Particle simulation based on the nonequispaced FFT
Particle simulation based on the nonequispaced FFT
Franziska Nestler Technische Universität Chemnitz
Faculty of Mathematics joint work with M. Pippig
Strobl16
Time-Frequency Analysis and Related Topics
Strobl, 06/2016
Particle simulation based on the nonequispaced FFT
1
Introduction: NFFT and fast summation
2
Coulomb interactions and fast Ewald summation
3
Numerical results
4
Extension to systems with dipoles
5
Summary
Introduction: NFFT and fast summation
The 3d-NFFT (FFT for nonequispaced data)
Notation
•
define the torus
T:=
R/
Z' [−
1/
2,
1/
2)
•
for M ∈ 2
Nset I
M:= {−
M/
2, . . . ,
M/
2− 1}
3⊂
Z3NFFT:
f(x
j) :=
Xk∈IM
f ˆ
ke
−2πik·xj xj∈
T3, j = 1, . . . , N
FFT: f(j) :=
Xk∈IM
f ˆ
ke
−2πik·j/M j∈ I
M, N := |I
M| = M
3adjoint NFFT:
h(k) :=
N
X
j=1
f
je
2πik·xj k∈ I
MComplexity: O(|I
M| log |I
M| + N )
[Dutt, Rokhlin 1993] [Beylkin 1995] [Potts, Steidl, Tasche 2001] [Greengard, Lee 2004]
Introduction: NFFT and fast summation
The 3d-NFFT (FFT for nonequispaced data)
Notation
•
define the torus
T:=
R/
Z' [−
1/
2,
1/
2)
•
for M ∈ 2
Nset I
M:= {−
M/
2, . . . ,
M/
2− 1}
3⊂
Z3NFFT:
f(x
j) :=
Xk∈IM
f ˆ
ke
−2πik·xj xj∈
T3, j = 1, . . . , N
FFT: f(j) :=
Xk∈IM
f ˆ
ke
−2πik·j/M j∈ I
M, N := |I
M| = M
3adjoint NFFT:
h(k) :=
N
X
j=1
f
je
2πik·xj k∈ I
MComplexity: O(|I
M| log |I
M| + N )
[Dutt, Rokhlin 1993] [Beylkin 1995]
[Potts, Steidl, Tasche 2001] [Greengard, Lee 2004]
Introduction: NFFT and fast summation
The 3d-NFFT (FFT for nonequispaced data)
Notation
•
define the torus
T:=
R/
Z' [−
1/
2,
1/
2)
•
for M ∈ 2
Nset I
M:= {−
M/
2, . . . ,
M/
2− 1}
3⊂
Z3NFFT:
f(x
j) :=
Xk∈IM
f ˆ
ke
−2πik·xj xj∈
T3, j = 1, . . . , N
FFT: f(j) :=
Xk∈IM
f ˆ
ke
−2πik·j/M j∈ I
M, N := |I
M| = M
3adjoint NFFT:
h(k) :=
N
X
j=1
f
je
2πik·xj k∈ I
MComplexity: O(|I
M| log |I
M| + N )
[Dutt, Rokhlin 1993] [Beylkin 1995]
[Potts, Steidl, Tasche 2001] [Greengard, Lee 2004]
Introduction: NFFT and fast summation
NFFT workflow
1
Deconvolution in Fourier space: O(|I
M|)
ˆ g
k:=
( fˆ
k
ck( ˜ϕ)
:
k∈ I
M,
0 : else.
2
Inverse FFT: O(|I
σM| log |I
σM|) g
l:= 1
|I
σM|
Xk∈IσM
ˆ
g
ke
−2πik·l/(σM),
l∈ I
σM.
3
Convolution in spatial domain: O(N )
Approximate f by a sum of translates of a 1-periodic
window-function.f(x
j) ≈
Xl∈IσM
g
lϕ(x ˜
j−
l/
σM), j = 1, . . . , N.
Oversampling factor σ ≥ 1: we can use more coefficients than given Fourier coefficients.
Introduction: NFFT and fast summation
NFFT workflow
1
Deconvolution in Fourier space: O(|I
M|)
ˆ g
k:=
( fˆ
k
ck( ˜ϕ)
:
k∈ I
M,
0 : else.
