Reliability of Interconnect Structures

Z. SUO

Mechanical and Aerospace Engineering Department and Princeton Materials Institute

Princeton University, Princeton, NJ 08544

suo@princeton.edu
Phone: 609-258-0250
Fax: 609-258-5877

This manuscript is prepared as a chapter in

Volume 8: Interfacial and Nanoscale Failure (W. Gerberich, W. Yang, Editors)
Comprehensive Structural Integrity (I. Milne, R.O. Ritchie, B. Karihaloo, Editors-in-Chief)
Due for publication early 2003
http://www.elsevier.com/locate/cosi

9/13/2002 1
8.1.7
Reliability of Interconnect Structures

Z. SUO
Princeton University, Princeton, NJ, USA

1. INTRODUCTION

2. CRACKING

2.1 Linear Elastic Fracture Mechanics

2.2 On Applying Fracture Mechanics to the Interconnect Structure

2.3 Energy Release Rate

2.4 Cannel Cracks in a Thin Film Bonded to a Substrate

2.5 Measuring Thin Film Toughness

2.6 The Effect of Underlayer on Channel Crack

2.7 Crack Kinetics: Subcritical Cracking and Substrate Creep

3. DEBONDING AND MIXED MODE CRACK

3.1 Mixed Mode Crack in a Homogeneous Material

3.2 Interfacial Fracture Mechanics

3.3 Measuring Interfacial Fracture Energy

3.4 Effect of Plasticity

4. RATCHETING

4.1 Temperature Cycling as a Qualification Test

4.2 Ratcheting toward a Steady Sate

4.3 Shakedown vs. Cyclic Plastic Deformation

4.4 Ratcheting Strain Rate

4.5 The Transient Process to Approach the Steady State

9/13/2002 2
 4.6 Ratcheting-induced Crack Growth

5. STRESS-INDUCED VOIDING

5.1 Stress-Induced Voiding in Encapsulated Metal Line

5.2 An Analysis of Stress-Induced Voiding

5.3 Nonhydrostatic Stress State in Encapsulated Metal Line

5.4 Voiding in Blanket Copper Film

5.5 Slit-like Voids

5.6 The Effects of Low k Dielectrics and Barriers

6. ELECTROMIGRATION

6.1 Electromigration in Encapsulated Metal Line

6.2 Basic Electromigration Characteristics

6.3 Balancing Electron Wind and Stress Gradient

6.4 No-Voiding Condition for Encapsulated Line

6.5 Saturated Void Volume (SVV)

6.6 Experimental Determination of Z* ,∆ α (T − T0 ), B , and D

6.7 No-Extrusion Condition for Encapsulated Line

6.8 The Immortal Interconnect

REFERENCES

9/13/2002 3
1. INTRODUCTION

On 23 January 1959, Robert Noyce wrote in his lab notebook: “… it would be desirable

to make multiple devices on a single piece of silicon, in order to be able to make

interconnections between devices as part of the manufacturing process, and thus reduce size,

weight, etc. as well as cost per active device.” With these words, the co-inventor (with Jack

Kilby) of the integrated circuit, and the co-founder (with Gordon Moore) of Intel Corporation a

decade later, spelled out the Monolithic Idea that would shape the microelectronic revolution

(Reid, 1984). To function, transistors, capacitors, and resistors must connect among themselves

and to the external world. Rather than first making these components individually and then

connecting them with wires, the microelectronic technology fabricates the interconnect structure,

as well as all the components, into a single solid piece, the chip.

Ever since, the technology has advanced by miniaturization, squeezing more and more

transistors on every single chip. By necessity the interconnect structure has complex

architectures, diverse materials, and small feature sizes. Figure 1 shows a transmission electron

microscope image of an interconnect structure. The structure is a three-dimensional network of

conductor lines embedded in a dielectric matrix, fabricated on the silicon surface that contains

the active devices. The conductor lines are on several levels, and linked by vias. The lines have

rectangular cross-sections and, on a given level, have a constant thickness but variable widths. A

silicon nitride film caps the interconnect structure, serving as a barrier of the environmental

molecules. Lithographically defined feature sizes, such as line width and via size, are as small as

100 nm. Films as thin as a few nanometers are used.

Traditional interconnect structures use aluminum as the conductor, and silica as the

dielectric. Tungsten vias link the lines between the levels, and titanium nitride thin layers lie

9/13/2002 4
between the aluminum lines and the dielectric. To make faster devices, the conductor must have

higher conductivity, and the dielectric must have lower dielectric constant. Copper has been

used as the conductor lines in more recent interconnect structures (Rosenberg et al., 2000). The

conductor lines at different levels are linked with copper vias. Liners such as tantalum lie

between the conductor and the dielectric to prevent copper diffusion into the dielectric and to

promote adhesion. Low dielectric constant materials of current use include an organic material

known as SiLK (Martin et al., 2000), and a carbon doped oxide. The former is ductile, and the

latter is brittle. Both are compliant compared to silicon dioxide. To further reduce the dielectric

constant, other materials, including porous materials, will be used in future interconnects

(Morgen et al., 1999).

The Monolithic Idea solves one problem, but creates another. Making reliable

interconnect structures has been a persistent challenge. Stress results from material deposition,

thermal expansion mismatch, and electromigration. Material deposition inevitably generates

stress; for example, deposition of refractory metals such as tungsten usually generates stress the

GPa range. Materials in interconnect structures, selected to function as conductors, dielectrics,

or barriers, have dissimilar thermal expansion coefficients. On cooling from the fabrication

temperature, the structures acquire stresses. To fit into the small space, each conductor line must

have a small cross-sectional area. The intense electric current motivates metal atoms to diffuse

in the line, generating tension where atoms deplete, and compression where atoms accumulate.

The study of interconnect reliability has a long history. Electromigration has been a

persistent concern ever since integrated circuits were mass-produced (Blech and Meieran, 1967).

As atoms diffuse, voids grow and sever metal lines. Stress-induced voiding was observed in

1980s, when the technology brought the line width comparable to the line thickness (Curry et al.,

9/13/2002 5
1984). Such an aluminum line, of approximately square cross-section, encapsulated in silica,

develops triaxial stresses of magnitude much above the yield strength of aluminum (Jones and

Basehore, 1987). Voids grow to relax the stresses, facilitated by atomic diffusion (Sauter and

Nix, 1992; Korhonen et al. 1993). Around the same time, thin film fracture mechanics was

advanced (Evans et al., 1988; Hutchinson and Suo, 1991; Thouless, 1991), providing tools to

quantify toughness of thin films (Ma et al., 1998; Cook and Liniger, 1999), and adhesion

between thin films (Dauskardt et al., 1998; Volinsky et al., 2002). The tools have been

incorporated into industrial practice in material selection, process control, and failure analysis.

