z_DK, cc-pVDZ(pt/sf/fw), cc-PVTZ(pt/sf/fw), cc-pVQZ(pt/sf/fw), cc-
pV5Z(pt/sf/fw), cc-pVDZ(fi/sf/fw), cc-pVTZ(fi/sf/fw), cc-pVQZ(fi/sf/fw), cc-
pV5Z(fi/sf/fw), cc-pVDZ(pt/sf/sc), cc-pVDZ(pt/sf/lc), cc-pVTZ(pt/sf/sc), cc-
PVTZ(pt/sf/lc), ccp-PVQZ(pt/sf/sc), cc-pVQZ(pt/sf/lc), cc-PV5Z(pt/sf/sc), cc-
PV5Z(pt/sf/lc), cc-pVDZ(fi/sf/sc), cc-PVDZ(fi/sf/lc), cc-PVTZ(fi/sf/sc), cc-
PVTZ(fi/sf/lc), cc-PVQZ(fi/sf/sc), cc-PVQZ(fi/sf/lc), cc-PV5Z(fi/sf/sc), cc-
pV5Z(fi/sf/lc), Pople-Style Diffuse, STO-3G* Polarization, 3-21G* Polarization, 6-
31G* Polarization, 6-31G** Polarization, 6-311G* Polarization, 6-311G**
Polarization, Pople (2d/2p) Polarization, Pople (3df,3pd) Polarization), HONDO7
Polarization, Huzinaga Polarization, Dunning-Hay Diffuse, aug-cc-pVDZ Diffuse,
aug-cc-pVTZ Diffuse, aug-cc-pVQZ Diffuse, aug-cc-pV5Z Diffuse, aug-cc-pV6Z
Diffuse, aug-pV7Z Diffuse, d-aug-cc-pVDZ Diffuse, d-aug-cc-pVTZ Diffuse, d-aug-
cc-pVQZ Diffuse, d-aug-cc-pV5Z Diffuse, d-aug-cc-pV6Z Diffuse, DHMS
Polarization, Dunning-Hay Rydberg, Dunning-Hay Double Rydberg, Binning-Curtiss
(1d Polarization), Binning-Curtiss (df) Polarization, Ahlrichs Polarization,
Glendenning Polarization, Blaudeau Polarization, Core/val. Functions (cc-pCVDZ),
Core/val. Functions (cc-pCVTZ), Core/val. Functions (cc-pCVQZ), Core/val.
Functions (cc-pCV5Z).
I hope you can see that this would become an increasingly unprofitable
exercise if I commented further. Let me conclude this section by stating how I think
CRYSTAL users might record their basis sets in published work. First of all (and I’m
generalizing only a little) it is a sad fact that true molecular quantum chemists will not
believe any work you do unless it is done with a ‘named and published’ basis set with
the name in question being one of the ten or so people in the above list. This is a cross
we all have to bear. Kill yourself or get over it. In my opinion, the simplest and most
sensible notation for CRYSTAL users who develop their own sets might be something
like “In this piece of terrifically important research we used a basis of contracted
Gaussian-type functions of the form s(9)sp(7)sp(6)sp(3)sp(1)d(4)d(1) for element A
and s(8)sp(6)sp(3)sp(1) for element B, where the letters give the shell type and the
numbers in brackets give the number of primitive Gaussians in each shell contraction.
The exponents and contraction coefficients are reported in appendix X/TableY/Web
Site Z.” This conveys all relevant information, there is no need to decide on some
semi-arbitrary core-valence partition and, being familiar with the periodic table, the
reader is able to work out for herself whether the set includes polarization functions or
not. And for God’s sake don’t invent any more acronyms.

Basis sets in CRYSTAL

The basis set information in the CRYSTAL input deck is reasonably
straightforward, and may be understood through the example given on the following
page:
 16

28 7 nickel basis with seven shells

0 0 8 2.0 1.0
367916.0 0.000227
52493.9 0.001929 basis set type for this shell:
11175.8 0.0111 0 = general basis set, given as input (like this)
2925.4 0.05 1 = STO-nG (Z=1-54)
882.875 0.1703 2 = Pople 3(6)-21G (Z = 1-54(18))
305.538 0.369
119.551 0.4035 shell type (0 = s, 1 = sp, 2 = p, 3 = d)
49.9247 0.1426
0 1 6 8.0 1.0
924.525 -0.0052 0.0086 number of primitive Gaussians in this shell
223.044 -0.0679 0.0609
74.4211 -0.1319 0.2135 scale factor
29.6211 0.2576 0.3944 NB:[(scale factor)2 × exponent in contraction]=true exponent
12.4721 0.6357 0.3973 formal electronic charge attributed to the shell
4.2461 0.2838 0.2586
0 1 4 8.0 1.0
56.6581 0.0124 -0.018 exponent of normalized primitive Gaussian
21.2063 -0.2218 -0.08
8.4914 -0.8713 0.2089 s contraction coefficient
3.6152 1.0285 1.255 p contraction coefficient
0 1 1 0.0 1.0
1.5145 1.0 1.0
0 1 1 0.0 1.0
0.6144 1.0 1.0
0 3 4 8.0 1.0
41.08 0.040500
11.4126 0.202200
3.856 0.433800
1.33 0.489700
0 3 1 0.0 1.0
0.411 1.0
84 oxygen basis with four shells
0 0 8 2.0 1.0
8020.0 0.00108
1338.0 0.00804
255.4 0.05324
69.22 0.1681
23.90 0.3581
9.264 0.3855
3.851 0.1468
1.212 0.0728
0 1 4 8.0 1.0
49.43 -0.00883 0.00958
10.47 -0.0915 0.0696
3.235 -0.0402 0.2065
1.217 0.379 0.347
0 1 1 0.0 1.0
0.4764 1.0 1.0
0 1 1 0.0 1.0
0.1802 1.0 1.0
99 0 end of basis set input
All-electron basis set for nickel oxide (NiO)
 17

Two sets of all electron basis sets are included as internal data in the
CRYSTAL code, neither of which are worth using any more. Nevertheless, they are:

(1) the minimal STO-nG basis sets of Pople and co-workers (atomic nos. 1-54)
These basis sets are designed to mimic the shape of Slater-type functions, and
are obtained by fitting STOs with n contracted primitive Gaussians (where n is
generally between 2 and 6). Such a fit can be done accurately and the main limitation
to the usefulness of these sets appears to be that the STO itself is not a perfect basis
function. They are still used occasionally in spite of the poor quality of the resulting
wave function, presumably because they are well-documented and generally provide,
due to fortuitous cancellation of errors, reasonable optimized geometries at low cost.

(b) split-valence 3-21 and 6-21 basis sets

In these sets, the core shells are described as a linear combination of three (up
to atomic number 54) or six (up to atomic number 18) Gaussian primitives with the
two valence shells containing two and one Gaussians. The exponents and contraction
coefficients have been variationally optimized for the isolated atoms, and s and p
functions of the same shell share the same exponent. A single set of polarization
functions (p, d) can be added without causing numerical problems. Standard
molecular polarization functions are usually also adequate for periodic compounds.

4. Basis set selection

In choosing a basis set the paramount but conflicting issues are accuracy and
computational cost. These are obviously inversely related, and there is little more to
be said about it. However, computational cost alone should not determine what basis
set is used. Selecting a smaller set purely on the basis of a lack of sufficiently
powerful computers or interest will often prove unsuitable for describing the system in
question, which rather defeats the object of performing the calculation in the first
place. The minimum basis set requirements of all properties to be computed should
always be considered.
A great deal of general experience has now been gained by computational
chemists in selecting appropriate basis sets for molecular problems14, and much of this
experience can be used in the selection of basis sets for studies of crystalline solids.
However, I think it is true to say that the molecular quantum chemists do not in
general like to optimize basis sets by varying the exponents or contraction coefficients
to minimize the energy (and this is not necessarily their fault since the most popular
codes such as GAUSSIAN do not include a facility for doing so, other than by
laborious hand optimization). Rather, as we have seen, there is a hierarchy of basis
sets with perceived qualities, and for a difficult problem where accuracy is important
one would use a ‘good quality’ standard basis set from a library without modification.
The literature is full of statements like ‘this calculation was carried out at the STO-2G
level’ (probable translation: this calculation is rubbish, but my molecule is just too
big) or ‘this property was calculated at the TZVP level’ (i.e. it’s probably quite good).
In crystalline systems by contrast, basis set optimization is usually necessary,
essentially for two reasons. Firstly, there is a much larger variety of binding than in
molecules and basis sets are thus less transferable. For example, carbon atoms may be
involved in strong covalent bonds, e.g. in polyacetylene or diamond, as well as in
highly ionic systems such as Be2C, where the Mulliken charge of carbon is close to –
4. Secondly, hierarchical libraries of basis sets comparable to those available for
 18

