poissonsolverswithinterpolatingscaling functions ...bigdft.org/images/e/ef/berlin-lg.pdfpoisson...
TRANSCRIPT
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Numerical Methods in Density Functional Theory
TU UNIVERSITY - BERLIN
Poisson Solvers with Interpolating ScalingFunctions
for Electronic Structure Calculations
Luigi Genovese
L_Sim - CEA Grenoble
23 July 2008
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Outline
1 Self-Consistent potential in DFTPoisson Solver
2 Poisson Solver for Free BCCalculation of the Poisson KernelPerformances
3 A Poisson solver for Surfaces BCPerformances
4 Summary and outlook
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Self-Consistent potential in DFT
In the DFT calculation in the Kohn-Sham formalism we have:
Find a set of orthonormal orbitals Ψi(r) that minimizes:
E =−12
N/2
∑i=1
ZΨ∗i (r)∇
2Ψi(r)dr +12
Zρ(r)VH(r)dr
+ Exc[ρ(r)] +Z
Vext(r)ρ(r)dr
where
ρ(r) = 2N/2
∑i=1
Ψ∗i (r)Ψi(r)
∇2VH(r) =−4πρ(r)
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Operations performed in BigDFT code
Different operatorsDaubechies:
Kinetic energy
Scalar products, Vnl
Interpolating SF:
Vloc (gaussians)
XC (ρ needed)
Hartree : ∇2VH = ρ
Numerical operationsConvolutions withshort filters
Scalar products
FFT (Poisson Solver)
Interpolating Daubechies
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
The Poisson Solver in electronic structure calculation
During the minimization procedure we need to perform
Poisson’s equationCalculation of the self-consistent potential:
∇2VH(r) =−4πρ(r)
Such equation should be solved at each minimisationiteration. Need of having an efficient and accurate formalism.
Plane waves approachThe most common approach. Uses the fourier components
f (x ,y ,z) = ∑px ,py ,pz
e−2πi
(pxLx
x+pyLy
y+ pzLz
z)fpx ,py ,pz
The Poisson equation is algebraic in the Fourier coefficients
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Poisson Solver with plane waves treatment
The Laplacian is diagonal in Plane waves representation
Immediate solution
Vpx ,py ,pz =1π
1(pxLx
)2+(
pyLy
)2+(
pzLz
)2 ρpx ,py ,pz ,
CharacteristicsSimple and fast, easy to parallelize (FFT)
Automatically implement Periodic BC on a finite volume
Do not fix the value of V0,0,0
⇒ May result in problems for systems with other BC
How to solve this equation for other BC?
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Problems with Plane Wave expansion
How to remove long-distance interactions?Model their effect and subtract it
Good for integrated quantities (e.g. total energy) but stillinefficient for the local values V (r)The size L of the effective system must be enlarged wrtthe original one
Modify the kernel operator K = Kshort + Klong
Does not implement well short-distance behaviour,resulting in errorsThese errors decrease when the size of the system islarge
In both casesWe must consider a size that is larger than the size of theoriginal system
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Isolated BC: the Green function treatment
Consider the Poisson equation for isolated BC. In this casethe solution is given by
Green’s function for Free BC (kernel)
∇2 1
r=−4πδ(r) =⇒ V (r) =
Zdr′
ρ(r′)|r− r′|
An unambiguous solutionSuch prescription is unique and is compatible with Free BC.How to implement it?Plane-wave based approaches:
Truncated kernel, numerically or analytically
Screening Functions
Approximate treatment, large box required.
