JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 115, B12205, doi:10.

1029/2010JB000843, 2010

Silicon self‐diffusion of MgSiO3 perovskite by molecular dynamics

and its implication for lower mantle rheology
Yosuke Ito1 and Mitsuhiro Toriumi1
Received 3 January 2010; revised 20 June 2010; accepted 2 July 2010; published 4 December 2010.
[1] The pressure effect of silicon self‐diffusion of MgSiO3 perovskite was investigated by
molecular dynamics (MD). The viscosity variation of MgSiO3 perovskite in the lower
mantle was derived by the Nabarro‐Herring (Herring‐Nabarro) model. For the MD
calculation, the spontaneous jumping of atoms by self‐diffusion was reproduced without
using artificial forces, and the consistency of migration enthalpy (Hm*) with
experimental data was improved. The results showed that migration enthalpy increases
monotonically with increasing pressure. The viscosity of MgSiO3 perovskite in the lower
mantle increases monotonically with increasing depth. The obtained depth profile is
distinguishable from that of MgO periclase and can be utilized to determine which
mineral dominates the lower mantle rheology. Depending upon the assumed shape of the
depth profile for lower mantle viscosity, we considered the dominant mineral as (1)
MgSiO3 perovskite for the monotonic shape case or (2) MgO periclase for the hill shape
case that has the highest‐viscosity zone in the middle of the lower mantle.
Citation: Ito, Y., and M. Toriumi (2010), Silicon self‐diffusion of MgSiO3 perovskite by molecular dynamics and its
implication for lower mantle rheology, J. Geophys. Res., 115, B12205, doi:10.1029/2010JB000843.

1. Introduction potential model used by Wright and Price [1993]. Karki

and Khanduja [2007] sought to examine the incon-
[2] Understanding the self‐diffusion of MgSiO3 perov-
sistencies by employing ab initio pseudopotential; how-
skite is important, largely because of the implications for
ever, the inconsistency still remained (657 kJ/mol at 30 GPa).
diffusion creep, which is considered to be the dominant
Ammann et al. [2008] attempted to reduce the inconsis-
mechanism of creep deformation in the lower mantle [Karato
tency by searching surrounding atoms’ configurations that
et al., 1995]. Diffusion creep is dominated by the diffusion
minimized the heights of energy barriers and reported an
of the slowest diffusion species. In MgSiO3 perovskite,
improved result (453 kJ/mol at 25 GPa), although there was
silicon is one of the most likely candidates for the slowest still a mismatch between experiments and simulations.
species since it has been confirmed as the slowest diffusion
[3] In order to resolve the inconsistency, we tried another
species in many silicates (reviewed by Béjina et al. [2003]).
approach. Atoms spontaneously jump into vacancies with
Reported experimental measurements for silicon diffusion
varying surrounding atoms’ configurations during natural
in MgSiO3 perovskite [Yamazaki et al., 2000; Dobson et al.,
self‐diffusion, though the previous simulations did not
2008] are limited to the uppermost pressure (25 GPa) area
reproduce the process. Instead, these simulations assumed
in the lower mantle. Because of the difficulty of performing
some atomic configurations of activated states (“saddle
experiments at high pressures, molecular simulations are
point configuration”) and tried to derive heights of energy
considered indispensable for investigating the rheology of
barriers by pushing atoms into vacancies based on the
the lower mantle. Wright and Price [1993] sought to
thermally activated process model. According to the model
determine the diffusive properties of MgSiO3 perovskite [e.g., Heitjans and Kaerger, 2004], the jump rate w with
using lattice dynamics. They derived activation enthalpy
which an atom jumps into an empty neighborhood site is
(Hm*) obtained from the height of energy barriers: the
described as w = v0 exp (−GM/RT), where GM is the height
increase in energy migrating atoms from their sites to saddle of energy barrier, R is the gas constant, and T is temperature.
points. However, there were inconsistencies that could not
The preexponential factor v0 denotes attempt frequency. The
be ignored between experiments (336 kJ/mol at 25 GPa
self‐diffusion coefficient D is described as D = (1/6)wl2,
[Yamazaki et al., 2000], 347 kJ/mol at 25 GPa [Dobson et al.,
where l is the jump distance. There are two difficulties in
2008]) and simulations (733 kJ/mol (0 GPa]) [Wright and
this approach. First, although definition of attempt fre-
Price, 1993]). These inconsistencies could be due to the
quency is given in the harmonic transition theory, the way to
apply the definition to more complex actual systems is
1
Department of Complexity Science and Engineering, Graduate School
unclear. Generally [e.g., Heitjans and Kaerger, 2004], the
of Frontier Science, University of Tokyo, Tokyo, Japan. attempt frequency is assumed to be in the order of the Debye
frequency. Ita and Cohen [1997] assumed that the attempt
Copyright 2010 by the American Geophysical Union. frequency is equal to the lowest frequency peak in the
0148‐0227/10/2010JB000843

