ewald (c49b1)
The Ewald Summation method
Invoking the Ewald summation for calculating the electrostatic interactions
can be specified any time the nbond specification parser is invoked. See
the syntax section for a list of all commands that invoke this parser.
Prerequisite reading: » nbonds
* Syntax | Syntax of the Ewald summation specification
* Defaults | Defaults used in the specification
* Function | Description of the options
* Discussion | More general discussion of the algorithm
Invoking the Ewald summation for calculating the electrostatic interactions
can be specified any time the nbond specification parser is invoked. See
the syntax section for a list of all commands that invoke this parser.
Prerequisite reading: » nbonds
* Syntax | Syntax of the Ewald summation specification
* Defaults | Defaults used in the specification
* Function | Description of the options
* Discussion | More general discussion of the algorithm
Top
[SYNTAX EWALD]
{ NBONds } { nonbond-spec }
{ UPDAte } { }
{ ENERgy } { }
{ MINImize } { }
{ DYNAmics } { }
The keywords are:
nonbond-spec::= [ method-spec ]
{ [ NOEWald ] }
{ }
method-spec::= { EWALd [ewald-spec] { [ NOPMewald [std-ew-spec] ] } }
{ { PMEWald [pmesh-spec] } }
ewald-spec::= KAPPa real [erfc-spec]
ljpme-spec::= DKAPpa real DFTX int DFTY int DFTZ int DORDer int
std-ew-spec::= { [ KMAX integer ] } KSQMAX integer
{ KMXX integer KMXY integer KMXZ integer }
pmesh-spec::= FFTX int FFTY int FFTZ int ORDEr integer [QCOR real (***) ]
erfc-spec::= { SPLIne { [EWMIn real] [EWMAx real] [EWNPts int] } }
{ INTErpolate { } }
{ }
{ ABROmowitz }
{ CHEBychev }
{ EXACt_high_precision }
{ LOWPrecision_exact }
{ ERFMode int }
[SYNTAX EWALD]
{ NBONds } { nonbond-spec }
{ UPDAte } { }
{ ENERgy } { }
{ MINImize } { }
{ DYNAmics } { }
The keywords are:
nonbond-spec::= [ method-spec ]
{ [ NOEWald ] }
{ }
method-spec::= { EWALd [ewald-spec] { [ NOPMewald [std-ew-spec] ] } }
{ { PMEWald [pmesh-spec] } }
ewald-spec::= KAPPa real [erfc-spec]
ljpme-spec::= DKAPpa real DFTX int DFTY int DFTZ int DORDer int
std-ew-spec::= { [ KMAX integer ] } KSQMAX integer
{ KMXX integer KMXY integer KMXZ integer }
pmesh-spec::= FFTX int FFTY int FFTZ int ORDEr integer [QCOR real (***) ]
erfc-spec::= { SPLIne { [EWMIn real] [EWMAx real] [EWNPts int] } }
{ INTErpolate { } }
{ }
{ ABROmowitz }
{ CHEBychev }
{ EXACt_high_precision }
{ LOWPrecision_exact }
{ ERFMode int }
Top
The defaults for the ewald summation are set internally
and are currently set to NOEWald, KAPPa=1.0, KMAX=5, KSQMax=27, and
NOPMewald, KAPPa=1.0, FFTX=FFTY=FFTZ=32, ORDEr=4, QCOR=1.0
Recommended values for Ewald are:
EWALD PMEWald KAPPa 0.34 ORDEr 6 -
FFTX intboxvx FFTY intboxvy FFTZ intboxvz -
CTOFNB 12.0 CUTNB 14.0 QCOR 1.0(***)
Where intboxv* is an integer value similar to or larger than the corresponding
unit cell dimension that has prime factors of 2,3, and 5 only (2,3 preferred).
grid point spacing should be between 0.8 and 1.2 Angstroms.
These recommended values should give relative force errors of roughly 10**-5.
To reduce the total PME cost at the expense of accuracy, decrease the cutoff
distances while increasing KAPPa (keep the product near 4) reduces the real
space cost. To reduce the K-space cost, either reduce ORDEr from 6 to 4 or
increase the grid spacing up to perhaps 1.5 Angstroms.
