Professional Documents
Culture Documents
Andrea Porro
Sorbonne Université
Università degli Studi di Milano
The fourth degree term P4 is directly related to satura- such a solution, the shape of a convex quartic polyno-
tion effects, since great values of the polarization would mial centered on the origin. This, however, is true only
result in a really high energetic cost, making the config- in the bulk, far from any perturbation or surface which
uration unsuitable; the term P0 stands for a threshold could alter the boundary conditions of the solution. More
between the linear and non-linear behavior. Once the complex and interesting cases will be studied in section
energy functional is written, many useful physical quan- IV.
tities can be computed. One of the easiest, and maybe
among the most important, is the linear longitudinal sus-
ceptibility function: its significance arises from being a B. Fixing the parameters
measurable quantity, both from experiments and com-
puter simulations, and, thus, an important term of com- The set of parameters chosen is not random and can-
parison for any theory aiming at describing the dynamic not be changed arbitrarily: parameters are fixed in order
behavior of a fluid. Moreover, it will be the quantity to reproduce some observables, known empirically both
we’ll compute by means of molecular dynamics simula- from computer simulations (molecular dynamics in par-
tions, in order to control the rightness of the model. In ticular) and real experiments. We can briefly review each
appendix A a complete computation of this quantity in of them.
linear approximation is reported. The final formula reads a. P0 , γ and K The value of the parameter K is
fixed by the known experimental value of the bulk sus-
(lin) 1
χ̃k (q) = (4) ceptibility for water: bulk properties, in fact, are related
1 + K + κl q 2 + α q 4 to the average values and macroscopic properties of a
Note that, for α > 0 and κl < 0, the function displays medium, which, in Fourier space, are related to the mo-
a maximum in Fourier space. Similarly, in the nonlinear mentum vector q = 0. Substituting this value in equa-
case, one gets: tion 4 one can diectly relate K and χbulk . The standard
value of K = 1/70 is chosen. In the same way, the term
1 2γPm 2
= 1+2γP02 , which represents the linear term in the
χ̃k (q) = 2 + κ q2 + α q4
2γPm l complete description (including both nonlinear and non-
local effects), can be related to the same physical quan-
Since bulk properties of water are well known this is also tity by the equation 2γPm 2
= 1 + K. Of course some
an essential quantity in order to correctly fix the values degeneracy arises from this choice, since only the prod-
of the parameters of the phenomenological model. Until uct between Pm and γ is fixed; γ’s value, however, can
now we have only obtained a functional expression for the be set in order to reproduce the nonlinearities’ scale in
energy of the system as a response of a given polarisation. the system.
However, we would like to recover an equation whose b. α and κl The maximum in equation 4 only arises
solution would give the expression for the polarisation. if κl < 0 and α > 0 (otherwise its shape would not
This comes from the minimization of the functional: the be Lorentzian). Maximum positioning, moreover, is well
polarization of the physical system is the one minimizing known both from experiments and numerical simulations
the energy of the configuration. In this way we’re able to −1
(3Å ), as we’ll see in the Molecular Dynamic section:
find the solution for the most probable configuration for
thus, computing the maximum point for equation 4 and
the polarization of the system.
setting its first derivative to zero, enables us to get a
In order to minimize the functional one needs to compute
relation between the maximum position, α and γ:
its functional derivative and put it to zero. The details
of the calculation are reported in appendix B. The final 2
κl = −2αqmax
equation for P(r), in a generic geometrical scenery is:
Again, some degeneracy arises from the equation (only
δH[P] 1 2γ one equation for fixing two parameters), but the correct
= P+ P(P2 + P02 )+ values are set in order to reproduce the known corre-
δP ε0 ε0
κl α lations lengths for dipolar correlation and overscreening
− ∇(∇ · P) + ∇∇ · (∇∇ · P) (5) effects.
ε0 ε0
This functional was already used with a set of parame-
ters that reproduces the position and the magnitude of III. MOLECULAR DYNAMICS
the maximum of χ(q), but it is quite poor as far as con-
cerns water properties, such as dipole correlations, which In order to compute the behaviour of the water sus-
gave a rather long range if compared with numerical sim- ceptibility function and to fix the field theory parame-
ulations. This is the reason why we made an intensive ters we performed a Molecular Dynamics simulation us-
use of Molecular Dynamics simulations to fix the model’s ing SPC/E model: in this model water molecules are
parameters in a better way. treated as rigid bodies composed by three point charges
The bulk stationary state for equation 3 is, of course, located in the nuclear sites of both Oxygen and Hydro-
the null solution hPimf = 0, since the energy takes, for gen atoms. The molecular forces are derived in terms of