More recently, a new failure mechanism has been identified, in which cracks grow in

brittle films under cyclic temperatures, aided by ratcheting deformation in a metal underlayer

(Huang et al., 2000). Stress-induced voiding has been observed in blanket copper films (Shaw et

al. 2001, Sekiguchi et al., 2001), and in copper vias in the copper/silica interconnect structure

(Hommel et al., 2001; Ogawa et al., 2002). Low dielectric constant materials have low stiffness;

their effects on interconnect reliability are the focus of current studies (Lee et al., 2002; Ambrico

et al., 2002; Liang et al., 2002; Rim et al., 2002).

This chapter is organized according to failure modes: cracking, debonding, ratcheting,

voiding, and electromigration. Such perversities are not only nuisances—problems to be tested

away by trial-and-error, but also opportunities for fundamental study of the mechanical behavior

of diverse materials, in integrated structures, at an unprecedented small size scale. The study of

electromigration has led to the discovery that adding a few percent of copper into aluminum

interconnects slows aluminum diffusion rate by nearly two orders of magnitude. The atomistic

origin of this behavior is still not fully understood today. The observation of stress-induced

voiding, perhaps more than any other single event, has motivated academic researchers to join

9/13/2002 6
their industrial colleagues to study atomistic processes of stress generation and relaxation in thin

films. The introduction of copper and low dielectric constant materials has given a new urgency

to study cracking and debonding in small structures. The fundamental study will impact other

major technologies, such as MEMS, integrated photonics, and thermal barrier coatings, where

small structure mechanical behavior is important. Vinci and Baker (2002) have edited a multi-

author review on mechanical properties in small dimensions.

The present method to design reliable small structures, if there is one, rests on continuum

mechanics, manifest in handbook solutions, commercial finite element codes, and related

measurement protocols. Their inadequacy is increasingly evident when the structures become

small, because (among other issues) plastic flows become discrete, and failure phenomena

become nucleation, rather than growth, controlled. In the rapidly changing economics of tools for

the small, the division of labor has yet been settled: what to understand, what to compute, and

what to measure. The tension between the relentless trend of miniaturization and the disquieting

lack of a method to design for reliability will inspire innovations for years to come.

9/13/2002 7
2 CRACKING

Figure 2 illustrates cracks in a brittle thin film bonded on a silicon substrate. The film is

under a state of biaxial tensile stress. Starting from a flaw in the film, a crack runs across the film

thickness and often arrests at the interface, leaving the interface and the substrate intact. The

crack, however, elongates laterally in the film, uninhibited until it meets a film edge or another

crack. The lateral crack length can be many times the film thickness. Such a crack is known as a

channel crack. Because the film may have many pre-existing flaws, under a large stress, many

channel cracks can form. A cracked film looks like a field of drying mud.

A brittle material fractures by breaking atomic bonds along the front of a crack. The

atomistic bond-breaking process remains unchanged when this brittle material is integrated in a

structure, surrounded by other materials, so long as the feature size is large compared to

individual atoms. It is, however, the close proximity of the dissimilar materials that gives rise to

the diverse fracture behaviors observed in the interconnect structure. The integration of

dissimilar materials introduces stresses. During fracture, even when the crack is confined in the

brittle material, the surrounding materials deform, sometimes inelastcially.

The pragmatic approach to fracture in the interconnect structure pays little attention to the

atomistic bond-breaking process itself, but instead focuses on deformation in the surrounding

materials. The division of labor is brilliantly made within the framework of linear elastic

fracture mechanics:

(a) The fracture toughness is a property specific to the given brittle material, unaffected

by the surrounding materials, and is measured experimentally.

(b) The crack driving force depends on the surrounding materials, and is related to their

deformation by calculation.

9/13/2002 8
 We begin with an outline of the linear elastic fracture mechanics, and its use in the

interconnect structure. Energy consideration is taken up next. We then describe a channel crack

in a thin film bonded to an elastic substrate. This simple structure leads to a method to measure

the fracture toughness. To illustrate the effects of deformation, we examine a channel crack in a

brittle film on an inelastic underlayer. The crack velocity is set by creep in the underlayer, as

well as by subcritical cracking in the brittle film.

2.1 Linear Elastic Fracture Mechanics

Lawn (1993) is an established textbook on brittle fracture. Here we outline the

fundamentals of immediate importance to the interconnect structure. Figure 3 illustrates a crack

in a body subject to a distribution of load that opens the crack. Model the body by a linear

elastic solid, the crack faces by two matching surfaces, and the crack front by a curve. Within the

theory of linear elasticity, a boundary value problem determines the stress field in the body. Of

course, the stress field depends on the boundary conditions: the size and the shape of the crack

and the body, as well as the magnitude and the distribution of the load. By modeling the crack

front as a mathematical curve, the linear elasticity theory does not account for the atomistic

bond-breaking process.

One feature of the stress field, common to all such elasticity boundary value problems,

plays a special role in the fracture mechanics. Let (r,θ ) be the polar coordinates, centered at a

particular point along the crack front, in the plane normal to the crack front. The (x1 ,x 2 )

coordinates are in the same plane. The linear elasticity theory shows that the stress field in the

body takes the form

K
σ ij (r, θ ) = Σij (θ ) . (1)
2π r

9/13/2002 9
The dependence on r is square-root singular. The stress field scales with K, known as the stress

intensity factor. The factor 2π is introduced by convention. The θ -dependent functions,

Σ ij (θ ) , are listed in Lawn (1993); for example,

Σ 22 (θ ) = cos(θ / 2)[1 + sin(θ / 2)sin(3θ / 2 )]. (2)

The opening stress a distance r directly ahead of the crack front is

K
σ 22 (r,0 ) = . (3)
2π r

For a crack in three dimensions, the stress intensity factor K may vary along the front of the

crack.

The r- and θ -dependence of the singular stress field are independent of the boundary

conditions. The boundary conditions, however, do determine the magnitude of the stress

intensity factor K. For example, a crack of length 2a, in an infinite elastic sheet, subject to a

tensile stress σ remote from the crack, has the stress intensity factor K = σ π a . Solutions of

the stress intensity factors for many crack configurations are collected in handbooks (e.g., Tada,

et al., 1985). Commercial finite element codes are routinely used to calculate K for complex

crack configurations. Dimensional considerations dictate that the stress intensity factor should

take the form

K = κσ L , (4)

where L is a representative length, σ a representative stress, and κ a dimensionless number.

