Studying Stress-Corrosion in Silica: Taking Multiscale Simulations To The Petascale

Studying stress-corrosion in
silica: taking multiscale

simulations to the petascale
Marco Caccin
Supervised by: A. De Vita, J.R. Kermode, O.A. Von Lilienfeld
Appl. Phys. 42 (2009) 214006
Open questions
H2O dissociation - crack advance in SiO2
Bond breaking mechanism
H2O diffusion to reaction site
Reaction kinetics
M Ciccotti
J. Phys. D: Appl. Phys. 42 (2009) 214006
M Ciccotti
When describing the macroscopic fracture properties of a

material, it is the relation of the external variables G, K which
is studied as a function of v. Even if assuming a uniquely
defined kinetic law v(G , K ), these local variables remain
hidden, and the presence of a process zone will cause the nonuniqueness of the macroscopic fracture laws v(G, K). This
induces a time- or crack-length dependent behaviour (R-curve)
that may lead to controversial interpretation (Lawn 1993).
These types of modelling will be shown later to be particularly
effective in rationalizing the propagation threshold behaviour
(cf section 3.4).
Effect of temperature on the crack propagation in
glass (from Wiederhorn and Bolz (1970)).
As a concluding remark to this section, it is generally
useful to identify the different space, time and energy scales of
he same reasons, the region I data can also be fitted by
t exponential equation proposed by Pollet and Burns
each symbolic block (cf figure 2) in order to estimate whether
d applied to glasses by Lawn (1993):
!
"
!
"
the different dissipation mechanisms will have time to activate
"Ea
(G G0 )
xp(G) = A exp
exp
,
and how they will influence the energetic balance or the fracture
kB T
kB T
(3)
propagation laws. Several spatial scales are present in the
e energetic balance has a more central role, as will be
problem, such as the sample lateral dimensions and thickness,
later concerning the threshold at G0 .
the loading displacement, the crack length, the basic grain
Wiederhorn equation (1) has been the most used in the
on of region I in glasses since it provides a more direct
size of the material (molecular rings for glass), the size of the
tion of the chemical reactions between a glass of a
process zone (or of each shell of a multiple process zone), the
mposition and the environment at the crack tip. In
ework, the basic mechanism of the stress-corrosion
enclave size, the length of the one or several cohesive zones in
was associated with the stress-enhanced thermal Figure 7. Basic mechanism of the stress-corrosion reaction (from
Michalske and Freiman (1983)).
the crack cavity. Typical time scales are given by the loading
n of a dissociative hydrolysis reaction represented
7 for the case of silica glass (Michalske and
or displacement rate, the crack velocity, the stimulation time
1983). Michalske and Bunker (1984) supported the 1) is related to the nearly first order of the stress-corrosion
of each equivalent block and its characteristic relaxation time, Figure 5. Effect of humidity on the crack propagation in soda-lime
effect of stress on the reaction rate by a molecular- reaction (Wiederhorn et al 1974).
(2009).
Stress-corrosion
silicate
glasses.
Journal
of mechanisms,
Physics D:the
Applied
Physics,
42(21),
214006.
For propagation in mechanisms
liquid environment, the
term
p/p
(from Wiederhorn
(1967)).
thein
rate
of0 thermal
activation
of different
rate glass
mulation Ciccotti,
of the reaction onM.
distorted
siloxane bonds.
r provided experimental support for their theory by is substituted by the activity of water molecules which
of
transport
of
different
relevant
chemical
species
in
the
bulk
M. silicate
(1967).
Influence
Water
Vapor
is strongly
dependent of
on the
composition
of the on
liquidCrack Propagation in Soda-Lime Glass. Journal of the American Ceramic Society, 50(8),
g the rateWiederhorn,
of hydrolysis of a series ofS.
strained
or in
crack cavity. We will try to keep these space and times
by the
the pH
tures (Michalske and Bunker 1993). For other glass environment. Notably, it is strongly affected
to the glass
ions (figure 8), several other reactions are possible, and ionic concentration of some species related
scales
in mind while proceeding in the description of the crack 3.1. Region I: stress-corrosion regime
407-414.
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
ment. When the probe tip comes in contact with the