(***) The QCOR value should be 1.0 for vacuum, solid, or finite systems.
For periodic systems in solution, it should be reduced (or set to zero) by an
amount that depends on how the net charge is distributed and on the effective
dielectric constant. For a treatise on this correction term, see:
S. Bogusz, T. Cheatham, and B. Brooks, JCP (1998) 108, 7070-7084 and references
contained therein (esp. Hummer and Levy).
The defaults for the ewald summation are set internally
and are currently set to NOEWald, KAPPa=1.0, KMAX=5, KSQMax=27, and
NOPMewald, KAPPa=1.0, FFTX=FFTY=FFTZ=32, ORDEr=4, QCOR=1.0
Recommended values for Ewald are:
EWALD PMEWald KAPPa 0.34 ORDEr 6 -
FFTX intboxvx FFTY intboxvy FFTZ intboxvz -
CTOFNB 12.0 CUTNB 14.0 QCOR 1.0(***)
Where intboxv* is an integer value similar to or larger than the corresponding
unit cell dimension that has prime factors of 2,3, and 5 only (2,3 preferred).
grid point spacing should be between 0.8 and 1.2 Angstroms.
These recommended values should give relative force errors of roughly 10**-5.
To reduce the total PME cost at the expense of accuracy, decrease the cutoff
distances while increasing KAPPa (keep the product near 4) reduces the real
space cost. To reduce the K-space cost, either reduce ORDEr from 6 to 4 or
increase the grid spacing up to perhaps 1.5 Angstroms.
(***) The QCOR value should be 1.0 for vacuum, solid, or finite systems.
For periodic systems in solution, it should be reduced (or set to zero) by an
amount that depends on how the net charge is distributed and on the effective
dielectric constant. For a treatise on this correction term, see:
S. Bogusz, T. Cheatham, and B. Brooks, JCP (1998) 108, 7070-7084 and references
contained therein (esp. Hummer and Levy).
Top
i) The EWALD keyword invokes the Ewald summation for calculation of
electrostatic interactions in periodic, neutral systems. The formulation of
the Ewald summation dictates that the primary system must be neutral. If
otherwise, the summation is not formally correct and some
convergence problems may result. The NOEWald (default) suppresses the Ewald
method for calculating electrostatic interactions. Van der waals
options VSHIFT and VSWITCH are supported with ewald. The algorithm
currently supports the atom and group nonbond lists and the CRYSTAL facilty
must be used. The PMEWald keyword invokes the Particle Mesh Ewald algorithm
for the reciprocal space summation. For details on the PME method, see
J. Chem. Phys. 103:8577 (1995). The EWALd algorithm is limited to CUBIC,
TETRAGONAL, and ORTHORHOMBIC unit cells. The PMEWald algorithm supports
all unit cells that are supported by the CRYSTAL facility.
ii) The KAPPa keyword, followed by a real number governs the width of the
Gaussian distribution central to the Ewald method. An approximate value
of kappa can be chosen by taking KAPPa=5/CTOFNB. This is fairly conservative.
Values of 4/CTOFNB lead to small force errors (roughly 10**-5). See
discussion section for details on choosing an optimum value of KAPPa.
iii) The KMAX key word is the number of kvectors (or images of the
primary unit cell) that will be summed in any direction. It is the
radius of the Ewald summation. For orthorombic cells, the value of
kmax may be independently specified in the x, y, and z directions with
the keywords KMXX, KMXY, and KMXZ. In the PME version, the number of
FFT grid points for the charge mesh is specified by FFTX, FFTY, and FFTZ.
iv) The KSQMax key word should be chosen between KMAX squared and 3 times
KMAX squared.
v) An appropriate, although not optimal, set of parameters can be
chosen by taking KAPPA=5/CTOFNB and KMAX=KAPPa*boxlength. The actual
values should then be performanced optimized for your particular system.
For the PME method, FFTX should be approximately the box length in Angstroms.
(for efficiency, FFTX should be a multiple of powers of 2,3, and 5).