5
B. Charge Density Correlation Function positions of the two Hydrogen atoms in the i -th molecule.
In SPC/E model we can change the reference frame for
For point-like charge each atom distribution is assumed a single molecule centering on the Oxygen atom:
to be a Dirac delta, so that the total charge distribution
is: rH(j) ,i = rO,i + uj,i dOH with j=1,2
N X
X
ρc (r) = qµ δ (3) (r − rµ,i ) (12) u1,i dOH · u2,i dOH = cos ϑ0
i=1 µ
where the uj,i are unitary vectors pointed in the direc-
with index µ referring to the atom into a given i molecule. tion of the O-H bonds and dOH is the distance between
However, not to keep things too much abstract, we’ll refer Hydrogen and Oxygen atoms within a molecule. The lo-
explicitly to the case of water. For SPC model water the cal frame systems are depicted in figure 4 for a better
density distribution of the system is comprehension. Moreover we can write
N h
X (u1,i − u2,i )dOH = u3,i dHH
ρc (r) = qO δ (3) (r − rO,i )+
i=1 In such a way it is quite easy to get the Fourier transform
of the charge density given in equation 12:
i
(3) (3)
+ qH δ (r − rH(1) ,i ) + qH δ (r − rH(2) ,i )
N h i
where the index i runs over all molecules, O stands for
X
ρ̃c (q) = qH eiq·rO,i − 2 + eiq·u1,i dOH + eiq·u2,i dOH
Oxygen and H(1) and H(2) are referred to the two dif- i=1
ferent Hydrogen atoms in a molecule, so that rO,i is the
position of the i -th Oxygen atom, rH(1) ,i and rH(2) ,i the Thus, equation 11 becomes,
N N
1 X iq·rO,i
h
iq·u1,i dOH iq·u2,i dOH
i X
−iq·rO,j
h
−iq·u1,j dOH −iq·u2,j dOH
i
S(q) = h qH e − 2 + e + e × qH e − 2 + e + e i (14)
V q 2 i=1 j=1
where V is the constant volume of the simulation box, where the simulations were performed in the NVE ensemble.
The product of the two sum can be divided into two different parts, an intramolecular part S (m) (q), with index j
equal to index i, standing for the correlations within the same molecule, and an intermolecular one S (d) (q), with j 6= i,
which describes correlations between different molecules. The structure factor, thus, reads:
Since we are taking the average of the correlation be- between the wave vector and the unitary vectors of the
tween the densities, we need to average over the direction
7
2
Hydrogen atoms: C. Protocol
±iq·udOH
he iϑ = j0 (qdOH ) (17)
We performed simulations both for the SPC/E model
with j0 the spherical Bessel function of order zero. and for a pure dipolar model based on a two-site descrip-
An analogous calculation can be done with the two last tion [12]. We will refer to the dipolar model as Two Site
terms of equation 16. Model (TSM). In both cases the followed protocol was
Thus, the intramolecular structure factor is the same, that we will now proceed to explain. As a
2nqH 2 first step, in order to generate the initial configuration
S (m) (q) =
2
3 − 4j0 (qdOH ) + j0 (qdHH ) (18) of the system, we placed the Oxygen atoms on the ver-
q
tices of a cubic lattice, whose size was chosen in order to
with n the molecular density of the medium and n = −3
−3
reproduce the density of 0.033Å , with a typical num-
0.033 molÅ in water at 300 K and 1 atm. ber of molecules of N = 1000. Hydrogen atoms were
In order to analyse the intermolecular part we go back randomly disposed around their Oxygen respecting the
to the previous reference frame. distance dOH and the angle H OH b of the SPC/E model.
N h
Oxygen positions were not taken randomly not to acci-
2
qH X iq·rH(1) ,i iq·rH(2) ,i
i
dentally superpose two of them, leading to high repulsion
S (d) (q) = h − 2e iq·rO,i
+ e + e
V q 2 i=1 energy and, consequently, to a slow thermalization. A
N h
2f s time step was used, while electrostatic interactions
were computed using the Ewald summation technique,
i
−iq·rH(1) ,j −iq·rH(2) ,j
X
× − 2e−iq·rO,j + e +e i (19)
in order to take into account the periodicity of the sys-
j=1
j6=i tem. The cutoff value for the Van der Waals interaction
was kept slightly below the half-cell size not to encounter
By using the definition problems due to the fluctuations of the cell’s volume dur-
hµν (r) = gµν (r) − 1 ing NPT phases of the simulation. A first short equili-
bration (0.01 ns) was performed in the NVE ensemble,
with gµν (r) radial distribution function taking into ac- followed by a long (2 ns) equilibration in the NPT en-
count the atomic species µ and ν, that in Fourier space semble. Volume, pressure and temperature averages were
reads: performed on the latter: their values were then used as
N parameters for the final simulation. This was cut in sev-
1 X iq·(rµ,i −rν,j )
hµν (q) = h e i eral smaller parts: after an equilibration phase, needed
N i,j=1
to impose the desired temperature by means of repeated
j6=i
scaling of the velocities, the final and real simulation was
and expanding the product in equation 19, recalling that performed in the NVE ensemble for an amount of 8 mil-
the ordering of the atoms taken into account is not im- lions steps, leading to a total time of 16 ns. Periodic
portant since we are summing over all the molecule in- printing of thermodynamic quantities and molecular tra-
dices i and j, as well as is irrelevant the numerical index jectory were done for further analyses. Temperature were
referring to the two different Hydrogen atoms in each kept at 293.15 K and pressure at 10−3 katm.
molecule, one gets:
2
4nqH
S (d) (q) =
2
hOO (q) + hHH (q) − 2hOH (q) (20) D. SPC/E Results
q
In such a way we got an analytical expression for the in- We can now expose the results of the performed nu-
tramolecular dynamical structure factor, which is always merical simulations for the SPC/E model: figure 5 shows
verified for a point charge model, and an expression re- the radial distribution function for the SPC/E model.
