F, L,. (L.V - FF, L,. (R.), I,. (R.) L R. R'L
F, L,. (L.V - FF, L,. (R.), I,. (R.) L R. R'L
F, L,. (L.V - FF, L,. (R.), I,. (R.) L R. R'L
ET = - ~ f,l,.~ [l.V~- ~
~
i=1
'f',
~ 2 1
n=1
Zn
I Rn-r ]",.dr.+.L
,' I 'f'l I 2 ~
i,j = 1
'f', l'f'))
1 ,I,..(r.),I,..(r.)dr.dr.
~ ff,l,.~(r.),I,.~(r.) Ir.-r.I'f'1
1 )
l'f')) 1 )
The variation of this total energy with respect to the The Hartree-Fock operator HHF(r) is defined by
one-eleetron wave funetions cPi(ri), subject to the eon-
straint of their orthonormality, leads to the Hartree-Fock (5)
equations
(4) with
h (r ) = - tV 2
-
z,
---
Z2 to be solved are the Schrödinger-type equations derived
Irl-rl Ir 2 - r l ' from Eq. (4) with potentials kept fixed, and the Poisson
(6) equation (7) relating the potential V C to the electronic
VC(r)= i~1 f 4>r(r') I r~r' I 4>i(r')dr' , density p. The solution of the Hartree-Fock-Slater equa-
tions (4) has to be achieved iteratively. '
and
111. CHOICE OF THE COORDINATE SYSTEM
with a a constant. The simplified form of a local ex- where the unknown one-electron wave functions are
change with a=O.7 is used here because the solution of (10)
the Hartree-Fock-Slater (HFS) equations with the FEM
discussed below depends only on the local form of the ap- m denotes the projection of the angular momentum onto
proximation. With this approach the total energy of the the internuclear axis. The interelectronic Coulomb po-
molecular system is calculated as tential can be written as folIows:
N
E tot = ~ Ei -t f p(r)Vc(r)dr-t f p(r)Vx(r)dr · (9)
(11)
;=1
Using the ansatz (10) for the wave functions <Pi(r) the
The Hartree-Fock-Slater equations (4), with the ex- Schrödinger-type equation (4) can be rewritten as a two-
change part according to Eq. (8), are a system of coupled dimensional differential equation in the axial-symmetric
integre-differential equations. The differential equations coordinate system
I
(12)
C
gll = [a~1 r [a~2 r
+
-21 asa [Kt(s,t)a;-
aVc] -iat"
1 a [K aV 1 2
at 1 +
2(s,t)---at
al l [a/2]2
g22= [ at '
= -4rrK 4(s,t)p(s,t) (13)
g33=/t·
for the Poisson equation (7). The coefficient functions We agree with Laaksonen et al.9 - 11 and Becke l2 in
K 1, K 2 , K 3 , and K 4 are calculated from the transforma- finding the prolate-spheroidal coordinates most suitable
tion functions 11 and 12through for the properties of the self-consistent calculation of dia-
tomic molecules. The transformation equations for this
coordinate system are given by the equations
K I (s,t)=K 2(s,t)=(R /2) sinhs sint , erties can be taken into account, e.g., one can use small
elements in regions of physical importance and large ele-
K 3(s,t)=(R /2)(sinh 2s +sin 2t )/(sinhs sinr ) , (16) ments in regions of lesser weight. Thus the point distri-
K 4(s,t),=(R 3 /2)(sinh 2s +sin2t ).sinhs sint . bution can be adapted to a given problem.