Need of an accurate and efficient algorithm
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Poisson Solver with Interpolating Scaling Functions
A convenient basis for an electrostatic problem
Interpolating Scaling FunctionsA set of localised functions centered on the nodes of auniform mesh
ϕj(x) = ϕ0(x− j)
Undergo multiscale relation
ϕ(x) =m
∑j=−m
hj︸︷︷︸filters
ϕ(2x− j)
-1.5
-1
-0.5
0
0.5
1
1.5
-6 -4 -2 0 2 4 6
LEAST ASYMMETRIC DAUBECHIES-16
scaling functionwavelet
4 The expansion coefficients are the real space valuesρjx ,jy ,jz = ρ(hx jx ,hy jy ,hz jz)
4 Represents exactly an order m polynomial
4 The first m discrete and continuous moments coincide
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
A finite three-dimensional convolution
The expression of the potential in this basis is thus intuitive:
V (i) = ∑j
Kij ρj
Where the central object is the
Poisson Kernel in the ISF basis
Kij = Ki−j , Ki =Z
K (|r|)ϕi(r)dr , K (r) =1r
Values of the potential are obtained via a convolution
V (i) = ∑j
Ki−j ρj
It can be treated via a zero-padded FFT algorithm
4 Exact, easy to parallelize (FFT)
4 For a box of N3 points, it reduces the scaling fromO(N6) to O(N3 logN)
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Characteristics of the approach
This approach is:
Explicit, guarantees the good BC
Real space-based, immediate interpretation of theexpansion coefficients
Can be combined with other real-space treatments ofthe density (e.g. XC) combined with ABINIT XC routines
Can be used independently from the BigDFT code
Preserves the first m (multipole) moments of theelectrostatic potential
Requires only the evaluation of the kernel
We need to evaluate N3 integrals Ki =R
K (|r|)ϕi(r)drwhere ϕi(r) = ϕix (x)ϕiy (y)ϕiz (z)
ISF basis is a tensor product
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Gaussian tensor product decomposition
It can be shown that (Beylkin et al.)
Approximation with gaussians
1r'∑
kωk e−pk r2
with k = 1, · · · ,89, pk , ωk suitably chosen
Accuracy of 10−8 for r ∈ (10−9,1)
We can rescale for R ∈ (0,L).
The computational cost is reduced N3→ 89×N
Kj =89
∑k=1
ωk Kjx (pk )Kjy (pk )Kjz (pk )
Kj(p) =Z
ϕ0(x)e−p(x−j)2dx
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Other properties of the scaling functions
The computational cost is reduced N3 → 89×N.
MoreoverThe scaling property of the interpolets
ϕ0(x/2) = ∑j
hjϕj(x) ,
Implies similar condition of the one-dimensional function
Kj(4p) =12 ∑
jhjK2i−j(p) .
Thus we can evaluate the integrals for low p (not too sharpgaussians), then rescaling.
ISF properties allows us to gain in accuracy
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Characteristics
Very fast with moderate memory occupation:
Elapsed Time for a 1283 grid on a Cray XT3
proc 1 2 4 8 16 32 64
s .92 .55 .27 .16 .11 .08 .09
More precise than other existing Free BC Poisson Solvers
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
1e-04
0.001
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Max
Err
or
Grid step
HockneyTuckerman
8th14th20th30th40th50th60th
100th
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Characteristics of the Solver
In summary, we have developed a technique
Free boundary conditions
Very high accuracy
Good computational performance, easy to parallelize
Can be used also in other contexts and/or combinedwith other treatment (e.g. XC)
Coupled with ABINIT XC routines
In BigDFT code, for big systems it represents a small amount(typical values are less than 2-3% for big systems) of theoverall computation
L. Genovese, T. Deutsch, A. Neelov, S. Goedecker, G. Beylkin
J. Chem. Phys. 125, 074105 (2006)
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
The Surfaces boundary conditions
The same formalism can be applied to other BC
Surfaces BCA domain isolated in one direction (say y ) and periodic in xand z, with periods Lx and Lz . A function f which lives insuch a domain can be expanded as
f (x ,y ,z) = ∑px ,pz
e−2πi( pxLx
x+ pzLz
z)fpx ,pz (y)
without any loss of generality.