lated to observable quantities for the intermolecular part. Figure 6, on the other hand, displays the functions h(q),
The latter, in fact, is expressed in terms of the Fourier as computed in the continuous approximation: that’s to
transform of the radial distribution function, that is a say, as the Fourier transforms of h(r) = g(r) − 1
quantity always available from molecular dynamics sim- Z ∞
ulations. sin qr
h(q) = 4π dr r2 (g(r) − 1) (21)
0 qr
n hOO (q)
gOO (r)
−0.5
1
0 −1
1.5
0
n hOH (q)
gOH (r)
−0.5
0.5
0 −1
1.5
0
n hHH (q)
gHH (r)
−0.5
0.5
0 −1
0 2 4 6 8 10 0 2 4 6 8 10 12 14
r (Å) q (Å
−1
)
FIG. 5. Radial Distribution Functions for SPC/E water FIG. 6. Fourier Transform of the radial distribution functions
for SPC/E water. The dashed parts indicate the low-precision
for low q’s.
made by enlarging the box’s size and the number of
molecules but without any significant enhancement. For lowing expression:
this reason some references tend to omit the troublesome
part for small q’s, putting a cutoff at a wavenumber of 4z 2 e2 X −iq·(rlH −rlH0 ) X −iq·(rlO −rlO0 )
−1
about 1Å from the origin. Serious advantages, on the S (d) (q) = h e + e +
V q2
l6=l0 l6=l0
other hand, derive from representation 19, since it does
l l0 l l0
not introduce any artificial effect on the system and takes
X X
−2 e−iq·(rO −rH ) − 2 e−iq·(rH −rO ) i
into account its periodicity. It must be kept into mind l6=l0 l6=l0
that not all the possible momenta q are allowed, but only
the ones lying into the simulation box: The formula can be made more manageable by substi-
tuting, for each sum,
2π X i j X
q= (nx , ny , nz ) e−iq·(r −r ) = 2 cos q · rij
L
i6=j i<j
being L the box size and nx , ny , nz ∈ N. Thus, instead This formalism is definitely way slower than the contin-
of the continuous formulation, we’ll make use of the fol- uous one, since it needs to be applied on the trajectories
9
40 χ(q) continuous FT
χ(q) discrete formalism 70
30
60
εq (0)
χ(q)
20 50
40
10
30
0 20
0 2 4 6 8 10 12 14
−1
q (Å ) 0.99
χq (0)
reported, with the continuous and the discrete formalism. It
can be easily seen how in this limit the continuous assump-
tion leads to some unphysical oscillations, whilst the correct
discrete development recovers the right behavior. 0.97
gOO (r)
molecular trajectories with the new fictitious trajectories 2
given by the real oxygen position and an imaginary po-
sition for an effective Hydrogen atom lying on the axis:
its charge and distance are free parameters of the dipolar 1
model, the only constraint is to reproduce the module
of the dipole moment of SPC/E water. The simple two
charges model we get to in this way is called, in refer- 0
ence [17] Dipolar Dumbbell model (DD), in opposition to
another dipolar model called Dipolar Hard Sphere model
(DS ), which is a pure dipolar model (as a gas of dipoles) 1.5
made of point dipoles. We chose to compare the sus-
ceptibility from the SPC/E model with both the results
gOH (r)
arising from the DD and the DS model in order to under- 1
stand the amount of information needed to fix the field
theory parameters, that’s to say, how far can we go with
simplifying the molecular dynamics model without loos- 0.5
ing too much accuracy for the dielectric and structural
properties of water. The DD model, in fact, with respect
to DS model, has somehow more information, because 0
it can choose for more degrees of freedom (charge and
2.5
distance). The results for our projections onto the DD
model are reported in figure 10, together with the results
2
from [3] for the DS model. The longitudinal dielectric
gHH (r)
G. Dipolar correlations
F. Pure Dipole Model
As hinted in sectionIII E a quantitative comparison be-
As we have seen, the information needed for tween the molecular dynamics simulations and the field
parametrizing the polarisation field theory model lies on theoretical approach was performed. In order to repro-
an intrinsically two site model. Thus, an attempt to use duce not only the dielectric properties of water (dielectric
a pure dipolar model in a molecular dynamics simulation susceptibility in particular), but the dipolar correlations
was done. In reference [12] a purely dipolar model for too, we compared the results obtained in a previous work
12
χ(q)
gOO (r)
1 5
0.5
0
0 0 2 4 6 8 10 12 14
−1
q (Å )
1.5
FIG. 12. Dielectric longitudinal susceptibility function for the
dipolar models taken into account. The dotted line stands for
gOH (r)
hP (0)P (r)i
L(r) =
0.5 µ2 ρ2
−3
Of course ρ = 0.033Å is the molecular density of water,
0 while µ = 2.35D is the dipole moment for SPC/E water
0 2 4 6 8 10 in Debye units.