The next step in the FEM is to choose a trial function
Other coordinate systems were also tested. For the for the element to approximate the solution of the
case of H 2+ we found the coordinates with the transfor- differential equation. In order to fulfill the continuity cri-
mation equations terion of the solution from one element to the next, the
trial functions are written in a special way. For each ele-
XI =(R /2)[( l-c Ins )2_1]( I-t 2 ) coeq: , ment a number of points, the nodal points, are chosen. If
x2 =(R /2)[( l-c Ins )2_1]( I-t 2 ) sine , (17) u "(s, t) is the approximation of the unknown function
over the element with the number e, this function is writ-
x~=(R/2)(I-clns)t , ten as a linear combination of the nodal values
p
to be most useful. By varying the transformation param- ue(s,t)= ~ u{N{(s,t) , (21)
eter c of the transformation equations (17) very accurate ;=1
results could be achieved with a very small number of
points.P However, for the H 2 problem we used the coor- where u{ is the ith nodal value of the eth element. The
dinate system defined by the transformation equations function N{(s,t) is the ith shape function. The sum runs
over all nodal values of the element. Two main types of
XI =(R /2)[(sinhs + 1)2_1]( I-t 2 ) covp , elements exist, namely, the Lagrangian- and the
Hermitian-type elements. (If the interpolation of an un-
X 2 = (R /2 )[ (sinhs + 1 )2 - 1]( 1- t 2 ) sinc , (18)
known function is calculated only due to the values of the
x3 =(R /2)(sinhs + l)t . function at specified points this is called a Lagrange inter-
polation. The interpolation of the unknown function
The reason for this was that the solution of the Poisson which takes also the derivatives of the function into ac-
equation within the coordinate system (17) proved to be count is called a Hermite interpolation.P) Lagrangian
unstable. elements have only one nodal value per nodal point,
namely, the value of the unknown function. Hermitian
IV. FINITE-ELEMENT METHOn (REFS. 1-3) elements have more than one nodal value per nodal point,
because here also the derivatives of the functions are no-
Originally the FEM was developed in engineering sei- dal values. One conclusion of Eq. (21) is that each shape
ence in order to calculate static and dynamic stresses of function equals 1 only at a single nodal point of the ele-
complicated constructions. Recently, it has been dem on- ment and equals 0 at all others. This restriction allows
strated l 4 - 20 that this method can also be used with great the construction of the shape functions from the given
success to solve quantum-mechanical problems. In order trial function.
to apply the FEM to the solution of the Hartree-Fock- We use Lagrange elements with two-dimensional poly-
Slater equations, one best starts from the variational nomials up to order 6 as trial function. To be able to con-
equivalent of the second-order partial differential equa- struct the shape functions we choose as many nodal
tions which, for the Schrödinger-type equation (12), takes points for each elements as there are free coefficients of
the form the two-dimensional polynomials. Figure 1 shows the no-
12 = ff [-f a~c
K, [ r -tK a~c
2 [ r
+41TK 4(S,t)P(S,tlVC ] ds dt . (20)
~ 0 6 g
2+2K
+(K 3 m 4VlN{NJ]dsdt (24)
FIG. 2. Structure of the global matrices resulting from the
and subdivision of a reetangular region given in Fig. 3. The matrix
elements of element 1 (2,3,4) are symbolized by 0 ( X, 0,6. ).
(25)
I
with the matrix elements for the Sehrödinger-type equation and to the matrix
equation
aN~ aN~ aN~ aN~
D;j=t f f [K 1 a;-- a/ +K af- a/2 ds dt (27)
12·u=d (33)
and the veetor elements for the Poisson equation. The order of this matrix equa-
tion is equal to the number Np of nodal variables. For
(28) the Lagrangian elements used by us, this is equal to the
total number of points.
The vector u" is the element nodal veetor of the eth ele- The ealeulation of tbe matrix elements (24,25,27,38) is
ment. done numerieally with a eonieal produet Gauss integra-
Adding the eontributions of all elements leads to the tion rule. 24 A number of only 7 X 7 integration points for
expressions eaeh element proved to be suffieient for all ealeulations.
The matrix eigenvalue problem is solved by an inverse
(29)
veetor iteration method 3, 25 with modifieations due to the
and iterative solution of the Hartree-Foek-Slater equations.
The matrix equation is solved by a Cholesky deeomposi-
(30) tion. 3, 26 All algorithms used at this point take eare of the
band strueture of the global matrices.:'
where u denotes the global nodal veetor.
The global matriees H, S, and n are band-struetured
matrices, beeause for any given nodal variable u, at point
(Si' t;) the sum over all elements runs only over those ele-
ments whieh share this point. As an example, Fig. 2 3 5
shows the strueture of the global matrix for the subdi-
vision of a reetangular region given in Fig. 3. The nodal
points are given by the vertiees of the triangles. Tbe no-
dal variables are numbered from 1 to 6 and the elements
from 1 to 4. The matrix elements of element 1 (2,3,4) are
marked with the symbols 0 (X, 0,6 ).
