Received: 27 July 2018

| Revised: 20 September 2018

| Accepted: 20 September 2018

DOI: 10.1111/jace.16106

ORIGINAL ARTICLE

A statistical model for the failure of glass plates due to nickel

sulfide inclusions

Antonio Bonati1 | Gabriele Pisano1 | Gianni Royer Carfagni1,2

1
Construction Technologies Institute –
Italian National Research Council (ITC-
CNR), San Giuliano Milanese, Italy
2
Department of Engineering and
Architecture, University of Parma, Parma,
Italy
Correspondence
Gianni Royer Carfagni, Department of
Engineering and Architecture, University
of Parma, Parma, Italy.
Email: gianni.royer@unipr.it
Funding information
Italian Civil Protection Department,
Presidency of the Council of Ministers,
Grant/Award Number: ReLUIS-DPC
2014-2018; Italian Ministry of University,
Grant/Award Number: MIUR-PRIN voce
COAN 5.50.16.01 code 2015JW9NJT
Abstract
Nickel Sulfide (NiS) inclusions can provoke the rupture of thermally treated glass
due to a phase transformation with volume increase that stresses the surrounding
glass. Starting from a Pareto statistics for the population of inclusion sizes, from
an assumed kinetics of the phase transformation, a micro‐mechanically motivated
model provides the statistical characterization of the probability of spontaneous
failure of glass during lifetime. A distinction based upon the composition of NiS
is used to discuss the effects of the heat soak test (HST), where glass remains at
high temperature for a certain time to speed‐up the phase transformation and
destroy those elements with critical inclusions. Three functions à la Weibull for
the probability of spontaneous rupture during lifetime are theoretically derived for
the case of no HST, short HST, and long HST. In particular, the probability of
collapse for long HSTs depends upon the holding time in the oven. An explana-
tory example shows the potentiality of the model for optimizing the HST parame-
ters toward a target probability of failure, but experimental campaigns are needed
for a proper calibration.

KEYWORDS
fracture mechanics/toughness, glass, heat soak test, NiS impurities, tempered glass, Weibull statistics

1 | INTRODUCTION Moreover, a subtle phenomenon, referred to as static fati-

Glass does not exhibit a plastic phase and is not able to
develop the diffused damage that can inelastically mitigate
the stress concentrations as in quasi‐brittle materials, so that
breakage is governed by the critical opening of cracks.
Such cracks are induced on the external surface of architec-
tural glass by the manufacturing process, transport, han-
dling, application, and other abrasion phenomena, while
volume cracks are very rare. They are generally not visible
with naked eyes and the nature of their size, shape, and
distribution is strongly probabilistic. Since they open
almost exclusively in mode I, fracture occurs when the
combination of crack depth and the tensile stress reaches a
critical value. This is why, although the theoretical tensile
strength of glass may reach 32 GPa,1 the measured tensile
bending strength generally does not exceed 100 MPa.
gue or subcritical crack propagation, is that cracks can
slowly grow even for opening stresses much lower than the
corresponding critical limit,2 so that failure is in general
time‐dependent.
The thermal‐induced tempering process is one of the
most common and efficient methods for strengthening glass
because, due to rapid cooling, a residual eigenstress state is
generated with compressive stresses in the external layers
and tensile stresses in the core part. Thus, as long as the
tensile surface stress, due to external actions, is smaller
than the residual compressive stress, surface cracks do not
open and cannot lead to failure.3 However, the potential
presence of Nickel Sulfide (NiS), which is a mineral con-
taminating the glass melt during manufacturing, may pro-
voke the spontaneous rupture of thermally treated glass
even after years from installation.

J Am Ceram Soc. 2018;1–17. wileyonlinelibrary.com/journal/jace © 2018 The American Ceramic Society | 1

2
| BONATI
Inclusions of NiS are characterized by two phases of
relevance, generally called α and β. The α‐phase is
stable above 379°C but, due to rapid cooling, in ther-
mally strengthened glass the NiS inclusions are trapped
in the high temperature α‐phase as a meta‐stable condi-
tion. The α and β phases have different crystal struc-
tures, and, as a result of the phase transformation, whose
speed is temperature dependent, a volume expansion
occurs that produces contact pressures in the surrounding
glass and causes crack nucleation. However, since NiS
has a higher thermal expansion coefficient than glass,
upon rapid cooling it contracts more than glass by leav-
ing a small gap that needs to be filled before the contact
pressure is developed. If NiS inclusions are present in
the core zone of thermally treated glass plates, where
residual stresses are tensile, micro‐cracks can potentially
propagate, and since the phase transformation takes time
and cracks can propagate subcritically, failure may occur
even some years after installation, during the service life.
A phenomenon of this kind does not usually occur in
annealed glass, since NiS inclusions have time to trans-
form into the β‐phase and the central tensile stress is
almost negligible. According to Swain,4 the α‐β transfor-
mation leads to a volumetric expansion of 4%, but in
recent years it has been observed5 that a value higher
than 2.5% is unlikely. Since there are no warning signals
and collapse can occur when external actions are almost
null, this phenomenon was originally defined as sponta-
neous fracture. The presence of NiS inclusion was firstly
identified by Ballantyne6 only at the beginning of the
1960s, to justify spontaneous breakage of tempered glass
plates of a facade in Australia.
It is generally assumed that inclusions originate from
Sulfur and Nickel in raw materials, which are the major
impurities found in glass in addition to very small quanti-
ties of other metals such as Iron and Copper.7 The nickel
source to NiS inclusions is not yet completely clear, but
probably it is due to contaminations of raw materials such
as feeders made of stainless steel, or other elements that
are used during storage or handling.8 Despite all mitigating
measures that the glass industry has adopted, the complete
elimination of NiS impurities is not possible today and,
remarkably, only 1 g of Nickel can produce up to 70 000
inclusions of 200 μm diameter.9
This is why curently all the glass plates bearing a secu-
rity risk have to be tested with the heat soak test (HST), in
which glasses are exposed to a certain temperature for a
certain time with the aim of speeding up the α‐β phase
transformation, so to destroy the glass panes affected by
critical NiS inclusions before installation. During last dec-
ades, different times and temperatures for the HST have
been proposed to optimize its efficiency, and regulations
have been consequently updated. Today, the process is
ET AL.
standardized by the European norm EN14179‐1,10 which
has been recently revised in 2016 following the study by
Yousfi,11 who found that for a NiS composition NiSx with
x>1:012, the α‐β transformation cannot be complete if tem-
perature exceeds 280°C. Hence, the HST temperature was
set 30°C lower than before, at 260 ± 10°C. However,
because of the temperature decrease, there are new open
questions about the correct holding time to guarantee a tar-
get level of safety.
All the basic variables that can lead to spontaneous
breakage have a strongly non‐deterministic nature, in
particular for what concerns the size, position, and
composition of NiS inclusions. In Section 2.2, a micro‐
mechanically motivated statistical model is presented for
the determination of the risk of “spontaneous” breakage
of glass not subject to HST. Then, in Section 3, how the
HST affects the statistical population of NiS inclusions
and, consequently, the risk of spontaneous breakage, is
discussed by proposing a further refinement of the statisti-
cal approach. A possible application is presented in Sec-
tion 4 on the basis of some experimental measurements
recorded in the literature,12,13 but since data are partial
and fragmented, the estimated values of the parameters
for the statistical analysis can only be regarded as tenta-
tive results. Anyway, this has been useful for showing the
potentialities of the proposed model, in its capability of
defining the safety level as a function of the HST holding
time.
2 | A MICRO‐MECHANICALLY
MOTIVATED STATISTICAL MODEL
FOR NiS INCLUSIONS
The volume expansion associated with the α‐β phase
transformation of NiS inclusions, in general spherical or
slightly ellipsoidal, may nucleate micro‐cracks that can
propagate under the internal tensile stress field of tem-
pered glass. Critical propagation occurs only when the
stress intensity factor (SIF) KI in mode I reaches a criti-
cal limit KIc, but in subcritical propagation, cracks can
slowly grow over time for values of the SIF comprised
between a lower bound KI0, below which cracks do not
grow, and KIc, above which propagation is instantaneous.2
The value of the SIF depends upon the crack size and,
consequently, by the type of NiS inclusion that generated
it.
2.1 | Representative volume elements
Imagine that the volume V of a glass under consideration
can be divided into representative volume elements
(RVEs), of volume ΔV. From a statistical point of view,
BONATI ET AL. | 3