B12205 1 of 9
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE
spectrum of the trajectories of the migrating atom projected
onto the shortest path of the vacancy. Next, a specific saddle
point configuration needs to be assumed. Ammann et al.
[2008] tried to search for a “better” saddle point configu-
ration, though some overestimations of activation enthalpies
still remained. On the other hand, our approach just re-
produces the spontaneous diffusion itself explicitly by
molecular dynamics (MD). The advantage of the approach
is that what we see is what we get: it directly derives the
self‐diffusion coefficient without assuming attempt fre-
quency and saddle point configuration. There is another
complication in this approach. Since the running time of
MD simulations is short and the frequency of spontaneous
jumping is low, the precision of results is usually limited.
However, Ito and Toriumi [2007] showed that sufficient
precision could be obtained with long MD running time.
[4] In summary, we applied molecular dynamics, which
reproduces the spontaneous jumping of atoms explicitly, to
the silicon self‐diffusion of MgSiO3 perovskite. We em-
ployed four pressure conditions (40, 60, 80, and 100 GPa)
corresponding to the lower mantle and derived self‐diffu-
sion coefficients. The depth profiles of the viscosities of
lower mantle minerals were estimated from self‐diffusion
coefficients [Yamazaki and Karato, 2001; Ito and Toriumi,
2007]. We also estimated the depth profile of MgSiO3
perovskite’s viscosity in the lower mantle from the derived
self‐diffusion coefficients.
2. Methodology
2.1. Molecular Dynamics
[5] In the MD simulation, the Buckingham model was
adopted to define interatomic potential. The form of the
model is as follows:
 

 Zi Zj Rij cij
Uij Rij ¼ þ bij exp   6; ð1Þ
Rij
ij Rij
where Rij is the interatomic distance. The parameters (ZMg =
+1.9104; ZSi = +2.9043; ZO = −1.6049; bMg‐O = 1041.435 eV;
rMg‐O = 0.2866 Å, cMg‐O = 0, bSi‐O = 1137.028 eV,
rSi‐O = 0.2827 Å, cSi‐O = 0, bO‐O = 2023.8 eV, rO‐O =
0.2674 Å, cO‐O = 13.83 eVÅ6) are taken from Oganov et al.
[2000]. The model’s reproduction of crystal structure, elastic
properties, and thermal expansion was examined by Oganov
et al. [2000]. We also examined compression, thermal
expansion, and orthorhombic distortions and confirmed that
the model reproduced these properties. Details of the ex-
aminations are given in the next section. Ab initio electronic
state calculations (reviewed by Karki et al. [2001]), which
would provide more precise results, require more calcula-
tions than the Buckingham model. Our study targeted
spontaneous migrations of atoms in MD calculations. Ito and
Toriumi [2007] reported that 107 steps (20 ns) were needed to
maintain the precision of self‐diffusion coefficient results.
Therefore, we considered that ab initio electronic state cal-
culations were not appropriate for this study.
[6] The MD system was composed of 5 × 5 × 4 unit cells,
which contained 4 Mg atoms, 4 Si atoms, and 12 O atoms
per unit cell with Pbnm symmetry. To induce silicon’s
diffusion, we introduced one silicon vacancy. Two oxygen
vacancies were also introduced to maintain the total charge
2 of 9
B12205
neutrality. As a result, the MD system contained 1997
atoms. Periodic boundary conditions were applied for all
(x, y, and z) directions to approximate the actual size of the
crystal. The potential that an atom feels was given by the
infinite summation of potentials between one atom and all
the other atoms. The Coulomb term (the first term of equation
(1)) was summed by the Ewald method [Ewald, 1921].
Summations of short‐range terms (the second and third
terms) were cut off at 7.5 Å. The Verlet algorithm integrated
Newton’s equations of motions with 2 fs intervals. Isothermal
and isobaric (NPT) ensembles were obtained by the scaling
method [Woodcock, 1971]. The MXDTRICL program
(http://www.sccj.net/download/JCPE) was used for MD
calculations.
[7] The method for obtaining self‐diffusion coefficients
was based on the work by Ito and Toriumi [2007]. The self‐
diffusion coefficient (D) was derived from the Einstein
equation, which has the following form:
1 dD E
D¼ jRn j2 ; ð2Þ
6 dt
where Rn indicates the positional vector of the nth atom while
the angle brackets indicate the average of all atoms. Positional
vectors were given by the trajectories of atoms recorded
during MD calculations. Initial relaxations were performed
for 2 ns (106 steps) to achieve thermal equilibrium. After the
initial relaxation, 2 ns (106 steps) MD calculations were
repeated 10 times and an average of 10 self‐diffusion coef-
ficients was adopted. During the MD calculation, the
vacancy concentration was fixed since there was no external
sink or source of vacancies. Note that the vacancy concen-
tration is artificial and the derived self‐diffusion coefficients
should be calibrated before applying to the Earth’s interior by
equation (5). Because the vacancy concentration is fixed,
activation enthalpies of the self‐diffusion coefficients ob-
tained by MD simulations correspond to migration enthalpies
(Hm*). The Arrhenius relationship is expressed as
 