2
Inverse FFT: O(|I
σM| log |I
σM|) g
l:= 1
|I
σM|
Xk∈IσM
ˆ
g
ke
−2πik·l/(σM),
l∈ I
σM.
3
Convolution in spatial domain: O(N )
Approximate f by a sum of translates of a 1-periodic
window-function.f(x
j) ≈
Xl∈IσM
g
lϕ(x ˜
j−
l/
σM), j = 1, . . . , N.
Oversampling factor σ ≥ 1: we can use more coefficients than given Fourier coefficients.
Introduction: NFFT and fast summation
NFFT workflow
1
Deconvolution in Fourier space: O(|I
M|)
ˆ g
k:=
( fˆ
k
ck( ˜ϕ)
:
k∈ I
M,
0 : else.
2
Inverse FFT: O(|I
σM| log |I
σM|) g
l:= 1
|I
σM|
Xk∈IσM
ˆ
g
ke
−2πik·l/(σM),
l∈ I
σM.
3
Convolution in spatial domain: O(N )
Approximate f by a sum of translates of a 1-periodic
window-function.f(x
j) ≈
Xl∈IσM
g
lϕ(x ˜
j−
l/
σM), j = 1, . . . , N.
Oversampling factor σ ≥ 1: we can use more coefficients than given Fourier coefficients.
Introduction: NFFT and fast summation
The 3d-NFFT (FFT for nonequispaced data)
Further implemented variants:
•
Gradient NFFT: Approximate
∇f(x
j) = −
Xk∈IM
2πik f ˆ
ke
−2πik·xj∀ j = 1, . . . , N.
•
Hessian NFFT: Approximate Hf(x
j) = −
Xk∈IM
4π
2kk>f ˆ
ke
−2πik·xj∀ j = 1, . . . , N.
•
Adjoint gradient NFFT (f
j∈
C3): Approximate
N
X
j=1
f>j
∇
xe
2πik·x x=xj
=
N
X
j=1
2πif
>jke
2πik·xj∀
k∈ I
M.
Complexity: O(|I
M| log |I
M| + N )
X [M. Pippig, PNFFT library, https://github.com/mpip/pfft]Introduction: NFFT and fast summation
ik and analytic differentiation
...for the gradient NFFT:
1
Differentiation in Fourier space (ik differentiation):
∇f(x
j) = −
Xk∈IM
2πik f ˆ
ke
−2πik·xjCompute three 3D-FFTs in the second step (one in each dimension).
2
Analytic differentiation: apply gradient to the window function. Modify the last step of the NFFT as follows:
∇f(x
j) ≈
Xl∈IσM
g
l∇ ϕ(x ˜
j−
l/
σM)
→ only one 3D-FFT has to be computed in the second step.
Introduction: NFFT and fast summation
ik and analytic differentiation
...for the gradient NFFT:
1
Differentiation in Fourier space (ik differentiation):
∇f(x
j) = −
Xk∈IM
2πik f ˆ
ke
−2πik·xjCompute three 3D-FFTs in the second step (one in each dimension).
2
Analytic differentiation: apply gradient to the window function. Modify the last step of the NFFT as follows:
∇f(x
j) ≈
Xl∈IσM
g
l∇ ϕ(x ˜
j−
l/
σM)
→ only one 3D-FFT has to be computed in the second step.
Introduction: NFFT and fast summation
Fast summation based on NFFT in 1d
[Potts, Steidl 2003]Computef(xj) :=
N
X
i=1
ciK(xi−xj)forxj∈[−L/2,L/2], j= 1, . . . , N.
0
−L L
K(x)
1
• Extend the interval at both ends.
• Construct asmooth transition(match derivatives up to some orderp∈N).
→Two-point-Taylor interpolation
• Theregularized kernel functionis smooth and periodic.
Introduction: NFFT and fast summation
Fast summation based on NFFT in 1d
[Potts, Steidl 2003]Computef(xj) :=
N
X
i=1
ciK(xi−xj)forxj∈[−L/2,L/2], j= 1, . . . , N.