the crucial property of the RVEs is that they do not interact   α

R
one another in the process of failure, ie, each RVE is like a P
ΔV ðRÞ ¼ ΔV ; (2.4)
η
ring included in a chain: breakage of one ring is not

affected by the strength of the other rings, although it where α ¼ α 1, η ¼ R0 ΔV 1=α , so as to emphasize the
implies failure of the whole chain. This means that we sup- dependence on ΔV. When the α‐β phase transformation of
pose that there is no interaction between cracks that possi- NiS is complete, the inclusion applies a pressure, denoted
bly develop in distinct RVEs, a property that is verified if with Π0 , to the surrounding glass. Following Swain,4 under
the RVE diameter is substantially greater than the length of an external tensile stress σ cr rupture occurs when R  Rcr ,
any crack that is triggered by NiS inclusions. The precise with
estimation of the size of the RVE would require refined π 2 1=2
considerations at the level of micromechanics. However, Rcr ¼ K ðΠ0 σ 3cr Þ ; (2.5)
7:1 LT
for the following considerations it is not important to pre-
cisely estimate ΔV, but rather to recognize that ΔV ≪ V.
where KLT is the SIF that is considered critical for the
In fact, in the limit ΔV=V ! 0, we will derive expressions
long-term life of the element. The value of KLT , which
defined by more general material parameters that bypass
should account for subcritical crack propagation, will be
the precise determination of ΔV, since such parameters can
specified later on.1 Substituting (2.5) into (2.4), one finds
be globally calibrated a posteriori by means of macroscopic
tests. 
σ cr
32α
The RVEs are characterized by a certain probability of PΔV ðσ cr Þ ¼ ΔV ; (2.6)
ηLT
hosting an inclusion, which can be approximately supposed
spherical of radius R. It is reasonable to assume that the where
lower is R, the higher is the probability of finding an inclu-  2 1=2 2=3
sion of radius equal or higher than R. For this, we use a π KLT Π0
ηLT ¼ : (2.7)
power‐law distribution à la Pareto, so that the probability 7:1 η
of finding in the RVE an inclusion of radius comprised For this derivation, observe that the probability of rup-
between R and R + dR reads ture at σ cr at the RVE level is equal to the probability of
pΔV ðRÞdR ¼ CR α dR; (2.1) finding an inclusion of radius R  Rcr .

where α is a scaling parameter, accounting for the 2.2 | The probability of spontaneous failure
dispersion of the distribution, and C is the normalization of a heat‐treated glass plate of volume V
constant, which is obtained by meeting the normalization
requirement The tempering process induces a permanent residual eigen-
Z 1 Z 1 stress characterized by compressions on the external layers
α 1 of the glass plate and tensions in the core part. It is gener-
pΔV ðRÞdR ¼ CR α dR ¼ 1 ) C ¼ 1 α : (2.2)
R0 R0 R0 ally assumed that it is uniform equibiaxial on layers at a
distance z from the mid‐plane of the plate, of the form
Therefore, the probability that the RVE hosts an inclusion Σeqb ¼ ½σ eqb ðzÞ; σ eqb ðzÞ; 0Š, where σ eqb ðzÞ follows the para-
of radius not smaller than R is bolic curve represented in Figure 1 of equation
1 "  2 #
1   α  α
2z
Z
α 1 r R
P
ΔV ðRÞ ¼ dr ¼ : (2.3) σ eqb ðzÞ ¼ σ max 1 3 ; (2.8)
R R0 R0 R0 h

The domain of (2.3) is ½R0 ; 1Š, whereas for R<R0 one

has P ΔV ¼ 1. The lower bound R0 may be interpreted as
the size of inclusions that are physiologically present in
any RVE, but it is also necessary to avoid that the function
(2.1) diverges as R ! 0, as usually done for distributions à
la Pareto. We will show that, since rupture is associated
with inclusions with R≫R0 , the parameter R0 is relevant
only because it affects the constant C of (2.2), which char-
acterizes the right‐hand‐side tail of the distribution.
The cumulative probability (2.3) can be rewritten in the
more convenient form
being σ max the maximum tensile stress at z = 0 and h the
plate thickness. To our knowledge, this is an almost univer-
sally accepted model but, of course, one could consider
spatial distributions more complicated than (2.8), possibly
affected by statistically described variations. However, in
our opinion, this would imply an unjustified further
1
One may assume, on the safe side, that KLT = KI0, but we will show in
Section 4.2 that this is too much conservative.
4
| BONATI ET AL.

the applied stress normal to the crack plane is σ ?  σ cr .

The solid angle Ω reads14

Z θcr
σ cr
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Ωðσ eqb ðzÞ; σ cr Þ ¼ 2 2π cos θdθ ¼ 4π 1 ;
0 σ eqb ðzÞ
(2.10)
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where θcr ¼ arccos σ cr =σ eqb ðzÞ. Hence, the probability
that the inclination of the micro‐crack is such that σ ?  σ cr
takes the form

Ωðσ eqb ðzÞ; σ cr Þ σ cr

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P? ðσ eqb ðzÞ; σ cr Þ ¼ ¼ 1 : (2.11)
4π σ eqb ðzÞ
F I G U R E 1 Parabolic distribution of the residual stress in the However, crack initiation is due to the synergistic action of
thickness and portion of the plate where the tensile stress is not
the volume expansion of the NiS inclusion and the tensile
smaller than σcr
stress from the thermal process. Consequently, it is likely
that the crack plane shall be orthogonal to the glass middle
plane, which collects the directions of maximum tensile
complication of the model and, therefore, it will not be stress. Therefore, on the safe side, we will assume that
considered here. P? ¼ 1.
Now, consider a volume V of thermally treated mate- In conclusion, the probability that a RVE of the glass
rial formed by a certain number of RVEs. A NiS inclu- plate fails because of an inclusion with critical stress com-
sion can lead to failure if and only if it lies in the prised between σ cr and σ cr þ dσ cr , reads
volume of glass where σ eqb  σ cr . We may assume that
the probability of finding an inclusion of critical size is
dPΔV ðσ cr Þ
uniform in the volume although, strictly speaking, due to dPf ;ΔV ðσ max ; σ cr Þ ¼ Ph ðσ max ; σ cr Þ dσ cr
dσ cr
gravitational settling during the float production process, pffiffiffirffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3 σ cr dPΔV ðσ cr Þ
inclusions should be expected to be slightly more numer- ¼ 1 dσ cr ; (2.12)
ous in the lower half (toward the tin‐bath side) than in 3 σ max dσ cr
the upper half of the plate. Since the probability of an
event is equal to the ratio between the number of favor- where Ph ðσ max ; σ cr Þ is given by (2.9) and PΔV ðσ cr Þ by
able events and the total number of events, for the case (2.6). From (2.12), the probability of survival of a RVE
at hand the probability that the inclusion is in a critical reads
position is equal to the ratio between the thickness of Z σmax pffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the layer where σ eqb  σ cr and the thickness h of the 3 σ cr dPΔV ðσ cr Þ
Ps;ΔV ðσ max Þ ¼ 1 1 dσ cr ;
plate, as illustrated in Figure 1. More precisely, from 0 3 σ max dσ cr
(2.8) the probability Ph that an inclusion is in a critical (2.13)
position reads where the upper limit of integration is σ max because the
pffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi probability of failure is null for σ cr >σ max . A thermally trea-
3 σ cr
Ph ¼ Ph ðσ max ; σ cr Þ ¼ 1 : (2.9) ted glass plate of volume V, made by a number V=ΔV of
3 σ max
RVEs, survives if and only if none of the RVEs reaches its
Strictly speaking, since micro‐cracks open in mode I, critical conditions, ie, its survival probability is given by
one should consider the probability that the maximum ten- the product of all the survival probabilities of the RVEs.
sile stress is at right angle with the axis of the micro‐crack Substituting (2.6) into (2.13), one has
generated by the volume expansion of the inclusion. If
Ps;V ðσ max Þ
there are not preferred orientations for micro‐cracks " pffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffi " 3  # #V=ΔV
induced by the volumetric expansion, the total number of σ max
σ cr 2α
Z
3 1 σ cr d
events is proportional to the solid angle subtended by a ¼ 1 ΔV dσ cr :
0 3 σ max dσ cr ηLT
sphere, equal to 4π, while the number of favorable events
(2.14)
is proportional to Ω, defined as the solid angle containing
the normals to all orientations for which the component of Posing ε ¼ ΔV=V, this expression can be rewritten as
BONATI ET AL. | 5