molecules do not really exist. For certain types of compound, such as molecular
crystals (e.g. urea) or many covalent materials, the molecular sets can sometimes be
used largely unmodified (although I don’t necessarily recommend it). However, for
strongly ionic crystals and metals the basis sets, particularly the valence states, need to
be redefined completely. In essentially all cases, the core states may be described
using the solutions of atomic calculations, as even in the presence of strong crystal
fields the core states are barely perturbed and may be described by the linear
variational parameters in the SCF calculation.
Redefining basis sets in this way is obviously time consuming and even more
obviously rather boring, and so over the last five years various people involved with
the CRYSTAL program have contributed to an effort to develop libraries of basis sets
for CRYSTAL to be made available on the internet. The URL of the official site is:

www.ch.unito.it/ifm/teorica/crystal/AEbasisset/mendel.html

The site shows a periodic table. Clicking on the symbol for the required element will
reveal a text file containing various different basis sets which may have been used in
different materials containing that element type. Accompanying each basis set is list of
authors, a list of materials where the set has been used, references to publications and
hints on optimization where relevant. This table, which is obviously not complete,
grew out of a set of text files compiled largely by me whilst working in Torino in
1995. Since my departure to Cambridge in early 1997, I have maintained my own
separate library whose content has now diverged significantly from that of the official
site. It is a little more ‘experimental’ in the sense that it contains sets which have
never been used in published calculations, and also sets which have simply been
optimized in atomic SCF calculations but then never developed further. The aim of
providing such untried sets is to starting points for reoptimization where one hopes
that at least the core functions are reasonable. In particular it contains almost all of the
heavier elements beyond zinc in the periodic table up to around lanthanum where we
are forced to stop because of the lack of f functions in CRYSTAL. The URL of the
Cambridge library is:

www.tcm.phy.cam.ac.uk/~mdt26/crystal.html

There is a link from this page to the Cambridge quantum Monte Carlo page, which
will shortly include a table of basis sets found to have been useful in QMC
calculations. Finally, if you want to obtain standard molecular basis sets to use as
starting points for solid calculations (or even to do molecular calculations), you can
find the very useful EMSL library at:

www.emsl.pnl.gov:2080/forms/basisform.html

For European users, this serviced is mirrored at Daresbury Laboratory at:

wserv1.dl.ac.uk:800/emsl-pnl/basisform.html

Presently I shall discuss reoptimization strategies for the basis sets given in the
standard CRYSTAL libraries and also discuss the adaptation of molecular bases for
various types of solid. First of all a number of general principles are given that should
be taken into account when choosing a basis set for a periodic problem.
 19

a. Diffuse functions
The pre-screening procedure used in CRYSTAL is based on overlaps between
Gaussian s functions associated with each shell whose exponents are set equal to the
lowest exponent of all the primitive Gaussians in the contraction. The number of
integrals to be calculated thus increases very rapidly with decreasing exponents of the
primitive Gaussians, an effect which is much less pronounced in molecular
calculations. The following table shows that the cost of HF calculations on silicon and
diamond, which for such small systems is determined almost exclusively by the
number of bielectronic integrals, can increase by a factor of 10 simply by changing the
exponent of the most diffuse single Gaussian from 0.168 to 0.078 (Si) and from 0.296
to 0.176 (C). The cost is largely dominated by this shell, despite the fact that large
contractions are used for the 1s, 2sp and the innermost valence shell. The last entries
in the table are examples of weird behaviour - see part c.

Diamond Silicon
_ N EHF _ N EHF

0.296 58 –75.6633 0.168 46 –577.8099

0.276 74 –75.6728 0.153 53 –577.8181
0.256 83 –75.6779 0.138 72 –577.8231
0.236 109 –75.6800 0.123 104 –577.8268
0.216 148 –75.6802 0.108 151 –577.8276
0.196 241 –75.6783 0.093 250 –577.8266
0.176 349 no convergence 0.078 462 no convergence

Table 4.1 - Total Hartree-Fock energy EHF per cell and number of bielectronic integrals in 106
units (N) to be evaluated as a function of the exponent (_) of the outer shell for diamond and
silicon. In both cases a ‘split-valence’ 6-21G basis set was used. You can repeat the silicon
calculations with ‘test10’ in the CRYSTAL98 distribution if you want.

In atoms and molecules a large part of the additional variational freedom provided by
diffuse functions is used to describe the tails of the wave function, which are poorly
described by the long-range decay of the Gaussian function. In crystalline compounds
by contrast, particular in non-metallic systems, the large overlap between neighbours
in all directions drastically reduces the contribution of low-exponent Gaussians to the
wave function. This has the consequence that a small ‘split-valence’ basis set such as
6-21G is closer to the Hartree-Fock limit in crystals than in molecules.

b. Number of primitives
As discussed previously, a typical basis set will have ‘core functions’ with
higher exponents and a relatively large number of primitives - these will have a large
weight in the expansion of the core states. The ‘valence functions’ with a large weight
in the outer orbitals will have lower exponents and contractions of only a very few
primitives. We can get away with putting a lot of primitives in the core since core
states have very little overlap with neighbouring atoms and thus the use of a large
number of primitives in the GTF contraction is of limited cost in CPU time. The use
of many primitives in the valence shells would add significantly to the cost of a
calculation.
 20

c. Numerical catastrophes
Under certain conditions a CRYSTAL calculation may fall into a non-physical
state during the SCF part characterized by an oscillating total energy significantly
higher than the true energy. Such calculations will not, in general, converge. It is
observed that the risks of numerical problems like this increases rapidly with
decreasing value of the most diffuse Gaussian exponent in the basis set. It happens,
for example, in the silicon calculation reported in Table 4.1 where the exponent of the
most diffuse basis function is 0.073. In general this behaviour may be attributed to
limitations in the accuracy of the Coulomb and exchange series evaluation. The
exchange is by far the more delicate of the two series since long-range contributions
are not taken into account and because the ‘pseudoverlap’ criteria associated with the
two overlap parameters ITOL4 and ITOL5 mimic the real behaviour of the density
matrix only in an approximate way. This means that calculations which require exact
Fock exchange (i.e. Hartree-Fock, hybrid DFT) are at much greater risk of showing
this behaviour. An LDA-DFT calculation of our errant Si calculation actually works
perfectly well.
In order to obtain physical solutions in situations like this, the usual remedy is
to increase the integral tolerances to give higher precision via the TOLINTEG
keyword. For non-metallic systems with medium-sized basis sets, the default integral
tolerances of 6 6 6 6 12 are adequate for the optimization of the exponents of the
valence shell and for systematic studies of the energy versus volume curves. However,
in metals, the optimization of the energy versus exponent curve at the Hartree-Fock
level is often not even possible, even with much higher integral tolerances. If you
insist on studying metals with HF, reasonable values of the valence shell exponent
(say 0.23 for beryllium and 0.10 for lithium) can be used for the study of the structural
and electronic properties of metallic systems even though they don’t correspond to a
variational minimum. Use DFT instead.

d. Basis set superposition error

A rather serious problem associated with Gaussian basis sets is basis set
superposition error (BSSE). A common response to this problem is to ignore it, since
it will go away in the limit of a complete basis. Sometimes this approach is justified,
but this requires investigation that is seldom performed, and some understanding of
BSSE is indispensable in order to perform accurate and reliable calculations. The
problem of BSSE is a simple one: in a system comprising interacting fragments A and
B, the fact that in practice the basis sets on A and B are incomplete means that the
fragment energy of A will necessarily be improved by the basis functions on B,
irrespective of whether there is any genuine binding interaction in the compound
system or not. The improvement in the fragment energies will lower the energy of the
combined system giving a spurious increase in the binding energy. It is often stated
that BSSE is an effect that one needs to worry about only in calculations on very
weakly interacting systems. This is not really true. BSSE is an ever-present
phenomenon and accurate calculations should always include an investigation of
BSSE. Examples of areas in which one should be particularly wary are the study of the
binding energy of molecules adsorbed on surfaces (see, for example Ref 15 for an
interesting discussion) or the calculation of defect formation energies..
The approach most commonly taken to estimate the effect of BSSE is the
counterpoise correction16: the separated fragment energies are computed not in the
individual fragment basis sets, but in the total basis set for the system including ‘ghost
 21

functions’ for the fragment that is not present. These energies are then used to define a
counterpoise-corrected (CPC) interaction energy, which by comparison with
perturbation theory, has been shown to converge to the BSSE-free correct value17. An
example of how to compute the counterpoise correction using CRYSTAL in a
calculation on a simple model system (CO adsorbed on MgO (100) surface) affected
by BSSE will be given in Exercise 2.