The number κ is determined by solving the boundary value problem. It depends on the ratio of

the crack length to L, as well as other ratios that characterize the integrated structure. Examples

will be given later.

As mentioned above, the theory of elasticity predicts that the stress field is square-root

singular. Of course we do not believe that the stress is infinite at the crack front. The physical

9/13/2002 10
significance of the singular stress field is understood as follows (Fig. 4). Imagine the plane

normal to the crack front. The stress field (1) is valid in an annulus, with the inner radius larger

than the zone in which the bond-breaking process invalidates the linear elasticity theory, and the

outer radius smaller than the size of the body and the crack. We call this annulus the K-annulus.

Neither the bond-breaking processes inside the inner circle, nor the boundary conditions outside

the outer circle affect the r- and θ - dependence of the stress field in the K-annulus: the stress

field (1) is valid in this annulus. For a brittle solid, the bond-breaking processes invalidate the

linear elasticity theory within a zone about 1 nm. A representative feature size in interconnect

structures is 100 nm. These lengths roughly set the size of the K-annulus.

So far the model does not account for the atomistic bond-breaking process. Imagine two

samples of the same solid, each containing a crack under an opening load. The two samples may

have different sizes and shapes. The loads on the two samples may also be different. Regardless

of such differences, because the stress field in the K-annulus depends on the boundary conditions

through the stress intensity factor, the bond-breaking processes in the two samples must be

identical if K is the same for the two samples. That is, K is the only messenger between the

external boundary conditions and the bond-breaking process.

The central idea of the linear elastic fracture mechanics has been so summarized: the same

K, the same crack-tip process. Within the fracture mechanics, the simplest way to account for

the bond-breaking process is to stipulate that the crack extends when the stress intensity factor K

in a body containing a crack reaches a critical value, Kc , known as the fracture toughness.

Once the stress intensity factor is calculated for a given crack configuration, in the from of

(4), the condition for the crack to grow is

κσ L = Kc . (5)

9/13/2002 11
For a given material, the fracture toughness is determined experimentally. Start with a specimen

of the material, introduce a crack into the sample, load the sample until the crack extends, and

record the critical load. Independently solve the elasticity boundary value problem, and obtain

the stress intensity factor in the form (4). The experimentally measured critical load then gives

the fracture toughness. The fracture toughness is a material property, which is independent of

the sample geometry and load. The stress intensity factor is a loading parameter, which does

depend on the sample geometry and the load.

2.2 On Applying Fracture Mechanics to the Interconnect Structure

Liu et al. (2000) have discussed the use of fracture mechanics to formulate design rules for

the interconnect structure. For example, assume that the representative stress is set by the thermal

mismatch between two materials. The length L can be set to be a feature size, such as film

thickness or via diameter. To compute κ , one needs to place a crack in the structure, and know

its location, orientation, and size. Such information is never available for a complex structure. A

pragmatic approach is to place cracks in the structure hypothetically. For the interconnect

structure, the fabrication process controls the structure to the feature size, so that the crack size

should not exceed the feature size. Ideally one wishes to find the worst crack that gives the

largest stress intensity factor. In practice, one does calculations for a few representative cracks.

Because such an approach is time consuming, broad conclusions drawn from idealized structures

are particularly valuable.

To make the above discussion concrete, consider the following example (Lu et al., 1991).

Figure 5 sketches the configuration. A cylinder of radius R is embedded in an infinite matrix.

9/13/2002 12
For simplicity, we assume that the cylinder and the matrix are both elastic, having similar

Young’s modulus E and Poisson’s ratio ν . Set the representative stress by

σ = E ∆α ∆T / (1 − ν ), (6)

where ∆ α is the difference in the coefficients of thermal expansion of the two materials, and ∆T

the temperature change. This stress, as shown by the elasticity solution, corresponds to the

magnitude of the longitudinal stress or twice that of the transverse stress in the cylinder. When

the cylinder is under compression, the matrix is under a tensile hoop stress, which decays far

away from the interface. We place a crack in the matrix, and ask if this crack will grow. It is

sensible to place the crack near the interface and in the radial direction, where the stress is tensile

and has a large magnitude. The choice of the crack length requires some thought. When the

crack length approaches zero, the stress intensity factor vanishes. When the crack length

approaches infinity, the stress intensity factor also vanishes, because of the decaying stress field.

Figure 5 sketches the trend of the stress intensity factor as a function of the crack length. This

elasticity boundary value problem has been solved analytically, giving the stress intensity factor:

π   a 
1+ 
1/ 2 −3 / 2

 8  R  R
K a
= . (7)
σ R

If we can estimate the largest flaw size possible due to a certain fabrication process, we can set

the crack size to this flaw size, and then calculate the stress intensity factor. If we know nothing

about the flaw, we may proceed as follows. The stress intensity factor (7) attains a maximum,

Kmax = 0.24σ R , at a / R = 0.5 . No matrix flaw can grow provided this maximum stress

intensity factor is below the fracture toughness, namely,

E∆ α∆T
0.24 R < Kc . (8)
1 −ν

9/13/2002 13
This condition invokes well defined quantities, and does not rely on any knowledge of the flaw.

Everything else being equal, a cylinder with a smaller radius is less likely to cause matrix

cracking. Assuming σ = 1GPa and Kc = 1MPa m , condition (8) gives the critical radius

Rc = 17 µm .

2.3 Energy Release Rate

Additional insight is gained by analyzing elastic energy change associated with crack

extension. Once again consider the crack, area A, in an elastic body, subject to a distribution of

load that opens the crack. Let F represent the magnitude of the load, and ∆ the displacement,

such that when ∆ changes by d∆ , the load does work Fd∆ to the body. Let U be the elastic

energy stored in the body. When the displacement varies, but the crack area remains fixed, the

work done by the load equals the change in the elastic energy of the body, namely, dU = Fd∆

when dA = 0 . The elastic energy U is represented by the area under the load-displacement curve

(Figure 6a).