Ps;V ðσ max Þ 2.3 | Parameters calibration

" pffiffiffi Z σ rffiffiffiffiffiffiffiffiffiffiffiffiffiffi " 3  # #1=ε
3 max
1 σ cr d σ cr 2α The parameters α and η of (2.4) are calibrated from exper-
¼ 1 ε V dσ cr :
3 0 σ max dσ cr ηLT imental measurements. The probability that a RVE hosts an
(2.15) inclusion of radius lower or equal of R is, from (2.4),

Assume that the volume V is constituted by a large   α

number of RVEs, ie, V=ΔV≫1, so that ε ! 0. Since R
P
ΔV ðRÞ ¼ 1 ΔV : (2.20)
limε!0 ½ð1 þ εaÞ1=ε Š ¼ expðaÞ, one can write η

Ps;V ðσ max Þ Let V0 be the reference volume, which statistically con-

" pffiffiffi Z " 3  # #
σ max
σ cr 2α
rffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3 1 σ cr d tains on average only one inclusion, formed by a number
¼ exp V dσ cr : V0 =ΔV of RVEs. Inclusions with radius higher or equal
3 0 σ max dσ cr ηLT
than R are not present in V0 if and only if none of the con-
(2.16)
stituent RVEs contains inclusions of this type. Hence, the
Observe, in passing, that the aforementioned limit corresponding probability reads
allows to derive expressions that include the size of the "   α #V0 =ΔV
RVE as part of more general material parameters, which in R
P
V0 ðRÞ ¼ 1 ΔV : (2.21)
turn can be calibrated from macroscopic tests. This means η
that, in this approach, it is not important to estimate the
precise size of the RVE, but rather to verify that V≫ΔV in Taking the limit for ΔV=V0 ! 0 and reasoning as in
order to achieve quantitative models whose parameters can (2.16) and (2.15), one has
be calibrated a posteriori. Then, after some calculations, "   α #
one has  R
PV0 ðRÞ ¼ exp V0 : (2.22)
2 pffiffiffi pffiffiffi  3    3 η
32α
3 πα Γ 2 α σ max
Ps;V ðσ max Þ ¼ exp4 V h i 5;
4 Γ 3ð1þα Þ ηLT Thus, the probability of finding in V0 an inclusion with
2

(2.17) radius  R becomes

"   α #
where α is defined in (2.4), ηNiS is given by (2.7) and Γ is  R
PV0 ðRÞ ¼ 1 exp V0 : (2.23)
the Euler Gamma function. Finally, the probability that a η
plate of volume V spontaneously fails in the long term
because of the presence of a NiS inclusion can be written The experimental campaign shall consist in considering
in the form a certain number of volumes V0 of glass, and in each of
" 3  # these measuring the size of the inclusion. Data of this type,
σ max 2α

Pf ;V ðσ max Þ ¼ 1 exp VΛNiS ; (2.18) albeit fragmented and partial, have been collected by glass
ηLT
manufacturers over the years.15 Disregarding small glass
where lots from pollution crisis, all types of float glass are
comparable in terms of nickel inclusions and can be used
pffiffiffi pffiffiffi  3  
3 πα Γ 2 α for the calibration. Since equation (2.23) represents a
ΛNiS ¼ i : (2.19)
retrocumulative function, the population of measured NiS
h
4 Γ 3ð1þα Þ
2
sizes shall be ranked in descending order and an experi-
The product VΛNiS synthetically takes into account the mental probability P V0 ;i ¼ i=ðN þ 1Þ, where N is the total
effects of the variability of tensile stress through the number of data, shall be assigned to each of them. Such
thickness, ie, the probability of failure of a plate of vol- data can then be plotted in the plane x ¼ ln½RŠ vs
ume V under parabolic tensile stress with maximum σ max , y ¼ ln½ln½1=ð1 P V0 ފŠ, where equation (2.23) becomes the
is the same of a plate of volume VΛNiS with constant‐ straight line
in‐the‐thickness tensile stress σ max . Hence, it represents y¼ α x þ ln V0 þ α ln η: (2.24)
the effective volume, which depends upon the scaling
parameter α of the Pareto distribution (2.4). The values The parameters α and η, which correspond to the best
of ΛNiS is plotted in Figure 2 as a function of α within a linear fit with the experimental data, can thus be estimated,
realistic range. either graphically or analytically.
6
| BONATI
0.35
0.3
0.25
0.2
0.15
1 1.5 2 2.5 3 3.5
3 | THE EFFECTS OF THE HEAT
SOAK TEST
Inclusions originate from Sulfur and Nickel impurities in
raw materials, which are the major spurious elements found
in glass in addition to very small quantities of other metals
such as Iron and Copper.7 The HST, prescribed by the
European norm EN14179‐1,10 aims at reducing the risk of
spontaneous breakage by provoking the α‐β phase transfor-
mation before glass is mounted on buildings.
3.1 | Different types of NiS inclusions
The chemical composition of the inclusions is variable.
During manufacturing, the raw materials are heated up to
1200‐1400°C but at this temperature Nickel, not soluble in
glass, forms small melt metal drops that react with O2 and
SO3, dissolved in molten glass, generating Nickel sulfides
and oxides.7 The nickel‐rich liquid phase crystallizes into
NiS and Ni7S6, which decomposes during cooling into
Ni9S8 + Ni3S2. Thus, NiS inclusions with different
composition (such as Ni7S6, Ni9S8, Ni3S2) can be found
at the beginning of crystallization. These present a
crystallographic structure known as α‐NiS, stable at high
temperature but meta‐stable at room temperature. The trans-
formation to β‐NiS phase is rapid in stoichiometric NiS,
whereas highly over‐stoichiometric inclusions may take a
time to transform much longer than the lifetime of the
building.16 The most critical are the over-stoichiometric
inclusions, which can transform a few years after the pro-
duction of glass.
The study on NiS inclusions by Yousfi11 is carried out
on 22 inclusions, half of which found where glass fractured
after the HST, and the second half collected from plates
ET AL.
4 4.5 5 F I G U R E 2 Values for ΛNiS as a
function of α*
which broke on site. Grain sizes of α‐phase were difficult
to estimate because they were already transformed into
β‐phase. Anyway, different microstructures were detected,
a finding that indicates that the initial composition of the
α‐phase often departs from the stoichiometric composition,
opening the possibility for different transformation mecha-
nisms. NiS zones in different stones present compositions
variable from 50.1 to 51.5 at %S. The majority of inclu-
sions are low over‐stoichiometric, being their average com-
position around 50.2 ± 0.15 at %S and 0.2 at %Fe. There
exist also inclusions with higher Sulfur content around
50.5 at %S and two inclusions showed highly over‐stoi-
chiometric composition with 50.82 and 52.70 at %S.
Yousfi11 proposed a critical composition of 51.14 at %S
above which inclusions do not transform during the glass
lifetime.
For near‐stoichiometric stones (50.4 at %S) subject to
HST, the complete volume transformation is reached in a
very few minutes, of the order of 1020 minutes. The
transformation time for over‐stoichiometric inclusion can
be very long, depending upon temperature. For Ni(Fe)S
inclusions of any composition, the time temperature trans-
formation (TTT) diagrams describe, as a function of tem-
perature, the time required to achieve a certain percentage
of volume transformation. An example is in Figure 3,
which reports the TTT diagram for Ni(Fe)S of different
compositions associated with a 90% volume transforma-
tion.
Being the percentage of highly over‐stoichiometric
inclusions (which are not critical) very low,11 their pres-
ence can be neglected on the safe side. Since near‐stoichio-
metric inclusions reach their complete transformation after
a few minutes of HST, the problem consists in statisti-
cally characterizing the failures associated with the
BONATI ET AL.
300
280
260
240
220
200
180
F I G U R E 3 Time temperature
transformation (TTT) diagram for Ni(Fe)S 160
of various composition, associated with a
0 50
90% volume transformation11
transformation of over‐stoichiometric inclusions as a func-
tion of the holding time tHST in the HST.
It seems reasonable to assume that the distributions of
near‐ and over‐stoichiometric inclusions have the same
shape parameter α . Thus, if 100ðξÞ% of the inclusions,
with ξ<1, is over‐stoichiometric (100(1 − ξ)% is near‐stoi-
chiometric), by reasoning as in Section 2.3, the probability
of finding in V0 an over‐stoichiometric inclusion of radius
≥R is still given by (2.23), but substituting V0 =ξ for V0 , ie,
"   α #
 V0 R
PV0 =ξ ðRÞ ¼ 1 exp ; (3.1)
ξ ηos
where α is the same as before, while ηos is the scale
parameter of the Pareto distribution (2.4), now for over‐
stoichiometric inclusions. Once the value of the coefficient
ξ has been defined, this parameter can be estimated from
equation (2.24), as described in Section 2.3.
The distribution (3.1) will be subject to a right‐trunca-
tion induced by the HST, ie, a certain number of over‐stoi-
chiometric inclusions will lead to plate collapse into the
oven, depending upon the operating time and temperature
of the HST. This is governed by the contact pressure with
the glass matrix generated by the volume expansion δv dur-
ing the α‐β transformation of the inclusion. The value of δv
can be obtained from the TTT diagram and, for a fixed
temperature, it is a function of the holding time tHST , ie,
δv ¼ δvðtHST Þ. The TTT curve is associated with a percent-
age of transformed volume: this has to be multiplied by the
variation of volume for a complete α β transformation in
order to obtain δv.
However, the contact pressure is also influenced by the
difference in thermal expansion coefficients of glass
(αG ¼ 9:0  10 6 K 1 ) and NiS (αN ¼ 16:3  10 6 K 1 ).4
| 7
100 150 200
At the last step of the thermal process of tempering, glass
is rapidly cooled to ambient temperature but, since αN is
higher than αG , the inclusion contracts more than the glass
so that there is a gap between the resulting void bubble in
glass and the inclusion. Hence, during the α‐β transforma-
tion, the inclusion will be in contact with glass only after a
certain relative volume expansion δv , which reads
   3
δv ¼ ðαN αG Þ Tg T þ1 1; (3.2)
where Tg ¼ 525 C is the glass transformation temperature
and T the environmental temperature. Since the inclusion is
small, one can consider the expansion in an infinite elastic
medium. Thus the hydrostatic contact pressure Πδv conse-
quent to the expansion δv can be calculated, for δv  δv ,
as
ð3λN þ 2μN Þ4μG
Πδv ¼ 1 ðδv δv þ 1Þ1=3 ; (3.3)
3λN þ 2μN þ 4μG
where λN ¼ 19:5 GPa, μN ¼ 29:2 GPa, λG ¼ 24:2 GPa,
and μG ¼ 28:5 GPa are the Lamè constants of the NiS
inclusion and glass, respectively. As an order of magnitude,
for a fully developed α‐β transformation, one has
δv ¼ 0:025;5 then, for an environmental temperature of
25°C, from (3.2) and (3.3) the pressure is of the order of
260 N/mm2.
Therefore, some of the inclusions will break the glass
after the HST treatment, following the expression (2.5) pro-
posed by Swain,4 but now by substituting Πδv for Π0 .
Moreover, the critical SIF KLT for long‐term phenomena
should be substituted by KHST , which represents the critical
value for the holding time in the HST, in general higher
than KLT because of the limited effect of subcritical crack
propagation. In conclusion, the HST leads to an upper
8
| BONATI ET AL.

