Paper abstract
An ab initio linear-scaling scheme
P. D. Haynes and M. C. Payne
Theory of Condensed Matter, Cavendish Laboratory, University of Cambridge
Abstract
In this paper we briefly survey the current state
of ab initio calculations in terms of the accuracy and range of
applicability of these methods for studying complex processes in real
materials. We highlight some of the successes and limitations of these
techniques and discuss the extent to which linear-scaling methods are
able to extend the scope and scale of ab initio
calculations. We argue that a combination of linear-scaling methods
and hybrid modelling schemes is required to overcome many of the
difficulties currently faced by conventional schemes, and present our
own contributions towards the development of a robust and reliable
linear-scaling method.
Keywords: computer-simulation, electronic-structure calculations,
ab-initio calculations, embedded-cluster, density-functional theory
INTRODUCTION
The power of ab initio calculations has increased rapidly in
recent years. In this paper we start by reviewing the present
capability of first-principles calculations and point out some of the
limitations of present conventional
techniques. We then focus on linear-scaling
techniques and discuss the extent to which they do represent a
significant step forward in first-principles approaches. We review hybrid modelling schemes which also provide a route
to much larger system-sizes and explain why linear-scaling approaches
are the method of choice for the first-principles element of such
schemes. Finally we briefly describe our work on the development of an
ab initio linear-scaling code based upon a penalty-functional method.
CONVENTIONAL TECHNIQUES
The speed of progress in first-principles calculations over the
fifteen years since Car and Parrinello's introduction of a unified
scheme for density-functional theory (DFT) and molecular dynamics [1] has been breathtaking. The combination of the
universal adoption of density-functional theory [2], significant improvements in numerical methods
[3] and the availability of powerful computers
has allowed first-principles studies of systems containing as many as
a thousand atoms from the whole periodic table, typically yielding
predictions of physical and chemical properties with an accuracy of a
few percent. A critical aspect of calculations on systems of this size
and complexity is the optimisation of atomic positions, since
experiment is not capable of providing this information to sufficient
accuracy. It is even possible to perform dynamical simulations in
which the forces on the atoms are calculated from first principles and
to study quantum effects in the ionic system using Feynman's path
integral formulation [4].
While the power of modern first-principles calculations is impressive,
it is important to appreciate that accessible system-sizes and
time-scales are actually very limited: a system containing a thousand
atoms typically has a volume of only 20 cubic angstroms. Although
structural optimisation is possible in most first-principles codes, in
general the true minimum energy ionic configuration will only be found
if the connectivity of the ionic system is correctly specified. While
dynamical simulations are in principle possible in any
first-principles scheme that calculates ionic forces, in practice only
the total energy pseudopotential technique has successfully used
dynamics to address real scientific questions.
LINEAR-SCALING TECHNIQUES
The computational cost of conventional wavefunction-based total energy
schemes scales at least as the cube of the number of atoms in the
system. Many linear-scaling first-principles schemes have been
proposed over the years [5,6,7,8]. While these schemes will allow calculations
for very large numbers of atoms, there are other technical problems
that may prevent the extraction of useful scientific information from
such calculations. In particular, the length-scales and, more
importantly, time-scales are still extremely limited. A system
containing a million atoms is still only 200 cubic angstroms in
volume, and the relevant time-scales (or equivalently the complexity
of the phase space) increase very rapidly with
system-size. Furthermore, in such a large simulation most of the atoms
will be very close to their bulk configuration and, even using a
linear-scaling algorithm, considerable computational effort will be
expended describing regions of the system where a full
first-principles calculation is not required.
HYBRID MODELLING SCHEMES
A schematic illustration of a hybrid modelling scheme applied to a
block of materials containing two cracks is shown in Fig. 1. Ab initio techniques are applied in
the lightest shaded regions containing the crack tip, where accurate
quantum-mechanical description of the bond-breaking process is
essential. In the surrounding region, empirical interatomic potentials
are used, and in the darkest shaded region continuum modelling is
used. This approach allows each region of a system to be described by
the modelling technique best-suited to the accuracy required in each
region and also addresses the problem of multiple length-scales.
Fig. 1: schematic illustration of a hybrid modelling scheme.
Hybrid modelling schemes incorporating empirical and semi-empirical
potentials in the atomistic region and a finite element description of
the continuum regions have been developed [9].