e. Pseudopotentials
It is well known that core states are not in general affected by changes in
chemical bonding. The idea behind pseudopotentials is therefore to treat the core
electrons as effective averaged potentials rather than actual particles. Pseudopotential
are thus not orbitals but modifications to the Hamiltonian and are used because they
can introduce significant computational efficiencies. In plane wave calculations,
pseudopotentials are essentially mandatory since the core orbitals have very sharp
features in the region close to the nucleus and too many plane waves would be
required to expand them if they were included. The most important characteristic of a
pseudo designed for such calculations is that it is as smooth as possible in the core
region. Pseudopotentials in Gaussian basis set calculations are not mandatory and
have different characteristics to those designed for plane waves since Gaussians
actually have sharp features in the core region. If the CPU time in CRYSTAL is
dominated by the integrals calculation, they will not even buy you very much since the
number of integrals is controlled by more diffuse functions which overlap strongly
with neighbouring atoms – something which basis functions with large weight in the
core orbitals are not very good at. However the use of pseudopotentials will decrease
the number of coefficients in the wave function and might give significant savings in
the SCF part. It is also quite easy to incorporate relativistic effects into
pseudopotentials which is increasingly important for heavy atoms. All electron
relativistic calculations are very expensive and not possible in CRYSTAL anyway.
Some people have used pseudopotentials to overcome the problem of the missing f
Gaussian basis functions in CRYSTAL and have done calculations on heavy atoms
containing f electrons. How does the use of pseudopotentials modify the basis set in
Gaussian calculations? Take an all-electron basis set for that atom. First of all one
might hope that basis functions which have a large weight only in the core orbitals
might be removed. Remove them. Make sure you are left with the correct number of
electrons. The remaining basis functions must then be optimized (see later) with
reference to the pseudopotential.

5. Practical optimization
So to summarize, there are a number of approaches to developing a basis set
for a periodic HF/DFT calculation. Obviously the easiest way is to download standard
sets from the online libraries or ask experienced CRYSTAL users and to use these sets
without modification. The second way is to start from one of these standard sets and
improve it. The third way is to suitably modify a molecular basis set for use in your
crystalline system. The fourth way is to develop a basis set from scratch using atomic
SCF calculations, probably using a nearby atom in the periodic table as a starting
point. Let us summarize first of all some ways ‘one might’ improve a basis set:

z Reoptimize the more diffuse exponents (and contraction coefficients if necessary).

z Decontract i.e. convert the more diffuse contractions into single Gaussian
primitives.
 22

z Convert sp functions into separate s and p functions.

z Add polarization functions if not already present
z Add more primitives (watch out for linear dependence problems).
z Use a better starting point
Reoptimization in this sense means varying an appropriate subset of the basis set
parameters until the energy is minimized. In principle this is a reasonably complex
multidimensional minimization, but there are various standard shell scripts available
to help you with this (see later). Be careful that the ratio between succesive exponents
doesn’t fall below 2-2.5, otherwise the basis may suffer from linear dependence
problems. Watch the CPU time, particularly when carrying out decontraction, or
adding polarization functions.

By means of some (very simple) examples I will now briefly consider the
adequacy of molecular basis sets for different types of crystalline compound. Note that
the basis sets discussed are hardly ‘state of the art’ and are meant to illustrate
particular principles only. Note also that I have ‘adapted’ some of these discussions
from those given in the CRYSTAL manual.

Covalent systems
Let’s consider again two (stereo)typically covalent systems, diamond and
silicon. I will use the CRYSTAL standard ‘split-valence’ 6-21G basis sets, that is, the
core shells are described with a contraction of six primitive Gaussians and the inner
and outer valence shells contain respectively two and one Gaussians. sp shells are
used throughout, in that the s and p functions of the same shell share the same
exponent, and all contraction coefficients are variationally optimized in the isolated
atoms. The best exponent of the outer shell of the atom is 0.196 for C and 0.093 for
Si. Reoptimization of the valence shell of C in two molecules gave 0.24 in CH4 and
0.189 in CO for these quantities18. Repeating the optimization in the two crystalline
compounds reveals that the most internal valence shell is essentially unaltered with
respect to the atomic solution, while for the outer single-Gaussian shell the best
exponent is 0.22 for diamond and 0.11 for silicon. These values are very similar to
those optimized in the isolated atoms. If a single-Gaussian d polarization shell (which
is five separate functions) is added to the 6-21G basis (i.e. to give 6-21G*) and the
exponents optimized one gets 0.8 for diamond and 0.4 for silicon. These values are
very close to those resulting from the molecular optimization, which are 0.8 for
diamond19 and 0.45 for silicon20. It seems therefore that small molecular split-valence
basis sets can therefore be used with confidence and essentially without modification
to describe covalent crystals. It is generally advisable however to reoptimize the
exponent of the most diffuse shell, which produces a slightly improved basis, while
reducing the cost of the calculation. That said, 6-21G* is not really all that good and a
larger better basis set with more variational freedom is quite easy to make for these
cases (see web libraries).

Ionic crystals
The classification of materials as covalent or ionic is a conventional one but
the division between the two is necessarily rather blurred. Examples of more or less
fully ionic compounds are LiH and MgO, and for these systems the cation valence
shell is almost completely empty. For such cations it often proves convenient to use a
basis set containing only ‘core’ functions plus an additional sp shell with a relatively
high exponent. As an example, in previous work using CRYSTAL such sp shells were
 23

used for Mg in MgO and for Li in LiH, Li2O and Li3N with respective exponents of
0.3-0.4 and 0.5-0.6 21. Total energies obtained either by using only core functions for
Li/Mg or by adding a ‘valence’ shell to the cation differed by less than 0.1eV/atom.
This figure was essentially the same for a relatively large range of exponents of the
outer shell, say 0.2-0.5 for Mg. It is usually difficult, and often impossible, to optimize
the exponents of functions which only have appreciable weight in almost empty
orbitals; one finds that the energy decreases almost linearly with the exponent. As
discussed in the previous section, very low exponent values require the calculation of
enormous numbers of integrals and may lead to numerical instabilities. Thus for ionic
crystals with nearly empty shells, and where the energy gain of optimization is
relatively small (say, a decrease in energy of less than 1 mHartree for a change in _ of
around 0.2) it is usually convenient to use a relatively large exponent for this shell.
Anions present a different problem. Reference to isolated ion solutions is
possible only for halides, because in such cases the ions are stable even at the Hartree-
Fock level. For other anions, which are stabilized by the crystalline field (such as H–,
O2–, N3– and C4–), the basis set must be redesigned with reference to the crystalline
environment. Consider, for example, the optimization of the O2– basis set in Li2O. The
difficulty is to allow the valence distribution to relax in the presence of two more
electrons. We begin from a standard STO-6G basis set i.e. six contracted primitive
Gaussians for the 1s shell, and six more to describe the 2sp shell. First of all, two
more Gaussians were introduced into the 1s contraction, in order to improve the virial
coefficient and total energy. The two outer Gaussians of the valence sp shell were then
removed from the contraction and allowed to vary independently. The exponents of
the two outer independent Gaussians and the coefficients of the four contracted ones
were optimized in Li2O. The best outer exponents of the ion were found to be 0.45 and
0.15 and are therefore considerably more diffuse than the neutral isolated atom, where
the best exponents are 0.54 and 0.24. The rest of the O2– valence shell is unchanged
with respect to the atomic situation. The introduction of d functions in the oxygen
basis set gives only a minor improvement in the energy of 1×10–4 Hartree per cell,
with a population of 0.02 electrons/atom/cell (d functions may be important in the
calculation of certain properties however - see later). Thus for anions, reoptimization
of the most diffuse valence shells is mandatory when starting from a standard basis
set.