IMPORTANT NOTE::: THE SUGGESTION THAT FFTX, FFTY, AND FFTZ HAVE
NO PRIME FACTORS OTHER THAN 2, 3, AND 5 SEEMS TO BE A REQUIREMENT.
LARGE ERRORS IN THE FORCE ARE OBSERVED WHEN THIS CONDITION IS NOT MET.
FUTURE VERSIONS OF CHARMM WILL FLAG THIS AS AN ERROR CONDITION.
ORDEr specifies the order of the B-spline interpolation, e.g. cubic is
order 4 (default), fifth degree is ORDEr 6. The ORDEr must be an even
number and at least 4.
vi) EWALd runs in parallel on both shared (PARVECT) and distributed
memory parallel computers. PME runs in parallel on distributed
memory computers.
vii) several algorithms are available for the calculation of the complimentary
error function, erfc(x). EXACt and LOWPrecision use an interative technique
described in section 6.2 of Numerical Recipies. ABRO and CHEB are polynomial
approximations. A lookup table (filled at the beginning of the simulation
using the EXACt method) can be used with either a linear (INTE) of cubic
spline (SPLINe) interpolation. SPLIne is recommended.
viii) Ewald with MMFF
A version of EWALD was developed for MMFF. The usual MMFF electrostatic
term: qq/(r+d) is split into two terms: qq/r - qq*d/(r*(r+d))
The first term is handled by the Ewald method in the usual manner
(real-space and k-space parts) and the second term is truncated
at the cutoff distance using a switching function (from CTONNB to CTOFNB).
Since the second term is quite small at the cutoff distance, the use of a
switching function should not introduce significant artificial forces.
i) The EWALD keyword invokes the Ewald summation for calculation of
electrostatic interactions in periodic, neutral systems. The formulation of
the Ewald summation dictates that the primary system must be neutral. If
otherwise, the summation is not formally correct and some
convergence problems may result. The NOEWald (default) suppresses the Ewald
method for calculating electrostatic interactions. Van der waals
options VSHIFT and VSWITCH are supported with ewald. The algorithm
currently supports the atom and group nonbond lists and the CRYSTAL facilty
must be used. The PMEWald keyword invokes the Particle Mesh Ewald algorithm
for the reciprocal space summation. For details on the PME method, see
J. Chem. Phys. 103:8577 (1995). The EWALd algorithm is limited to CUBIC,
TETRAGONAL, and ORTHORHOMBIC unit cells. The PMEWald algorithm supports
all unit cells that are supported by the CRYSTAL facility.
ii) The KAPPa keyword, followed by a real number governs the width of the
Gaussian distribution central to the Ewald method. An approximate value
of kappa can be chosen by taking KAPPa=5/CTOFNB. This is fairly conservative.
Values of 4/CTOFNB lead to small force errors (roughly 10**-5). See
discussion section for details on choosing an optimum value of KAPPa.
iii) The KMAX key word is the number of kvectors (or images of the
primary unit cell) that will be summed in any direction. It is the
radius of the Ewald summation. For orthorombic cells, the value of
kmax may be independently specified in the x, y, and z directions with
the keywords KMXX, KMXY, and KMXZ. In the PME version, the number of
FFT grid points for the charge mesh is specified by FFTX, FFTY, and FFTZ.
iv) The KSQMax key word should be chosen between KMAX squared and 3 times
KMAX squared.
v) An appropriate, although not optimal, set of parameters can be
chosen by taking KAPPA=5/CTOFNB and KMAX=KAPPa*boxlength. The actual
values should then be performanced optimized for your particular system.
For the PME method, FFTX should be approximately the box length in Angstroms.
(for efficiency, FFTX should be a multiple of powers of 2,3, and 5).
IMPORTANT NOTE::: THE SUGGESTION THAT FFTX, FFTY, AND FFTZ HAVE
NO PRIME FACTORS OTHER THAN 2, 3, AND 5 SEEMS TO BE A REQUIREMENT.
LARGE ERRORS IN THE FORCE ARE OBSERVED WHEN THIS CONDITION IS NOT MET.