condensate, a sharp change in retardation is observed
and the condensate length can be measured very accurately, Fig. 7. Measured in this way, a Laplace pressure
of !36 " 5 MPa was determined for a crack tip in silica glass at 40% relative humidity. Also, from the closing forces, an adhesive energy, Go = 180 " 20 mJ/m2
was obtained, which is reasonably close to the energy
required to create two water surfaces, 144 mJ/m2.
This technique allows important information to be
obtained on the nature of the crack tip condensate,
which constitutes the local environment for stress corrosion crack propagation in a moist atmosphere. In silica glass, the condensate composition was shown to be
very close to pure water and to be in stationary equilibrium with the moist atmosphere.42 By contrast, the
crack tip condensate size in soda lime silicate glass can
evolve in time due to alkali ion diffusion and exchange
with hydronium ions at the crack tip.43
The coupled modeling of crack tip shielding,
induced by water penetration into glass and the evolutions of the local crack tip environment, can be very
useful to predict the shape of the vK curves in the
stress corrosion cracking of silicate glasses as a function
of their composition and external environmental.49
water penetrates because of the dilatation of the silica

glass due to the high tensile stresses at the head of the
crack tip. As the crack propagates through the waterpenetrated glass, the water is left behind on the fracture
surface. This water could be detected using the technique of nuclear reaction analysis,50 which measures
the concentration of hydrogen atoms within the glass.
The details of the experimental technique are given in
reference (50). The resolution of the technique was sufficient to show that the concentration of hydrogen in
the water-penetrated surface exceeded that of the oil/
etched specimen (cf. caption in Fig. 8) for which
adsorbed water was only at the surface.
To obtain a theoretical estimate of the experimental curve in Fig. 8a, the rate of diffusion of water into
the tip of a crack from the crack opening has to be calculated. This was performed recently by Wiederhorn
et al.47 The diffusion equation for water into glass was
solved in cylindrical coordinates and included the effect
of the mechanical swelling stresses on the diffusivity. A
data curve was estimated from the theoretical penetration curve (the curve calculated in reference 47), and
the instrument curve given in the paper by Tomozawa
et al. (the curve obtained from the oil/etched specimen,
Fig. 8). The predicted experimental curve from
Open questions
Role of H2O ahead of crack

front
5 nm
Fig. 7. (a) Typical atomic force microscopy (AFM) height image of a crack tip. Note the depression in the vicinity of the crack tip. (b)
Typical AFM phase image of a crack tip. The zone of the phase shift at the front of the crack marks the region of water condensation.
The size of the image is 400 nm2. The scale to the right of the images is 5 nm for the height image and 5 for the phase image. Taken
from reference (49).
POOR QUALITY FIG
Near - crack tip stress field and

plasticity
Colour online, B&W in print
15
Wiederhorn, S. M., Fett, T., Guin, J.-P., & Ciccotti, M. (2013). Griffith Cracks at the Nanoscale. International Journal of Applied Glass Science, 4(2), 7686.
Approach
M Ciccotti
elevated
surface
f plastic
s, which
s around
re of the
ly when
here the
fracture
bcritical
ized and
< Kc
generally
Region I
here the
reaction,
region II
kinetics
tip. In
o induce
ions, the
hing the
itions, a
what is
erhorns
of crack
for most
onsistent
aradigm
orrosion
bated for
al 1991,
be found
hat may
different
complex
(mainly
Atomistic model
F
T
O
L
e
l
a
ltisc MM)
Figure 1. Graphic representation of the mechanical elements for
u
M
/
fracture mechanics.
M
(Q
requires nanometre scale resolved investigation techniques and
questioning about the relevant physical laws at the nanoscale.
Such studies should be very promising to solve the debates
by direct observation and to relate the phenomenological
parameters to the specific composition and structure of glasses.
The development of several nanoscale investigation
techniques in the 1980s has led to significant insights into the
comprehension of the different mechanisms (cf sections 4.3
4.6). In particular SEM, TEM and AFM measurements
of both post-mortem crack surfaces and direct in situ
observations of the crack tip neighbourhood have permitted
the investigation of the space and time scales of occurrence of
these phenomena under specific conditions. The progressive
increase in resolution down to the micrometre scale of several
structural and compositional analysis techniques (such as
Raman, IR, Brillouin spectroscopies, x-ray, electron and
neutron scattering, NMR, XPS, SIMS) have also permitted
QM region K-way partition
Approach: why partitioning?
n QM core atoms, m
n2 buffer atoms
N + M valence electrons (N
n, M
DFT code time to solution and cost
m)
(N + M )p
K-way partitioning of QM core region t.t.s.
((N/K + M ))
Targets: reduce t.t.s. (better weak scaling),

reduce CPU time per calculation (better strong scaling)
Approach: distributed multi-scale MD

BG/Q Execute Nodes
Front End Node

N clusters
N forces
(1)
{F i}, {F(2)i}, , {F(M)i}
Cluster 1
Forces {F(1)i}
Cluster 2
Forces {F(2)i}
64 nodes
QM/MM MD
with N QM atoms
QM Box 1
64 nodes
QM calculation on Cluster 1
QM Box 2
64 nodes
QM calculation on Cluster 2
...
QUIP
proxy
QUIP
Cluster M
Forces {F(M)i}
QM Box M
64 nodes
QM calculation on Cluster M
Partitioning scheme
Input: atoms connectivity map

(bonds) graph
Target: optimal load balance.