D ¼ D0 exp Hm* =RT ; ð3Þ
where D is the self‐diffusion coefficient derived by the MD
simulation, D0 is the preexponential term, R is the gas con-
stant, and T is temperature.
2.2. Estimation of Viscosity
[8] In a material that flows by creep dominated by inner-
crystal diffusion, viscosity and self‐diffusion coefficients are
related by the Nabarro‐Herring (Herring‐Nabarro) model
[e.g., Poirier, 2000]:


 ¼ Að RT =V Þ G2 =Deff ð4Þ
where h is viscosity, A is the geometrical constant, R is the
gas constant, T is temperature, V is molar volume, G is
grain size, and Deff is the effective self‐diffusion coeffi-
cient, which is approximately equal to that of the slowest
diffusing species. To relate Deff to the self‐diffusion
coefficient D of the MD calculation, we first assumed the
slowest diffusing species was Si. Note that Mg would also
be examined further because it was shown that Fe‐Mg
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE B12205

where Ceff and C are Si vacancy concentrations of MgSiO3

Figure 1. Comparison of isothermal (300 K) compressions
of MgSiO3 perovskite. The solid squares indicate results of
this study. Experiments were performed by diamond anvil
cell (DAC) (Yagi et al. [1979], open squares; Mao et al.
[1991], plusses; and Fiquet et al. [2000], crosses) and shock
compression (natural enstatite [Gong et al., 2004] shown as
open triangles).
perovskite in the lower mantle and the MD simulation,
respectively. Variations in grain size certainly influence
viscosity. The Nabarro‐Herring (Herring‐Nabarro) form
(equation (4)) suggests a square linear relationship between
grain size (G) and viscosity (h). However, because of a lack
of information for grain size distribution in the lower mantle,
we assumed it to be uniform. Yamazaki et al. [1996] reported
a significant slow grain growth for MgO periclase and
MgSiO3 perovskite aggregates (G / t1/10.6 for MgSiO3
perovskite and G / t1/10.8 for MgO periclase, where G and t
are grain size and time, respectively) at the pressure and
temperature condition near that of the lower mantle (1700 K
and 25 GPa), implying that nonuniformities for grain sizes in
the lower mantle are small. Two unknown constants, Ceff and
A, remained. Because of the difficulty of estimating their
correct values, we introduced a proportional relationship.
From equations (4) and (5) and the assumption, we derived a
relationship between viscosity h and the self‐diffusion
coefficients of MD simulations D as follows:
 / ðV =RT Þð1=DÞ: ð6Þ
Since equation (6) derives viscosity variations, results of
analysis based on equation (6) were shown as depth varia-
tions (see Figure 10), not absolute values.

interdiffusion is as slow as Si diffusion at the uppermost 3. Results

temperature and pressure conditions of the lower mantle
[Holzapfel et al., 2005]. Next, we assumed that the Si [9] First, we performed MD calculations for (1) isother-
vacancy was formed by impurities (i.e., extrinsic), and the mal (300 K) compressions, (2) isobaric (25 GPa) thermal
vacancy concentration does not depend on temperature and expansions, and (3) orthorhombic distortions to examine the
pressure. Another possibility for the mechanism of vacancy potential model employed in this study. The results were
formation is a thermal one (i.e., intrinsic). Two (intrinsic compared with experimental results in Figures 1, 2, and 3,
and extrinsic) mechanisms can be characterized by acti- respectively. In Figure 1, we considered experiments on
vation enthalpies (H*). For the intrinsic case, the activation pure synthetic MgSiO3 perovskite [Yagi et al., 1979; Fiquet
enthalpy is given by the total sum of the migration (Hm*) et al., 2000] and (Mg,Fe)SiO3 perovskite of Fe/Mg = 0–0.2
and formation (Hf*) enthalpies as H* = Hf*+ Hm*,
whereas in the extrinsic case the activation enthalpy is
simply given by the migration enthalpy Hm* as H* = Hm*.
The reported formation enthalpy (Hf*) of SiO2 Pseudo‐
Schottky (600 kJ/mol at 25 GPa [Karki and Khanduja,
2006]) is much higher than the activation enthalpies ob-
tained in experiments [Yamazaki et al., 2000; Dobson et al.,
2008], so we considered that the intrinsic case is unlikely.
We also assumed that the enhancement of diffusion by grain
boundaries (i.e., grain boundary diffusion) could be ne-
glected. Grain boundary diffusion enhances the effective
diffusion, especially in cases where grain sizes are small. In
cases where grain boundary diffusion dominates in a mate-
rial, the other creep model (Coble creep) rather than the
Nabarro‐Herring (Herring‐Nabarro) model dominates the
flow. A reported comparison between lattice and grain
boundary diffusions [Yamazaki et al., 2000] showed that the
enhancements were not so significant; lattice diffusion is
considered to be dominant in cases when grain sizes are
larger than a few tens of micrometers. Using these assump-
tions, the Deff of MgSiO3 perovskite in the lower mantle and Figure 2. Comparison of isobaric (25 GPa) thermal expan-
the self‐diffusion coefficient derived by MD simulations in sions of MgSiO3 perovskite. The solid squares show data of
this study (D) are related by this study, and the crosses show data of Funamori et al.
[1996]. The data of Funamori et al. [1996] were picked
Deff ¼ ðCeff =C ÞD; ð5Þ
up at 25 ± 0.5 GPa.