FUTURE VERSIONS OF CHARMM WILL FLAG THIS AS AN ERROR CONDITION.
ORDEr specifies the order of the B-spline interpolation, e.g. cubic is
order 4 (default), fifth degree is ORDEr 6. The ORDEr must be an even
number and at least 4.
vi) EWALd runs in parallel on both shared (PARVECT) and distributed
memory parallel computers. PME runs in parallel on distributed
memory computers.
vii) several algorithms are available for the calculation of the complimentary
error function, erfc(x). EXACt and LOWPrecision use an interative technique
described in section 6.2 of Numerical Recipies. ABRO and CHEB are polynomial
approximations. A lookup table (filled at the beginning of the simulation
using the EXACt method) can be used with either a linear (INTE) of cubic
spline (SPLINe) interpolation. SPLIne is recommended.
viii) Ewald with MMFF
A version of EWALD was developed for MMFF. The usual MMFF electrostatic
term: qq/(r+d) is split into two terms: qq/r - qq*d/(r*(r+d))
The first term is handled by the Ewald method in the usual manner
(real-space and k-space parts) and the second term is truncated
at the cutoff distance using a switching function (from CTONNB to CTOFNB).
Since the second term is quite small at the cutoff distance, the use of a
switching function should not introduce significant artificial forces.
Top
The Ewald Summation in Molecular Dynamics Simulation
The electrostatic energy of a periodic system can be expressed by a lattice
sum over all pair interactions and over all lattice vectors excluding
the i=j term in the primary box. Summations carried out in this simple
way have been shown to be conditionally convergent. The method developed by
Ewald, in essence, mathematically transforms this fairly straightforward
summation to two more complicated but rapidly convergent sums. One summation
is carried out in reciporcal space while the other is carried out in real
space. Based on the formulation by Ewald, the simple lattice sum can be
reformulated to give absolutely convergent summations which define the
principal value of the electrostatic potential, called the intrinsic potential.
Given the periodicity present in both crystal calculations and in dynamics
simulations using periodic boundary conditions, the Ewald formulation becomes
well suited for the calculation of the electrostatic energy and force. If we
consider a system of point charges in the unit or primary cell, we can specify
its charge density by
ro(r) = sum_i [ q_i * delta(r-r_i)]
In the Ewald method this distribution is replaced by two other distributions
ro_1(r) = sum_i [ q_i ( delta(r-r_i) - f(r-r_i)]
and
ro_2(r) = sum_i [q_i f(r-r_i)
such that the sum of the two recovers the original. The distribution,
f(r), is a spherical distribution generally taken to be Gaussian, the
width of the gaussian dictated by the parameter, KAPPa. The charge
distributions are situated on the ion lattice positions, but integrate
to zero. The potential from the distribution ro_1(r) is a short range
potential evaluated in a direct real space summation (truncated at
CTOFNB). The diffuse charge distribution placed on the lattice sites
reduces to the potential of the corresponding point charge at large r.
ro_2(r), being a continuous distribution of Gaussians situated on the
periodic lattice positions, is a smoothly varying function of r and thus
is well approximated by a superposition of continuous functions. This
distribution is, therefore, expanded in a Fourier series and the
potential is obtained by solving the Poisson equation. The point of
splitting the problem into two parts, is that by a suitable choice of
the parameter KAPPa we can get very good convergence of both parts of
the summation.
For the real space part of the energy, we choose kappa so that the
complementary error function term, erfc(kappa*r) decreases rapidly
enough with r to make it a good approximation to take only nearest
images in the sum and neglect the value for which r > CTOFNB. The
reciprocal space sums are rapidly convergent and a spherical cutoff in k
space is applied so that the sum over k becomes a sum over {l,m,n}, with
(l**2+m**2+n**2) < or = to KSQMAX A large value of KAPPa means that the
real space sum is more rapidly convergent but the reciprocal space sum
is less rapid. In practice one chooses KAPPa to give good convergence
at the cutoff radius, CTOFNB. KMAX is then chosen to such that the
reciprocal space calculation converges. The equation (KMAX/(box
length)=KAPPa may be used as a rough guide. Optimization with respect
to the timing trade offs, ie. how much time is spent in real space vs
k-space should be performed before a lengthy production run.