r (Å) When parametrized by means of the SPC/E water
model, the field theoretical approach does not succeed
FIG. 11. Radial distribution functions for the purely dipolar at reproducing the dipole correlations, as clearly shown
models used in this work: in blue the model reproducing the in figure 13: the magnitude has only a qualitative ac-
experimental dipole moment of gaseous water, in red the one cordance with the data arising from molecular dynamics,
reproducing the same for liquid water (MC). and the model does not reproduce the position and the
height of the maxima (phase shift). On the other hand,
when parametrized by means of the projection onto a DD
[18] by means of very accurate molecular dynamics sim- model with q = qO and d = d/2 (see figures 13 for expla-
ulations with the SPC/E model with the dipolar cor- nations) the maxima are better reproduced, leading to
relations we got in this work. The formal definition of a quantitative agreement with the numerical data from
longitudinal dipole-dipole spatial correlation L(r) is [18] molecular dynamics. Thus, the DD projection revealed
to be a good way for recovering the structural properties
X ni (r) of SPC/E water. Some enhancements, however, could
1 1 X
L(r) = (µ̂i · rij µ̂j · rij ) (25) be realised: when looking at distances greater than 1
N i
ni (r) j=1 nm, the small long correlations are only ameliorated, but
not really well reproduced, since the field theoretical ap-
where µ̂i,j are are the normalized dipole moments of proach overestimates them; moreover, the longitudinal
molecules i and j respectively and rij is a unit vector susceptibility in Fourier space is really different for the
along the direction from molecule i to j. However, this field theory approach than the susceptibility of the pro-
definition, only refers to single dipole correlations. In or- jected case. A possible solution could be to introduce
der to compare the results given by molecular dynamics some new parameter in the model in order to manage
13
0.7
40 0.020
SPC/E 0.6 FT
0.015
30 FT 0.5 0.010
MD
0.4 0.005
L(r)
χ(q)
0.000
20 0.3 - 0.005
0.2 10 12 14 16 18 20 22 24
10 0.1
0.0
0
0 2 4 6 8 10 12 14 3 4 5 6 7 8 9
q(Å-1) r(Å)
(a) (b)
25
DD 0.6 0.010
FT
0.008
20 0.5
FT 0.006
MD
15 L(r) 0.4 0.004
χ(q)
0.002
0.3
10 0.000
0.2 10 12 14 16 18 20 22 24
5 0.1
0.0
0
2 4 6 8 10 12 14 3 4 5 6 7 8 9
q(Å-1) r(Å)
(c) (d)
0.010
DD 0.6 0.008 FT
15
0.5 0.006
FT MD
0.4 0.004
10
L(r)
χ(q)
0.002
0.3
0.000
0.2 10 12 14 16 18 20 22 24
5
0.1
0.0
0
2 4 6 8 10 12 14 3 4 5 6 7 8 9
q(Å-1) r(Å)
(e) (f)
FIG. 13. On the left: in panel a, in red the longitudinal susceptibility in Fourier space as computed with SPC/E model
4
numerical simulations, in black the longitudinal susceptibility as reported in equation 4, with the parameters α = 0.0107Å
2
and κl = −0.2Å reproducing the position and the height of the maximum; in panel c, in red the projection of the results
for the SPC/E model onto a DD model with partial charge q = qO and dOH = d/2 (see figure 3), in black the longitudinal
4 2
susceptibility as reported in equation 4, with the parameters α = 0.0213Å and κl = −0.288Å reproducing the position and
the height of the maximum of the projection; in panel e the same as in panel c but for a DD model with partial charge q = qH
4 2
and dOH = d and with the parameters α = 0.0341Å and κl = −0.361Å . On the right: in red the longitudinal dipole-dipole
spatial correlation function, reproduced from Fig. 2b of [18] for the SPC/E model, in black the same as computed in this work
by means of the field theory approach: in panel b the parameters are fixed using the SPC/E model, in panels d and f using
its projection onto the DD model c and e respectively.
14
and the remaining two (we are dealing with a fourth or-
der equation) need to be chosen on some large distance
from the origin, where the polarization is low and the
non-linear behavior of the solution could be, as explained
above, safely neglected in favour of the linear one. The
boundary conditions for equation 27, thus, read:
P 3 (z) P (z) P 00 (z) P (4) (z) where we introduced the notation Plin (z) to denote the
+ −κl χ bulk +αχbulk = 0 (28)
Pm 3 Pm Pm Pm linear solution in order to keep the notation clear and
we imposed null boundary conditions at infinity to rejoin
Practically, it is an equation in the new variable the bulk solution hPimf , which we expect to be reached,
P (z)/Pm : it is clear that for Pm P (z), for low po- since at infinity the surface effects should be negligible.
larizations, the first term in equation 28 tends to zero, The analytical solution is given by the solution of the as-
loosening more and more the nonlinear character of the sociated characteristic polynomial, whose complex roots
solution. In the opposite case, for Pm P (z), the equa- have both a real and an imaginary part. In this way the
tion will tend to increase the nonlinear effects. This is solution to equation 29 will have the form
why Pm can be used as a benchmark dividing the linear
and the nonlinear case. Recalling Appendix B, moreover, Plin (z) = C1 e−az cos(bz) + C2 e+az sin(bz)+
to solve the equation in the right boundary condition, we
have to impose two conditions in the origin of the axes + C3 e−az sin(bz) + C4 e+az cos(bz) (30)
15
6 1
4
z0 (λ units)
0.5
0.6
PD0 (z) − P (z)
χ(z) = lim (33)
D0 →0 D0 0.4
being P (z) the solution to the homogeneous nonlinear
0.2 D0 = Pm /102
equation 27 and PD0 (z) its non homogeneous counter-
D0 = Pm /103
part. In the absence of a surface the constant bulk
2 −1 0
value would be χ = χbulk = (2γPm ) . Since the non- 0 2 4 6 8 10 12 14
homogeneous solutions PD0 (z) in the D0 → 0 limit are al- z (Å)
most identical to the homogeneous ones (or, better, their
difference is not clearly visible by means of their plots) we
will not represent them; however, the numerical discrep- FIG. 18. Longitudinal susceptibility computed using formula
33 for two different values of D0 for Ps /Pm = 2. The param-
ancy between PD0 (z) and P (z) results in the longitudinal
eters are the same as in figure 17.
susceptibility function.