0
−L L
−(L+δ) L+δ
∂j
∂xjKB(L) =∂x∂jjK(L)
K(x)
KB(x) KB(x)
1
• Extend the interval at both ends.
• Construct asmooth transition(match derivatives up to some orderp∈N).
→Two-point-Taylor interpolation
• Theregularized kernel functionis smooth and periodic.
Introduction: NFFT and fast summation
Fast summation based on NFFT in 1d
[Potts, Steidl 2003]Computef(xj) :=
N
X
i=1
ciK(xi−xj)forxj∈[−L/2,L/2], j= 1, . . . , N.
0
−L L
−(L+δ) L+δ
K(x)
KR(x)
KB(x) KB(x)
1
• Extend the interval at both ends.
• Construct asmooth transition(match derivatives up to some orderp∈N).
→Two-point-Taylor interpolation
• Theregularized kernel functionis smooth and periodic.
Introduction: NFFT and fast summation
Fast summation based on NFFT in 1d
[Potts, Steidl 2003]Computef(xj) :=
N
X
i=1
ciK(xi−xj)forxj∈[−L/2,L/2], j= 1, . . . , N.
K(x) =KR(x) ∀x∈[−L, L]
f(xj) =
N
X
i=1
ciKR(xi−xj)
0
−L L
−(L+δ) L+δ
K(x)
KR(x)
KB(x) KB(x)
1
Far field computations (use FFT and NFFT): KRis periodic with period2(L+δ) =:h. f(xj) =
N
X
i=1
ciKR(xi−xj)≈
N
X
i=1
ci
X
`∈IM
ˆb`e2πi`(xi−xj)/h
=X
`∈IM
ˆb` N
X
i=1
cie2πi`xi/h
!
| {z }
adj. NFFT
e−2πi`xj/h
| {z }
NFFT
*Use the FFT to approximate the FKˆb`.
Introduction: NFFT and fast summation
Fast summation based on NFFT in 1d
[Potts, Steidl 2003]Computef(xj) :=
N
X
i=1
ciK(xi−xj)forxj∈[−L/2,L/2], j= 1, . . . , N.
K(x) =KR(x) ∀x∈[−L, L]
f(xj) =
N
X
i=1
ciKR(xi−xj)
0
−L L
−(L+δ) L+δ
K(x)
KR(x)
KB(x) KB(x)
1
Far field computations (use FFT and NFFT):
KRis periodic with period2(L+δ) =:h.
f(xj) =
N
X
i=1
ciKR(xi−xj)≈
N
X
i=1
ci
X
`∈IM
ˆb`e2πi`(xi−xj)/h
=X
`∈IM
ˆb` N
X
i=1
cie2πi`xi/h
!
| {z }
adj. NFFT
e−2πi`xj/h
| {z }
NFFT
*Use the FFT to approximate the FKˆb`.
Introduction: NFFT and fast summation
Fast summation based on NFFT in 1d
[Potts, Steidl 2003]Computef(xj) :=
N
X
i=1
ciK(xi−xj)forxj∈[−L/2,L/2], j= 1, . . . , N.
K(x) =KR(x) ∀x∈[−L, L]
f(xj) =
N
X
i=1
ciKR(xi−xj)
0
−L L
−(L+δ) L+δ
K(x)
KR(x)
KB(x) KB(x)
1
Far field computations (use FFT and NFFT):
KRis periodic with period2(L+δ) =:h.
f(xj) =
N
X
i=1
ciKR(xi−xj)≈
N
X
i=1
ci
X
`∈IM
ˆb`e2πi`(xi−xj)/h=X
`∈IM
ˆb` N
X
i=1
cie2πi`xi/h
!
| {z }
adj. NFFT
e−2πi`xj/h
| {z }
NFFT
*Use the FFT to approximate the FKˆb`.