The CCP5 notes in several articles in 1993 cover some possible
optimization strategies and criteria although a simple line search will
suffice. Complete optimiztion of the ewald method for a particular
application requires optimizing CTOFNB, KAPPa, and KMAX. A discussion
of optimization and error analysis can be found in Kolfka and Perram,
Molecular Simulation, 9, 351 (1992). For PME, see Feller, Pastor,
Rojnuckarin, Bogusz, and Brooks. J. Phys. Chem., 100, 42, 17011 (1996)
and some of Tom Darden's published work.
Using LJ-PME
The ljpme-spec closely mirrors the pmesh-spec of its electrostatic counterpart.
The DKAPpa keyword defaults to the value set by KAPPa and a value around 0.4
(/Angstrom) is recommended. The dispersion grid dimensions DFT{X,Y,Z} are
analogous to FFT{X,Y,Z} and are defaulted to the same values. As with
electrostatic PME, the values should be around 1.2 times the box dimensions for
tight energy and force convergence with the recommended DKAPpa. Similarly the
spline order for the dispersion term is controlled by the DORDer, which
defaults to the value defined by ORDEr and can usually be relaxed to 4 or 5,
still yielding precise energies and forces.
The LJ-PME code is implemented for the slow routines as well as for DOMDEC.
One restriction of the DOMDEC implementation is the requirements FFTX==DFTX,
FFTY==DFTY and FFTZ==DFTZ for parallel runs with more than one direct node;
this is not much of a limitation in practice. Only potential shifting (VSHIft)
and potential switching (VSWItch) methods are implemented, and the latter is
recommended. The use of atom based lists (VATOm) is currently required for
LJ-PME.
Currently the LJ-PME feature is not activated in a default build. To activate
this feature, pass the "--with-ljpme" argument to the configure script before
compiling. The implementation of LJ-PME uses C++11, which prevents the use of
very old compilers (e.g. GCC versions pre-4.9). On some systems with native
GCC toolchains that don't meet this requirement, newer GCC modules should be
loaded. The Intel C++ compiler uses GCC under the hood and on systems with an
old GCC toolchain, this could potentially cause problems. In this case, the
Intel C++ simply needs to be made aware of the location of a suitable GCC
installation on the machine by setting the CXXFLAGS environmental variable to
"-gxx-name=/full/path/to/newer/gcc/version/bin/g++". At runtime the
corresponding /full/path/to/newer/gcc/version/lib folder should be added to the
LD_LIBRARY_PATH environmental variable also.
The Ewald Summation in Molecular Dynamics Simulation
The electrostatic energy of a periodic system can be expressed by a lattice
sum over all pair interactions and over all lattice vectors excluding
the i=j term in the primary box. Summations carried out in this simple
way have been shown to be conditionally convergent. The method developed by
Ewald, in essence, mathematically transforms this fairly straightforward
summation to two more complicated but rapidly convergent sums. One summation
is carried out in reciporcal space while the other is carried out in real
space. Based on the formulation by Ewald, the simple lattice sum can be
reformulated to give absolutely convergent summations which define the
principal value of the electrostatic potential, called the intrinsic potential.
Given the periodicity present in both crystal calculations and in dynamics
simulations using periodic boundary conditions, the Ewald formulation becomes
well suited for the calculation of the electrostatic energy and force. If we
consider a system of point charges in the unit or primary cell, we can specify
its charge density by
ro(r) = sum_i [ q_i * delta(r-r_i)]
In the Ewald method this distribution is replaced by two other distributions
ro_1(r) = sum_i [ q_i ( delta(r-r_i) - f(r-r_i)]
and
ro_2(r) = sum_i [q_i f(r-r_i)
such that the sum of the two recovers the original. The distribution,
f(r), is a spherical distribution generally taken to be Gaussian, the
width of the gaussian dictated by the parameter, KAPPa. The charge
distributions are situated on the ion lattice positions, but integrate
to zero. The potential from the distribution ro_1(r) is a short range
potential evaluated in a direct real space summation (truncated at
CTOFNB). The diffuse charge distribution placed on the lattice sites
reduces to the potential of the corresponding point charge at large r.