3 of 9
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE
Figure 3. Pressure dependences of orthorhombic distor-
tions (ratio of basic cell lengths b/a and c/a) of MgSiO3
perovskite at ordinary temperature (300 K). The crosses
and the plusses show experimental data [Mao et al., 1991;
Fiquet et al., 1998], and the circles and the squares show
theoretical data [Wentzcovitch et al., 1995; Marton et al.,
2001], respectively. Pressures of Fiquet et al. [1998] are
given by averages of multiple pressure scales (Pt and ruby
fluorescence). The 50 GPa data of Marton et al. [2001] are
given by 900 K data instead because of a lack of 300 K data.
Temperatures of energy calculations [Wentzcovitch et al.,
1995] were corresponding to 0 K because this calculation
does not treat finite temperatures.
[Mao et al., 1991], performed using diamond anvil cells
(DAC). Another experiment [Gong et al., 2004] consisted of
the shock compression of natural enstatite. Observed dif-
ferences of molar volumes between this study and the ex-
periments [Yagi et al., 1979; Mao et al., 1991; Fiquet et al.,
2000] were under 3% at 25–70 GPa and under 5% at 40–
4 of 9
B12205
120 GPa [Gong et al., 2004]. Since the differences are
smaller than the difference between the experiments them-
selves, we believe that the model reproduced the compres-
sion very well. In Figure 2, results from this study are
compared with those of Funamori et al. [1996], who used a
multianvil apparatus. Better agreement (agreement within
3% of the values and their first derivatives (i.e., thermal
expansibility)) was confirmed. Orthorhombic distortions
were expressed by ratios of orthorhombic unit cell lengths
(b/a and c/a). Symmetry increases
pffiffiffi from orthorhombic to
cubic when b/a = 1 and c/a = 2 ffi 1.414; so differences in
values of b/a and c/a from these values correspond to the
amounts of orthorhombic distortions. In Figure 3, the
pressure dependencies of the ratios were compared with
those of experiments [Mao et al., 1991; Fiquet et al., 1998]
and other theoretical calculations [Marton et al., 2001;
Wentzcovitch et al., 1995] at various pressures (0–120 GPa).
The results of b/a in this study (1.02–1.015) are lower than
those of experiments (1.03). Results of the variational
induced breathing (VIB) model potential [Marton et al.,
2001] and zero‐temperature ab initio energy calculations
[Wentzcovitch et al., 1995] were also compared. The results
of the VIB model and this study are a little lower than those
of the experiments. The b/a distortion of ab initio calcula-
tions [Wentzcovitch et al., 1995] was more significant
(1.047–.081) than those of the experiments, and its c/a was
also larger than those of experiments. Despite the simplicity
of the potential model used in the study, it gave results
comparable to those of experiments.
[10] Silicon self‐diffusion coefficients for MgSiO3
perovskite were measured at four pressure conditions (40,
60, 80, and 100 GPa). At first, we examined the temperature
variations of self‐diffusion coefficients at 100 GPa, shown
in Figure 4. The temperature range is 4200–5900 K with a
Figure 4. Temperature dependences of silicon (Si) self‐
diffusion coefficients at 100 GPa, 4200–5900 K. The ver-
tical line indicates the boundary of two (normal (lower),
abnormal (higher)) temperature regions.
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE B12205

Figure 5. Determination of the boundary of the lower and

higher regions at 100 GPa. (a) Original data of oxygen self‐
diffusion coefficients, (b) smoothed data, and (c) second de-
rivatives of the smoothed data. In Figure 5c, the solid arrow
shows the searching direction of the boundary, and the
dashed arrow shows the searched boundary.
Figure 6. Temperature dependences of silicon (Si) self‐
diffusion coefficients at 80 GPa (4200–5700 K), 60 GPa
(3800—5100 K), 40 GPa (3800—4600 K). The vertical
lines indicate boundaries of the regions.