From a strictly computational point of view it is not
clear what should be the meaning of the limit D0 → 0: if
D0 is too small we could encounter some precision prob- 1
lems, since the homogeneous and the non-homogeneous
equation would be, numerically, almost identical, and we 0.8
would rely excessively in the numerical precision of the
machine. The results for two different displacement fields
χ(z)
0.6
are reported in figure 18. Seen the convergence of the
plots near the surface and the agreement with a con- 0.4
stant value of χbulk far from the surface, we decided to
take D0 = Pm /103 as the standard value for all the solu- 0.2 6λ
tions shown. Further improvement is needed in order to 8λ
ensure convergence even in intermediate points between 0
these two limits. Supplementary doubts rising from the 0 2 4 6 8 10 12 14 16 18 20
goodness of the far field approximation mentioned above z (Å)
are dispelled by comparing the susceptibility for two dif-
ferent values of the far field limit: in figure 19 the two FIG. 19. Longitudinal susceptibility for different far field ap-
result are really well superimposed, meaning that the ap- proximations for Ps /Pm = 2.The parameters are the same as
proximation method is correct and that the two solutions in figure 17.
converge to the same result. Final results for different
17
Pel (z)
χ(z)
0.6
Pel (z) = Pel (0) = Pel (0) f (z)
Pel (0)
Ps
0.4 Pm = 1
Ps
Pm = 2
it is easy to see that, in the absence of the external dis-
Ps placement field D0 ,
0.2 Pm = 4
Ps
Pm = 10 k 1
0 Htot (Pel (0)) = (Pel (0)−Ps )2 + Pel (0)2 A[f (z)] (34)
0 1 2 3 4 5 6 7 8 2 2ε0
z (Å) being
values of Ps /Pm are drawn in figure 20. Increasing the The harmonic potential stands for a cost function linked
value of Ps creates and widens the zone of vanishing sus- to the surface polarisation. The numerical computation
ceptibility near the surface, as expected from the exper- of equation 35, however, leads to convergence problem
imental results shown in figure 1 from [2], opposed to due to the first Coulombic integral, which is both non-
the bulk behavior far from the surface, in which case the local and improper. The way to get rid of this issue
susceptibility would be constant and equal to χb . This is explained in Appendix D, where, making use of the
means an increase in the size of the portion of fluid not finite temperature formalism, we manage to write the
responding to the external perturbation. From a physical total energy in the form
point of view, that’s because the more the external field 2
ε0 δUP [P]
Z
is intense, the greater becomes the number of molecules H[P] = dr + UP [P] (36)
involved into forced polarization that won’t be able to 2 δP
linearly respond, as if they were frozen. They are, in this
model, effectively frozen at the surface, since we assumed with
that for z = 0 polarisation was fixed to a certain value 1
Z
Ps : it’s so explained the reason of a null susceptibility at UP [P] = dr KP2 (r)+
the surface. 2ε0
2 2
+ κl ∇ · P(r) + α ∇ ∇ · P(r)
D. Elastic Interaction with the Surface: effects on
P and χ More explicitly
1
Z
As stated in the previous parts, imposing a fixed polar- H[P] = dr α∇∇ · (∇∇ · P)+
ization at the surface represents a quite strong assump- 2ε0
2
tion. Thus, we will try to relax this condition introduc-
ing an elastic potential involving the polarization value − κl ∇∇ · P + KP(r) + UP [P]
at the surface, endowing this latter with the possibility of
fluctuating around an equilibrium point. We propose to and it becomes transparent that equation 36 does not
model the interaction between the fluid and the surface include nonlocal terms in 1/r as before. In planar sym-
using an harmonic potential that reads: metry, thus,
k 2
(Pel (0) − Ps )2
Z
Hsurf =
2 A[f (z)] = dz Kf (z) − κl ∂z2 f (z) + α∂z4 f (z) +
ε0
the module of the functional A[f (z)] when performing
minimization of expression also containing this term. In Note that, in equation 39c, we found an expression for
this way the minimum for the expression 34 is easily com- the electric field in the bulk:
puted:
D0 K D0
k Ps = = Ebulk
Pel (0) = (37) ε0 1 + K ε0 εbulk
2|A| + k
If we then define the functionals
(with k = 2ε0 k). In the limit k → ∞ we find Pel (0) = Z ∞
Ps : this is in agreement with the previous frozen picture, B[f (z)] = dzf (z)
since an infinitely steep potential corresponds to fixing 0
a surface polarisation. In the opposite case, for k → 0,
we recover the bulk solution Pel (0) = 0, which stands for and
the equilibrium point of the equation 36. ∞
ε0 δUP [P (z)]
Z
In presence of an external perturbation the equation C[f (z)] = dz
Pz=0 0 δP (z)
to solve in order to find the polarization is
D (2) D (4)
we can easily perform the derivation of Htot with respect
(1 + K)PelD0 (z) − κl Pel 0 (z) + αPel 0 (z) = D0 (38) to Pel (0) in order to find its minimum:
with D0 the external displacement field. Of course this
is not but the non homogeneous version of equation 29, D0
k + Pel (0) − Ps + 2A[f (z)]Pel (0)+
thus, its solution is the solution to the relative homo- 1+K
geneous equation plus a particular solution, that can be KD0
+ {B[f (z)] + 2C[f (z)]} − 2D0 B[f (z)] = 0
cast among the constant solution satisfying equation 38: 1+K
D0 (with k = 2ε0 k). The solution is
PelD0 (z) = + Pel (z)
1+K
kPs 2D0 B
Note that, now, Pel (z) has not the same value at the sur- Pel (0) = + +
face Pel (0) of equation 37. The addition of the constant k + 2|A| k + 2|A|
term modifies the functional form Htot : 1 2
kD0 KD0
Htot = Hbulk + Hsurf + Hinter 1+K 1+K (B + 2C)
− − (40)
k + 2|A| k + 2|A|
We analyze separately the different terms with respect to 3 4
the unperturbed case.