truncation of the population of existing inclusions in the

  α
α 1 R
survived specimens. To evaluate this, one should take into ptr;os
ΔV ðR; zÞ ¼ αþ1 α
: (3.6)
R0;os
 
R0;os Rtr ðzÞ R0;os
account the variability of the residual stresses in the thick-
ness, as indicated in Figure 1; this implies that the upper By integrating (3.6) with respect to R, with limits of
bound Rtr for the truncated distribution of over‐stoichio- integration R and Rtr , one obtains the cumulative probabil-
metric inclusions depends upon the distance z from the ity function, which reads
middle plane of the plate and the maximum tensile stress   α 

σ max at z ¼ 0. From (2.5) one obtains ;tr;os Rtr ðzÞ R α

PΔV ðR; zÞ ¼   α ; (3.7)
Rtr ðzÞ α
R0;os
( "  2 #) 3=2
π 2 1=2 2z
Rtr ðσ max ; zÞ ¼ K Π σ max 1 3 ; where α ¼ α 1 again. Then, by substituting (2.5) and
7:1 HST δv h
(3.4) into (3.7), the probability of finding an over‐stoichio-
(3.4) metric inclusion in the RVE of volume ΔV associated with
σ cr at a distance z takes the form

n h 3=2 α α
2 1=2 2z2 io h
2 1=2
i
KHST Πδv σ max 1 3 h KLT Π0 σ cr3=2
Ptr;os
ΔV ðσ max ; σ cr ; zÞ ¼ α ; (3.8)
n h 2z2 io 3=2 h i α
2 Π 1=2
KHST σ max 1 3 2 Π 1=2 σ 3=2
KLT
δv h 0 R0

pffiffiffi pffiffiffi
where z varies between h=ð2 3Þ and þh=ð2 3Þ, ie, the where σ R0 is the critical stress for inclusions of radius R0;os
interval in which the residual stresses are tensile. The pres- as per (2.5), defined as
sure Πδv is given by equation (3.3), where δv carries the 2=3
πK 2

dependence upon tHST . σ R0 ¼ pffiffiffiffiffiLT
ffi : (3.9)
7:1 Π0 R0;os

3.2 | The probability of failure after the HST Equation (3.8) can be rewritten in the more convenient
form
The statistical population of the over-stoichiometric
inclusions can be obtained from the same population in the 3 
n h  2 io32α
ϒHST σ cr 2α σ max 1 3 2zh
pristine material, expressed by (2.1), but taking into Ptr;os
ΔV ðσ max ; σ cr ; zÞ ¼ ;
3 
2α
n h 2z2 io32α
account the truncation defined by (3.4). This implies a new ϒHST σ R0 σ max 1 3 h
form for the constant C according to the normalization
(3.10)
requirement
where the quantity ϒHST >1, which is of the form
Z Rtr ðzÞ
α
CR α dR ¼ 1 ) CðzÞ 2 1=2
 
KLT Π0
R0;os ϒHST ¼ α ; (3.11)
2 Π 1=2
α 1 KHST δv
¼h   αþ1 α
i ; (3.5)
α
R0;os Rtr ðzÞ R0;os R0;os synthetically collects the effects of the HST upon the distri-
bution of over‐stoichiometric inclusions. Here Πδv of (3.3)
where R0;os is the counterpart of R0 of (2.3) for the over‐ carries the dependence upon tHST through δv. From (3.9)
1
stoichiometric population. This leads to the truncated prob- and observing that ηos ¼ R0;os ΔV α , equation (3.10) can
ability density function of the type be rewritten in the form

3 
n h 2z2 io32α
ϒHST σ cr 2α σ max 1 3 h
Ptr;os
ΔV ðσ max ; σ cr ; zÞ ¼ ΔV : (3.12)
α
n h 2z2 io32α
Π0 1=2 ηos KHST
7:1 2