5 of 9
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE
Figure 7. Relative absolute differences between neighbor-
hood data points (R = ∣log10(D1) − log10(D0)∣/∣log10(D0)∣) of
40 GPa data.
50 K step interval. The observed result had the following
features: (1) a linear increase in the lower‐temperature area
and (2) a discontinuous rapid increase in the higher‐tem-
perature area. Since melting occurred at a higher temper-
ature of the higher area, we consider that the behavior of
the higher‐temperature area was that of premelting and not
solid‐state diffusion, which this study investigated. We
hence adopted the activation enthalpy of the lower area as
the value of migration enthalpy (Hm*). Because of the
margin of error of the data, the boundary of the two regions is
indistinct and hard to determine. We adopted the following
procedure to give the boundary. First, we chose the data of
oxygen self‐diffusion coefficients to give the boundary.
Oxygen diffusion is two orders of magnitude faster than Si
self‐diffusion [Dobson et al., 2008]. Note that activation
enthalpies of the oxygen diffusion data cannot be compared
with those of experiments since this study does not treat the
most energetically favorable [Karki and Khanduja, 2006]
Mg‐O vacancy combinations. The error is reduced with
faster self‐diffusion coefficients [Ito and Toriumi, 2007].
The faster data gives a more distinct boundary. The routine to
derive the boundary is shown in Figure 5. The original data
of oxygen self‐diffusion coefficients (Figure 5a) were
smoothed by the filter (fnew(x) = f(x + Dx)/4 + f(x)/2 + f(x −
Dx)/4), where Dx represents finite increment of x (Figure
5b), and then second derivatives of log10D were derived
(Figure 5c). In Figure 5c, the boundary was searched from
the maximum point to the lower temperatures (shown by the
solid arrow), and given by the first negative point (shown
by the dashed arrow). On the basis of the routine, the
temperature of the boundary was given as 5200 K, and the
activation enthalpy of the lower‐temperature area (4200–
5200 K) was calculated as 400 kJ/mol. Belonoshko [1994]
calculated the melting temperature of MgSiO3 perovskite
as 5300 K at 110 GPa. The temperature of the boundary is
consistent with the melting temperature, supporting the idea
that the origin of the boundary is premelting.
[11] The temperature dependences of Si self‐diffusion
coefficients were also examined at other pressures (40, 60,
and 80 GPa). The results are shown in Figure 6. The features
observed at 100 GPa were also seen at 80 GPa and 60 GPa. By
6 of 9
B12205
the same procedure, the temperatures of the boundaries were
derived as 4950 K (80 GPa) and 4550 K (60 GPa). At the
lower‐temperature areas (4200–4950 K and 3800–4550 K),
activation enthalpies (Hm*) were calculated as 376 kJ/mol
and 364 kJ/mol, respectively. Unlike the higher pressures,
the number of 40 GPa data was limited and the boundary
was difficult to derive by a deterministic procedure. We
hence placed the boundary at 4450 K by eye since there was
a small jump. The nondeterministic procedure contains
some ambiguity, and as a result we expected larger error in
the activation enthalpy of 40 GPa. Scattering at lower
temperatures was observed at 40 GPa. Although relative
absolute differences between neighborhood data points (R =
∣log10(D1) − log10(D0)∣/∣log10(D0)∣) were small (−0.25) at
4100–4450 K, the differences suddenly became large (−0.5)
at 4000–4050 K and lower temperatures (Figure 7). The
large differences at the lower temperatures exceeded 25%,
the acceptable upper limit of error [Ito and Toriumi, 2007].
Hence, to avoid a large error, we did not adopt the data of
the scattered lower area. This treatment derives migration
enthalpy Hm* as 346 kJ/mol for 40 GPa at 4050–4450 K.
[12] From the available data (at 40, 60, 80, and 100 GPa),
pressure dependencies of the two coefficients (D0, Hm) in
equation (3) were expressed by functions of pressures.
Figure 8 plots observed migration enthalpies (Hm*) at the
four pressures. Although the linearity of pressure variations
was difficult to examine since the number of available data
points was only four, the data increased monotonically with
increasing pressure. We then derived a linear relationship by
fitting the equation as follows:
Hm *ð PÞ½kJ=mol ¼ 0:875P½GPa þ 310: ð7Þ
[13] As a result, the activation volume of migration Vm* =
dHm*/dP was estimated to be 0.875 cm3/mol. The pre-
exponential term D0 varied greatly with an adopted Hm*, so
we used the linearly fitted Hm* (equation (7)) instead of the
Figure 8. Temperature dependences of migration enthal-
pies (Hm*) of silicon (Si) self‐diffusion coefficients. Solid
diamonds indicate this study. The open circle and cross
show experimental data [Yamazaki et al., 2000; Dobson
et al., 2008] at 25 GPa, respectively.
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE
Figure 9. Pressure dependence of preexponential terms
(D0) estimated by minimizing the least squares difference
(equation (8)).
observed Hm* to avoid the effect of the error of the observed
Hm* on D0. D0 was estimated using the following proce-
dure. For each pressure condition, (1) Hm* was obtained by
equation (7); (2) D0,min was derived by minimizing fol-
lowing the square difference as follows:
Xh  i 2
Di  D0;min exp Hm* =RTi ; ð8Þ
i the authors suppose the assumption‐free saddle point con-
where Di and Ti are the observed self‐diffusion coefficients
and the corresponding temperature, respectively; and (3)
the derived D0,min was adopted as D0. With this procedure,
D0 was derived as 2.92 × 10−9 [m2/s] at 40 GPa, 4.17 ×
10−9 [m2/s] at 60 GPa, 6.83 × 10−9 [m2/s] at 80 GPa, and
9.60 × 10−9 [m2/s] at 100 GPa. The parameters increased
log linearly with increasing pressure as seen in Figure 9. A
log linear fitting of the data derived the pressure (P)
dependence of D0 as
3