Imagine two copies of the body, identical in all respect except for a small difference in the

crack areas. The copy with the larger crack is more compliant and, at the same displacement,

stores lower elastic energy (Figure 6b). Regard the elastic energy stored in the body as a function

of the displacement and the crack area, U(∆ , A). Define energy release rate, G, as the reduction

of the elastic energy associated with the crack increasing per unit area, when the external load

does no work, namely,

∂U (∆, A)
G=− . (9)
∂A

9/13/2002 14
The partial derivative signifies that the displacement ∆ is held fixed when the crack area A

varies. For a given crack configuration, one can solve the elasticity boundary value problem,

compute the elastic energy stored in the body, U(∆ , A), and then obtain the energy release rate G.

The above definition is still sloppy. In general, the energy release rate varies along the

crack front. Denote the curved coordinate along the crack front by s, and the energy release

rate is a function G(s ). Imagine a body subject to a fixed displacement ∆ . The elastic energy

stored in the body is a function of the crack shape, as well as the crack area. Let the crack

advance by a small length δa, which can vary along the crack front. Associated with the change

in the crack, the elastic energy stored in the body varies by δU . Define the energy release rate

such that

δU = − ∫ G(s)δads . (10)

The integral extends along the entire crack front. In what follows, we will mostly overlook this

detail, and assume that the energy release rate is uniform along the crack front.

When both the displacement and the crack area vary, the elastic energy varies according to

dU = Fd∆ − GdA . (11)

The elastic energy is a function of the displacement and the crack area, U(∆ , A). The load F and

the energy release rate G are the differential coefficients. Just as F is the force conjugate to the

displacement ∆ , the energy release rate G is the force conjugate to the crack area A. Such a

force is known as a configurational force.

Define the potential energy as Π = U − F∆ . Write Equation (11) in a different form:

dΠ = − ∆dF − GdA . (12)

9/13/2002 15
Now the potential energy is a function of the load and the crack area, Π(F , A). The

displacement ∆ and the energy release rate G are the differential coefficients. Equation (12)

suggests an equivalent definition of the energy release rate:

∂Π (F, A)
G=− . (13)
∂A

The partial derivative signifies that the load F is held fixed when the crack area A varies. When

the body is linearly elastic, U = F∆ / 2 and Π = −U , so that (13) becomes

∂U (F, A)
G=+ . (14)
∂A

The partial derivative signifies that the load F is held fixed when the crack area A varies. The

sign difference in Equations (9) and (14) reflects a simple physical fact. When the crack area is

larger, the body is more compliant, so that the body stores less elastic energy at a fixed

displacement, but stores more elastic energy at a fixed load.

It has been shown that the energy release rate relates to the stress intensity factor as

K2
G= , (15)
E

where E = E under the plane stress condition, and E = E / (1 − ν 2 ) under the plane strain

conditions. Here E is Young’s modulus, and ν Poisson’s ratio. Fracture mechanics has been

formulated either in terms of the stress intensity factor, or the energy release rate. The two

formulations are equivalent. Both contribute to the understanding of the fracture mechanics.

The simplest fracture condition is that the crack advances when the energy release rate attains a

critical value, Γ , known as the fracture energy. Obviously, the fracture energy relates to the

fracture toughness as Γ = Kc2 / E .

9/13/2002 16
2.4 Channel Cracks in a Thin Film Bonded to a Substrate

In an interconnect structure, often a crack runs in one material, but the neighboring material

remains intact and exerts a constraint on the crack. Figure 7 illustrates two structures for

comparison. In Fig. 7a, a crack of length a, in a freestanding sheet of thickness h, is subject to a

tensile stress σ remote from the crack. Assume that the displacement at the load point is fixed,

so that the load does no work when the crack extends. When the crack is introduced, the stress

near the crack faces is partially relieved. The volume in which the stress relaxes scales as a 2 h ,

so that relative to the uncracked, stressed sheet, the elastic energy in the cracked sheet changes

by ∆U ~ −a2 hσ 2 / E . Consequently, G ~ aσ 2 / E , namely, the energy release rate for a crack in

a freestanding sheet increases with the crack length.

Figure 7b illustrates a thin elastic film bonded to an elastic substrate. When the crack

length a is much larger than the film thickness h, the stress field in the crack wake becomes

invariant as the crack extends. The volume in which the stress relaxes scales as ah 2 , so that the

introduction of the crack changes the elastic energy by ∆ U ~ −ah2 σ 2 / E f . The energy release

rate is

σ 2h
G=Z . (16)
Ef

The dimensionless number Z depends on the elastic constants of the film and the substrate. The

number must be determined by solving the boundary value problem. When the thin film and the

substrate have similar elastic constants, Z = 2.0. Beuth (1992) calculated the Z values for a thin

film on an infinite substrate with dissimilar elastic constants. When the substrate is stiffer than

the film, Z is between 1 and 2. When the substrate is much more compliant than the film, Z can

be very large.

9/13/2002 17
 For the channel crack, the energy release rate scales with the film thickness, and is

independent of the crack length when the crack length exceeds several times the film thickness

(Amrico and Begley, 2002). Take Z = 2, σ = 109 Pa , E f = 1011 Pa , and Γ = 10J/m 2 . Equating

the energy release rate G in (16) to the fracture energy Γ , we find a critical film thickness

hc = 0.5µm . Channel cracks can propagate in films thicker than hc , but not in films thinner than

hc .

Ye et al. (1992) have shown that the channel crack energy release rate increases if the crack

can dip into the substrate, or the film and substrate can debond. Ambrico et al. (2002) have

studied channel cracks multilayers with lateral patterning, with particular reference to

interconnects consisting of copper and low k dielectrics. Using the extended finite element

method (XFEM), Huang, et al. (2003) have studied the effects of substrate thickness and channel

spacing, with particular attention to compliant substrates.

2.5 Measuring Thin Film Toughness

For a brittle solid, toughness is independent of the sample size, so that one may expect to

extrapolate toughness measured using bulk samples to thin films. However, thin films used in

the interconnect structure are processed under very different conditions from bulk materials, or

are unavailable in bulk form at all. Consequently, it is necessary to develop techniques to

measure thin film toughness. Such a technique is attractive if it gives reliable results, and is

compatible with the interconnect fabrication process. Ma et al. (1998) have developed such a

technique, on the basis of channel cracks in a thin film bonded to a silicon substrate. The test

procedure consists of two steps (Fig. 8): (a) generating pre-cracks, and (b) propagating the

cracks using controlled stress. A convenient way of generating pre-cracks is by scratching the

9/13/2002 18
surface using a sharp object. A gentle scratch usually generates multiple cracks on the two sides

of the scratch. Care must be taken to generate cracks just in the film, but not in the substrate.

Use a bending fixture to load the sample, and a digital camera to record the crack growth events.