However, first-principles calculations have been incorporated in such
schemes to a far lesser extent [10]. The
reason is primarily that the standard cluster and supercell based
techniques are not easily included in such hybrid modelling
schemes. Cluster models for extended systems are known to be very
sensitive to the termination of the cluster, and supercells have
serious topological restrictions due to the need to impose periodicity
e.g. it is impossible to model a single dislocation or surface in a
supercell. In contrast, linear-scaling schemes need have no toplogical
restrictions and provide a controlled termination of the quantum
mechanics and thus may be incorporated into hybrid modelling
schemes.
PENALTY-FUNCTIONAL METHOD
The cubic scaling of traditional ab initio methods arises in
general either from the cost of diagonalizing the Hamiltonian or from
orthogonalizing the wavefunctions in iterative schemes. Most
linear-scaling techniques have been developed by reformulating the
quantum-mechanical equations in terms of generalized Wannier functions
or the single-particle density-matrix, instead of the extended
single-particle wavefunctions. Enforcing the localization of the
Wannier functions or truncating the density-matrix beyond a certain
range results in a linear-scaling method. However, although many such
schemes have been proposed [5,6,7,8], progress towards a truly general purpose
scheme has been rather slow. Some of the problems encountered in
implementing linear-scaling schemes are the existence of multiple
minima in the energy surface, and slow convergence towards the
electronic ground-state, even when the correct minimum is located. The
result is a method with a computational cost that is linear in the
number of atoms but with a very large prefactor so that the
linear-scaling calculation only becomes faster than conventional
schemes for extremely large systems.
In order to address the problems mentioned above, we are attempting to
formulate a linear-scaling scheme with the following features:
- Single minimum in the search space.
- Fast, guaranteed convergence to this minimum from any starting
configuration.
- Computational cost and memory as small as possible making the
technique competitive with conventional techniques even for
moderately sized systems.
Ultimately, we wish to have a
linear-scaling code with the same reliability and efficiency as
present wavefunction-based ab initio codes.
As mentioned above, the cubic scaling of conventional techniques
results from the cost of solving Schrödinger's equation for the
single-particle wavefunctions which are the eigenfunctions of the
Kohn-Sham Hamiltonian. Within a self-consistent ab initio
scheme it is also necessary to find the ground-state electronic
density, but since this stage of the calculation already scales
linearly with system-size, we will only address the solution of
Schrödinger's equation for a fixed electronic density (and thus
Hamiltonian). In a conventional scheme this may be achieved by
diagonalizing the Hamiltonian (directly or iteratively) in the
representation of some set of basis functions. Plane-waves and
Gaussians have been popular choices for this purpose. We have chosen
to use truncated spherical-waves [11] for this
purpose for the following reasons:
- Truncated spherical-waves naturally obey the boundary conditions
imposed to truncate the density-matrix (or localize the Wannier
functions).
- The completeness of the basis set may be controlled by means of
a kinetic energy cut-off, as for plane-waves.
- Meaningful comparison may thus be made to the results of
plane-wave calculations.
- Quantities such as the overlap matrix and matrix elements of the
kinetic energy and non-local pseudopotential operators may be
calculated analytically.
Instead of solving for the extended wavefunctions, we choose to solve
for the single-particle density-matrix which contains the same
information but is not subject to the constraint of orthogonality and
is short-ranged so that it may be truncated yielding a linear-scaling
technique. In conventional methods, the lowest energy wavefunctions
are occupied according to Pauli's exclusion principle, but in
density-matrix-based schemes, the occupation numbers of these
wavefunctions become variable parameters and it is necessary to impose
the following constraints:
- Normalization: the occupation numbers must sum to the correct number
of electrons in the system.
- Idempotency: the occupation numbers should be zero or unity, in
order to obey the Pauli exclusion principle.
The first constraint is straightforward to impose, but the second,
namely that the density-matrix and its square are identical, is
non-linear and hence rather problematic. Some methods [6] have used the purifying transformation [12] to impose this constraint implicitly, but it
has been argued [13] that this scheme may
suffer from a rate of convergence that is less than optimal. Kohn [7] first proposed the use of a penalty-functional
for imposing the idempotency constraint, but his use of a non-analytic
functional to obtain a variational principle rules out the application
of current efficient algorithms to this scheme [14]. We have proposed the use of an analytic
penalty-functional [8] to approximately impose
the idempotency constraint. The advantages of this method are an
improved rate of convergence and a single minimum in the search
space. However, the disadvantage of the scheme is that this minimum is
not obtained for the idempotent ground-state density-matrix, but a
nearly-idempotent approximation to it, which approaches the correct
limit as the strength of the penalty-functional is increased. The
solution to this problem has been the development of an analytic
correction to the estimated total energy which gives accurate values
for the true ground-state energy from minimizing density-matrices
which need not be idempotent. The method has so far been applied to
crystalline silicon and yields results in good agreement with
experiment and conventional plane-wave calculations. In Fig. 2 we plot the CPU time per iteration versus the
system-size for the penalty-functional method running on a DEC 500au
computer with 96Mb of memory, and thus demonstrate the linear scaling
of the method.