2 where term 1 is the unperturbed part, term 2 the inter-
k D0 2 k D0 action contribution with the medium, term 3 the inter-
(P (0) − Ps ) −→ + Pel (0) − Ps (39a)
2 el 2 1+K action of the perturbation with the surface and term 4
19
D0 z (Å)
PelD0 (z) = Pel (0)f (z) + (41)
1+K
FIG. 21. The polarisation between two surfaces, expressed in
4
with Pel (0) given by equation 40. Moreover, an analytic units of the ratio PPm
s
. The used parameters are α = 0.0107Å
expression for the dielectric susceptibility in linear ap- and κl = −0.2Å .
2
χ(z)
2B − k + K(B − 2C)
1 1
χlin (z) = χb + f (z)
k + 2|A| 1+K
(42)
A figure relative to this quantity in a simpler case in given 0
in the following section.
0 2 4 6 8 10 12 14 16 18 20
z (Å)
E. Polarization confined between two surfaces
FIG. 22. Longitudinal dielectric susceptibility for a system
Some interesting results, for the linear solution, regard- confined between two surfaces being 20 nm far for different
ing the surface harmonic potential arise from the sym- values of k. The parameters’ values are the same as in figure
metric case, in which water is confined between two sur- 21.
faces, where the polarization, for hypothesis, takes the
same value but in opposite directions (thus, with oppo-
site sign). In the linear approximation the equation to A[f (z)], by definition, is always even. The complete sus-
solve is still equation 29; however, the boundary con- ceptibility, as shown in Appendix C in equation C1, can
ditions are no more the same, having no conditions at now be written in the simple form
infinity but at both the surfaces. If we place the surfaces
in z = 0 and z = z1 they read: k
χlin (z) = χb 1 − f (z) (43)
k + 2|A|
00
Plin (z)|z=0 = Ps Plin (z)|z=0 = 0
The plots of χlin for three typical values of k are shown in
figure 22. See how the susceptibility tends to the constant
00
Plin (z)|z=z1 = −Ps Plin (z)|z=z1 = 0 bulk value as k → 0, when the surface term becomes
less and less significant. Note that for k → ∞ we find
The solution, of course, will always have the form ex- equation
pressed in equation 30, but the constants C1 , C2 , C3 and
C4 will be more difficult to determine, since there are χlin (z) = χb [1 − f (z)] (44)
no vanishing boundary conditions. However they can be
easily found using Mathematica. A solution for the sur- that’s the susceptibility for the linear solution with a
faces being 20 nm far is shown in figure 21. In this case, fixed polarization at the surface, while for k → 0 we
moreover, we can reach a simple solution for the dielec- find the constant value χ(z) = χb , that’s, as expected,
tric susceptibility even assuming an harmonic potential the limit in the linear in bulk approximation. The pic-
at the surfaces. Functionals B[f (z)] and C[f (z)], in fact, ture of an harmonic potential at the surface, thus, seems
are now integrals of an odd function (equation 30 and to be coherent with the known results for the linear ap-
its second and fourth derivatives) over an even integra- proximation. Results for a single surface are reported in
tion domain, leading both, thus, to a null value, while Appendix C.