After a crack propagates some distance away from the scratch, the crack grows at a steady

velocity.

By recording crack growth at slightly different bending loads, one can measure the crack

velocity as a function of the stress. The steady velocity is very sensitive to the applied stress.

Consequently, the critical stress is accurately measured by controlling the velocity within a

certain range that is convenient for the experiment. The critical stress then determines the

fracture energy according to the mechanics result (16). The technique has been used to measure

the fracture energy of silica ( Γ = 16.5J/m 2 ) and silicon-nitride ( Γ = 8.7J/m 2 ) (Ma et al., 1998),

as well as a carbon doped oxide ( Γ = 1.5J/m 2 ) (Jun He, private communication).

After deposition, the film has a residual stress, σ R . The residual stress can be measured by

the wafer curvature method (Nix, 1989). When the structure is bent by a moment M (per unit

thickness), the film stress becomes

6E f M
σ = σR + , (17)
Es Hs2

where Hs is the thickness of the substrate, and E f and Es are the plane strain modulus of the

film and the substrate. The bending moment also causes a tensile stress in the substrate,

σ s = 6M / Hs2 . To measure the toughness of the film, one has to propagate the channel crack in

the film without fracturing the substrate. This, in turn, requires that the substrate should have

very small flaw size, and that the scratch should not produce flaws in the substrate. When

properly cut, the silicon substrate can sustain tensile stresses well over 1 GPa. Inspecting (16)

9/13/2002 19
and (17), one concludes that the technique is inapplicable when the film has a very large fracture

energy, large residual compressive stress, small thickness, or low modulus. In the experiment of

Ma et al. (1998), a thin metal layer is deposited on the silicon substrate, and the brittle film is

deposited on the metal. The metal layer serves as a barrier preventing the crack from entering

the substrate. The metal layer, upon yielding, also increases the energy release rate for a given

bending moment (Section 2.6).

One may also greatly amplify the energy release rate for a given stress by using the

structure in Fig.9. The cavity underneath the brittle film may be made by MEMS fabrication

method. When the crack grows, the stress in the film is relaxed, and the energy release rate,

σ2L
G= , (18)
Ef

now scales with the length of the cavity, rather than the film thickness.

2.6 The Effect of Underlayer on Channel Crack

When a brittle film is bonded to a plastically deformable substrate, the constraint of the

substrate on the crack in the film is partially lost. In the limiting case, when the yield strength of

the substrate is much below the stress in the film, the film is equivalent to a freestanding sheet.

Consequently, the plastic deformation in the substrate encourages the channel cracks in the film.

This effect has been studied by Hu and Evans (1989) by using a shear lag model, by Beuth and

Klingbeil (1996) for a brittle film on a semi-infinite elastic-plastic substrate, and by Ambrico and

Begley (2002) for a brittle film on a ductile underlayer. In this section, using a shear lag model,

we (Suo and He, 2002) examine the structure depicted in Fig. 10, consisting of an overlayer, an

underlayer, and a substrate. The crack cuts across the overlayer, with the crack root arresting at

the interface between the overlayer and the underlayer. The overlayer is taken to be elastic, and

9/13/2002 20
the substrate is rigid. The property of the underlayer will be varied to study its effect on the

channel crack.

The overlayer may consist of multiple brittle films. The stress in the overlayer is a function

of the position through the thickness, σ (z ). Define the membrane force in the overlayer by

N = ∫ σ (z )dz . (19)

The integral extends over the thickness of the overlayer. Similarly define the membrane stiffness

by

B = ∫ E (z )dz , (20)

where E (z ) is the plane strain modulus of the material at z.

We now consider the wake of the crack in the overlayer on a ductile underlayer. When the

overlayer is uncracked, subject to a membrane force N∞ , uniform in the plane of the layer, the

displacement field in the overlayer is set to be zero. Relative to this reference state, the overlayer

deforms when a crack is introduced. Let u(x ) be the displacement field averaged across the

thickness of the overlayer. On the interface, the overlayer and the underlayer interact with each

other through the shear stress field τ (x ). We model the underlayer as a time-independent elastic

plastic material, with a nonlinear relation between the shear stress τ and the displacement u.

When a crack cut across the thickness of the overlayer at x = 0, the membrane force in the

overlayer, N (x ) , varies with the position, having the boundary conditions N (0) = 0 and

N (∞ ) = N∞ . Hooke’s law relates the membrane force to the displacement gradient:

du
N (x ) = N ∞ + B . (21)
dx

The force balance of an element of the overlayer requires that

9/13/2002 21
 dN
=τ. (22)
dx

Inserting (21) into (22), we obtain the ordinary differential equation for u(x ):

d 2u
B = τ (u). (23)
dx 2

Multiplying (23) by du / dx , and integrating from x = 0 to x = ∞ , we obtain that

= ∫ τ (u)du .
u
N∞2 0

(24)
2B 0

In reaching the above equation, we have used the following boundary conditions: u = u0 and

du / dx = − N∞ / B at x = 0, and u = 0 and du / dx = 0 at x = ∞ . When the underlayer property

τ (u) is known, (24) provides the relation between the crack opening displacement u0 and the

membrane force N∞ in the uncracked overlayer.

For a multi-material overlayer, the energy release rate depends on the position, G(z ). The

integral

f = ∫ Gdz (25)

gives the elastic energy change associated with the crack advancing per unit length in the

overlayer. The quantity f has the unit of force, and is the driving force for the channel crack.

The channel driving force is given by the complementary energy integral (Beuth and

Klingbeil, 1996):

f = 2 ∫ u0 dN∞ .
N∞

(26)
0

Once the u0 − N∞ relation is obtained from (24), the driving force for the channel crack is

calculated from (26)

9/13/2002 22
 We next list solutions for idealized material law, τ (u) . We assume that the underlayer has

the uniform property, and the shear stress is constant across the thickness of the underlayer.

Consequently, the shear strain γ is also constant across the thickness of the underlayer. The

shear strain relates to the overlayer displacement as γ ( x ) = u(x ) / H , where H is the thickness of

the underlayer.

Elastic underlayer. The underlayer is linear elastic with shear modulus µ , so that

τ = µu / H . Integrating (24), we obtain the u0 − N∞ relation:

H
u0 = N∞ . (27)
µB

Integrating (26) gives the driving force for the crack in the overlayer:

H
f = N∞2 . (28)
µB

This result reproduces that obtained by Xia and Hutchinson (2000). As expected, the crack

driving force is quadratic in the membrane force. At a given membrane force, the crack driving

force is large when the underlayer is thick and compliant. This result is applicable to, for

example, a nitride passivation film on a complaint dielectric layer. As the microelectronic

industry implementing low dielectric constant materials, with accompanying low stiffness, the

likelihood of cracking in the brittle passivation film will increase.