ϒHST π ΔV σ max 1 3 h
BONATI ET AL. | 9

For values of the parameters within reasonable ranges, Ptr;os

f ;ΔV ðσ max Þ
the second term at the denominator in (3.12) is negligible pffiffi (Z
σ max ½1 3ðð2zÞ=hÞ2 Š
)
1 þh=ð2 3Þ
Z
with respect to the first one. Thus, (3.12) can be simplified d tr
¼ pffiffi P ðσ cr ; zÞ dσ cr dz:
as h h=ð2 3Þ σ cr;min ðzÞ dσ cr ΔV
n h  2 io32α (3.17)
3 
ϒHST σ cr 2α σ max 1 3 2zh
Ptr;os After some calculations, (3.17) becomes
ΔV ðσ max ; σ cr ; zÞ ¼ ΔV 3 
α
;
ϒHST η2LT;os
Ptr;os
f ;ΔV ðσ max Þ
(3.13) !32α Z pffiffi "  2 #32α
þh=ð2 3Þ
ΔV ϒHST 1 σ max 2z
where ηLT;os is given by (2.7), when η is substituted by ηos . ¼ pffiffi 1 3 dz
h ϒHST ηLT;os h=ð2 3Þ h
Differently from the case described in Section 2.2, it is
!32α
not straightforward to define the counterpart of (2.9) for ϒHST 1 σ max0
Ph ðσ max ; σ cr Þ, since PtrΔV of (3.13) depends upon z. Assum- ¼ ΔVΛNiS ;
ϒHST ηLT;os
ing again that the axes of the generated cracks are always
at right angle with the maximum tensile stress direction (3.18)
(P? ¼ 1), the probability that a RVE placed at z becomes
critical for a stress comprised between σ cr and σ cr þ dσ cr where ΛNiS , still given by (2.19), has been plotted as a
for an over‐stoichiometric inclusion reads function of α in Figure 2.
The probability of survival of a RVE is Ptr;os s;ΔV ðσ max Þ
¼ 1 Ptr;os ðσ
f ;ΔV max Þ. Since a volume V survives if all the
d tr
dPftr;os
;ΔV ðσ max ; σ cr ; zÞ ¼ P ðσ max ; σ cr ; zÞdσ cr : (3.14) RVE constituting V survives, one can write the probability
dσ cr ΔV
of survival of a plate of volume V in the form

To obtain the averaged‐in‐the‐thickness probability

2 !32α 3V=ΔV
ϒHST 1 σ max
densitypffiffiffi function, (3.14)
pffiffiffi must be integrated for Ptr;os
s;V ðσ max Þ ¼ 1
4 ΔVΛNiS 5 :
h=ð2 3Þ  z  þh=ð2 3Þ, ie, in the part under tensile ϒHST ηLT;os
stress, and divided by h. (3.19)
The integration with respect to σ cr deserves further con-
siderations. The maximum size of the inclusion that can be Taking the limit ΔV=V ! 0 for large V, one obtains
found at z is given by Rtr ðσ max ; zÞ of (3.4). For long‐term 2 !32α 3
conditions, the relationship between critical radius and criti- tr;os ϒ HST 1 σ max
cal stress is given by (2.5). Therefore, the minimum critical lim Ps;V ðσ max Þ ¼ exp4 VΛNiS 5:
ε!0 ϒHST ηLT;os
stress that can affect the inclusion at z, say σ cr;min ðzÞ, is
given by (3.20)
π 2  3=2
K Π0 1=2 σ cr;min ðzÞ

¼ Rtr ðσ max ; zÞ ) σ cr;min ðzÞ However, there is another cause of potential rupture:
7:1 LT
2=3 although their α‐β transformation will be certainly com-
2
Π0 1=2

πKLT
¼ : pleted after the HST, near-stoichiometric inclusions can
7:1Rtr ðσ max ; zÞ still cause delayed spontaneous rupture. In fact, due to sub-
(3.15) critical crack propagation, the β‐phase volume increase
may not be critical for a time of the order of tHST , for
which the critical SIF is KHST , but become unstable in the
By substituting (3.4) into (3.15), one obtains long‐term lifetime, for a SIF equal to KLT .
The probability of survival can be obtained with
"   2 #
KLT 4=3 Πδv 1=3 exactly the same arguments leading to (3.20), provided
  
2z
σ cr;min ðzÞ ¼ σ max 1 3 : that one use for the scale parameter the value ηns , which
h KHST Π0
can be estimated similar to ηos as in (3.1) with the
(3.16)
graphical procedure of Section 2.3, by using 1 − ξ in
Equation (3.14) shall thus be integrated with respect place of ξ. Moreover, the coefficient ϒHST is obtained
hto σ cr betweeni σ cr;min ðzÞ, given by (3.16), and σ max from (3.2) by using for Πδv the contact pressure Π0 ,
1 3ðð2zÞ=hÞ2 . The probability of spontaneous collapse associated with δv ¼ 0:025 in (3.3), which refers to a
due to the presence of an over‐stoichiometric NiS inclusion complete transformation. In conclusion, the probability of
at the RVE level is then given by survival after the HST reads
10
| BONATI ET AL.

from the technical literature in order to show the potentiali-

2 !3α 3
ϒHST jΠ0 1 σ max 2
Ptr;ns
s;V ðσ max Þ ¼ exp
4 VΛNiS
ϒHST jΠ0 ηLT;ns
Since a glass element survives if and only if neither the
near‐stoichiometric, nor the over‐stoichiometric inclusions
reach their critical condition, the survival probability
Ptrs;V ðσ max Þ after HST will be given by the product of the
corresponding survival probabilities. Hence, emphasizing
the dependence of ϒHST upon Pδv and, consequently, upon
the holding time tHST by writing ϒHST ¼ ϒHST ðtHST Þ, the
probability that an element of volume V collapses during
its service life after the HST is Ptrf;V ðσ max Þ ¼ 1 Ptrs;V ðσ max Þ
and reads
Ptrf;V ðσ max ; tHST Þ ¼ 1 Ptr;ns
s;V ðσ max Þ
2 !32α 3
ϒ HST ðtHST Þ 1 σ max

exp4 VΛNiS 5:
ϒHST ðtHST Þ ηLT;os
To complete the analysis, one can consider the limit
case of a short HST, of the order of 15 minutes. Recall
from the discussion on (3.2) that the contact between NiS
and glass will occur only after a certain volume expansion
because the NiS initially lies inside a void bubble. There-
fore, in short HST one can assume that over‐stoichiometric
inclusions do not stress the glass, ie, Πδv ! 0 and conse-
quently, from (3.2), ϒHST ! þ1. In conclusion, one has
from (3.22) that
Ptrf;V ðσ max ; shortÞ ¼1 Ptr;ns
s;V ðσ max Þ
2 !32α 3
σ max

exp4 VΛNiS
ηLT;os
Remarkably, in this case, the probability of survival
Ptr;os
s;V ðσ max Þcoincides with the expression obtained for the
case of no HST, given by equation (2.18), provided that
ηLT is substituted by ηLT;os . In fact, we have made the
hypothesis that in short HST the over‐stoichiometric inclu-
sions remain unaffected. On the other hand, the near‐stoi-
chiometric ones are completely transformed, so that the
term Ptr;ns
s;V ðσ max Þ does not change.
4 | A PRACTICAL APPLICATION
ties of the proposed model in a paradigmatic practical case.
5:
4.1 | Statistical population of inclusion sizes
(3.21) for glass
To calibrate the parameters α and η of the Pareto distribu-
tion (2.4) with the method described in Section 2.3, the
experimental data need to be statistically associated with a
reference volume V0 , defined as the volume that statisti-
cally contains on average only one inclusion. According to
Kasper,15 in standard transparent float glass one breakage
every six tons of material occurs during the HST: since it
is estimated that in this process approximately 25% of the
existing NiS inclusions leads to failure, one derives the
presence of one inclusion every 1.5 tons of material so
that, for a density of 2500 kg/m3, the reference volume is
V0 = 0.6 m3. Concerning the inclusion sizes, the most
comprehensive database seems to be that collected by the
China and Korea laboratories of Saint‐Gobain, but only
(3.22)
partial data have been presented in conferences12 or are in
course of publication.13 Inclusions were optically detected
in three lots of several tons of glass: 135 NiS inclusions
were found in green glass (0.6% Fe2O3 colored, 5 mm
thick), while the other lots, composed by float transparent
glass were significantly less contaminated. Unfortunately,
the distribution of inclusion sizes has been approximately
measured only for green glass,12,13 but not for transparent
glass, whereas the volume V0 is known only for transparent
glass but not for green glass.2
We then conjecture that the population of inclusion
sizes for transparent glass has the same shape as for the
green glass, ie, it is described by Pareto distributions of the
form (2.4) that have the same shape parameter α , but dif-
(3.23) fer in the scale parameter η, which depends upon the refer-
5:
ence volume V0 as described in Section 2.3. Using the data
by Kasper,13 the observed inclusion sizes are grouped in
classes of 50 μm width. The relative frequency pi for the
ith class is given by the ratio between the number of the
inclusions assigned to that class and the total number of
inclusions. The histogram of the so‐obtained data‐set is
shown in Figure 4.
Then, the retrocumulative probability, ie, the probability
of finding a NiS inclusion of radius  Ri associated with
the ith class, is obtained from
i
P ðRi Þ ¼ 1 ∑ pj ; (4.1)
j¼1

where the central value of any ith interval is assumed as

Unfortunately, the recorded data for the frequency and size
of NiS inclusions are partial and fragmentary. Therefore,
while waiting for a proper experimental campaign, we
attempt, albeit tentatively, an interpretation of the results
the value of the inclusion radius Ri . The parameters of the
2
Certainly, V0 should be smaller in green glass, much more contaminated,
than in transparent glass.
BONATI ET AL. | 11