D0 m2 =s ¼ 108:82510 P½GPa8:892 : ð9Þ
Although the temperatures of the MD simulations were
much higher than those of the lower mantle, the results
could be extrapolated using equation (3). We derived the
depth profile of the viscosity of MgSiO3 perovskite in the
lower mantle by extrapolation. Regarding the geotherm, we
assumed an adiabatic temperature gradient (0.3 K/km) and
the temperature of decomposition of ringwoodite to
perovskite and periclase at the top of the lower mantle
(1900 K at 660 km [Ito and Katsura, 1989]). Figure 10
compares the derived viscosity profile with that of MgO
periclase [Ito and Toriumi, 2007]. The viscosity of
MgSiO3 perovskite increased log linearly with increasing
depth up to 16 times from the top of the lower mantle. On
the other hand, MgO periclase increased slightly (2.3
times) in the upper area of the lower mantle and began to
decrease with increasing depth.
4. Discussion
[14] We obtained an equation for migration enthalpies
Hm* of silicon self‐diffusion of MgSiO3 perovskite as
7 of 9
B12205
equation (7). By employing this equation, Hm* of MgSiO3
perovskite at the pressure at the top of the lower mantle
(25 GPa) was derived to be 332 kJ/mol. This result can be
compared with the activation enthalpies of the experi-
mental studies [Yamazaki et al., 2000; Dobson et al., 2008]
directly. In the experiments, we consider that Si diffusion
was extrinsic as stated in section 2.2. We also consider that
Si vacancy concentrations can be assumed to be constant
with temperature and pressure variations since elements
with variable oxidation states (e.g., Fe) were not included
in the experimental samples. This result was consistent
with the results of experimental studies (336 kJ/mol at
25 GPa [Yamazaki et al., 2000], 347 kJ/mol at 25 GPa
[Dobson et al., 2008]). Hence, we first consider that the
reproduction of spontaneous jumping, which this study em-
ployed, is a valid approach for investigating self‐diffusion.
On the other hand, the previous approaches (733 kJ/mol at
0 GPa in the work by Wright and Price [1993]; 657kJ/mol
at 30 GPa in the work by Karki and Khanduja [2007];
453 kJ/mol at 25 GPa in the work by Ammann et al.
[2008]) somewhat overestimated the experimental results.
According to the thermally activated process model,
activation enthalpies are related with saddle point con-
figuration. In this study, atoms spontaneously jump into
vacancies, and the saddle points are naturally configured
without any assumptions. Hence, although detail of the
difference of the saddle point configurations is still unclear,
figurations as a cause of the consistency. The potential model
and parameters employed in this study were optimized for
equilibrium properties of experiments [Oganov et al., 2000].
There is a weak point of the potential model. Although
coordination numbers and bond angles of atoms vary during
their migrations, the variations were not taken into account
during the optimization. However, from the agreement
between activation enthalpies of this study and those of the
experiments, it is implied that the weak point does not exert
much effect.
[15] The migration enthalpies (Hm*) increase monotoni-
cally with increasing pressure. This behavior is different
from those of O and Mg of MgO periclase [Ito and Toriumi,
Figure 10. Depth variations of relative viscosities (h/
h660km) of the lower mantle minerals.
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE
2007], which show negative activation volumes at higher
pressures. Karato [1978, 2008] indicated the possibility of
transitions of diffusion mechanisms from vacancies to
interstitial atoms with increasing pressure, causing negative
activation volumes at high pressures. Karato [1978] also
indicated “soft” interatomic potential favors causing such
transitions. Since Hm* of silicon of MgSiO3 perovskite
(400 kJ/mol at 100 GPa) is higher than those of O and Mg
of MgO periclase (261 and 245 kJ/mol at 100 GPa,
respectively), the interatomic potential of MgO periclase is
considered to be softer than that of MgSiO3 perovskite. So
we consider that our results [this study; Ito and Toriumi,
2007] support the suggestions made by Karato [1978].
Although these studies did not confirm the hypothesis
directly, further direct observation of the transition by MD
simulation might be possible.
[16] The obtained viscosity profile of MgSiO3 perov-
skite shows a monotonic increase with increasing depth
(Figure 10). On the other hand, the viscosity profile of
MgO periclase is hill shaped; it increases in the upper area of
the lower mantle (660–1160 km) and decreases in the deeper
area (1160–2700 km) with increasing depth. Yamazaki and
Karato [2001] approximated the lower mantle by a two‐
phase system composed of MgSiO3 perovskite and MgO
periclase. The model rheology for the entire lower mantle is
based on two microstructures: (1) a load‐bearing framework
(LBF) of a strong phase containing isolated pockets of a
weak phase and (2) an interconnected layer of a weak phase
(IWL), separating boudins and clasts of a strong phase.
MgSiO3 perovskite is the dominant mineral in the LFB case,
and MgO periclase is the dominant mineral in the IWL case.
The depth profiles of the two minerals are clearly different
and well distinguished as shown in Figure 10. On the basis of
the difference, we could use the profiles to determine which
mineral dominates the lower mantle rheology by assuming
the depth profile of lower mantle viscosity. Monotonic
nearly constant depth profiles were considered by studies of
geodynamic modeling [e.g., Mitrovica and Forte, 1997]. The
shapes of the profiles are similar to those of MgSiO3
perovskite, and we consider MgSiO3 perovskite as dominant
in that case. The possibility of a hill shape, in which the
highest‐viscosity zone is in the middle of the lower mantle,
has previously been considered [Morra et al., 2009]. Since