Semi-ionic crystals
Intermediate situations should be considered individually and the adequacy of
selected basis sets must be carefully tested. Examples of semi-ionic compounds are _-
quartz (SiO2) and corundum (Al2O3). The exponents of the outer shell for the two
cations in the 6-21G basis are 0.093 (Si) and 0.064 (Al). In both cases, this function
proves to be too diffuse, even causing numerical catastrophes at the HF level in the Al
case. For quartz, reoptimization in the bulk gives _=0.15 for Si (the dependence of the
total energy on _ is much smaller than in pure silicon (Table 4.1) and the cost at
_=0.15 is only 50% of the one at _=0.09). In contrast the best molecular and
crystalline exponent for oxygen (_=0.37) coincide. Corundum is more ionic than
quartz, and about two valence electrons are transferred to oxygen. In this case it is
better to eliminate the most diffuse valence shell of Al, and to use as independent
functions two Gaussians of the inner valence shells (_=0.94 and 0.3 respectively).
 24

Metals
It is often stated22 that Gaussian basis sets are somehow inappropriate for
describing simple metals and that plane-waves, for example, are a more ‘natural’ basis
via some sort of analogy with the orbitals of the free electron gas (which are plane-
waves). However until recently very few studies had been done to argue properly one
way or the other. The reason for this has its origin in the fact that to reproduce the
nearly uniform density characterizing simple metallic systems such as lithium and
beryllium one needs to use very diffuse Gaussians. You will recall from an earlier
section that this is a very bad idea in Hartree-Fock because of the Fock exchange
pathology and it is generally impossible to optimize the basis set in such cases. Until
DFT calculations became possible (with the 1995-1996 release of CRYSTAL) it was
thus quite difficult to separate the effects of basis set and Hamiltonian. The few
Gaussian DFT studies that have been done since then seem to indicate that GTFs are
able to provide a reliable and efficient description of simple metallic systems23,24.
An interesting example of a CRYSTAL study of a metallic system is that of
Doll, Harrison and Saunders24 who investigated the effect of computational
parameters (including the basis set) on the calculated surface and bulk properties of
metallic lithium. This system will be used in Exercise 3, and therefore a certain
amount of apparently irrelevant detail will be presented here. Doll et al. began with
the following core s function taken from an earlier study25:

0 0 6 2.0 1.0
840.0 0.00264
217.5 0.00850
72.3 0.00335
19.66 0.1824
5.44 0.6379
1.5 1.0

Keeping this core function fixed, they then proceeded to add functions to this basis of
increasing complexity and examined the convergence of various properties:

z BASIS SET 1 [s(6)sp(1)sp(1)]: The sp exponents were optimized with LDA (or
PWGGA) to 0.5 and 0.08. However, as an exponent of 0.08 gives rise to a diffuse
basis function close to numerical instability and was quite expensive, exponents of
0.5 and 0.1 were in fact used.
z BASIS SET 2 [s(6)sp(1)sp(1)sp(1)]: ‘even tempered’ exponents (i.e. the ratio
between the exponents is kept fixed - to 2.5 in this case). Exponents 0.5, 0.2, 0.08.
This ratio is close to the lowest which can be tolerated before on-site (atomic)
linear dependence is seen. It is however also known from previous work to
converge the atomic energy to within less than 0.0001 Ha of the exact Hartree-
Fock ground state energy.
z BASIS SET 3: [s(6)sp(1)sp(1)sp(1)d(1)] Just like BASIS SET 2 but with a d
polarization function, whose exponent was optimized with a PWGGA functional
to be 0.15. However, the d function leads only to a minor change in the total
energy. Changing the exponent to 0.5 changes the energy only by around 0.00005
Ha.

As expected from previous discussions about the Fock exchange pathology, an

optimization of the basis set exponents was not possible either at the Hartree-Fock
 25

level or using the B3LYP hybrid DFT functional: the outermost exponent became
more and more diffuse until finally the solution became unstable.
One of the properties computed by Doll et al. was the cohesive energy for
which accurate energies of the free atom are also required. A richer basis set is
required to compute this accurately because of the need to describe the long-range
behaviour of the atomic wave function (note the absence of p functions):

z BASIS SET 4: [s(6)s(1)s(1)s(1)s(1)s(1)] with outer exponents of 0.6, 0.24, 0.0096,

0.04 and 0.0016.

In the LDA and PWGGA calculations it is also required to expand the exchange and
correlation potentials in an auxiliary basis set. Doll et al. went beyond the defaults
and used a set consisting of 13 even-tempered s-functions with exponents from 0.1 to
2000, 3 even-tempered p-functions with exponents from 0.1 to 0.8, and 2 d-functions
with exponents of 0.12 and 0.3. This is sufficient to integrate the charge density to an
accuracy of 10-7 |e|. For the free atom, they used an auxiliary basis set with 18 even-
tempered s-functions with exponents from 0.0037 to 4565.
In metals the reciprocal space sampling is also a critical and rather delicate
issue. In CRYSTAL the sampling is performed on a Monkhorst-Pack net where the
density of points is determined by a shrinking factor. The Fermi energy and shape of
the Fermi surface are determined by interpolation onto a ‘Gilat’ net. This net is simply
related to the Monkhorst-Pack net by an additional subdivision factor. To further
improve convergence, the finite temperature generalization of density functional
theory can be used to apply Fermi surface smearing (with the SMEAR keyword). Note
that a higher number of sampling points in the Gilat net leads to a systematic
improvement at zero temperature. At finite temperature, the number of Gilat points
does not influence the results so long as the MP net is sufficiently dense. Properties
were investigated using smearing of 0.001Ha and 0.02 Ha. It was found that a
shrinking factor of 16 for the MP net and a temperature of 0.001Ha gave good results
(i.e. convergence of the energy to at least 0.0001 Ha with respect to reciprocal space
sampling.
The conclusions of Doll et al. were essentially as follows. The cohesive energy
and lattice constant were stable even with the smallest of the basis sets. The elastic
constants and surface energies were more sensitive to basis set. The converged values
of all properties were in full agreement with experiment and calculated values from
the literature. The results in best agreement with experiment were obtained with the
Perdew-Wang GGA funtional. Hartree-Fock and hybrid functionals were very difficult
because of the Fock exchange pathology. Finite temperature calculations could be
used to improve convergence and an extrapolation to zero temperature was both
possible and accurate. [Now why not try Exercise 3?].

Transition elements
A particularly interesting new field to which CRYSTAL has been applied over
the last six or seven years is that of magnetic compounds containing transition
elements. These are examples of what physicists refer to as ‘strongly correlated
materials’. Such materials have been the subject of controversy due to the great
difficulties that density functional theory calculations (based on LDA or GGA
functionals) have had in this area, with magnetic insulators being predicted to be
metals and so on. Lest previous sections of these notes have convinced you that
Hartree-Fock calculations are actually rather useless, it turns out that UHF actually
 26

gets the ground state qualitatively correct in magnetic insulators (I will briefly explain
this shortly). A collaboration between the Torino and Daresbury (and other) groups
beginning in 1993 was able to demonstrate this and helped to highlight the problem
with DFT. Relevant papers from that time can be found in References26,27,28,29,30,31,32,33
and since then a great deal of other work has been done by a wide variety of groups 34.
A talk attempting to explain this topic in simple terms for first-year graduate students
is available on the web at www.tcm.phy.cam.ac.uk/~mdt26/tmo/scm_talk.html.
So, what is a strongly correlated material? ‘Strongly-correlated’ is a term used
in many-body physics to mean that a particular parameter in a ‘toy’ model (referring to
‘on-site’ intratomic Coulomb interactions) is bigger than other parameters in the
model (related to the band width, or kinetic energy). It should not be confused with
‘correlation’ in the quantum chemistry sense, which is the energy difference between
the exact non-relativistic energy and the Hartree-Fock energy in the limit of a
complete basis set. What characteristics of a material, from the point of view of its
electronic structure, make it strongly correlated? This is a good question. To begin
with, it is generally assumed that electrons in strongly-correlated materials are
‘localized’ in some sense (see Professor Resta’s talk for a definition of this) and their
constituent atoms retain much of their free-atom-like characteristics. It is thus
convenient to consider such problems in real, as opposed to reciprocal space. To a
first approximation you might think of the crystalline orbitals corresponding to the 3d
states in such materials as periodic arrays of particular atomic d orbitals (such as the
dxz or dyz or whatever) multiplied by a phase factor with some k. As a function of k
these will form relatively narrow bands. To create an insulating state, some of these
bands must be full and some must be empty. What mechanism exists for splitting the
sub-bands within the d manifold?

z Crystal field splitting: in the presence of a cubic crystal field (for example) due to
the presence of neighbouring atoms the d manifold will be split into eg (dx2-y2 and
dz2) and t2g (dxy,dxz,dyz) subsets. The energy scale of this is often rather small.
z Exchange splitting: Electrons of the same spin tend to stay out of each other's way
because of the exchange interaction, and so the interelectronic Coulomb repulsion
will be smaller for the majority-spin electrons The majority-spin bands will
therefore be lowered in energy with respect to the minority spin bands (this can
split up-spin eg bands and down-spin eg bands for example).