Rigid, perfect plastic underlayer. Neglect elasticity of the film, and assume that the

underlayer deforms plastically at a constant stress τ = τ Y . Integrating (24), we obtain the

u0 − N∞ relation:

N ∞2
u0 = . (29)
2Bτ Y

Inserting into (26) gives driving force for the crack in the overlayer:

9/13/2002 23
 N∞3
f = . (30)
3B τ Y

This reproduces the result of Hu and Evans (1989). The crack driving force is cubic in the

membrane force, inversely proportional to the underlayer yield strength, and independent of the

underlayer thickness.

Elastic, perfectly plastic underlayer. The film now has the shear modulus µ, the yield

strength τ Y , and the yield strain γ Y = τ Y / µ . When u0 < γ Y H , the film is elastic, and (27)

provides the u0 − N∞ relation. When u0 > γ Y H , a segment of the film near the crack deforms

plastically, and the film fare from the crack is still elastic. Using (27), we write the yield

condition in terms of the membrane force:

N∞
= 1. (31)
τ Yγ Y HB

Integrating (24), we obtain the opening displacement of the crack wake

1 1 N∞ 
= + 
2
u0
γ Y H 2 2  τ Yγ Y HB 
. (32)

This is the u0 − N∞ relation when N∞ / τ Y γ Y HB > 1 ; for smaller N∞ , the film is elastic, and the

u0 − N∞ relation is given by (27). Inserting the whole u0 − N∞ relation into (26), we find the

driving force for the crack in the overlayer:

  N∞   N∞  
1 −  τ γ HB  + 3 τ γ HB   .
−3 −2
N3
  Y Y   Y Y  
f = ∞ (33)
3B τ Y

When N∞ / τ Y γ Y HB = 1, the underlayer is elastic, and (33) recovers (28). When

N∞ / τ Y γ Y HB »1 , the two terms in the bracket is negligible, and (33) recovers the result for the

rigid perfect plastic case (30).

9/13/2002 24
 Power-law plastic underlayer. Now we assume that the film deforms according to the

power law:

τ  u 
n

τ Y  Hγ Y 
= , (34)

where τ Y is an arbitrarily selected reference stress, and γ Y and n are parameters used to fit

experiment data. Following the same procedure as before, we obtain the u0 − N∞ relation

 (n + 1) N∞2  n +1
1

γ Y H  2τ Y γ Y HB 
u0
= , (35)

and the crack driving force

2γ Y H  (n + 1)n + 2 N∞n + 3  n +1
1

n + 3  2BHτ Yγ Y 
f = . (36)

This equation recovers the linear elastic case Eq. (28) when n = 1, and the rigid perfect plastic

case Eq. (30) when n = 0.

Figure 11 plots the normalized channel driving force, f / (Hγ Y τ Yγ Y HB ), as a function of

the normalized membrane force, N∞ / τ Y γ Y HB . As expected, for the same membrane force, the

more deformable the underlayer, the larger the channel driving force.

2.7 Crack Kinetics: Subcritical Cracking and Substrate Creep

As mentioned above, the crack extends in the thin film at some finite velocity. The

magnitude of the velocity is sensitive to the stress, but in general is much below the speed of the

elastic waves in the solids. Consequently, the observed crack velocity should not be an inertia

effect. If the brittle solid does not creep, the time dependence originates from either the bond-

breaking process, or the time-dependent deformation in the surrounding materials.

9/13/2002 25
 Slow crack growth has been extensively studied in bulk brittle materials, where the bond-

breaking process is assisted by environmental molecules, such as water molecules (Lawn, 1993).

The time dependence results from the transport of the environmental molecules to the crack

front, and the reaction at the crack front. The crack velocity can be measured experimentally as

a function of the energy release rate. Ma et al. (1998) and Cook and Liniger (1999) have

provided data for several thin films. The measured V-G curve is sometimes fit to a variety of

empirical relations. One such a relation, suggested by the consideration of microscopic

mechanisms, is

  G 
V = V0 sinh β  − 1  ,
  Gth  
(37)

where V0 , Gth and β are parameters to fit experimental data.

I owe Dr. Jun He, of Intel Corporation, for the following idea. The V-G curve is specific to

the brittle material and the environment, but is independent of the surrounding materials, if they

do not prevent the environmental molecules from reaching the crack front. The V-G curve for a

given material can be measured using a simple structure, such as a film on a silicon substrate.

When the brittle material is in a more complex structure, the same V-G still applies. Because the

crack velocity is readily measurable, the measured crack velocity, together with the known V-G

curve, determines the energy release rate. The method is accurate because the crack velocity

sensitively depends on the energy release rate. As we have seen in Figure 11, the energy release

rate is affected by the deformation behavior of the surrounding materials. Direct measurements

of G provide a means to study the deformation behavior of the surrounding materials.

The crack velocity is also affected if the surrounding materials creep. As an example,

consider the structure illustrated in Fig. 10 again, now assume that the underlayer is viscous

(Liang et al. 2002a). Initially, the film is in a uniform biaxial tensile stress state σ . When the

9/13/2002 26
underlayer creeps, the stress field in the film relaxes in the crack wake, but intensifies around the

crack tip. When the crack tip moves slowly, the crack wake has a long time to relax, and the

stress intensity around the crack tip increases. When the crack tip moves rapidly, the crack wake

has a short time to relax, and the stress intensity around the crack tip decreases. Consequently,

the crack can attain a steady velocity, given by

Hhσ 2
V = 0.6 . (38)
ηG

As expected, the larger the steady velocity, the smaller the energy release rate. If the film is

susceptible to subcritical crack growth, one can separately determine the V-G curve of the

material, e.g., using the film on an elastic substrate. Once the crack velocity V is measured for

the film on the viscous underlayer, one can read the energy release rate G form the V-G curve.

The measured V and G, together with Equation (38), provide a means to determine the viscosity

of the underlayer.

3 DEBONDING AND MIXED MODE CRACK

The interconnect structure consists of many dissimilar materials. Debonding has been a

primary concern. Figure 12 is a micrograph of a thin film debonding from a substrate. Within

the linear elastic fracture mechanics, we model the debond process as a crack extending on the

interface between two elastic materials. Impurities, of minute amount by volume fraction, may

accumulate on the interface to a large coverage, significantly changing the interfacial fracture

energy. Adhesion sensitively depends on processing conditions, and on materials deliberately

added to the interface. Measurements of the interfacial fracture energy provides tools for process

control and material selection to promote adhesion in interconnect structures.