the profile corresponds to that of MgO periclase, we consider
that the lower mantle rheology is dominated by MgO peri-
clase in that case.
[17] Acknowledgment. Some computations in this study were per-
formed on Earth Simulator at Earth Simulator Center, Japan Agency for
Marine‐Earth Science and Technology (JAMSTEC).
References
Ammann, M. W., J. P. Brodholt, and D. P. Dobson (2008), DFT study of
migration enthalpies in MgSiO 3 perovskite, Phys. Chem. Miner., 36,
151–158, doi:10.1007/s00269-008-0265-z.
Bejina, F., O. Jaoul, and R. C. Liebermann (2003), Diffusion in minerals at
high pressure: A review, Phys. Earth Planet. Inter., 139, 3–20,
doi:10.1016/S0031-9201(03)00140-7.
Belonoshko, A. B. (1994), Molecular‐dynamics of MgSiO3 perovskite at
high‐pressures: Equation of state, structure, and melting transition, Geo-
chim. Cosmochim. Acta, 58, 4039–4047, doi:10.1016/0016-7037(94)
90265-8.
8 of 9
B12205
Dobson, D. P., R. Dohmen, and M. Wiedenbeck (2008), Self‐diffusion of
oxygen and silicon in MgSiO3 perovskite, Earth Planet. Sci. Lett., 270,
125–129, doi:10.1016/j.epsl.2008.03.029.
Ewald, P. P. (1921), Die Berechnung optischer und electrostatischer Gitter-
potentiale, Ann. Phys., 369, 253–287, doi:10.1002/andp.19213690304.
Fiquet, G., D. Andrault, A. Daweale, T. Charpin, M. Kunz, and D.
Haussermann (1998), P‐V‐T equation of state of MgSiO3‐perovskite,
Phys. Earth Planet. Inter., 105, 21–31, doi:10.1016/S0031-9201(97)
00077-0.
Fiquet, G., A. Dewaele, D. Andrault, M. Kunz, and T. Bihan (2000), Ther-
moelastic properties and crystal structure of MgSiO3 perovskite at lower
mantle pressure and temperature conditions, Geophys. Res. Lett., 27(1),
21–24, doi:10.1029/1999GL008397.
Funamori, N., T. Yagi, W. Utsumi, T. Kondo, T. Uchida, and M. Funamori
(1996), Thermoelastic properties of MgSiO3 perovskite determined by in
situ X ray observations up to 30 GPa and 2000 K, J. Geophys. Res., 101,
8257–8269, doi:10.1029/95JB03732.
Gong, Z., Y. Fei, Y. Dai, L. Zhang, and F. Jing (2004), Equation of state
and phase stability of mantle perovskite up to 140 GPa shock pressure
and its geophysical implications, Geophys. Res. Lett., 31, L04614,
doi:10.1029/2003GL019132.
Heitjans, P., and J. Kaerger (2004), Diffusion in Condensed Matter:
Methods, Materials, Models, Springer, Berlin.
Holzapfel, C., D. Rubie, D. Frost, and F. Langenhorst (2005), Fe‐Mg inter-
diffusion in (Mg,Fe)SiO3 perovskite and lower mantle reequilibration,
Science, 309(5741), 1707–1710, doi:10.1126/science.1111895.
Ita, J., and R. E. Cohen (1997), Effects of pressure on diffusion and
vacancy formation in MgO from nonempirical free‐energy integrations,
Phys. Rev. Lett., 79, 3198–3201, doi:10.1103/PhysRevLett.79.3198.
Ito, E., and T. Katsura (1989), A temperature profile of the mantle transi-
tion zone, Geophys. Res. Lett., 25, 1095–1098.
Ito, Y., and M. Toriumi (2007), Pressure effect of self‐diffusion in periclase
(MgO) by molecular dynamics, J. Geophys. Res., 112, B04206,
doi:10.1029/2005JB003685.
Karato, S. (1978), The concentration minimum of point defects in solid at
high‐pressures and the plasticity of the lower mantle (in Japanese), paper
presented at Annual Meeting, Seismol. Soc. Jpn., Tokyo.
Karato, S. (2008), Deformation of Earth Materials: An Introduction to the
Rheology of Solid Earth, Cambridge Univ. Press, London.
Karato, S., S. Zhang, and H. Wenk (1995), Superplasticity in Earth’s lower
mantle: Evidence from seismic anisotropy and rock physics, Science,
270, 458–461, doi:10.1126/science.270.5235.458.
Karki, B. B., and G. Khanduja (2006), Computer simulation and visualiza-
tion of vacancy defects in MgSiO3 perovskite, Model. Simul. Mater. Sci.
Eng., 14, 1041–1052, doi:10.1088/0965-0393/14/6/011.
Karki, B. B., and G. Khanduja (2007), A computational study of ionic
vacancies and diffusion in MgSiO3 perovskite and post‐perovskite, Earth
Planet. Sci. Lett., 260, 201–211, doi:10.1016/j.epsl.2007.05.031.
Karki, B., L. Stixrude, and R. Wentzcovitch (2001), High‐pressure elastic
properties of major materials of Earth’s mantle from first principles, Rev.
Geophys., 39(4), 507–534, doi:10.1029/2000RG000088.
Mao, H. K., R. J. Hemley, Y. Fei, J. F. Shu, L. C. Chen, A. P. Jephcoat, and
Y. Wu (1991), Effect of pressure, temperature, and composition on lattice
parameters and density of (Fe,Mg)SiO‐perovskites to 30 GPa, J. Geo-
phys. Res., 96, 8069–8079, doi:10.1029/91JB00176.
Marton, F. C., J. Ita, and R. E. Cohen (2001), Pressure‐volume‐temperature
equation of state of MgSiO3 perovskite from molecular dynamics and con-
straints on lower mantle composition, J. Geophys. Res., 106, 8615–8627,
doi:10.1029/2000JB900457.
Mitrovica, J. X., and A. M. Forte (1997), The radial profile of mantle vis-
cosity: Results from the Joint Inversion of convection and postglacial
rebound observables, J. Geophys. Res., 102, 2751–2769, doi:10.1029/
96JB03175.
Morra, G., D. A. Yuen, C. Philippe, P. Koumoutsakos, L. Boschi, and P.J.
Tackley (2009), The role of a high viscosity hill in the lower mantle on
the fate of subducting slabs, Geophys. Res. Abstr., 11, Abstract
EGU2009‐13553.
Oganov, A. R., J. P. Brodholt, and G. D. Price (2000), Comparative study
of quasiharmonic lattice dynamics, molecular dynamics and Debye
model applied to MgSiO3 perovskite, Phys. Earth Planet. Inter., 122,
277–288, doi:10.1016/S0031-9201(00)00197-7.
Poirier, J. P. (2000), Introduction to the Physics of the Earth’s Interior,
Cambridge Univ. Press, Cambridge, U. K.
Wentzcovitch, R. M., N. L. Ross, and G. D. Price (1995), Ab initio study of
MgSiO3 and CaSiO3 perovskites at lower‐mantle pressures, Phys. Earth
Planet. Inter., 90, 101–112, doi:10.1016/0031-9201(94)03001-Y.
Woodcock, L. V. (1971), Isothermal molecular dynamics calculations for
liquid salts, Chem. Phys. Lett., 10, 257–261, doi:10.1016/0009-2614
(71)80281-6.
B12205 ITO AND TORIUMI: SILICON SELF‐DIFFUSION OF PEROVSKITE B12205