With the substitution of the trial funetions the fune-
tionals I 1 and I 2 are seen to be funetions of the nodal
variables U i : The eonditions for I 1 and /2 to be minimal 2 4 6
are
FIG. 3. Subdivision of a reetangular region with four tri-
all a/ 2 angular elements. The enumeration of the nodal points leads to
-=0, --=0, i=I, ... ,Np. (31)
au; au; a bandwidth of 3 for the corresponding global matrices.
4998 D. HEINEMANN, B. FRICKE, AND D. KOLB 38
V. BOUNDARY CONDITIONS
S=('1 +'2)/R ,
(34)
11=('1 -'2 )/R ,
where the relation to the coordinates sand t used above
(15) is given by S=coshs and 1]=cost. Within these coor-
dinates the asymptotic behavior!' for the wave functions
4J(S, 1]) is
~~exp[ -( -2E)I/2(R /2)S], S-+ 00 (35)
where E is the one-electron energy eigenvalue. For the
boundary conditions of the wave functions we use the O -.:::I~ _ __1__ _...1.__ _. 1 . __ ___.JI.....--~
Solve. No Test
.tty=e~y onvergenee TABLE I. Convergence properties of total energy and energy
eigenvalue of the system H 2 with increasing number of points.
All values are given in a.u,
H 2, internuclear distance R = 1.40 a.u,
Points ET E
Calculate:
Density s Is, t) 6X6 - 1.128737 69 -0.59583041
11 XiI -1.13362024 -0.59466119
16X 11 -1.13362884 -0.594658 13
21 x n -1.13362949 -0.59465860
26X 11 -1.13362956 -0.59465856
FIG.4. Flowchart of the finite-element program solving the
41 X 16 -1.133629571 7(2) -0.5946585694(3)
Hartree-Fock-Slater equations.
38 SOLUTION OF THE HARTREE-FOCK-SLATER EQUATIONS ... 4999
rAßLE 11. Total energy and energy eigenvalues of the system N 2 for different grid sizes and different orders of the polynomials
over the elements. The last figure for the largest grids is uncertain. All values are given in a.u.
N 2, internuclear distance R =2.07 a.u,
Laaksonen et al. a This work, fifth-order polynomials
Points 5989 2601 3136 3721 4356
Er -108.346622 -108.346605 -108.3466076 -108.3466086 - 108.3466090
E(lug) -13.981070 -13.9810680 -13.981068'28 -13.98106837 -13.98106840
E(lu u) -13.979661 -13.979658 1 - 13.979658 38 -13.97965847 -13.97965850
E(2ug ) -1.007215 -1.00721471 -1.00721472 -1.00721471 -1.00721471
E(2uu ) -0.460725 - 0.460725 06 -0.46072505 -0.46072505 -0.46072505
E( I1T u) -0.404235 -0.40423461 - 0.40423462 -0.40423462 -0.40423462
E(3ug ) -0.350058 -0.35005851 0.35005852 -0.35005852 -0.35005852
ae and Ii.E set to 10- 8 a.u, and Ii. v = 10-8~ Typically the order polynomials as trial functions this results in a total
change of the total energy Ii. E· was the most restricting number of 26 X 26 grid points, and in 31 X 31 grid points
criterion for the convergence. for the sixth-order polynomials. Such an equidistant
The self-consistent calculation is initialized by the vari- point distribution is not at all optimal, and it will be
able screening potential of Eichler and Wille. 27 With this shown that point distributions, which are physically more
starting potential about 20 iterations were needed to adequate, will increase the accuracy.
achieve convergence to 10- 8• Table I shows the convergence of the total energy and
the lUg level of the system H 2 with increasing grid size.