9/13/2002 27
 It is instructive to compare a crack running on a bimaterial interface to a crack running in

a homogenous, isotropic material. A crack in a homogenous, isotropic material extends along

such a path that the crack tip is always under the tensile condition. Consequently, the crack tip

often follows a curved path, and the fracture property of a given material is prescribed by a

single quantity, the mode I fracture toughness. By contrast, the crack confined on a bimaterial

interface often propagates under the combined tensile and shear conditions. The fracture energy

depends on the relative proportion of the tensile and shear loads.

This section will begin with a discussion of a stationary, planar crack in a homogenous,

isotropic material, subject to the mixed mode conditions. The crack grows to seek a mode I path.

We then outline the interfacial fracture mechanics, the methods to measure the fracture energy as

a function of the mode angle, and the applications to interconnect structures. We finally examine

the effect of plastic deformation on debonding.

3.1 Mixed Mode Crack in a Homogeneous Material

First examine a stationary, planar crack in a homogenous, isotropic material. The body is

subject to a distribution of load that can both open and shear the crack. Let (r,θ ) be the polar

coordinates centered at a given point on the crack front (Fig. 13). We will consider the situation

that the shear load is in the plane (r,θ ). The stress field near the crack front has the form

KI K
σ ij (r, θ ) = Σij (θ ) + II Σ ij (θ ).
I II
(39)
2π r 2πr

The θ -dependent functions, Σ ijI (θ ) and Σ ijII (θ ), can be found in Lawn (1993). Mode I field is

symmetric with the crack plane, with Σ 22I ( 0) = 1 and Σ 12I (0) = 0. Mode II field is antisymmetric

9/13/2002 28
with the crack plane, with Σ 22II ( 0) = 0 and Σ 12II (0) = 1. In particular, the stresses on the plane

θ = 0 a distance r ahead the crack front are

KI K II
σ 22 (r,0 ) = , σ 12 (r,0 ) = . (40)
2π r 2πr

The stress intensity factors, KI and KII , scale the amplitudes of opening and shearing loads,

respectively. We will be interested in situations that the crack front is under tension, rather than

compression, namely, KI ≥ 0 .

Once a crack configuration is prescribed, the stress intensity factors KI and KII are

determined by solving the elasticity boundary value problem. For example, for a crack of length

2a in an infinite sheet, subject to remote tensile stress σ and shear stress τ , the two stress

intensity factors are KI = σ πa and KII = τ π a . Solutions for many crack configurations have

been collected in handbooks (e.g., Tada et al., 1985). Finite element methods have been used

routinely to determine the stress intensity factors under the mixed mode conditions (Li et al.,

1985).

Again the stress field (39) is valid in an annulus, with the inner radius larger than the

zone of the bond-breaking process, and the outer radius smaller than a length representative of

the external boundary conditions. The mixed mode conditions are characterized by two loading

parameters, KI and KII . The two “messengers” transmit the boundary conditions to the crack tip

process. The relative amount of mode II to mode I is specified by the mode angle ψ , defined by

tan ψ = KII / K I . (41)

With KI ≥ 0 , the mode angle ranges between − π / 2 ≤ ψ ≤ π / 2 . A pure mode I crack

corresponds to ψ = 0 . A pure mode II crack corresponds to either ψ = + π / 2 or ψ = − π / 2 .

9/13/2002 29
 The energy release rate G is still defined as the elastic energy reduction as the crack

extends unit area, while the load does no additional work. The energy release rate relates to the

two stress intensity factors by

G=
E
(
1 2
K I + K II ).
2
(42)

Instead of using KI and KII to represent the external loads, we use G to represent the amplitude

of the loads, and ψ to represent the mode of the loads.

Under the mixed mode conditions, upon growing, the planar crack often kinks at an angle

from its original plane. The kink angle depends on the relative amount of mode II to mode I

load. Erdogan and Sih (1963) showed that the experimentally measured kink angles in a

plexiglass are well predicted by the criterion that the crack kinks to the plane with the maximum

hoop stress. For the stationary crack, the hoop stress near the crack tip is

θ θ θ
cos3   − II 3cos 2   sin   .
KI K
σ θθ (r,θ ) = (43)
2πr 2 2πr 2 2

The hoop stress maximizes at an angle θ * , given by

θ *
tan  = −
2 tanψ
. (44)
2 1+ 1+ 8tan2 ψ

When KII > 0 , ψ > 0 and the crack kinks down (i.e., θ * < 0 ). The converse is true when

KII < 0 . When the plane crack is pure mode I, this criterion predicts that the crack extends

straight ahead. When the plane crack is pure mode II, this criterion predicts that the crack kinks

at an angle θ * = 70.5°.

The moving crack tip seeks a trajectory that is locally mode I; any nonzero KII will cause

the crack to deflect. To illustrate this point, consider two examples. It is well known that a crack

in a double-cantilever beam is unstable: the crack tends to curve one way or the other (Fig. 14).

9/13/2002 30
By symmetry, the crack on the mid-plane of the sample is pure mode I. This mode I path,

however, is unstable. A crack, lying slightly off the mid-plane, has a mode II component that

tends to deflect the crack further away from the mid-plane.

As a second example, Figure 15 illustrates a thin film spalling from an adherent substrate

(Thoulsss et al., 1987). The film is well bonded to the substrate, and is under a residual tensile

stress. It is sometimes observed that a crack starts from the edge of the film, dives into the

substrate, and then grows parallel to the interface. In the crack wake, the residual stress in the

film is partially relieved, and the film and a thin layer of the substrate material underneath form a

composite plate, bending up. This observation is peculiar in that, were the crack to run on the

film-substrate interface, the residual stress in the film would be fully relieved in the crack wake.