Pareto distribution (2.4) are thus calibrated graphically as
in Section 2.3 using V0 = 0.6 m3. The probability plot of
(2.24) is shown in Figure 5A, from which the parameters

α ¼ 2:056 and η ¼ 71:78 μm m 1=α are obtained. The
resulting retrocumulative function of inclusion size of
(2.23) is shown in Figure 5B.
4.2 | Statistical analysis for
KLT ¼ K0 ¼ 0:3 MPa m1=2 and
KHST ¼ KIc ¼ 0:75 MPa m1=2
Assuming from5 that a complete α‐β transformation leads
to a volume expansion of the order of 2.5%, recalling (3.2)
and using δv ¼ 0:025 in (3.3), one obtains
Πδv ¼ Π0 ≅ 260 MPa. One may assume on the safe side,
as indicated in footnote 1, that KLT ¼ KI0 ¼ 0:3 MPa m1=2 ,
which is the limit below which subcritical crack
F I G U R E 4 Histogram of the inclusions size (radius of the
spherical inclusions in μm) for green glass, as derived from available
data 12,13
propagation does not occurs. In this way, (2.7) provides

ηLT ¼ 5:012 MPa m2=α .
Consider the case of a standard glass window of size
1000 mm × 1000 mm × 4 mm, hence of volume V = 4 ×
10−3 m3. For α* = 2.056, one has ΛNiS ¼ 0:261 from
equation (2.19). The surface compression for fully tem-
pered glass varies in general between 95 MPa and
105 MPa, while the maximum tensile stress in the middle
is, from (2.8), half of the surface compression, ie, it varies
between 47.5 MPa and 52.5 MPa. Assume then that
σ max ¼ 50 MPa. The risk of spontaneous breakage during
lifetime, given by equation (2.18), is Pf ;V ðσ max Þ ¼ 0:1181.
This means that one breakage every Nb ¼ 8  9 windows
is expected (Nb ¼ 1=Pf ;V ≅ 8:5). This value may be con-
sidered unrealistically high, because it comes from the very
conservative assumption KLT ¼ KI0 . In any case, one
should consider that the results from the proposed model
strongly depend upon the assumed volume expansion con-
sequent to the complete α‐β phase transformation of NiS,
which is here assumed to be 2.5%, although other studies
provide an estimate between 2.5%5 and 4%.4 This certainly
represents a key‐point to be clarified by further studies.
If the HST is short, ie, holding time is of the order of
15 minutes, the near‐stoichiometric inclusions can undergo
the complete α‐β transformation, whereas the over‐
stoichiometric ones are almost unchanged. Hence, the prob-
ability of spontaneous collapse during the life‐time of a
window is given by equation (3.23). If ξ is the percentage
of over‐stoichiometric inclusions, for these the reference
volume becomes V0 =ξ, while it is V0 =ð1 ξÞ for near‐
stoichiometric inclusions. In agreement with the qualitative
estimations by Yousfi,11 it is assumed ξ ¼ 0:33. Consider-
ing again the data presented by Kasper,12,13 for a reference
volume V0 =ξ ¼ 0:6 m3/0.33 = 1.82 m3, the Pareto parame-
ter, besides α ¼ 2:056 which is independent of V0, is

2 1
(A) (B)
0.9 Experimental points
1 Theoretical function
0.8
Retro-cumulative probability
0.7
0
0.6

-1 0.5

0.4

-2
0.3

0.2
-3
0.1

-4 0
3 3.5 4 4.5 5 5.5 6 0 100 200 300 400 500 600

F I G U R E 5 (A) Probability plot from equation (2.24) of the inclusion sizes R. (B) Obtained retrocumulative probability function ((2.4) with
α* = 2.056 and η = 71.78 μm m−1/α*)
12
| BONATI ET AL.

ηos ¼ 41:86 μm m−1/α*. Supposing again that the maximum
volume expansion due to the α‐β transformation is 2.5%,

from (2.7) one obtains ηLT;os ¼ 15:15 MPa m2=α . For what
concerns the population of near‐stoichiometric inclusions,

one has V0 =ð1 ξÞ ¼ 0:89 m3 and ηns ¼ 59:08 μm m 1=α ,

from which ηLT;ns ¼ 12:04 MPa m2=α . For a maximum
residual stress σ max ¼ 50 MPa, from (3.21) one finds
tr;ns
Ps;V ðσ max Þ ¼ 0:921. Therefore, the risk of spontaneous
breakage, given by equation (3.23), becomes
Ptrf;V ðσ max ; shortÞ ¼ 0:1164, only slightly lower than for the
case of no HST.
The effect of the HST upon the distribution of over‐stoi-
chiometric inclusions depends upon its duration time tHST
for the standardized temperature of 260°C. One of the
major difficulties is to choose the type of over‐stoichio-
metric inclusion to be considered. Following Yousfi,11
assume that Ni48.86S51.14 is the reference composition: Fig-
ure 6 indicates how the transformed volume fraction fv;os
varies with holding time tHST at 260°C. The values of fv;os
are recorded in Table 1.
From equation (3.22), with α ¼ 2:056, ΛNiS ¼ 0:261,
tr;ns
ηLT;os ¼ 15:15 MPa m2 =α , σ max ¼ 50 MPa, Ps;V ðσ max Þ ¼
0:921 and ϒHST ðtHST Þ given by (3.2), the probability of
spontaneous failure Ptrf;V ðσ max ; tHST Þ correlated with the
transformed volume fraction fv;os at holding time tHST at
260°C is obtained. The values of Ptrf;V ðσ max ; tHST Þ and the
expected number of windows for one breakage Nb are
recorded in Table 1 and plotted as a function of tHST in
Figure 7. Observe that the distribution of over‐stoichio-
metric inclusions does not vary for tHST < ∼ 20 min, corre-
sponding to a transformed volume fraction fv;os < ∼ 0:5.
This is due, as discussed in (3.2), to the formation of a
void bubble consequent to the different thermal expansion
of NiS and glass.
It is evident from Figure 7A that Ptrf;V very weakly
depends upon tHST . Comparing the probability of collapse
for the case of no HST (Pf ;V ðσ max Þ ¼ 0:1181) with those
of Table 1 and Figure 7, it is evident that the HST pro-
vides almost no benefit. This is certainly due to the fact
that the assumption KLT ¼ KI0 ¼ 0:3 MPa m1=2 is exces-
sively conservative, providing unrealistically high values of
ϒHST , given by (3.2) and shown in Figure 8. At the same
time, the SIF KHST should be lower than the critical SIF
KIC ¼ 0:75 MPa m1=2 ,17 because glass plates do not col-
lapse only in the course of the test, but also during
transportation and application.