Introduction: NFFT and fast summation
For d ≥ 2 dimensions, radial kernels
Computef(xj) :=
N
X
i=1
ciK(kxi−xjk)forj= 1, . . . , N. Assumekxi−xjk ≤L.
dj
drjKB(L+δ) = 0 dj
drjKB(L) =drdjjK(L)
L
−L L+δ
−(L+δ)
K(x)
• 1d regularization on[−h/2,h/2]with h:= 2(L+δ)
• claim vanishing derivatives at the boundary
• rotate and extend the function to the torushT2 (constant value over the striped area)
• result: periodically smooth function in 2 variables
• approximate by bivariate trigonometric polynomials (2d FFT)
Franziska Nestler TU Chemnitz, Faculty of Mathematics
Coulomb interactions and fast Ewald summation
Definition of the Coulomb interaction energy
Let N charges q
j∈
Rat positions
xj∈
R3be given.
Coulomb interaction energy (x
ij:=
xi−
xj):
U := 1 2
N
X
i,j=1 i6=j
q
iq
jkx
ijk = 1 2
N
X
j=1
q
jφ(j) with φ(j) :=
N
X
i=1 i6=j
q
ikx
ijk .
3d-periodic boundary conditions choose S :=
Z3,
xj∈ L
T3and set
φ(j) := φ
S(j) :=
Xn∈S N
X
i=1 i6=jifn=0
q
ikx
ij+ Lnk
→ crystals, . . .
Fast NFFT based algorithm: P
2NFFT, O(N log N)
[Pippig, Potts 2011]Coulomb interactions and fast Ewald summation
Definition of the Coulomb interaction energy
Let N charges q
j∈
Rat positions
xj∈
R3be given.
Coulomb interaction energy (x
ij:=
xi−
xj):
U := 1 2
N
X
i,j=1 i6=j
q
iq
jkx
ijk = 1 2
N
X
j=1
q
jφ(j) with φ(j) :=
N
X
i=1 i6=j
q
ikx
ijk .
3d-periodic boundary conditions choose S :=
Z3,
xj∈LT3and set
φ(j) := φ
S(j) :=
Xn∈S N
X
i=1 i6=jifn=0
q
ikx
ij+
Lnk→ crystals, . . .
L
Fast NFFT based algorithm: P
2NFFT, O(N log N)
[Pippig, Potts 2011]Coulomb interactions and fast Ewald summation
Definition of the Coulomb interaction energy
Let N charges q
j∈
Rat positions
xj∈
R3be given.
Coulomb interaction energy (x
ij:=
xi−
xj):
U := 1 2
N
X
i,j=1 i6=j
q
iq
jkx
ijk = 1 2
N
X
j=1
q
jφ(j) with φ(j) :=
N
X
i=1 i6=j
q
ikx
ijk .
3d-periodic boundary conditions choose
S:=Z3,
xj∈ L
T3and set
φ(j) := φ
S(j) :=
Xn∈S N
X
i=1 i6=jifn=0
q
ikx
ij+ Lnk
→ crystals, . . .
Fast NFFT based algorithm: P
2NFFT, O(N log N)
[Pippig, Potts 2011]Coulomb interactions and fast Ewald summation
Definition of the Coulomb interaction energy
Let N charges q
j∈
Rat positions
xj∈
R3be given.
Coulomb interaction energy (x
ij:=
xi−
xj):
U := 1 2
N
X
i,j=1 i6=j
q
iq
jkx
ijk = 1 2
N
X
j=1
q
jφ(j) with φ(j) :=
N
X
i=1 i6=j
q
ikx
ijk .
3d-periodic boundary conditions choose S :=
Z3,
xj∈ L
T3and set
φ(j) := φ
S(j) :=
Xn∈S N
X
i=1 i6=jifn=0
q
ikx
ij+ Lnk
→ crystals, . . .
Fast NFFT based algorithm: P
2NFFT, O(N log N)
[Pippig, Potts 2011]Coulomb interactions and fast Ewald summation
Ewald splitting
Idea of Ewald summation
[Ewald 1921]:
Ewald splitting 1
r =
erf(αr) r| {z } long ranged, continuous
+
erfc(αr) r| {z } singular in 0, short ranged
0 0.2 0.4 0.6 0.8 1
0 2 4 6 8 10
r
•
erf(x) :=
√2πRx0
e
−t2dt (error function), lim
r→0erf(αr) r=
√2απ•
erfc(x) := 1 − erf(x) (complementary error function)
•
α > 0 (scaling parameter)
Coulomb interactions and fast Ewald summation
Open boundary conditions
φ(j) =
N
X
i=1 i6=j
q
ikx
ijk =
N
X
i=1 i6=j
qi
erfc(αkxijk) kxijk
+
N
X
i=1
qi
erf(αkxijk) kxijk
− 2α
√ π q
j• short range part:
direct evaluation (truncation)
• long range part:
use NFFT based fast summation for radial kernels (d = 3) φ
long(j) ≈
Xk∈IM
ˆ b
k NX
i=1
q
ie
2πik·xi/h!