20
Z
1
Z Z Z
0 0
1 × 4π = dr dr0 ∇r · dq1 eiq1 ·r P̃(q1 ) ∇r0 · dq2 eiq2 ·r P̃(q2 ) × dq3 eiq3 ·(r−r ) 2 2 =
2π q3
1
Z
0
= − dr dr0 dq1 dq2 dq3 q1 · P̃(q1 ) q2 · P̃(q2 )eir·(q1 +q3 ) eir ·(q2 −q3 ) 2 2 =
2π q3
q1 · P̃(q1 ) q2 · P̃(q2 ) q · P̃(−q) q · P̃(q)
Z Z
= − dq1 dq2 dq3 δ(q1 + q3 )δ(q2 − q3 ) 2 2 = dq 4π (A5)
2π q3 q2
Z Z Z Z
0
3= dr κl (∇ · P(r))2 = dr κl ∇r · dqeiq·r P̃(q) ∇r · dq0 eiq ·r P̃(q0 ) =
Z
0
= dr dq dq0 κl iq · P̃(q) iq0 · P̃(q0 ) eir·(q+q ) =
Z Z
0 0 0 0
= − dq dq κl δ(q + q ) q · P̃(q) q · P̃(q ) = dq κl q · P̃(q) q · P̃(−q) (A7)
Z Z Z Z
0
dr α (∇2r · P(r))2 = dr α ∇r ∇r · dqeiq·r P̃(q) ∇r ∇r · dqeiq ·r P̃(q0 ) =
4=
Z Z Z
0
= dr α − dq q q · P̃(q) eiq·r − dq0 q0 q0 · P̃(q0 ) eiq ·r =
Z Z
= α dq dq0 δ(q + q0 ) q · q0 q · P̃(q) q0 · P̃(q0 ) = α dq q 2 q · P̃(q) q · P̃(−q)
(A8)
longitudinal one and the transverse one: Finally, we can write the explicit formula for the longi-
tudinal susceptibility:
qi qj qi qj
χ̃ij (q) = χ̃k (q) 2 + χ̃⊥ (q) δij − 2 (A14)
q q 1
χ̃k (q) = (A16)
Thus, 1 + K + κl q 2 + α q 4
4 qi qj qi qj
h i
χ̃−1 2
ij (q) = 1 + K + κl q + α q + K δ ij − =
q2 q2 Appendix B: Functional Derivatives
qi qj qi qj
= χ̃−1 −1
k (q) q 2 + χ̃⊥ (q) δij − q 2 (A15)
We need to compute the functional derivative of
∇r · P(r)∇r0 · P(r0 )
2 2
2
Z Z
0 2 2
2ε0 H[P] = drdr + dr γ P (r) + P0 + κl ∇ · P(r) + α ∇ ∇ · P(r) (B1)
4π|r − r0 |
1 2 3 4
iq0 ·(r−r0 )
δH1 1 ∇r0 · P(r0 ) (F T ) 1 0e
Z Z Z Z
= − ∇r · dr0 = − ∇r · dr 0
∇ r 0 · dq e iq·r
P̃(q) · dq =
δP ε0 4π|r − r0 | ε0 (2π)3 q 02
iq0 ·r
1 ir0 ·(q−q0 ) e
Z
0 0
= − ∇r · dr dq dq iq · P̃(q) e =
ε0 (2π)3 q 02
0
1 eiq ·r 1 1
Z Z
= dq dq0 δ(q − q0 )q · q0 P̃(b) 02 = dqP̃(q)eiq·r = P(r) (B2)
ε0 q ε0 ε0
Less generally, in the chosen symmetry: It is clearly visible how the two limits tend one to the
other: on one hand we find the nonlinear solution with
Z
∂(P + η)
2 Z
∂P
2 frozen polarization at the surface in the Ps /Pm → 0 limit;
dz − dz = on the other we have the linear solution with harmonic
∂z ∂z
Z ∞ potential at the surface in the k → ∞ limit. The fact
∂P ∂η that both the cases tend to the same limit is a proof of
=2 dz · =
0 ∂z ∂z the rightness of our assumptions and of the formalism
∞ Z ∞
∂2P coherence.
H ∂P
= ·η − 2 dz 2 · η (B14)
H
H
∂z H
0 0 ∂z 4
linear for k → ∞
H
Since η need to be chosen with the constraints Ps /Pm = 0.01
Ps /Pm = 0.05
η[0] = η[z → ∞] = 0 (B15) 2 Ps /Pm = 0.1
χ(z)
in order not to modify the boundaries of the problem,
one gets the same as in equation B4. More interesting
conditions arise from the fourth term in equation B1: 0
2 2
∂ 2 (P + η) ∂2P
Z Z
dz − dz = 0 2 4 6 8 10 12
∂z 2 ∂z 2
Z ∞ z (Å)
∂2P ∂2η
=2 dz 2 · =
0 ∂z ∂z 2 FIG. 23. The longitudinal dielectric susceptibility for: in
2 ∞ 3 ∞ Z ∞ blue the linear solution with harmonic potential at the surface
∂ P ∂η H ∂HP ∂4P
=2 2
· − 2 H·H
3
η + 2 dz 4 · η for k → ∞ (equation C2); in the other shades the complete
∂z ∂z 0 ∂z HH0 0 ∂z nonlinear solution with constrained conditions at the surface
(B16) for Ps /Pm → 0. The used parameters are α = 0.0107Å ,
4
2 −2
κl = −0.2Å and Pm = 0.31CÅ .
If we want the boundary terms to go to zero we need to
impose the condition
∂ 2 P
=0 (B17) Appendix D: Coulombic Integral
∂z 2 z=0
so that the mean field equation is satisfied: In order to solve equation 35 it is useful to recall the
finite temperature formalism, as shown in reference [8]:
δH the partition function, constrained by the Gauss’s law,
=0 (B18) reads
δP Pmf Z Y
Z = D[P]D[D]e−β U [D,P] δ(∇ · D − ρ)
r
Appendix C: Susceptibility: linear case with an
harmonic potential on a single surface with
ε0
Z
U [D, P] = drE2 (r) + UP [P]
In this Appendix we’ll briefly show the behavior of the 2
nonlinear solution in the limit PPms → 0. In such a case we
Equation 1b and
found very different behavior as regards from the results Z
shown in figure 20, as it is shown in figure 23. In the same
δ(∇ · D − ρ) = D[iϕ]eiϕ(∇·D−ρ)
figure the analytic behavior of the linear approximation
is shown for the case harmonic potential at the surface in
enable to write the partition function as
the k → ∞ limit: this is easily obtainable recalling equa-
D0 Z
tion 33, where Plin (z) and Plin (z) are given, respectively, R
Z = D[P]D[D]D[iϕ]e−β U [D,P]+ r iϕ(∇·D−ρ)
by equations 41 and 37. The final result reads
1
2B − k + K(B − 2C)
Equation 1b and the rescaling ϕ → βϕ give
χlin (z) = χb + f (z)
k + 2|A| 1+K Z
β R 2
R
(C1) Z = D[P]D[iϕ]e−βUP [P]− 2ε0 r P −β r iϕρ ×
and for k → ∞ we get Z
β R 2
× D[D]e− 2ε0 r [D −2D·(P−iε0 ∇ϕ)]
χlin (z) = χb [1 − f (z)] (C2)
24
The integral in D field can now be performed, since by a. Gradient Descent The gradient descent method,
completing the square one obtains a simple Gaussian in- first proposed by Cauchy [28] in 1847, is an algorithm
tegral. Defining φ = iϕ one gets the final results whose aim’s to find the local minimum of a given func-
Z R ε0 (∇φ)2 tion. We’ll briefly explain its mechanism in order to
Z = D[P]D[φ]e−β{UP [P]+ r [ 2 +φρ+P·∇φ]} understand the underlying principle governing the used
code, where instead of looking for a minimum we’re look-
The mean field equation are then obtained by performing, ing for the zeros of a function knowing only some local
as usual, the functional derivative of the exponent with properties.