Wright, K., and G. D. Price (1993), Computer simulations of defects and Yamazaki, D., T. Kato, and H. Yurimoto (2000), Silicon self‐diffusion in
diffusion in perovskites, J. Geophys. Res., 98, 22,245–22,253, MgSiO3 perovskite at 25 GPa, Phys. Earth Planet. Inter., 119, 299–309,
doi:10.1029/93JB01098. doi:10.1016/S0031-9201(00)00135-7.
Yagi, T., H. K. Mao, and P. M. Bell (1979), Lattice parameters and specific
volume for perovskite phase of orthopyroxene composition (Mg,Fe)
Y. Ito and M. Toriumi, Department of Complexity Science and
SiO3, Year Book Carnegie Inst. Washington, 78, 612–613. Engineering, Graduate School of Frontier Science, University of
Yamazaki, D., and S. Karato (2001), Some mineral physics constraints Tokyo, 5‐1‐5 Kashiwanoha, Kashiwa, Chiba, 277‐0882, Tokyo, Japan.
on the rheology and geothermal structure of Earth’s lower mantle, (ito@gaea.k.u‐tokyo.ac.jp)
Am. Mineral., 86(4), 385–391.
Yamazaki, D., T. Kato, E. Ohtani, and M. Toriumi (1996), Grain growth
rates of MgSiO3 perovskite and periclase under lower mantle conditions,
Science, 274, 2052–2054, doi:10.1126/science.274.5295.2052.

9 of 9