Mixed representationFor such functions the Poisson’s equation become(
∂2y −µ2
px ,pz
)Vpx ,pz (y) = ρpx ,pz (y) ,
where µ2px ,pz
= 4π2(px/Lx )2 + (pz/Lz)2.
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
A Green’s function formalism
The Green’s function for the one-dimensional Helmoltzequation can be used(
∂2y −µ2)G(µ;y) = δ(y) ;
G(µ;y) =
{− 1
2µ e−µ|y | µ > 012 |y | µ = 0
,
the components of the potential can be carried out:
A Green’s function for each Fourier component
Vpx ,pz (y) =Z
dy ′G(µpx pz ;y− y ′)ρpx ,pz (y ′) .
We can use Interpolating Scaling Functions for the Isolateddirection.
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Surfaces BC with Mixed representation
By defining the one-dimensional kernels
K (µpx ,pz ; j) =Z
G(µpx pz ;y)ϕj(y)dy
We can have a treatment similar to the Free BC case
1-dim convolution for each reciprocal space component
Vpx ,pz (i) = ∑j
K (µpx ,pz ; i− j)ρpx ,pz (j)
From mixed representation to full real spaceThe calculation can be performed with a (semi-)zero-paddedFFT algorithm. The I/O are function of the real domain.
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Speed-up the calculation
The calculation of the kernel can be improved
SpeedThe analytic form of the Green functions allows for recursionrelations−→ improvement in speed of a factor of (Nz + 1)/3
AccuracyThe scaling relation will again help in the kernel
K (2µ; j) =12 ∑
jhjK (2µ,2i− j)
A “Free-Lunch” caseWe have the same advantages of the Free BC treatment withan explicit formalism conceived for surfaces BC
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Poisson Solver for surface boundary conditions
Elapsed Time on a Cray XT3, 1283 grid
#proc 1 2 4 8 16 32 64sec .43 .26 .16 .10 .07 .05 .04
More precise than other treatments
1e-16
1e-14
1e-12
1e-10
1e-08
1e-06
1e-04
0.01
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Abs
olut
e re
lativ
e er
ror
grid spacing
Mortensen8th
16th24th40th
h8 curve
L. Genovese, T. Deutsch, S. Goedecker
J. Chem. Phys. 127, 054704 (2007)
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
A Poisson solver for surface problems
Like the Free BC caseWe developed a technique
4 Accurate and fast, easy to parallelize
4 Can be applied both in real or reciprocal space codes
4 Explicit treatment No supercell or screening functions
4 More precise than other existing approaches
4 Allows comparisons betweeen different backgrounds(charged systems)
Example of the plane capacitor:Periodic
x
y
V
x
Hockney
x
y
V
x
Our approach
x
y
V
x
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Flexibility: the Poisson Solver in BigDFT
Flexible Boundary ConditionsDirect approach(non-iterative)
Explicit, high precision
Clean and precise treatment of the electrostaticSeveral applications:
Impurities, dopants
Surface properties (adsorbtion, cathalysis)
Charged defects, in bulk or in surfaces
These Solvers can be used independently from BigDFT codeAlready integrated in other DFT codes (CP2K, ABINIT, OCTOPUS)
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese
PoissonSolver with
ISF
DFT HartreepotentialPoisson Solver
Free BCPoisson Kernel
Performances
Surfaces BCPerformances
Outlook
Summary and outlook
Interpolating SF can be use to solve the Poisson’s equation
∇2VH =−4πρ
with a Green’s function treatment, in different environements:
Isolated BC
VH(~j) =R
d~x ρ(~x)
|~x−~j|
Surfaces BC
Vpx ,pz (y) =Z
dy ′G(µpx pz ;y− y ′)ρpx ,pz (y ′)
From ρ(~j) on a uniform grid
Explicit treatment of the boundary conditions
Can be generalised to other environmentsWires BC
Spherical heterostructures (multipole expansion)
Laboratoire de Simulation Atomistique http://inac.cea.fr/L_Sim Luigi Genovese