Title: Linear scaling electronic structure methods in chemistry and physics Computing in Science
1Linear scaling electronic structure methods in
chemistry and physicsComputing in Science
Engineering, vol. 5, issue 4, 2003 (1)Stefan
Goedecker, Gustavo E. ScuseriaLinear Scaling
Density Functional Calculations with Gaussian
OrbitalsJournal of Physical Chemistry A, vol.
103, no. 25, 1999 (2) Gustavo E. Scuseria
Linear Scaling Electronic Structure
MethodsReviews of Modern Physics, Vol 71, No. 4,
July 1999 (3) Stefan Goedecker
Linear scaling electronic structure methods
DFT Journal Club METU Physics Department
Nazim
Dugan dugannaz_at_GMail.com
2Stefan Goedecker is a professor of computational
physics at the University of Basel. His research
interests include linear scaling algorithms for
electronic structure calculations and other
atomistic simulation methods. He received a PhD
from the Swiss Federal Institute of Technology in
Lausanne. Contact him at Stefan.Goedecker_at_unibas.c
h Gustavo E. Scuseria is the Robert A. Welch
Professor of Chemistry at Rice University. His
research interests include the development of
low-order scaling electronic structure methods
and their application to molecules and solids.
His undergraduate and PhD degrees are in physics
from the University of Buenos Aires. He is a
member of the American Chemical Society and a
Fellow of the American Physical Society, the
American Association for the Advancement of
Science, and the Guggenheim Foundation. Contact
him at guscus_at_rice.edu
DFT Journal Club
METU Physics
Department
3 Outline
Exposition of the problem Strategies for linear
scaling Benchmark Calculations
DFT Journal Club
METU Physics
Department
4Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Exposition of the problem
DFT Journal Club
METU Physics
Department
5Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Complexity of algorithms TCPU const N
k complexity O( N k ) Linear scaling means
O(N)
Most physical quantities are extensive - that is,
they grow linearly with system size. We might
therefore expect that the computational effort
will grow linearly with system size as well. An
even slower increase in computing time is
certainly not possible unless we ignore the basic
physics of the electronic system. (1)
DFT Journal Club
METU Physics
Department
6Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Two body interactions in many particle systems
(electrons, atoms, planets)
combinations of
particles Computation time
(quadratic) In DFT even though the
complexity of finding ground state energy has
linear scaling
DFT Journal Club
METU Physics
Department
7Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Why do we make approximations One electron
system ?
Matrix (2D) Evaluation of elements N 2 Matrix
diagonalization N 3 2 electron system
? 4 th
rank Tensor In General Dimensionality of
Hamiltonian 2Ne Evaluation of elements Ne
2Ne Diagonalization method not known
DFT Journal Club
METU Physics
Department
8Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Scaling of other electronic structure
methods Full configuration interaction
exponential Coupled-cluster
O( N 6 ) Quantum Monte Carlo O(
N 3 )
DFT Journal Club
METU Physics
Department
9Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Origin of cubic scaling in DFT Coulomb and XC
potential evaluation N 2 Matrix
diagonalization to solve Kohn-Sham equation
N 3 Orthonormalization of Kohn-Sham
orbitals N(N-1)/2 orbital pairs cost of each
integral is proportional to N NN(N-1)/2
N 3 TCPU cc N 2 cx N 2 cm N 3 co N 3
DFT Journal Club
METU Physics
Department
10Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Applicability of cubic scaling methods 10
electrons 1 seconds 100 electrons
16 minutes 1000 electeons 11.5 days
10000 electrons 32 years With linear
scaling methods up to 25000 atoms on 24 nodes of
Earth Simulator
DFT Journal Club
METU Physics
Department
11Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Crossover point Cubic scaling TCPU c3 N
3 Linear scaling TCPU c1 N
José M. Soler Universidad Autónoma de Madrid
DFT Journal Club
METU Physics
Department
12Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Strategies for linear scaling
DFT Journal Club
METU Physics
Department
13Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Well known low complexity algorithms Fast
Fourier Transform (FFT) FT O(N 2) ? FFT O(N
log(N) ) Cooley and Tukey algorithm published in
1965 Quick Sort Bubble Sort O(N 2) ? QS
O(N log(N) ) Divide and Conquer !!!