But what mechanism exists for splitting the subbands corresponding to the two
different eg orbitals of the same spin?

z On-site Coulomb interactions: we populate one of the eg orbitals with an up-spin

electron and leave the other empty. Imagine there are n electrons on this particular
transition metal ion. Thus an electron in the occupied eg sub-band feels the
potential of n-1 electrons and an ‘added’ electron in the virtual orbital would feel
the potential of n electrons. The difference is the on-site Coulomb interaction U
(which is the ‘strong correlation’ i.e. a sort of screened intra-atomic Hartree
interaction).

It turns out that such behaviour can be replicated at the UHF level with a
single determinant wave function. The reason for the failure of LDA/GGA
calculations to do the same is interesting. Within the LDA, the potential felt by each
electron is computed from a functional of the total electron densities. For such simple
 27

density functionals this leads to eigenvalues which are relatively weak functions of the
particular orbital occupancy. Ultimately this behaviour stems from the spurious
inclusion of 'self-interaction’ effects in the exchange-correlation potential (the
interelectron Coulomb energy in a one-electron atom is non-zero using the LDA!). In
HF theory, the non-local exchange exactly cancels the self-interaction and introduces a
strongly orbitally-dependent potential which splits the manifold of d states in precisely
the manner expected from a simple empirical (‘Hubbard model’) estimate of the on-
site interactions between electrons in different orbitals. Indeed a variety of new ‘DFT’
schemes (e.g. LDA+U, SIC-LDA) which emulate important features of the Hartree-
Fock Hamiltonian have now been developed which give better descriptions of the on-
site interactions than regular DFT. New 'exact-exchange' DFT formulations are
currently the focus of intense research and also hold a great deal of promise. So the
point is that although DFT is in principle exact and you can compute the total energy
as a functional of the density, the eigenvalue spectrum does not necessarily correspond
to anything physical (since the orbitals are merely auxiliary functions used to
parameterize the density). Unfortunately having the wrong eigenvalue spectrum means
that KS-DFT calculations will sometimes converge to an incorrect ground state with
the wrong density.
To treat systems like these with any degree of accuracy at the UHF level in
CRYSTAL (or at the hybrid DFT level which is also promising) reasonably good
basis sets for the transition elements are required. These are not that widely available
even to molecular quantum chemists since until relatively recently most of the effort
in developing molecular GTF basis sets has been for first- and second-row atoms. One
reason for this may be that molecules containing transition metal atoms tend to be very
badly described at the Hartree-Fock level. Molecular bonds tend to have a fairly high
degree of ‘covalency’ and the existence of partially-occupied d states leads to a great
many nearly-degenerate levels, and thus to a large ‘static correlation’ (i.e. the weight
of the HF determinant in a CI expansion would be small, and a multi-determinant
treatment is more appropriate). Basis sets to describe correlation using quantum
chemistry correlated wave function techniques need to be much richer than those for
systems well-described at the Hartree-Fock level since they need to treat all of the
unoccupied levels. It may seem surprising that single-determinant HF could be so
successful in periodic crystalline magnetic insulators containing transition elements,
but this is an important characteristic of these ionic materials. As we have seen, the
highly symmetric environment and long-range Coulomb forces tend to separate the
orbitals into well-defined subsets with a significant gap between occupied and
unoccupied states. Hence, the ground state of NiO (for example) is rather well
described by a single determinant. In this sense, a strongly correlated magnetic
insulator is in many ways a ‘simpler system’ than many molecules. The success of
UHF calculations in these materials (and also hybrid DFT schemes) has now been
well documented in a variety of publications.
In order to study materials containing first and second transition series
elements, I carried out an intermittent program of work between 1993 and 1995 aimed
at developing entirely new contracted atomic basis sets for elements beyond atomic
number 20 (Ca). These are not necessarily that good by molecular standards, but in
many cases (particularly for the 3d elements Sc-Zn) these have since been reoptimized
in the solid state and used in published work. These sets are available on the Torino
and Cambridge Gaussian basis set library web sites referred to earlier. Since the
Torino site has a fairly strict and entirely laudable no paper-no web policy, the atomic
basis sets beyond Zn as far as lanthanum including the 4d transition elements are
 28

largely only available from the Cambridge library. It is encouraging to note however
that where these sets have been used in real materials, only relatively minor
reoptimizations of the most diffuse functions were required.

Programs

There are two programs (that we have access to) which you can use to optimize basis
sets (and incidentally, geometries). Both are Unix shell scripts which work by
repeatedly calling CRYSTAL whilst varying the requested basis set/geometry
parameters in a CRYSTAL input file. You can do this by hand of course but dying of
boredom is, I imagine, rather unpleasant.

Billy
The first and older program was originally written by me in around 1991 (the first
program I ever wrote. Aaaah...). For some reason it is called ‘billy’ and it is a Unix
csh script. In most respects however, it has been superseded by LoptCG (see below).
Billy is very easy to use and is pretty robust, but generally requires more calls to
CRYSTAL than LoptCG since it doesn’t use any gradient information in the
multidimensional minimization. Rather, it uses repeated line minimizations. Since
basis set parameters typically vary only in narrow ranges and roughly keep certain
ratios to each other, this is actually not as bad as it sounds and billy usually finds the
same minimum as LoptCG (in basis set minimizations anyway), just more slowly. It is
actually better at riding over badly behaved input decks where CRYSTAL runs fail to
converge for certain values of the parameters. To use it, simply precede any parameter
in the CRYSTAL input file you wish to optimize with an asterisk (‘*’) and type
something like ‘billy input_filename 5 ‘ where 5 is an initial percentage scan range.
See the supplied documentation for additional command line flags for multiple
minimization cycles, and for controlling GUESSP restarts. If your calculation is not
particularly time-consuming and you don’t want to optimize too many parameters,
billy is still perfectly serviceable.
Find billy at : www.tcm.phy.cam.ac.uk/~mdt26/downloads/billy.tar.gz

LoptCG
LoptCG was written in around 1996 by Claudio Zicovich-Wilson whilst working in
Torino with the CRYSTAL group, and is now available with the CRYSTAL98
distribution on request. It is a Unix ksh script and is different from billy in that it
calculates numerical energy gradients by finite differencing and uses this information
to help it carry out a multidimensional conjugate gradient (Polak-Ribiere)
minimization. It requires a certain investment of time to learn how to use it properly,
and requires more complicated input, but its minimization algorithm is quite a bit
more intelligent than billy and I recommend its use. It also has many other nice
features such as the ability to change the minimization strategy halfway through an
optimization. See the supplied documentation for details. More information about
LoptCG is also available at www.ch.unito.it/ifm/teorica/LoptCG.html .

35
5. Examples
 29

In this final section we shall discuss a few brief but important examples of the
effects of basis set selection on the total energy and other related properties. First of
all we consider the perovskite KMnF3, described with the basis set reported in Table
5.1. As this system is almost fully ionic, the basis sets for bulk calculations have been
derived from basis sets optimized for the isolated K+, Mn2+ and F– ions. The exponents
of the most diffuse single-Gaussian sp and d shells were reoptimized in the bulk. The
basis set in Table 5.1 is expected to be reasonably good - there are three valence sp
shells on the anion and two on the cation (the 4s orbitals of K and Mn are almost
completely empty). The d electrons are described by two shells, a contraction of four
Gaussians for the inner part, and a single Gaussian for the outer part. The calculation
with this basis is cheap, taking only a few minutes on a medium-sized workstation.
This is because (a) the unit cell contains only five atoms, (b) the system has high
symmetry and (c) the external Gaussians of the two cations have reasonably large
exponents (0.50 and 0.22 for Mn and K respectively) and that of the anion is not too
diffuse (0.18).
The basis set of Table 9 can be improved in a number of ways:
(1) Polarization functions (d functions) may be added to the K and F bases.
(2) Additional diffuse sp shells can be added to Mn.
(3) the 4-1G contraction for the d electrons on Mn can be substituted with a 5-1G.
Some results of these basis set improvements are shown in Table 5.2. The effect both
on the run time and in calculations of a number of properties (the binding energy,
lattice parameter, bulk modulus, ferro-antiferromagnetic energy difference 6E, and
two elastic constants) are given. Most of these quantities are very stable with respect
to basis set improvements, and only in the case of 6E is a maximum variation of about
20% observed. This stability is a consequence of the fully ionic nature of KMnF3. and
of its high symmetry. The electron charge distributions of K+ and F– are essentially
spherical, so that polarization functions are nearly useless. The 4sp shell of Mn is
empty and so the additional sp shell with exponent 0.25 is not used.
In order to understand the relationship between basis set flexibility,
polarizability of the ions and the crystalline symmetry, we will consider the influence
of d polarization functions on two ionic compounds, Na2O and K2O. In Table 5.3
some calculated bulk properties of these materials are reported, obtained with and
without d functions on the cation. The d functions are seen to have negligible
influence on all but one of these properties, namely the C44 elastic constant, where
reductions of 10% and 42% are observed when d functions are introduced. These
results can be interpreted in the following way. The cations are in a high symmetry
position, and when the unit cell is modified according to the deformations required for
the evaluation of B, C11 and C11–C12, their local symmetry remains cubic. The d
functions are thus used essentially for describing the ‘breathing’ of the ion. This may
also be described by the s and p functions of the valence shell however. For these
quantities the difference between K2O (with a variation of the order of 3%) and Na2O
(1%) is due to the larger polarizability of the potassium ion, or in other words to the
smaller energy difference between the valence s/p and virtual d levels. On the contrary
the deformation required for the evaluation of the C44 elastic constant drastically
reduces the atomic point symmetry. The cation is thus no longer in a centrosymmetric
position and the ion can undergo a dipolar relaxation, for which a combination of p
and d functions is required. This effect is less marked for Na2O, because the sodium
ion is less polarizable.
 30