respect to P and φ. In this way we get Given a scalar function f (x) and a beginning point x0
δUP we evaluate the former in x0 and then in the same point
−∇φ = plus a small increment ε, so that we are enabled to com-
δP
where pute the first derivative of the function at that point.
Assuming a linear dependence (consciously knowing to
1
Z
be wrong) of the function not only in the neighbourhood
UP [P] = dr KP2 (r)+
2ε0 of x0 we can write
2 2
f (x) = f (x0 )+∂x f (x)x=x (x−x0 )+O (x−x0 )2 (E1)
+ κl ∇ · P(r) + α ∇ ∇ · P(r) (D1) 0
The total energy can now be written as In this approximation, if we are looking for the zero x1
2 such that f (x1 ) = 0, we get
ε0 δUP [P]
Z
H[P] = dr + UP [P] (D2)
2 δP f (x0 )
x1 = x0 − (E2)
f 0 (x0 )
Appendix E: Convergence Issues Of course, unless we are extremely in luck, when evalu-
ating the function f (x) in x1 we’ll find a nonzero value:
As mentioned before, even trying to numerically solve thus, the point x1 will be again used instead of x0 in
the complete problem, integration algorithms fail to solve equation E2 in order to get a new point x2 and so on,
differential equations on large intervals, due to their un- until the proceeding have reached a desired precision; the
derlying mechanism for attacking boundary values prob- ith passage reads:
lems, whose core relies on the interpolation of exponential
functions and gives rise to diverging solution if attempted f (xi )
xi+1 = xi − (E3)
to be used beyond their capabilities. f 0 (xi )
Thus, an effective method would be to split the integra-
tion domain into a chosen number of small sub-intervals, Of course this algorithm can be generalized to multi-
where the number and the size of the intervals are to be variable functions f (x) and vector functions F(x) by us-
chosen in order the integration algorithm not to diverge ing gradient and Jacobian matrix as an abstraction of the
in any of them. derivative concept:
However, this procedure is far from being without any
xi+1 = xi − ∇f −1 (xi )f (xi ) (E4a)
related issue, the most significant of them being that we
don’t know at all the boundary condition needed in the
intermediate intervals to solve the differential equation. xi+1 = xi − J−1 (F(xi )) · F(xi ) (E4b)
As a first attempt we could start by imposing in these
unknown intermediate points the function and its first Whereas in the one dimensional case the only choice the
derivative to be equal to the values of Plin (z) and of its algorithm had to take was to go against the derivative
first derivative in such points. Of course this condition is direction, in multi-dimensional case the use of the gra-
completely arbitrary, however could help us starting from dient (hence the name of the algorithm) and, even more
a good point point in order to limit convergence time to abstractly, of the Jacobian matrix ensures the algorithm
the real solution. This latter can be reached if we recall to find the direction of steepest descent to the minimum
that other conditions are to be imposed. In particular, (zero) of the function. This latter case will be the one
the differential equation being of fourth order, the solu- we’ll recall to for applying the algorithm to our problem.
tion need to be C 3 , that’s to say that its derivatives have b. Application How does this algorithm apply to our
to be continuous up to the third order. necessities? As mentioned before our problem is the dis-
The solution of the equation by numerical approach, in continuity of the second and third derivatives in the junc-
our case, doesn’t ensure this condition to be satisfied in tions of the various segments. If we store in a vector of
the junction points between the various fragments: this dimension 2(N − 1) (being N the number of segments
apparent problem could help us to correctly link the nu- we divided the whole domain into and N − 1 the number
merical solutions; in particular we’ll appeal to the gradi- of internal points in the interval) the differences between
ent descent algorithm. the left and right derivatives in those points, we can look
25
at this vector as the vector function to minimize, or, bet- proceed with the gradient descent algorithm. At the fol-
ter, to put to zero, whereas the vector variable of such a lowing step the solution will be computed with the new
function are the unknown 2(N − 1) boundary conditions boundary condition stored in the x2 vector and so on,
to impose for solving the equation on the fragments. until all the entries of the vector function are null (under
Slightly varying the boundary conditions of an amount ε some tolerance threshold).
we are enabled to get the Jacobian matrix of the “error In this way we have found the true numerical solution,
vector function” and, thus, using equation E4b, we can which is independent of the initial intermediate boundary
conditions and continuous up to the third derivative.