| {z }
adj. NFFT
e
−2πik·xj/h| {z }
NFFT
Computation of the forces:
F
(j) := −q
j∇φ(j) = −q
j∇φ
short(j)
| {z }
direct
−q
j∇φ
long(j)
| {z }
via gradient NFFT
Coulomb interactions and fast Ewald summation
Open boundary conditions
φ(j) =
N
X
i=1 i6=j
q
ikx
ijk =
N
X
i=1 i6=j
qi
erfc(αkxijk) kxijk
+
N
X
i=1
qi
erf(αkxijk) kxijk
− 2α
√ π q
j• short range part:
direct evaluation (truncation)
• long range part:
use NFFT based fast summation for radial kernels (d = 3) φ
long(j) ≈
Xk∈IM
ˆ b
k NX
i=1
q
ie
2πik·xi/h!
| {z }
adj. NFFT
e
−2πik·xj/h| {z }
NFFT
Computation of the forces:
F
(j) := −q
j∇φ(j) = −q
j∇φ
short(j)
| {z }
direct
−q
j∇φ
long(j)
| {z }
via gradient NFFT
Coulomb interactions and fast Ewald summation
3d-periodic boundary conditions
φ(j) =
Xn∈Z3 N
X
i=1 i6=jifn=0
qierfc(αkxij+Lnk) kxij+Lnk
+
Xn∈Z3 N
X
i=1
qierf(αkxij+Lnk) kxij+Lnk
− 2α
√ π q
j• short range part:
direct evaluation (truncation)
• long range part:
Fourier coefficients are known analytically.
φ
long(j) =
Xk∈Z3
e
−π2kkk2/(α2L2)kkk
2N
X
i=1
q
ie
2πik·xi/L!
| {z }
adj. NFFT
e
−2πik·xj/L| {z }
NFFT
• coefficients tend to zero exponentially fast
• truncate:Z37→ IM
Coulomb interactions and fast Ewald summation
Mixed periodicity
φ(j) :=
Xn∈S N
X
i=1 i6=jifn=0
q
ikx
ij+ Lnk
2d-periodic
xj
∈ L
T2×
R, S :=
Z2× {0}
→ thin liquid films, . . .
L
L
1
•
Compute analytic solution with respect to the 2 periodic dimensions.
•
Use fast summation in 1d for the non-periodic dimension.
1d-periodic
xj
∈ L
T×
R2, S :=
Z× {0}
2→ nano channels, . . .
L
1
•
Compute analytic solution with respect to the 1 periodic dimension.
•
Use fast summation in 2d for the
non-periodic dimensions.
Numerical results
Parameter choice
Parameter tuning: quite well understood for the 3d-periodic case.
Choose the NFFT mesh size in the mixed periodic case as follows:
2d-periodc 3d-periodic
periodic dims. non periodic dim. periodic dims.
box length L
h:= 2L+ 2δ L# grid points M M
3:= 2M + P M
0
−L L
−(L+δ) L+δ
regularization regularization
periodic
non periodic
P/2 M M P/2
grid points
Franziska Nestler TU Chemnitz, Faculty of Mathematics
Numerical results
Parameter choice
If the parameters are chosen appropriately, the achieved rms errors are comparable:
0.5 1 1.5 2
10−14 10−11 10−8 10−5 10−2
splitting parameterα
∆FZ2×{0}
M= 16, P= 8, h= 25 M= 32, P= 16, h= 25 M= 64, P= 30, h≈24.67 M= 128, P= 44, h≈23.44 M= 256, P= 76, h≈22.97
0.5 1 1.5 2
10−14 10−11 10−8 10−5 10−2
splitting parameterα
∆FZ3
Figure:Achieved rms force errors for the 2d-periodic (left) compared to the 3d-periodic (right) case. (L= 10,N= 300)
Large particle systems:
• Tune parameters for a small system.