Shell Mn K F
exponents coeffs. Exponents coeffs. exponents coeffs.
s(d) p s(d) p s(d) p
s 292601.0 0.000227 172500.0 0.00022 13770.0 0.000877
42265.0 0.0019 24320.0 0.00192 1590.0 0.00915
8947.29 0.0111 5140.0 0.01109 326.5 0.0486
2330.32 0.0501 1343.9 0.04992 91.66 0.1691
702.047 0.1705 404.5 0.1702 30.46 0.3708
242.907 0.3691 139.4 0.3679 11.50 0.41649
94.955 0.4035 54.39 0.4036 4.76 0.1306
39.5777 0.1437 22.71 0.1459

sp 732.14 -0.0053 0.0086 402.0 -0.00603 0.0084 19.0 -0.1094 0.1244

175.551 -0.0673 0.0612 93.5 -0.805 0.0602 4.53 -0.1289 0.5323
58.5093 -0.1293 0.2135 30.75 -0.0109 0.2117 1.387 1.0 1.0
23.129 0.2535 0.4018 11.92 0.258 0.3726
9.7536 0.6345 0.4012 5.167 0.684 0.4022
3.4545 0.2714 0.2222 1.582 0.399 0.186

sp 38.389 0.0157 -0.0311 17.35 -0.0074 - 0.44 1.0 1.0

0.0321
15.4367 -0.2535 - 7.55 -0.129 -0.062
0.0969
6.1781 -0.8648 0.2563 2.939 -0.6384 0.1691
2.8235 0..9937 1.6552 1.19 1.08 1.5
0.674 1.03 1.06

sp 1.2086 1.0 1.0 0.389 1.0 1.0 0.179 1.0 1.0

sp 0.4986 1.0 1.0 0.216 1.0 1.0

d 22.5929 0.0708
6.1674 0.3044
2.0638 0.5469
0.7401 0.5102

d 0.249

Atom
sp = = = 0.4017 1.0 1.0 = = =

sp = = = 0.2216 1.0 1.0 0.15 1.0 1.0

sp 0.067 1.0 1.0 0.0281 1.0 1.0

Table 5.1 - Exponents and coefficients of the Gaussian-type basis functions adopted for the
study of KMnF3. The first and second part of the table refer to the bulk and atomic basis sets
respectively. In the lower part of the table the first two rows give the exponents and
coefficients of the functions modified in the atom with respect to the ion in the bulk; the last
row refers to the functions added for the description of the atomic tails. The atomic basis set is
used for the evaluation of the binding energies given in Table 5.2. The symbol ‘=’ stands for
‘unmodified’.
 31

Case Basis set N t BE a0 B 6E C44 C11-C12

a) Table 5.1 83 1198 0.604 4.280 64.6 0.293 30.9 91.8

b) 5-1d not 4-1d 83 1409 0.605 4.280 64.1 0.292 30.8 90.9

c) 87 1858 0.606 4.284 63.3 0.334 30.3 91.3

b) + sp Mn(_=0.25)
d) 102 2541 0.608 4.280 63.5 0.319 30.2 89.8
c) + d on F (_=0.7)
e) 107 3040 0.609 4.276 63.9 0.325 29.6 94.8
d) + d on K (_=0.4)

Table 5.2 - Effect of basis set on bulk properties of KMnF3. BE, a0, 6E, B, C44, C11 and C11–
C12 are the binding energy (hartree), the lattice parameter (Å), the energy difference between
ferromagnetic and antiferromagnetic phases (millihartree), the bulk modulus and two of the
elastic constants (GPa). M is the number of functions in the basis set; t is the total CPU time.

Na2O K2O
no d d 6% no d d 6%

ET -398.693 -398.695 -0.002 -1273.184 -1273.193 -0.010

a0 5.498 5.487 -0.2 6.550 6.466 -1.3
B 58.7 58.7 0.0 33.3 34.6 +3.5
C11 127.3 126.14 -0.9 71.8 74.1 +3.1
C12 23.9 23.8 -0.4 14.2 14.8 +4.1
C44 37.8 34.4 -10.2 19.7 13.9 -41.7

Table 5.3 - Effect on bulk properties of simple oxides of adding d polarization functions to the
cations. 6% is the percentage difference between the calculation with and without d functions
(for ET it is the absolute difference)
 32

1. Some simple properties of Gaussians

(a) Compute the overlap integral of two normalized s-type Gaussian functions with
exponents _ and `, and with a common centre.
(b) Compute the overlap integral between two s-type functions centred at A and B,
and with exponents _ and ` respectively.
(c) As (b), but for p functions.

NB: Note the following standard integral:

' 2 /
00 e <_x = 1
2 _

2. Basis set superposition error (BSSE)

This exercise is concerned with the evaluation of the counterpoise correction
to the basis set superposition error in a simple system, namely CO molecules adsorbed
on the (100) surface of MgO. Experimentally, one finds that the CO molecules are
vertically adsorbed via the C atom over the Mg ions at the (001) face with an
adsorption heat of around 3.6 kcal/mol. Before we do any computations, we must first
examine the energetics of surface/adsorbate interactions.
The interaction energy between an ad-molecule M and a surface site S within a
surface complex MS is obtained by calculating the total energies of the three systems
involved and then finding

6E = E ( MS ) < E ( M ) < E ( S ) .

Each of these quantities must be evaluated at the appropriate equilibrium geometry.

Note that the interaction energy is defined to be negative for attractive interactions and
that the term ‘binding energy’ refers to the negative of the interaction energy. When
adsorption layers are treated by periodic methods, as in CRYSTAL, a more specific
definition is required in order to take into account the ‘lateral interaction energy’,
which is the interaction energy per unit cell between the molecules in their periodic
array adsorbed on the surface. This quantity can be either positive (repulsion) or
negative (attraction) depending on the nature of the molecules. In the limit of very low
coverage the distance between molecules becomes very large and the lateral
interaction energy tends to zero.
To compute the binding energy per unit cell per adsorbate molecule the
following quantities must therefore be computed:
(1) Eslab/ads : the energy per unit cell of a crystal slab with an interacting periodic array of
adsorbed molecules.
(2) Eslab : the energy per unit cell of the clean crystal slab.
 33

(3) Eads : the energy per unit cell of the periodic array of adsorbed molecules in the
absence of the surface.
(4) Emol : the energy of a single isolated adsorbate molecule.

The interaction energy per unit cell between the whole adsorbate layer and the surface
is

6E slab / ads = E slab/ ads < E slab < E ads .

The lateral interaction energy is

L
6E ads = E ads < N u E mol

where N is the number of molecules in the unit cell.

The experimental energy of adsorption 6Eexp corresponds to a process where
the molecules move from an ideal gas state into an adsorbed state where attractive
interactions cause them to become attached to the host surface:

6E exp = E slab / ads < E slab < N u E mol

One thus establishes the link between the computed interaction energy and the
experimental 6Eexp:

L
6E exp = 6E slab / ads + 6E ads

In the limit of low coverage, this is effectively

6E exp 5 6E slab / ads

So anyway, having calculated this quantity, we might suspect that basis set
superposition error has affected the result. This error arises in the following way. As
the basis sets used in the calculation are generally far from complete, both the
adsorbate layer and the surface layer may use the additional variational freedon
offered by each others basis functions to lower their energy. This gives a non-physical,
stabilizing contribution to the energy of the surface-adsorbate complex, and may also
lead to artificial charge transfer if the basis set description of the two subsytems is
unbalanced. Hence there may be an error in the interaction energy which is connected
with the superposition of the basis functions of the two subsystems.
The stability of the result with respect to this error may be checked by the
counterpoise correction. To do this one must recalculate Eslab and Eads supplementing
the basis set of each subsystem with all the basis functions of the other but without
their electrons and nuclei. These additional basis functions are referred to as ‘ghost
functions’. The energies obtained at the equilibrium geometry of the complex for each
subsystem e.g. E(M{S}) are lower than the energies calculated at the same geometry
with the basis functions of the respective subsystems alone e.g. E(M) , and the
difference ¡ is defined as the BSSE:
 34

¡ ( M ) = E ( M ) < E ( M {S})

¡ ( S ) = E ( S ) < E ( S { M })

The BSSE values ¡(M) and ¡(S) are then used to define a counterpoise-corrected
interaction energy

c
6E exp = 6E exp + ¡ ( M ) + ¡ ( S ) .