VII. RESULTS For a number of only 26 X 11 grid points an accuracy
better than 10- 8 was reached with fifth-order polynomi-
We present the results for some of the systems calculat- als. Increasing the number of grid points by a factor of
ed by us. The systems here were chosen to be identical to about 3 to 41 X 16 points enabled us to add two more
the systems already calculated by Laaksonen et al.,9-11,22 significant digits to a new benchmark with ten-figure ac-
in order to compare the numerical quality of the results. curacy. This shows the stability of the FEM even for
For the simplicity the points were distributed equidis- highly accurate calculations.
tantly in the selected coordinate system. As an example, The next system was N 2• In Table 11 we compare the
Fig. 5 shows a triangularization of the two-dimensional convergence of the results with increasing grid size for '
region with a total number of 50 elements. For the fifth- both fifth- and for sixth-order polynomials. For 2401
TABLE III. Total energy and energy eigenvalues of the system CO and BH for the largest grids used
so far. The last figure given is uncertain. All values are given in a.u.
Laaksonen et alea This work, sixth order
BH CO BH CO
Points 4356 4356
R 2.366 2.13 2.366 2.13
ET -24.808852 -112.129925 -24.80885148 -112.12991528
E( 10') -6.5323604 -18.744146 -6.53236004 -18.74414325
E(20') -0.4078652 -9.911347 -0.407 8~5 19 -9.911 34609
E(30') -0.1731323 -1.044 171 0.17313242 -1.044 17077
E(40') -0.489071 -0.48907075
E( 17T) -0.413613 -0.412612 71
E(50') -0.303029 - 0.303029 91
"Reference 11.
5000 O. HEINEMANN, B. FRICKE, AND D. KOLB 38
TAHLE IV. Total energy and energy eigenvalues of the sys- no special vector routines were used. About 90% of the
tem N 2 for two different point distributions. Mesh A refers to CPU time was needed to solve the matrix eigenvalue
an equidistant grid with all elements of equal size, whereas mesh problem (32) and the matrix equation (33). The CPU
B refers to a logarithrnic point distribution in the s coordinate. time per iteration is about 20 seconds for the 31 X 31
All values are given in a.u. point grid with sixth-order polynomials. The increase of
N 2, internuclear distance R =2.07 a.u, CPU time with grid size is about
Sixth-order polynornials (40)
Mesh A Mesh B
Here n denotes the number of grid points, n1 is the num-
Points 961 961
ber of levels to be calculated, and mb is the bandwidth of
Er -108.34645 -108.34659
E( lag) -13.98104 -13.981066
the global matrices. For the bandwidth mb we can write
E( lau) -13.97963 -13.979656 (41)
E(2a g )' -1.0072143 -1.0072150
E(2a u ) -0.4607255 -0.4607253 when nord is the order of the polynomials of the trial
E( I1T u ) -0.4042346 -0.4042347 function. The last formula is only true for the regular
E(3a g ) -0.3500577 0.3500587 grids used up to now.
1G. Strang and G. Fix, An Analysis 0/ the Finite Element 7R. Gaspar, Acta Phys. Acad. Sei. Hung, 3, 263 (1954).
Method (Prentice-Hall, Englewood Cliffs, 1973). 8W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965).
20. H. Norrie and G. de Vries, An Introduction to Finite Ele- 9L. Laaksonen, D. Sundholm, and P. Pyykkö, ehern. Phys. Lett.
ment Analysis (Academic, New York, 1978). 96, 1 (1983).
3H. R. Schwarz, Methode der Finiten Elemente (Teubner, 10L. Laaksonen, D. Sundholm, and P. Pyykkö, Int. J. Quantum
Stuttgart, 1980). Chern. 27, 601 (1985).
4J. C. Slater, Quantum Theory 0/ Atomic Structure, (McGraw- 11L. Laaksonen, D. Sundholm, and P. Pyykkö, Comp. Phys.
HilI, New York, 1960). Rep. 4, 313 (1986).
sC. Froese-Fischer, The Hartree-Fock Method for Atoms 12A.D. Hecke, J. Chem. Phys. 76, 6037 (1982).
(McGraw-Hill, New York, 1977). 13W. Schulze and D. Kolb, Chem. Phys. Lett. 122, 271 (1985).
6J. C. Slater, Phys. Rev. 81, 385 (1951). 14A. Askar, A. Cakrnak, and R. Rabits, Chern. Phys. 33, 267
38 SOLUTION OF THE HARTREE-FOCK-SLATER EQUATIONS ... 5001