• Use tuned parameters also for lager systems (same particle density), chooseM3∼N.
• Complexity:O(NlogN).
103 104 105 106
10−2 10−1 100 101 102 103
#charges
time[s]
∼N
∼NlogN
Figure:Comparison of the computation times (2dp:*, 3dp:4).
Numerical results
Parameter choice
If the parameters are chosen appropriately, the achieved rms errors are comparable:
0.5 1 1.5 2
10−14 10−11 10−8 10−5 10−2
splitting parameterα
∆FZ2×{0}
M= 16, P= 8, h= 25 M= 32, P= 16, h= 25 M= 64, P= 30, h≈24.67 M= 128, P= 44, h≈23.44 M= 256, P= 76, h≈22.97
0.5 1 1.5 2
10−14 10−11 10−8 10−5 10−2
splitting parameterα
∆FZ3
Figure:Achieved rms force errors for the 2d-periodic (left) compared to the 3d-periodic (right) case. (L= 10,N= 300)
Large particle systems:
• Tune parameters for a small system.
• Use tuned parameters also for lager systems (same particle density), chooseM3∼N.
• Complexity:O(NlogN).
103 104 105 106
10−2 10−1 100 101 102 103
#charges
time[s]
∼N
∼NlogN
Figure:Comparison of the computation times (2dp:*, 3dp:4).
Extension to systems with dipoles
Systems with charges and dipoles
Given
•
N
1charges q
j∈
Rat positions
xj(j = 1, . . . , N
1)
•
N
2dipoles with dipole moments
µj∈
R3at positions
xj(j = N
1+ 1, . . . , N
1+ N
2) Replace the charges q
jby the operators ξ
j:
q
j7→ ξ
j:=
(
q
j: j ∈ {1, . . . , N
1},
µ>j
∇
xj: j ∈ {N
1+ 1, . . . , N
1+ N
2}.
Electrostatic energy and potentials:
U := 1 2
N1+N2
X
j=1
ξ
jφ(j) with φ(j) :=
Xn∈Z3 N1+N2
X
i=1 i6=jifn=0
ξ
ikx
ij+ Lnk .
Extension to systems with dipoles
NFFT based computation of the long range part
Potentials (for charges and dipoles):
φ
long(j) ≈
Xk∈IM
ˆ b
k
N1
X
i=1
q
ie
2πik·xi/L| {z }
adj. NFFT
+
N1+N2
X
i=N1+1
µ>i
∇
xie
2πik·xi/L| {z }
adj. gradient NFFT
e
−2πik·xj/L| {z }
NFFT
Forces of the charges:
F(j) = −q
j∇
xjφ(j),
gradient NFFTinstead of NFFT.
Forces of the dipoles:
F(j) = −[∇
xj∇
>xjφ(j)] ·
µjFlong
(j) ≈ −
∇xj
∇
>xjX
k∈IM
ˆ b
k(S
c(k) + S
d(k)) e
−2πik·xj/L
| {z }
Hessian NFFT
·µ
jSummary
P
2NFFT for charged particle systems
Long range parts of the potentials: same structure for all types of periodic boundary conditions.
[N., Pippig, Potts, J. Comput. Phys. 2015], [N., Pippig, Potts, Comput. Trends in Solvation and Transport in Liquids 2015]
• 3d-periodic: coefficientsˆbkare known analytically (Ewald summation)
• 0d-periodic: NFFT based fast summation (3D)
• 1d-periodic: Ewald summation (1D) and fast summation (2D)
• 2d-periodic: Ewald summation (2D) and fast summation (1D)
Long range parts of the forces: Use gradient NFFT.
Extended method for computation of interactions with dipoles.[N., Appl. Numer. Math. 2016]
New NFFT modules: Hessian NFFT, adjoint gradient NFFT.[M. Pippig, PNFFT library, https://github.com/mpip/pfft] P2NFFT is publicly available at www.scafacos.de
Numerical results for systems with charges and dipoles: going to be published.