Fig. 2: scaling of the penalty-functional method with system-size.
CONCLUSIONS
In conclusion, we have shown that although linear-scaling techniques
are a significant step forward in the development of ab initio
quantum-mechanical calculations with respect to the range of
applications to which such methods may be applied, they do not in
themselves overcome all of the limitations of conventional techniques.
We have argued that the combination of ab initio linear-scaling
techniques with hybrid modelling schemes is essential in order to
efficiently and accurately model large-scale processes in real
materials which are of interest to scientists and engineers today. We
have outlined some of our work towards achieving this goal, namely the
development of a penalty-functional-based ab initio
linear-scaling scheme.
REFERENCES
- R. Car and M. Parrinello, "Unified Approach for Molecular
Dynamics and Density-Functional Theory",
Phys. Rev. Lett. 55, 2471 (1985).
- P. Hohenberg and W. Kohn, "Inhomogeneous Electron Gas",
Phys. Rev. 136, 864 (1964);
W. Kohn and L. J. Sham, "Self-Consistent Equations Including
Exchange and Correlation Effects", ibid.
140, 1133 (1965);
for a review see, for instance,
R. O. Jones and O. Gunnarsson, "The density functional
formalism, its applications and prospects",
Rev. Mod. Phys. 61, 689 (1989).
- M. C. Payne, M. P. Teter, D. C. Allan, T. A. Arias and
J. D. Joannopoulos, "Iterative minimization techniques for
ab initio total-energy calculations: molecular dynamics and
conjugate gradients", Rev. Mod. Phys. 64, 1045
(1992).
- See, for instance,
M. P. Allen and D. J. Tildesley, Computer Simulation of
Liquids pp. 270ff. (Oxford University Press, 1987).
- W. Yang, "Direct Calculation of Electron Density in
Density-Functional Theory", Phys. Rev. Lett.
66, 1438 (1991);
W. Yang and T.-S. Lee, "A density-matrix
divide-and-conquer approach for electronic structure
calculations of large molecules", J. Chem. Phys.
103, 5674 (1995);
S. Baroni and P. Giannozzi, "Towards Very Large-Scale
Electronic-Structure Calculations", Europhys. Lett.
17, 547 (1992);
D. A. Drabold and O. F. Sankey, "Maximum Entropy Approach
for Linear Scaling in the Electronic Structure Problem",
Phys. Rev. Lett. 70, 3631 (1993);
L.-W. Wang, "Calculating the density of states and
optical-absorption spectra of large quantum systems by the
plane-wave moments method", Phys. Rev. B 49,
10154 (1994);
O. F. Sankey, D. A. Drabold and A. Gibson, "Projected random
vectors and the recursion method in the electronic-structure
problem", ibid. 50, 1376 (1994);
R. N. Silver and H. Röder, "Calculation of the densities
of states and spectral functions by Chebyshev recursion and
maximum entropy", Phys. Rev. E 56, 4822
(1997);
S.-Y. Qiu, C. Z. Wang, K. M. Ho and C. T. Chan, "Tight-binding
molecular dynamics with linear system-size scaling",
J. Phys.: Condens. Matter 6, 9153 (1994);
A. Canning, G. Galli, F. Mauri, A. de Vita and R. Car,
"O(N) tight-binding molecular dynamics on massively parallel
computers: an orbital decomposition approach",
Comp. Phys. Comm. 94, 89 (1996);
A. P. Horsfield, A. M. Bratkovsky, D. G. Pettifor and M. Aoki,
"Bond-order potential and cluster recursion for the
description of chemical-bonds -- efficient real-space methods
for tight-binding molecular-dynamics", Phys. Rev. B
53, 1656 (1996);
Y. Wang et al., "Order-N Multiple Scattering Approach
to Electronic Structure Calculations", Phys. Rev. Lett.