This quantity should always be calculated as a check of the quality and balance of the
basis set.
In CRYSTAL, the counterpoise correction for such a surface/adsorbate system
is calculated using the GHOSTS option, by means of which selected atoms may be
turned into ‘ghosts’ by deleting their nuclear and electronic charges, and setting their
conventional atomic number to zero.

Prepare input files for the following (using basis set libraries and appropriate defaults)
mgo_co: single layer MgO slab with adsorbed CO molecules.
mgo_ghost: single layer MgO slab with ghost CO basis functions.
co: periodic array of CO atoms without MgO slab (real or ghost).
co_ghost: periodic array of CO molecules with ghost MgO basis functions.
co_mol: single isolated CO molecule.
mgo: single layer MgO slab

The geometry for the MgO(100)+adsorbed CO system may be specified in CRYSTAL

format as follows (you should be able to work out all other geometries using this
example):

CRYSTAL
000
225
4.21
2
12 0.0 0.0 0.0
8 0.5 0.5 0.5
SLAB
001
11
BREAKSYM
ATOMINSE
2
6 1.488 –1.488 4.605
108 1.488 –1.488 5.729
ENDG

Can you explain why we use 108 for the atomic number of oxygen when adding the
CO molecule? Now modify the inputs containing CO molecules using the
SUPERCELL option to ensure an adsorbate coverage of one half, that is, half the
magnesium ions on the surface have an adsorbed CO molecule. Run CRYSTAL
 35

Hartree-Fock calculations from your six files. They should run very quickly. Note the
total energies of these six calculations.

(a) Calculate the interaction energy (kcal/mol) per unit cell between the MgO surface
cell
and the CO overlayer 6E slab / ads , ignoring BSSE. Is the interaction attractive or
repulsive?
(b) Calculate the lateral interaction energy. Do the CO molecules on the surface attract
or repel each other? Hence correct the interaction energy of adsorption to form the
quantity 6Eexp which may be meaningfully compared with experiment. The
experimental heat of adsorption is –3.6 kcal/mol. How well are we doing?
(c) Compute the counterpoise correction to 6Eexp. What is the magnitude and sign of
the counterpoise-corrected interaction energy? How big is the BSSE compared to the
uncorrected interaction energy?
(d) Repeat the whole exercise using a posteriori correlation corrections to the Hartree-
Fock energy using the PWGGA functional, and also regular DFT with the PWGGA
functional. How do the results compare with experiment and each other? How would
you suggest the calculations presented here (which are reasonably crude) might be
improved?

NB: To avoid typing mistakes on calculators, it is better to copy energies onto the
Unix calculator with the mouse and do the sums with that (type ‘bc -l’ to get the
calculator).
Also, to convert from Hartree to kcal/mol multiply by 627.50754.

3. Basis set development and computation of properties in lithium metal

Read again the description of the lithium metal calculations of Doll, Harrison
and Saunders in section 5 of the notes (originally reported in Ref. 24). Prepare
CRYSTAL input files for bulk Li using the three basis sets specified (noting also the
details of the reciprocal space sampling and auxiliary basis sets, and using sensible
defaults for the rest of the input parameters). Prepare an input file also for atomic
lithium using the suggested Li atomic set. See if you can come up with ways of
improving these basis sets (you might familiarize yourself with the optimizer scripts
billy and/or LoptCG while you’re at it) and examine how these ‘improvements’ affect
the CPU time. Calculate some simple bulk properties of your choice (suggestions:
lattice constant, bulk modulus, cohesive energy, surface energies.) and see how they
change using different Hamiltonians (HF and different DFT functionals) and the
various basis sets (including your ‘improved’ one). Ask for the experimental values
when you’ve finished.

Note that the geometry of bulk Li can be specified in CRYSTAL format as follows:

CRYSTAL
000
229
3.44
1
3 0.0 0.0 0.0
ENDG
 36

1(a). A useful property of Gaussian functions is that the product of Gaussians are
other Gaussians, and also that so many integrals factorize. In cases like this, we use

2 2 2
e <_r e < `r = e <(_ + ` ) r

to convert to a single Gaussian, and combine with the formula

3
' ' ' <_r 2 ' <_x 2 ' <_y 2 ' <_z 2 £/ ¥2
0<' 0<' 0<' e dxdydz = 0<' e dx 0 e
<'
dy 0 e
<'
dz = ² ´
¤_ ¦

Including real normalization constants N1 and N2, we get

3
£ / ¥2
q1 q 1 = N 12 ² ´ =1
¤ 2_ ¦

3
£ / ¥2
q 2 q1 = N1 N 2 ² ´
¤_ + ` ¦

3
£ / ¥2
q2 q2 = N 22 ² ´ =1
¤2` ¦

Eliminating, we find

3 3
3 (_` ) 4 £ _` ¥2
q 2 q1 = 2 2 =² ´
¤ (_ + ` ) / 2 ¦
3
(_ + ` ) 2
The fraction in parentheses in the last expression on the RHS is the ratio of the
geometric average to the arithmetic average of the positive numbers _ and `. This is
less than 1, except when the numbers are equal, and then it is 1.
(b) Another useful fact is that the product of two Gaussians is a new Gaussian, even if
they are not on the same centre.
To show this, we first need the fact that

_ ( x < A) 2 + ` ( x < B) 2 = (_ + ` ) x 2 < 2(_A + `B) x + _A 2 + `B 2 = (_ + ` ) ( x < C ) 2 + D

 37

_A + `B _`
which is easily solved to give C = and D = ( A < B) 2 . This implies
_+` _+`
that

_`
< ( A <B ) 2
<_ ( r < A ) 2 < ` ( r<B )2 2
N1e N 2e = N1 N 2 e _ +` e < a ( r < C)

with a = _ + ` . For s functions, the overlap integral is thus

_` 3
< ( A<B )2 £ / ¥2
S = N1 N 2 e _ +` ² ´
¤a ¦

With the known normalization factors from (a), we finally obtain

3
_`
£ _` ¥2 < ( A <B ) 2
S =² _ +`
´ e
¤ (_ + ` ) / 2 ¦

which is also obvious from the solution of part (a). Note that the new centre
_A + `B
C= is simply the weighted mean of the original centres.
_+`

(c) From the previous solution, it is obvious we can now generalize:

2
q 1 = N 1 p1 (r < A )e <_ ( r < A )

2
q 2 = N 2 p2 ( r < B ) e < ` ( r < B )

‰
2
q 1q 2 = N 1 N 2 Fp(r < C)e <a ( r <C)

where p1 is a polynomial of total degree n1, expressing the angular dependence around
centre A, p2 has total degree n2 around centre B, and

_A + `B
C=
_+`

£ _` ¥
F = exp² < ( A < B) 2 ´
¤ _+` ¦

a =_ + `

In particular, for two sp shells, we obtain x < Ax = ( x < C x ) + ( C x < Ax ) etc., so

 38

( x < Ax )( x < B x ) = ( x < C x ) 2 < 2( x CA + x CB )( x < C x ) + x CA x CB

and so on, where xCA is short for Cx–Ax etc. The following integrals are standard:
3
£/ ¥2
I 0 = 000 exp(<a (r < C) 2 )dp =² ´
¤a ¦

3
£ 1 ¥£ / ¥ 2 I
I 2 = 000 ( x < C x ) 2 exp(<a (r < C) 2 )dp = ² ´² ´ = 0
¤ 2a ¦¤ a ¦ 2a

The overlap integrals are thus

£ I 2 + x CA x CB I 0 x CA y CB I 0 xCA zCB I 0 ¥
² ´
S = N 1 N 2 F × ² y CA x CB I 0 I 2 + y CA yCB I 0 y CA zCB I 0 ´
² ´
¤ zCA x CB I 0 x CA y CB I 0 I 2 + z CA zCB I 0 ¦

This is an important general point. All the integrals involving the various components
of complete shells of Cartesian (or spherical harmonic) Gaussians are obtained from a
few values in common for all integrals, combined with simple expressions involving
the relative co-ordinates of the centre.