Thank you for your attention!
Summary
P
2NFFT for charged particle systems
Long range parts of the potentials: same structure for all types of periodic boundary conditions.
[N., Pippig, Potts, J. Comput. Phys. 2015], [N., Pippig, Potts, Comput. Trends in Solvation and Transport in Liquids 2015]
• 3d-periodic: coefficientsˆbkare known analytically (Ewald summation)
• 0d-periodic: NFFT based fast summation (3D)
• 1d-periodic: Ewald summation (1D) and fast summation (2D)
• 2d-periodic: Ewald summation (2D) and fast summation (1D) Long range parts of the forces: Use gradient NFFT.
Extended method for computation of interactions with dipoles.[N., Appl. Numer. Math. 2016]
New NFFT modules: Hessian NFFT, adjoint gradient NFFT.[M. Pippig, PNFFT library, https://github.com/mpip/pfft] P2NFFT is publicly available at www.scafacos.de
Numerical results for systems with charges and dipoles: going to be published.
Thank you for your attention!
Summary
P
2NFFT for charged particle systems
Long range parts of the potentials: same structure for all types of periodic boundary conditions.
[N., Pippig, Potts, J. Comput. Phys. 2015], [N., Pippig, Potts, Comput. Trends in Solvation and Transport in Liquids 2015]
• 3d-periodic: coefficientsˆbkare known analytically (Ewald summation)
• 0d-periodic: NFFT based fast summation (3D)
• 1d-periodic: Ewald summation (1D) and fast summation (2D)
• 2d-periodic: Ewald summation (2D) and fast summation (1D) Long range parts of the forces: Use gradient NFFT.
Extended method for computation of interactions with dipoles.[N., Appl. Numer. Math. 2016]
New NFFT modules: Hessian NFFT, adjoint gradient NFFT.[M. Pippig, PNFFT library, https://github.com/mpip/pfft]
P2NFFT is publicly available at www.scafacos.de
Numerical results for systems with charges and dipoles: going to be published.
Thank you for your attention!
Summary
P
2NFFT for charged particle systems
Long range parts of the potentials: same structure for all types of periodic boundary conditions.
[N., Pippig, Potts, J. Comput. Phys. 2015], [N., Pippig, Potts, Comput. Trends in Solvation and Transport in Liquids 2015]
• 3d-periodic: coefficientsˆbkare known analytically (Ewald summation)
• 0d-periodic: NFFT based fast summation (3D)
• 1d-periodic: Ewald summation (1D) and fast summation (2D)
• 2d-periodic: Ewald summation (2D) and fast summation (1D) Long range parts of the forces: Use gradient NFFT.
Extended method for computation of interactions with dipoles.[N., Appl. Numer. Math. 2016]
New NFFT modules: Hessian NFFT, adjoint gradient NFFT.[M. Pippig, PNFFT library, https://github.com/mpip/pfft]
P2NFFT is publicly available at www.scafacos.de
Numerical results for systems with charges and dipoles: going to be published.
Thank you for your attention!
Summary
P
2NFFT for charged particle systems
Long range parts of the potentials: same structure for all types of periodic boundary conditions.
[N., Pippig, Potts, J. Comput. Phys. 2015], [N., Pippig, Potts, Comput. Trends in Solvation and Transport in Liquids 2015]
• 3d-periodic: coefficientsˆbkare known analytically (Ewald summation)
• 0d-periodic: NFFT based fast summation (3D)
• 1d-periodic: Ewald summation (1D) and fast summation (2D)
• 2d-periodic: Ewald summation (2D) and fast summation (1D) Long range parts of the forces: Use gradient NFFT.
Extended method for computation of interactions with dipoles.[N., Appl. Numer. Math. 2016]
New NFFT modules: Hessian NFFT, adjoint gradient NFFT.[M. Pippig, PNFFT library, https://github.com/mpip/pfft]
P2NFFT is publicly available at www.scafacos.de
Numerical results for systems with charges and dipoles: going to be published.