75, 2867 (1995);
I. A. Abrikosov et al., "Order-N Green's Function
Technique for Local Environment Effects in Alloys",
Phys. Rev. Lett. 76, 4203 (1996);
I. A. Abrikosov et al., "Locally self-consistent
Green's function approach to the electronic structure problem",
Phys. Rev. B 56, 9319 (1997);
E. Smargiassi and P. A. Madden, "Orbital-free kinetic-energy
functionals for first-principles molecular dynamics",
ibid. 49, 5220 (1994);
M. Foley and P. A. Madden, "Further orbital-free kinetic-energy
functionals for ab initio molecular dynamics", ibid.
53, 10589 (1996);
S. Goedecker and L. Colombo, "Efficient Linear Scaling Algorithm
for Tight-Binding Molecular Dynamics", Phys. Rev. Lett.
73, 122 (1994);
S. Goedecker and M. Teter, "Tight-binding electronic-structure
calculations and tight-binding molecular dynamics with localized
orbitals", Phys. Rev. B 51, 9455 (1995);
R. Baer and M. Head-Gordon, "Chebyshev expansion methods for
electronic structure calculations on large molecular systems",
J. Chem. Phys. 107, 10003 (1997);
U. Stephan and D. A. Drabold, "Order-N projection method for
first-principles computations of electronic quantities and
Wannier functions", Phys. Rev. B 57, 6391
(1998);
D. M. C. Nicholson and X.-G. Zhang, "Approximate occupation
functions for density-functional calculations", ibid.
56, 12805 (1997);
F. Gagel, "Finite-Temperature Evaluation of the Fermi Density
Operator", J. Comp. Phys. 139, 399 (1998);
R. N. Silver, H. Röder, A. F. Voter and J. D. Kress,
"Kernel Polynomial Approximations for Densities of States and
Spectral Functions", ibid. 124, 115 (1996);
A. F. Voter, J. D. Kress and R. N. Silver, "Linear-scaling
tight binding from a truncated approach", Phys. Rev. B
53, 12733 (1996);
G. Galli and M. Parrinello, "Large Scale Electronic Structure
Calculations", Phys. Rev. Lett. 69, 3547
(1992);
F. Mauri, G. Galli and R. Car, "Orbital formulation for
electronic-structure calculations with linear system-size
scaling", Phys. Rev. B 47, 9973 (1993);
F. Mauri and G. Galli, "Electronic-structure calculations and
molecular-dynamics simulations with linear system-size scaling",
ibid. 50, 4316 (1994);
P. Ordejón, D. A. Drabold, M. P. Grumbach and
R. M. Martin, "Unconstrained minimization approach for
electronic computations that scales linearly with system size",
ibid. 48, 14646 (1993);
P. Ordejón, D. A. Drabold, R. M. Martin and
M. P. Grumbach, "Linear system-size scaling methods for
electronic-structure calculations", ibid. 51,
1456 (1995);
J. Kim, F. Mauri and G. Galli, "Total-energy global
optimizations using nonorthogonal localized orbitals",
ibid. 52, 1640 (1995);
K. C. Pandey, A. R. Williams and J. F. Janak, "Localized
orbital theory of electronic structure: A simple application",
ibid. 52, 14415 (1995);
P. Ordejón, E. Artacho and J. M. Soler,
"Self-consistent order-{$N$} density-functional calculations
for very large systems", ibid. 53, 10441
(1996);
for reviews see, for instance,
D. R. Bowler et al., "A comparison of linear scaling
tight-binding methods", Modelling Simul. Mater. Sci.
Eng. 5, 199 (1997);
G. Galli, "Linear scaling methods for electronic structure
calculations and quantum molecular dynamics simulations",
Current Opinion in Solid State and Materials Science
1, 864 (1996).
- X.-P. Li, R. W. Nunes and D. Vanderbilt, "Density-matrix
electronic-structure method with linear system-size scaling",
Phys. Rev. B 47, 10891 (1993);
M. S. Daw, "Model for energetics of solids based on the
density matrix", ibid. 47, 10895 (1993);
E. B. Stechel, A. R. Williams and P. J. Feibelman, "N-scaling
algorithm for density-functional calculations of metals and
insulators", ibid. 49, 10088 (1994);
W. Hierse and E. B. Stechel, "Order-N methods in
self-consistent density-functional calculations", ibid.
50, 17811 (1994);
E. Hernández and M. J. Gillan, "Self-consistent
first-principles technique with linear scaling", ibid.
51, 10157 (1995);
E. Hernández, M. J. Gillan and C. M. Goringe,
"Linear-scaling density-functional-theory technique:
The density-matrix approach", ibid. 53,
7147 (1996).