0.8

0.6

0.4

0.2

0 F I G U R E 6 Effect of holding time tHST

0 10 20 30 40 50 60 70 80 on the transformed volume fraction fv,os for
Ni48.86S51.1411

T A B L E 1 Correlation among transformed volume fraction fv,os of over‐stoichiometric inclusions, tHST at 260°C, probability of failure in
service life Ptrf;V ðσ max ; tHST Þ, and expected number of windows for one breakage Nb. Inclusions composition Ni48.86S51.14
(KLT = KI0 = 0.3 MPa m1/2 and KHST = KIC = 0.75 MPa m1/2)

fv,os 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tHST [min] 12.62 15.08 16.91 18.64 20.35 22.36 24.68 28.26 34.35 50.75
Ptrf;V 0.11664 0.11664 0.11664 0.11664 0.11634 0.1162 0.11604 0.11589 0.11574 0.11558
Nb 8.57 8.57 8.57 8.57 8.59 8.61 8.62 8.63 8.64 8.65
BONATI ET AL. | 13

0.1168 8.8
(A) (B)
8.75
0.1166

8.7
0.1164

8.65
0.1162

8.6

0.116
8.55

0.1158
8.5

0.1156
8.45

0.1154 8.4
0 5 10 15 20 25 30 35 40 45 50 55 0 5 10 15 20 25 30 35 40 45 50 55

F I G U R E 7 Correlation between tHST and (A) the probability of spontaneous failure Ptrf;V from (3.22); (B) the expected number of windows
for one breakage Nb (KLT ¼ 0:3 MPa m1=2 and KHST ¼ 0:75 MPa m1=2 )

450

400

350

300

250

200

150

100

F I G U R E 8 Values of ΥHST as a 50
function of the hydrostatic pressure induced
by the α‐β transformation with 0
KLT ¼ 0:3 MPa m1=2 and
0 50 100 150 200 250 300
KHST ¼ 0:75 MPa m1=2

4.3 | Analysis for KLT ¼ 0:55 MPa m1=2 and
KHST ¼ 0:6 MPa m1=2
The statistical analysis is here repeated with
KLT ¼ 0:55 MPa m1=2 and KHST ¼ 0:6 MPa m1=2 . These
changes do not affect the calibration of the parameters of
the Pareto distribution, so that α ¼ 2:056, η ¼
 
71:78 μm m 1=α , ηos ¼ 41:86 μm m 1=α and ηns ¼

59:08 μm m 1=α . From equation (2.7), one obtains ηLT ¼
 
23:73 MPa m2=α , ηLT;os ¼ 33:99 MPa m2=α and ηLT;ns ¼
2=α
27:02 MPa m . Then, from (3.21), one has Ptr;ns s;V ðσ max Þ
¼ 0:9979.
To sum up, the probability of spontaneous collapse
Pf,V(σmax) (no HST) is given by (2.18) with
V = 4 × 10−3 m3, ΛNiS = 0.261, σmax = 50 MPa, ηLT =
23.73 MPa m2/α*, and α* = 2.056. For short HST, it is
given by equation (3.23), with ηLT,os = 33.99 MPa m2/α*
and Ptr;ns
s;V ðσ max Þ ¼ 0:9979. After a HST of holding time
tHST, it is given by (3.22), where the dependence upon
tHST is brought by the quantity ΥHST defined in (3.2),
which is affected by the values for the critical SIFs KLT
and KHST.
The risk of breakage with no HST test is Pf,V(σmax) =
1.035 × 10−3, which means one breakage for every 97
windows. After a short HST it becomes Ptrf;V ðσ max ; shortÞ
¼ 5:51  10 3 , ie, one breakage for every 181 windows.
By comparing these results with those of the
previous section, it is evident that, by increasing KLT
and by reducing KHST, even a very short HST offers a
remarkable beneficial contribution. Table 2 takes the
place of Table 1, and shows how the collapse
probability varies with the holding time tHST. The coun-
terparts of Figures 7 and 8 are, respectively, the Fig-
ures 9 and 10.
14
| BONATI ET AL.

By comparing Figures 7 and 9, it is evident that the
higher is the value of KLT (0.55 vs 0.3 MPa m1/2) and the
lower is KHST (0.6 vs 0.75 MPa m1/2), the more relevant is
the dependence of the probability of spontaneous failure
induced by NiS inclusions upon the holding time tHST of
the HST. In particular, when KHST is close to KLT, the risk
of breakage during the long‐term life of the element due to
the subcritical crack growth is minimized while the effi-
ciency of the HST is maximized, whereas the opposite is
true when KHST ≫ KLT.
This is evident most of all by comparing Figures 8 and
10. The values reached by ΥHST(tHST) are much closer to

T A B L E 2 Correlation between the transformed volume fraction fv,os of over‐stoichiometric inclusions, tHST at 260°C, probability of failure
during service life Ptrf;V ðσ max ; tHST Þ, and expected number of windows for one breakage Nb. Inclusions composition Ni48.86S51.14
(KLT = 0.55 MPa m1/2 and KHST = 0.6 MPa m1/2)

fv,os 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tHST [min] 12.62 15.08 16.91 18.64 20.35 22.36 24.68 28.26 34.35 50.75
Ptrf;V 0.00551 0.00551 0.00551 0.00551 0.00528 0.00486 0.00443 0.004 0.00357 0.00313
Nb 181.49 181.49 181.49 181.49 189.51 205.87 225.7 250.02 280.45 319.52

(A) 6 10-3 (B) 340

320
5.5
300

5 280

260
4.5
240

4 220

200
3.5
180

3 160
0 5 10 15 20 25 30 35 40 45 50 55 0 5 10 15 20 25 30 35 40 45 50 55

F I G U R E 9 Correlation between tHST and (A) the probability of spontaneous failure Ptrf;V , from (3.22); (B) the expected number of windows
for one breakage Nb (KNiS = 0.55 MPa m1/2 and KHST = 0.6 MPa m1/2)

15

10

F I G U R E 1 0 Values of ΥHST as a
function of the hydrostatic pressure induced
0 by the α‐β transformation with
0 50 100 150 200 250 300
KLT = 0.55 MPa m1/2 and
KHST = 0.6 MPa m1/2
BONATI ET AL.
unity and, consequently, the effects on the population of
over‐stoichiometric inclusions, whose truncation depends
upon the holding time of the HST, are higher.
The examples just presented confirm the importance of
a precise characterization of the critical SIFs KLT and KHST.
One may tentatively assume that the critical SIF KIC(t) for
a phenomenon of characteristic duration t, is equal to
Kmod(t) · KIC, where KIC is the critical SIF for a load his-
tory defined by standards,18 whereas Kmod(t) is the factor
for load duration that defines the macroscopic strength of
float glass that is affected by static fatigue. However, such
values are usually calculated by using several approxima-
tions, which include neglecting the existence of a lower
bound KI0 below which subcritical crack growth does not
occur.19 For example, for t = 50 years most standards sug-
gest a value of Kmod of the order of 0.3, but 0.3 KIC ≅ KI0.
On the other hand, the illustrative example recorded in Sec-
tion 4.2 has shown that if one considers KLT = KI0 the
effects of subcritical crack growth are maximized, whereas
the effects of the HST are unrealistically almost annihi-
lated. Therefore, for this particular application, it is neces-
sary to determine more precisely the dependence of the
critical stress factor upon time, because a conservative esti-
mate, as usually done in the structural design, would imply
that the HST is almost useless, in contradiction with the
practical experience. A similar problem also arises for the
determination of KHST.
5 | DISCUSSION AND
CONCLUSIONS
A micro‐mechanically motivated statistical model, which
correlates the statistical expectation of finding a NiS
inclusion of critical size in the inner part of a glass sheet
with the breakage consequent to its volumetric expansion
due to the α‐β phase transformation, has been proposed
for assessing the risk of spontaneous collapse of thermally
treated glass. The population of sizes of NiS inclusions is
interpreted by a Pareto probability function, with shape
and scale parameters, respectively α* and η, which can be
experimentally evaluated as indicated in Section 2.3. The
critical conditions that lead to failure are provided follow-
ing Swain.4 These assumptions imply a probability of
spontaneous failure during the element lifetime governed
by a two‐parameter Weibull distribution of the type
" 3  #
σ max 2α