Split system in K parts
Convex (minimal edgecut)
Equal size
Proposed solution: combined

method
Iterative spectral bisection
Merging of 2 smallest
neighbouring regions
Subgraph digestive ripening
Fiedler, M. (1975). A property of eigenvectors of nonnegative symmetric matrices and its application to graph theory. Czech. Math. J., 25(4), 619-633.
Results: partition quality
Results: DFT scaling

Strong scaling - crack front in Silicon
Weak scaling - crack front in a-SiO2
The Machine Learning front
Big calculations generate big data.
How to draw relation atomic structure - observable? Machine learning
Which properties can be learnt?
Any scalar function of atomic coordinates (potentially)
Vectorial property of local environment: forces
How accurately?
The bigger the database, the better
Importance of accurate estimation of prediction error
Rupp, M., Tkatchenko, A., Mller, K. R., & von Lilienfeld, O. A. (2012). Fast and accurate modeling of molecular atomization energies with machine learning. PRL, 108(5), 058301.
Bartk, A. P., Payne, M. C., Kondor, R., & Csnyi, G. (2010). Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons. PRL, 104(13), 136403.
Li, Z. (2014). Submitted.
Predicting error on predictions

Kij = k(Di , Dj )
Kernel ridge regression
: regression coefficients
= (K + I) 1 t
X
tnew =
i k(Di , Dnew )
i
Strong evidence of correlation

error - regression coefficient
can we build a 2nd layer ML to

predict of new configuration?
Ongoing and perspective work

Powerful computational framework applicable to:
Dry vs. wet dynamical fracture in silica (crystalline, amorphous)
Bond breaking reaction mechanism
Influence of penetrated water on elastic stress field
Study of 3D crack fronts: are crack fronts curved?
Flexoelectricity in crack cavities
Database generation of DFT calculations:

enabling ML atomistic configuration physical observable
Thank you for your attention!
Substitute automate with use

bigger computer. Argument
still holds.

Studying Stress-Corrosion in Silica: Taking Multiscale Simulations To The Petascale

Uploaded by

Copyright:

Available Formats

Studying Stress-Corrosion in Silica: Taking Multiscale Simulations To The Petascale

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Studying Stress-Corrosion in Silica: Taking Multiscale Simulations To The Petascale

Uploaded by

Copyright:

Available Formats

Studying stress-corrosion in

silica: taking multiscale

Supervised by: A. De Vita, J.R. Kermode, O.A. Von Lilienfeld

Appl. Phys. 42 (2009) 214006

Bond breaking mechanism

H2O diffusion to reaction site

J. Phys. D: Appl. Phys. 42 (2009) 214006

When describing the macroscopic fracture properties of a

ment. When the probe tip comes in contact with the

water penetrates because of the dilatation of the silica

Role of H2O ahead of crack

POOR QUALITY FIG

Near - crack tip stress field and

QM region K-way partition

Approach: why partitioning?

DFT code time to solution and cost

K-way partitioning of QM core region t.t.s.

Targets: reduce t.t.s. (better weak scaling),

Approach: distributed multi-scale MD

Front End Node

Input: atoms connectivity map

Target: optimal load balance.

Convex (minimal edgecut)

Proposed solution: combined

Iterative spectral bisection

Subgraph digestive ripening

Results: partition quality

Results: DFT scaling

Weak scaling - crack front in a-SiO2

The Machine Learning front

Big calculations generate big data.

How to draw relation atomic structure - observable? Machine learning

Which properties can be learnt?

Any scalar function of atomic coordinates (potentially)

Vectorial property of local environment: forces

The bigger the database, the better

Importance of accurate estimation of prediction error

Predicting error on predictions

Kernel ridge regression

Strong evidence of correlation

can we build a 2nd layer ML to

Ongoing and perspective work

Dry vs. wet dynamical fracture in silica (crystalline, amorphous)

Bond breaking reaction mechanism

Influence of penetrated water on elastic stress field

Study of 3D crack fronts: are crack fronts curved?

Flexoelectricity in crack cavities

Database generation of DFT calculations:

Thank you for your attention!

Substitute automate with use

You might also like