ro_2(r), being a continuous distribution of Gaussians situated on the
periodic lattice positions, is a smoothly varying function of r and thus
is well approximated by a superposition of continuous functions. This
distribution is, therefore, expanded in a Fourier series and the
potential is obtained by solving the Poisson equation. The point of
splitting the problem into two parts, is that by a suitable choice of
the parameter KAPPa we can get very good convergence of both parts of
the summation.
For the real space part of the energy, we choose kappa so that the
complementary error function term, erfc(kappa*r) decreases rapidly
enough with r to make it a good approximation to take only nearest
images in the sum and neglect the value for which r > CTOFNB. The
reciprocal space sums are rapidly convergent and a spherical cutoff in k
space is applied so that the sum over k becomes a sum over {l,m,n}, with
(l**2+m**2+n**2) < or = to KSQMAX A large value of KAPPa means that the
real space sum is more rapidly convergent but the reciprocal space sum
is less rapid. In practice one chooses KAPPa to give good convergence
at the cutoff radius, CTOFNB. KMAX is then chosen to such that the
reciprocal space calculation converges. The equation (KMAX/(box
length)=KAPPa may be used as a rough guide. Optimization with respect
to the timing trade offs, ie. how much time is spent in real space vs
k-space should be performed before a lengthy production run.
The CCP5 notes in several articles in 1993 cover some possible
optimization strategies and criteria although a simple line search will
suffice. Complete optimiztion of the ewald method for a particular
application requires optimizing CTOFNB, KAPPa, and KMAX. A discussion
of optimization and error analysis can be found in Kolfka and Perram,
Molecular Simulation, 9, 351 (1992). For PME, see Feller, Pastor,
Rojnuckarin, Bogusz, and Brooks. J. Phys. Chem., 100, 42, 17011 (1996)
and some of Tom Darden's published work.
Using LJ-PME
The ljpme-spec closely mirrors the pmesh-spec of its electrostatic counterpart.
The DKAPpa keyword defaults to the value set by KAPPa and a value around 0.4
(/Angstrom) is recommended. The dispersion grid dimensions DFT{X,Y,Z} are
analogous to FFT{X,Y,Z} and are defaulted to the same values. As with
electrostatic PME, the values should be around 1.2 times the box dimensions for
tight energy and force convergence with the recommended DKAPpa. Similarly the
spline order for the dispersion term is controlled by the DORDer, which
defaults to the value defined by ORDEr and can usually be relaxed to 4 or 5,
still yielding precise energies and forces.
The LJ-PME code is implemented for the slow routines as well as for DOMDEC.
One restriction of the DOMDEC implementation is the requirements FFTX==DFTX,
FFTY==DFTY and FFTZ==DFTZ for parallel runs with more than one direct node;
this is not much of a limitation in practice. Only potential shifting (VSHIft)
and potential switching (VSWItch) methods are implemented, and the latter is
recommended. The use of atom based lists (VATOm) is currently required for
LJ-PME.
Currently the LJ-PME feature is not activated in a default build. To activate
this feature, pass the "--with-ljpme" argument to the configure script before
compiling. The implementation of LJ-PME uses C++11, which prevents the use of
very old compilers (e.g. GCC versions pre-4.9). On some systems with native
GCC toolchains that don't meet this requirement, newer GCC modules should be
loaded. The Intel C++ compiler uses GCC under the hood and on systems with an
old GCC toolchain, this could potentially cause problems. In this case, the
Intel C++ simply needs to be made aware of the location of a suitable GCC
installation on the machine by setting the CXXFLAGS environmental variable to
"-gxx-name=/full/path/to/newer/gcc/version/bin/g++". At runtime the
corresponding /full/path/to/newer/gcc/version/lib folder should be added to the
LD_LIBRARY_PATH environmental variable also.