- W. Kohn, "Density Functional and Density Matrix Method Scaling
Linearly with the Number of Atoms", Phys. Rev. Lett.
76, 3168 (1996).
- P. D. Haynes and M. C. Payne, "Corrected penalty-functional
method for linear-scaling calculations within density-functional
theory", Phys. Rev. B 59, 12173 (1999).
- V. B. Shenoy, R. Miller, E. B. Tadmor, R. Phillips and
M. Ortiz, "Quasicontinuum models of interfacial structure
and deformation", Phys. Rev. Lett. 80,
742 (1998);>br>
R. Miller, E. B. Tadmor, R. Phillips and M. Ortiz,
"Quasicontinuum simulation of fracture at the atomic scale",
Modelling Simul. Mater. Sci. Eng. 6, 607
(1998);
V. B. Shenoy, R. Miller, E. B. Tadmor, D. Rodney, R. Phillips
and M. Ortiz, "An adaptive finite element approach to
atomic-scale mechanics - the quasicontinuum method",
J. Mech. Phys. Solids 47, 611 (1999);
E. B. Tadmor, G. S. Smith, N. Bernstein and E. Kaxiras,
"Mixed finite element and atomistic formulation for
complex crystals", Phys. Rev. B 59,
235 (1999);
F. F. Abraham, J. Q. Broughton, N. Bernstein and
E. Kaxiras, "Spanning the continuum to quantum length
scale in a dynamic simulation of brittle fracture",
Europhys. Lett. 44, 783 (1998);
R. E. Rudd and J. Q. Broughton, "Coarse-grained
molecular dynamics and the atomic limit of finite
elements", Phys. Rev. B 58, 5893 (1998).
- U. Eichler, C. M. Kolmel and J. Sauer, "Combining ab
initio techniques with analytical potential functions for
structure predictions of large systems: Method and application
to crystalline silica polymorphs", J. Comput. Chem.
18, 463 (1997);
U. Eichler, M. Brandle and J. Sauer, "Predicting absolute
and site specific acidities for zeolite catalysts by a
combined quantum mechanics interatomic potential function
approach", J. Phys. Chem. B 101, 10035
(1997);
M. Brandle, J. Sauer, R. Dovesi and N. M. Harrison,
"Comparison of a combined quantum mechanics/interatomic
potential function approach with its periodic
quantum-mechanical limit: Proton siting and ammonia
adsorption in zeolite chabazite", J. Chem. Phys.
109, 10379 (1998);
M. Brandle and J. Sauer, "Acidity differences between
inorganic solids induced by their framework structure.
A combined quantum mechanics molecular mechanics ab initio
study on zeolites", J. Am. Chem. Soc. 120,
1556 (1998);
M. Svensson, S. Humbel, R. D. J. Froese, T. Matsubara,
S. Sieber and K. Morokuma, "ONIOM: A multilayered
integrated MO+MM method for geometry optimizations and
single point energy predictions. A test for
Diels-Alder reactions and Pt(P(t-Bu)(3))(2)+H-2
oxidative addition", J. Phys. Chem. 100,
19357 (1996);
T. Matsubara, S. Sieber and K. Morokuma, "A test of the
new 'integrated M0+MM' (IMOMM) method for the
conformational energy of ethane and n-butane",
Int. J. Quantum Chem. 60, 1101 (1996);
R. D. J. Froese, D. G. Musaev and K. Morokuma,
"Theoretical study of substituent effects in the
diimine-M(II) catalyzed ethylene polymerization reaction
using the IMOMM method", J. Am. Chem. Soc.
120, 1581 (1998).
- P. D. Haynes and M. C. Payne, "Localised spherical-wave basis
set for O(N) total-energy pseudopotential calculations",
Comp. Phys. Comm. 102, 17 (1997).
- R. McWeeny, "Some Recent Advances in Density Matrix Theory",
Rev. Mod. Phys. 32, 335 (1960).
- S. Goedecker, "Linear Scaling Electronic Strcuture Methods",
to appear in Rev. Mod. Phys. (1999).
- P. D. Haynes and M. C. Payne, "Failure of density-matrix
minimization methods for linear-scaling density-functional
theory using the Kohn penalty functional",
Solid State Communications 108, 737 (1998).
ACKNOWLEDGMENTS
This work was supported by grant GR/L55346 from the Engineering and Physical
Sciences Research Council, UK.
Paper abstract