Pf ;V ðσ max Þ ¼ 1 exp VΛNiS ; (5.1)
ηLT
where the coefficients, illustrated in Section 2.2, are the
shape and scale parameters 32 α and ηLT, respectively, the
| 15
volume V of the considered element and the parameter
ΛNiS, accounting for the variability in the thickness of the
residual stresses. The scale parameter ηLT depends upon η,
upon the contact pressure Π0 between NiS and glass, gen-
erated by the volume increase in the α‐β transformation of
NiS, and upon the critical SIF KLT for long‐term subcritical
crack propagation.
To evaluate the effects of the HST, in particular for
what concerns the holding time tHST, a distinction has
been made between near‐stoichiometric and over‐stoichio-
metric NiS inclusions, associated with either a short or
long phase‐transition time, respectively. We have sup-
posed that both populations can be interpreted by Pareto
distributions with the same shape parameter α*, but dif-
ferent scale parameters. Even a very short HST, with
tHST of the order of a few minutes, produces the com-
plete α‐β phase transformation in near‐stoichiometric
inclusion, whereas the over‐stoichiometric ones do not
change. The consequent probability of spontaneous fail-
ure during lifetime is
Ptrf;V ðσ max ; shortÞ ¼ 1 Ptr;ns
s;V ðσ max Þ
2 !32α 3
σ max (5.2)

exp4 VΛNiS 5:
ηLT;os
With respect to (5.1), the scale parameter ηLT is substituted
by ηLT,os, obtained from (2.7) with ηos in place of η; more-
over, Ptr;ns
s;V ðσ max Þ, given by (3.21), represents the probabil-
ity that near‐stoichiometric inclusions, completely
transformed in their β‐phase in the HST, do not break
glass. Observe that, from this analysis, an inclusion that
has reached its complete α‐β transformation during HST
may become critical in the long term because of subcritical
crack growth.
The over‐stoichiometric inclusions gradually transform
with time increasing their volume, so that the contact
pressure with glass can lead to failure. One has to con-
sider that, due to the differential thermal expansion with
glass during cooling, NiS inclusions lie in void bubbles,
and consequently the contact occurs only after a certain
volume expansion. This is why short HSTs do not mod-
ify the distribution of over‐stoichiometric inclusions,
which can change only for reasonable holding time tHST.
From a statistical point of view, this leads to an upper
truncation of their statistical distribution in the surviving
specimens and, consequently, to a reduction in the risk
of breakage during service life. From Section 3.2, the
collapse probability becomes
16
| BONATI
tr;ns
Ptrf;V ðσ max ; tHST Þ ¼ 1 Ps;V ðσ max Þ
2 !32α 3
ϒ ðt Þ 1 σ
HST HST max

exp 4 VΛNiS 5;
ϒHST ðtHST Þ ηLT;os
(5.3)
where the difference with the previous expression is in the
coefficient ΥHST, with ΥHST > 1. This term, defined in
(3.2), carries the dependence upon tHST.
To appreciate the potentialities of the proposed model, a
practical application has been attempted in Section 4. The
required input data are the volume fraction of over‐stoi-
chiometric and near‐stoichiometric inclusions, their volume
expansion due to the complete α‐β transformation, the dis-
tribution of inclusion sizes for a reference volume of glass,
and the critical SIFs for glass, associated with either long‐
term phenomena in the service life (KLT) or times compara-
ble with tHST (KHST). These data have been tentatively esti-
mated from what found in the literature. Graphs have been
presented that show the probability of spontaneous rupture
during lifetime in the case of no HST, short HST and long
HST. In the latter two situations, the values of KLT and
KHST strongly affect the dependence of the probability of
rupture on tHST.
This study aims at providing a contribution to solve
the long‐standing problem of defining the risk of sponta-
neous breakage in glass due to NiS inclusions. Further-
more, the proposed model would allow to calibrate the
holding time of the HST on the basis of the risk reputed
acceptable for the particular glass element under consider-
ation, which defines the accepted probability of collapse.
Unfortunately, there are still open issues that should be
considered by further studies. For example, the values of
KLT and KHST should be precisely assessed from micro‐
mechanically motivated measurements to take into account
the effect of static fatigue, because even a slightly conser-
vative estimate, accepted in structural design, would
underestimate the benefic effects of the HST, as illustrated
in the presented examples. In any case, calibration of the
model starts from the statistical characterization of the
population of NiS inclusions in commercial glass but,
unfortunately, although many data have been collected in
the last 50 years by manufacturers, these appear to have
been only fragmentarily and partially reported in the tech-
nical literature. We hope that our study will promote fur-
ther research, suggesting an organized collection and
organization of existing data and encouraging new tar-
geted experimental campaigns.
ET AL.
ACKNOWLEDGMENTS
GRC acknowledges the support of the Italian Ministry of
University under grant MIUR‐PRIN voce COAN 5.50.16.01
code 2015JW9NJT and of Italian Civil Protection Depart-
ment, Presidency of the Council of Ministers, under project
ReLUIS‐DPC 2014‐2018.
ORCID
Gianni Royer Carfagni http://orcid.org/0000-0003-4879-
9846
REFERENCES
1. Shelby J. Introduction to glass science and technology. Cam-
bridge, UK: Royal Society of Chemistry; 2005.
2. Wiederhorn SM, Bolz LH. Stress corrosion and static fatigue of
glass. J Am Ceram Soc. 1970;53(10):543–8.
3. Pisano G, Royer Carfagni G. Statistical interference of material
strength and surface prestress in heat‐treated glass. J Am Ceram
Soc. 2017;100(3):954–67.
4. Swain MV. Nickel sulphide inclusions in glass: an example of
microcracking induced by a volumetric expanding phase change.
J Mater Sci. 1981;16:151–8.
5. Schneider J, Hilcken J. Nickel sulphide (NiS-) induced failure of
glass: fracture mechanics model and verification by fracture data.
In: Proceedings of International Conference at Glasstec. Egi-
neered Tansparency. Düsseldorf, DE; 2010:125–36.
6. Ballantyne ER. Report 061-5. Technical report, CSIRO. Division
of Building Research, Melbourne (AUS); 1961.
7. Kasper A, Stadelmann H. Chemical behavior of nickel sulfide in
soda lime glass melts. Glastech Ber Glass Sci Technol. 2002;75
(1):1–11.
8. Karlsson S. Spontaneous fracture in thermally strengthened glass
‐ a review and outlook. Ceram‐Silik. 2017;61(3):188–201.
9. Kasper A, Bordeaux F. Nickel sulphide: new results to optimise
the heat soak test for tempered building glasses. Glastech Ber
Glass Sci Technol. 2000;75(5):130–42.
10. EN 14179-1: 2016. Glass in building - Heat soaked thermally
toughened soda lime silicate safety glass - Part 1: Definition and
description. European Norm.
11. Yousfi O. Transformations de phase des Sulfures de Nickel dans
les verres tremps [Phase transformation of Nichel Sulphide in
glass]. PhD thesis, Institut National Plytechnique de Grenoble
(FR); 2009.
12. Kasper A. News from an old theme: spontaneous cracking of
thermally toughened safety glass. In: Proceedings of Glass Per-
formance Days. Tampere, FI; 2017:154–6.
13. Kasper A. Spontaneous cracking of thermally toughened safety
glass. Part two: Nickel sulphide inclusions identified in annealed
glass. GLAS Struct Eng (Forthcoming).
14. Batdorf SB, Crose JG. A statistical theory for the fracture of brit-
tle structures subjected to nonuniform polyaxial stresses. J Appl
Mech. 1974;41(2):459–64.
BONATI ET AL. | 17

15. Kasper A. Spontaneous cracking of thermally toughened safety 19. Wan KT, Lathabai S, Lawn BR. Crack velocity functions and
glass. Part one: properties of nickel sulphide inclusions. Glass thresholds in brittle solids. J Eur Ceram Soc. 1990;6(4):259–68.
Struct Eng. 2018. https://doi.org/10.1007/s40940-018-0083-8.
16. Kasper A, Moschek S, Stadelmann H, Zeihe R. Composition and
structure of NiS inclusions in floatglass, and their impact on the
heat soak test. In: Proceedings of Glass Performance Days. Tam-
How to cite this article: Bonati A, Pisano G, Royer
pere, FI; 2003:692–5. Carfagni G. A statistical model for the failure of
17. Wiederhorn SM. Fracture surface energy of glass. J Am Ceram glass plates due to nickel sulfide inclusions. J Am
Soc. 1969;52(2):99–105. Ceram Soc. 2018;00:1–17. https://doi.org/
18. Collini L, Royer Carfagni G. Flexural strength of glass‐ceramic for 10.1111/jace.16106
structural applications. J Eur Ceram Soc. 2014;34(11):2675–85.