2. Adsorption of CO on the MgO (001) surface.

This is what I get in the HF and a posteriori PWGGA cases (you will have to do pure
DFT PWGGA yourself..).
(energy units Hartree unless otherwise stated)

STANDARD HARTREE-FOCK TREATMENT

total energies:
CO –112.6265204421
CO with MgO ghost –112.6287274346
MgO –549.2096009758
MgO with CO ghost –549.2153973595
MgO with adsorbed CO –661.8392701443
interaction energy per unit cell (Eslab/ads) –0.0031487264
i.e. –1.976 kcal/mol (attractive)
lateral interaction energy
energy of isolated CO molecule –112.6270474548
energy of CO layer with no MgO –112.6265204421
L
lateral interaction energy 6E ads +0.0005270127 (repulsive)
L
[Note this is half coverage - for full coverage, a 6E ads of +0.00624 - twelve times
higher - was obtained]
Therefore 6E exp is –0.0026217137
i.e. –1.645 kcal/mol (attractive)
which may be compared with around –3.6 kcal/mol experimentally. Not too bad.
However,
counterpoise correction:
 39

¡(CO) +0.0022069925
¡(MgO) +0.0057963837
Therefore total BSSE +0.0080033762
c
Counterpoise-corrected energy 6E exp +0.0053816625 (repulsive)
Therefore all the binding observed in the original calculation was due to basis set
superposition error. Such a calculation using CRYSTAL, ignoring BSSE, was
published in 1986. It seems we must improve our treatment. For such a weak bond,
correlation may be important. One way to improve the results could therefore be to
estimate correlation-corrected interaction energies using density functionals of the
Hartree-Fock density:
A POSTERIORI CORRELATION
correlation-corrected total energies:
CO –113.116080
CO with MgO ghost –113.117780
MgO –550.756528
MgO with CO ghost –550.760809
MgO with adsorbed CO –663.882652
interaction energy per unit cell (Eslab/ads) –0.010044
i.e. –6.302 kcal/mol (attractive)
lateral interaction energy
energy of isolated CO molecule –113.116612
energy of CO layer with no MgO –113.116080
L
lateral interaction energy 6E ads +0.000532(repulsive)
Therefore 6E exp is –0.009512
i.e. –5.968 kcal/mol (attractive)
counterpoise correction:
¡(CO) +0.001700
¡(MgO) +0.004281
Therefore total BSSE +0.005981
c
Counterpoise-corrected energy 6E exp –0.003531 (attractive)
i.e. –2.215 kcal/mol
to be compared with around –3.6 kcal/mol experimentally. This is a reasonable result.
It suggests that much of the binding for CO on MgO is due to correlation effects
beyond the Hartree-Fock level. Of course however, we have no reason to expect such
good agreement since we are still using a rather poor model, and the next correction
may shift the answer considerably and in the wrong direction. We should probably
investigate:
(a) Use of better basis sets on all atoms.
(b) Slabs thicker than one layer.
(c) Rumpling of slabs at the surface.
(d) Careful optimization of CO bond length and position above surface.
(e) More sophisticated correlation treatments than correlation-corrected Hartree-Fock
– quantum Monte Carlo being a good bet!

3. There aren’t any right answers to this exercise. Just get some hands on experience
at playing with input files and running CRYSTAL/billy/LoptCG.
 40

1
CRYSTAL 98 user documentation, University of Torino and CLRC Daresbury Laboratory (1998).
2
See, for example, I. Shavitt, in Methods of Electronic Structure Theory, edited by H.F. Schaefer
(Plenum, New York, 1977).
3
R. Shepard, Adv. Chem. Phys. 69, 63 (1987) ; B.O. Roos, Adv. Chem. Phys. 69, 399 (1987).
4
R.J. Bartlett, J. Phys. Chem 90, 4356 (1989).
5
For a comprehensive review, see: C.A. Weatherford and H.W. Jones, ETO multicentre integrals,
(Reidel, Dordrecht, 1982).
6
J. Almlöf, in Lecture Notes in Chemistry, 64, edited by B.O. Roos (Springer, Berlin, 1994)
7
M. Challacombe and J. Cioslowski, J. Chem. Phys. 100, 464 (1994).
8
See, for example, R.C. Raffenetti, J. Chem. Phys. 60, 918 (1974) ; M.W. Schmidt and K. Ruedenberg,
J. Chem. Phys. 71, 3951 (1979).
9
L.E. McMurchie and E.R. Davidson, J. Comput. Phys. 26, 218 (1978)
10
V.R. Saunders in Methods in Computational Molecular Physics, NATO ASI Ser. D, Vol 113, edited
by G.H.F. Diercksen and S. Wilson, (Reidel, Dordrecht, 1983), p.1.
11
See, for example, the simplified discussion at www.tcm.phy.cam.ac.uk/~mdt26/tmo/scm_talk.html .
12
S. Goedecker, Rev. Mod. Phys 71, 1085 (1999).
13
R.J. Needs, G. Rajagopal, M.D. Towler, P.R.C.Kent, A.J. Williamson, CASINO 1.0 User Manual,
University of Cambridge (2000).
14
See, for example, E.R. Davidson and D. Feller, Chem. Rev. 86, 681 (1986) and references therein.
15
J. Sauer, P. Ugliengo, E. Garrone and V.R. Saunders, Chem. Rev. 94, 2095 (1994).
16
S.F. Boys and F. Bernardi, Mol. Phys. 19, 553 (1970).
17
M. Gutowski, J.G.C.M. van Duijneveldt-van de Rijdt, J.H. van Lenthe, F.B. van Duijneveldt, J.
Chem. Phys. 98, 4728 (1993).
18
J.S. Binkley, J.A. Pople, W.J. Hehre, J. Am. Chem. Soc. 102, 939 (1980).
19
P.C. Hariharan, J.A. Pople, Theor. Chim. Acta. 28, 313 (1973).
20
W.J. Pietro, MM. Francl, W.J. Hehre, D.J. Defrees, J.A. Pople, J.S. Binkley, J. Am. Chem. Soc. 104,
5039 (1982).
21
R. Dovesi, C. Roetti, C. Freyria-Fava, M. Prencipe and V.R. Saunders, Chem. Phys. 156, 11 (1991);
M. Causà, R. Dovesi, C. Pisani and C. Roetti, Phys. Rev. B 33, 1308 (1986).
22
Everybody in the Theory of Condensed Matter Group, Cavendish Laboratory, University of
Cambridge (private communication).
23
I. Baraille, C. Pouchan, M.Causà and F. Marinelli, J. Phys. C 10, 10969 (1998).
24
K.Doll, N.M. Harrison and V.R. Saunders, J. Phys. Cond. Mat 11, 5007 (1999).
25
M. Prencipe, A. Zupan, R. Dovesi, E. Aprà and V.R. Saunders, Phys. Rev. B 51, 3391 (1995).
26
W.C. Mackrodt, N.M. Harrison, V.R. Saunders, N.L. Allan, M.D. Towler, E. Aprà and R. Dovesi,
Philos. Mag. A, 68, 653 (1993)
27
M.D. Towler, N.L. Allan, N.M. Harrison, V.R. Saunders, W.C. Mackrodt and E. Aprà, Phys. Rev. B,
50, 5041 (1994) ; M.D. Towler, Ph.D. thesis, University of Bristol (1994).
28
J.M. Ricart, R. Dovesi, V.R. Saunders and C. Roetti, Phys. Rev. B, 52, 2381 (1995).
29
M.D. Towler, R. Dovesi and V.R. Saunders, Phys. Rev. B, to be published (October 1995).
30
W.C.Mackrodt, N.L. Allan, N.M. Harrison, V.R. Saunders and M.D. Towler, Phys. Rev. Lett.,
submitted (1995).
31
M.D. Towler, N.L. Allan, N.M. Harrison, V.R. Saunders and W.C. Mackrodt, J. Phys.: Cond. Mat.,
7, 6231 (1995).
32
M. Catti, G. Valerio and R. Dovesi, Phys. Rev. B, 51, 7441 (1995).
33
G. Valerio, M. Catti, R. Dovesi and R. Orlando, Phys. Rev. B, 52, 2422 (1995).
34
See www.tcm.phy.cam.ac.uk/~mdt26/sc_refs.html for a list of references to ‘strongly correlated
materials’ calculations using CRYSTAL.
35
This section is based on work done by R. Dovesi and F. Freyria-Fava, and the text, apart from certain
grammatical aspects, is largely reproduced from RD’s notes on ‘Total Energy and Related Properties’
from the 1995 Torino Summer School.