DFT Journal Club
METU Physics
Department
14Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Locality (Nearsightedness) in QM makes it
possible to apply CUTOFF
Because the extended eigenorbitals diagonalizing
the independent particle Hamiltonian, usually
referred to as canonical orbitals, do not reflect
this locality principle, they are not suitable as
the basic quantities in O(N) calculations. Linear
scaling also rules out the use of basis functions
extending over the whole computational volume,
such as plane waves. (1)
Blip Functions E Hernández, MJ Gillan, CM
Goringe, Phys Rev B 55, 13485(1997)
W. Kohn, Phys. Rev. Lett. 76, 3168(1996)
DFT Journal Club
METU Physics
Department
15Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Two-body problem
Neighborhood Keep neighbor information for each
particle Calculate interactions with neigbors
only ( O(N) ) Update neighbors ( O(N 2) ) Mesh
technique Divide space into subspaces Calculate
interactions only with particles of owner and
neighbor subspaces ( O(N) ) Check subspace of
particles ( O(N) )
DFT Journal Club
METU Physics
Department
16Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Tree-code and the Fast Multipole Method (FMM)
- Search starting from the root of the tree - If
the particle is far enough to a goup of
particles, treat the group as a single
particle at the center of mass (monopole
approximation). - If it is not, go one step
further in the tree and check again. - Use FMM
for Far Field only
Tancred Lindholm, N-body Algorithms, 1999
DFT Journal Club
METU Physics
Department
17Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Density Matrix approach in DFT
Traditional sequence
DM sequence
Eigenvectors of the effective Hamiltonian which
are obtained through the diagonalization step are
in practice only needed to construct the density
matrix. However, this is not the only way of
obtaining the density matrix, and one can instead
adopt direct search methods like CGDMS. (2)
W. Kohn, Phys. Rev. Lett. 76, 3168(1996)
DFT Journal Club
METU Physics
Department
18Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Density Matrix Minimization
- Express Total Energy in terms of density
matrix - Minimize wrt density matrix
E Hernandez, MJ Gillan, CM Goringe, Phys. Rev. B
55, 7147(1996)
DFT Journal Club
METU Physics
Department
19Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Sparse Density Matrix
In this regard, the size of the HOMO-LUMO gap is
connected to localization and, consequently,
sparsity in the system. It is well known that
systems showing metallic character (i.e., small
HOMO-LUMO gap) yield denser Hamiltonians and
density matrices than insulators. (2) Insulaters
exponential decay Metals at finite temperature
exponential decay Metals at zero temperature
algebraic decay
Progressive convergence Dynamical adjustment of
treshold
DFT Journal Club
METU Physics
Department
20Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Divide and conquer method
The idea is to calculate certain regions of the
density matrix by considering subvolumes and then
to generate the full density matrix by adding up
these parts with the appropriate weights. (3)
W. Yang, Phys. Rev. Lett. 66, 1438(1991)
DFT Journal Club
METU Physics
Department
21Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
The Chebyshev Fermi operator expansion
The rational Fermi operator expansion
The desired linear scaling can be obtained by
introducing a localization region for each
column, outside of which the elements are
negligibly small. For the kth column, this
localization region will be centered on the kth
basis function. (3)
DFT Journal Club
METU Physics
Department
22Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Benchmark Calculations
DFT Journal Club
METU Physics
Department
23Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
System Water clusters in two and three
dimensions Size Up to 1152 molecules in 2D,
1000 moelcules in 3D Functionals LSDA Becke
88 exchange, Lee-Yang-Parr correlation (BLYP)
GGA Perdew-Burke-Erznehof (PBE) Software
Gaussian 99 (Development version) Basis sets
3-21G and 6-31G (up to 15000 basis
functions) Hardware SGI Origin-2000 195 MHz 4
MB cache Up to 10 GB disk space, 180
megawords RAM
DFT Journal Club
METU Physics
Department
24Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Energies (Hartrees) and CPU Times (min) per SCF
Cycle Obtained by Conjugate Gradient Density
Matrix Search (CGDMS) and Diagonalization in a
Series of Two-Dimensional Water Cluster
Calculations at the LSDA/3-21G Level of Theory
DFT Journal Club
METU Physics
Department
25Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
2D 3-21-G
2D 6-31-G
CPU Time (min)
3D 3-21-G
3D 6-31-G
number of basis functions
DFT Journal Club
METU Physics
Department
26Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
RNA fragment 1026 atoms, 6767 basis
function LSDA/3-21G
. . . all three DFT steps for the RNA piece are
computationally more expensive than those for the
water clusters, especially CGDMS which is about 5
more costly than the 3D cluster case. These
results simply indicate that typical biomolecules
may have density matrices and Hamiltonians which
are denser than 3D water clusters, but they are
still amenable to efficient treatment by the
methods and algorithms discussed in this work.
DFT Journal Club
METU Physics
Department
27Exposition of the problem
Strategies for linear scaling
Benchmark Calculations
Concluding Remarks Significant factors
dimensionality, HOMO-LUMO gap, desired
accuracy, basis set The more compact a molecular
system is, the less sparse all matrices are, and
the more demanding the O(N) DFT calculation will
turn out to be. (larger prefactor) . . . if DFT
fails to deliver a next generation of
significantly more accurate functionals, it would
then be reasonable to assume that much work will
be devoted to developing fast (i.e., small
prefactor) O(N) wave function methods . .
. Linear scaling DFT codes SIESTA, CONQUEST,
ONETEP, GAUSSIAN
DFT Journal Club
METU Physics
Department
28References
Practical Methods for Ab Initio Calculations on
Thousands of Atoms D.R. Bowler, I.J. Bush, M.J.
Gillan International Journal of Quantum
Chemistry, Vol. 77, 831842 (2000) Large-Scale
Electronic Structure Calculations Using Linear
Scaling Methods G. Galli Physica Status. Solidi
(b) 217, 231 (2000) Recent progress in linear
scaling ab initio electronic structure
techniques D.R. Bowler, T. Miyazaki, M.J.
Gillan Journal Of Physics Condensed Matter, 14
(2002) 27812798 ONETEP linear-scaling
density-functional theory with plane-waves P.D.
Haynes, A.A. Mostofi, C.K. Skylaris, M.C.
Payne Journal of Physics Conference Series 26
(2006) 143148
DFT Journal Club
METU Physics
Department
29We may regard the present state of the universe
as the effect of its past and the cause of its
future. An intellect which at any given moment
knew all of the forces that animate nature and
the mutual positions of the beings that compose
it, if this intellect were vast enough to submit
the data to analysis, could condense into a
single formula the movement of the greatest
bodies of the universe and that of the lightest
atom for such an intellect nothing could be
uncertain and the future just like the past would
be present before its eyes. Marquis Pierre
Simon de Laplace, 1814
DFT Journal Club
METU Physics
Department