Such a global energy consideration has no physical basis: it is the local process of bond breaking

that selects the crack path. The experimental observation has been interpreted that the crack tip

moves along a mode I trajectory. To see how this works, consider a special case that the thin

film and the substrate have similar elastic modulus. Let h be the film thickness, and d be the

depth of the crack parallel to the interface. When the crack is long compared to d, the effect of

the residual stress on the crack is well described by the equivalent axial force and bending

moment:

1
P = σh, M = σh(d − h). (45)
2

The stress intensity factors have been calculated, given by

KI =
1
2
(Pd
−1/ 2
cosω + 2 3 Md
−3/ 2
sin ω ), (46a)

KII =
1
2
(Pd −1/ 2 sinω − 2 3Md −3 / 2 cos ω ), (46b)

9/13/2002 31
with ω ≈ 52°. Suppose that the crack selects the depth d * by the condition that KII = 0 , and we

obtain that d * = 3.8h . One can further confirm that the mode I path is stable in that, if a parallel

crack at a depth d different from d * , then KII is nonzero and is in the direction that tends to

deflect the crack back toward the depth d * .

Curved crack paths can be simulated using numerical methods. For a given crack

configuration, one solves the elasticity boundary value problem, and computes the stress

intensity factors KI and KII . One then advances the crack by a small length in the direction, say,

set by the criterion of maximum hoop stress (Equation 44). The path so selected should be

essentially a mode I path. For a single crack tip, the precise length for each increment is

unimportant, so long as it is much smaller than the representative size of the sample. To

simulate simultaneous growth of multiple cracks, however, one has to know how much to

advance each crack. An ingredient of time-dependence has to be introduced into the model. For

example, if the solid is susceptible to subcritical cracking (Section 2.7), the V-G relation

provides the needed information. Once the energy release rate is calculated for every crack tip in

a given configuration, one advances each crack according to the kinetic law for a small time step.

Similarly, one can simulate the growth of a crack in three dimensions with a curved front by

advancing each point on the crack front according to the kinetic law and its local energy release

rate. Because the crack extends under the mode I conditions, the V-G curve can be obtained

experimentally using a specimen containing a single straight mode I crack.

The regular finite element method meshes the geometry of the crack and uses a fine mesh

near the crack tip. When the crack grows, remeshing is required. To circumvent these

difficulties, an extended finite element method (XFEM) has been advanced (Moes et al., 1999).

For the nodes around the crack tip, one adds enriching functions derived from the singular crack-

9/13/2002 32
tip stress field. For nodes on the crack faces, one adds the Heaviside function to represent the

displacement jump. Consequently, the mesh can be coarse near the crack tip, and the elements

need not to conform to the crack geometry. As the cracks grow, one updates the nodes to be

enriched. No remeshing is necessary. Figure 16 shows the simulated mud-crack pattern in

tensile film bonded to a substrate (Liang et al., 2002b).

3.2 Interfacial Fracture Mechanics

Now consider a crack on the interface between two elastic materials. Williams (1959)

discovered that the singular stress field at the interface crack tip is not square root singular. The

stresses on the interface a distance r ahead of the crack tip take the form

Kr iε
σ 22 + iσ 12 = . (47)
2πr

The notation of complex numbers is used: i = −1 and r iε = e iε ln r = cos(ε ln r) + i sin(ε ln r ). The

bimaterial constant ε is defined by

1  (3 − 4ν1 )/ µ1 + 1/ µ2 
2π  (3 − 4ν 2 ) / µ2 + 1/ µ1 
ε= ln , (48)

where ν is Poisson’s ratio, µ the shear modulus, and the subscripts 1 and 2 refer to the two

materials. The constant is bounded, ε < (1/ 2π )ln 3 ≈ 0.175.

The stress intensity factor K is complex-valued, and has the dimension

K = [stress][length]
1/ 2 −iε
. (49)

The amplitude of the stress intensity factor has the familiar dimension, K = [stress ][length] .
1 /2

Indeed, K relates to the energy release rate as

1  1 1  K2
2  E1 E2  cosh2 (πε )
G= + . (50)

9/13/2002 33
For a given crack configuration, the stress intensity factor K is determined by solving the

elasticity boundary value problem. Finite element method and other numerical methods have

been developed to determine the stress intensity factor (Matos et al., 1990). Hutchinson and Suo

(1991) have compiled solutions for layered materials.

Following Rice (1988), we define the mode angle ψ by

exp(i ψ ),
−iε
K = Kl (51)

where l is an arbitrary length. To see the consequence of this definition, combine (47) and (51),

yielding

K i[ψ + ε ln (r / l)]
σ 22 + iσ 12 = e . (52)
2πr

The ratio of the shear stress to the tensile stress is

σ 12 / σ 22 = tan[ψ + ε ln (r / L)]. (53)

When ε = 0, as for a crack in a homogenous material, the ratio σ 12 / σ 22 is independent of the

distance r in the K-annulus, and the mode angle ψ characterizes the relative portion of shear to

tension. When ε ≠ 0, the ratio σ 12 / σ 22 varies with the distance r, and equals tan ψ at r = l . The

variation is not rapid, because ε is small and because a logarithm is a slowly varying function.

Thus, tan ψ approximates the ratio σ 12 / σ 22 so long as r is not far from l. For a brittle interface,

a natural choice is l = 1 nm, representative of the bond breaking zone size. With this choice, the

mode angle ψ represents the relative portion of shear to tension at the size scale of bond-

breaking zone.

In a homogeneous material, crack faces come into contact under compression. By

contrast, interface crack faces may come into contact even the remote load has a tensile

component. When the contact zone is large, one has to take into account of the forces on the

9/13/2002 34
crack faces in solving the boundary value problem. In many situations, however, the contact

zone is small compared to the overall dimension. Consequently, the K-annulus exists, with the

inner radius enclosing the contact zone, as well as the bond-breaking process zone. Rice (1988)

has examined the condition for small-scale contact.

Williams’s elastic solution shows that the displacement jump at a distance r behind the

crack tip is

1 1
δ2 + iδ1 =  + 
Kr iε 2r
 E1 E2  2(1+ 2iε )cosh(πε ) π
. (54)

The jump in the displacement component normal to the crack faces is

δ2 = δ cos[ψ + ε ln (l / r) − tan −1 (2ε )], (55)

where δ = (δ12 + δ 22 )
1 /2
is the magnitude of the displacement jump. If the crack is required to be

open within l < r < 100l , the mode angle must be restricted within

− π / 2 + 2ε < ψ < π / 2 + 2.6ε , for ε > 0

− π / 2 − 2.6 ε < ψ < π / 2 + 2ε , for ε < 0 (57)

The number 100 is arbitrary, but the condition in (57) is insensitive to this number. When ε = 0,

the above condition simply says that the contact does not occur when the crack is under tension,

which is known for homogeneous materials.

As an example, consider a crack, length 2a, on the interface between two semi-infinite

materials, subject to remote tensile stress σ and shear stress τ . Write σ + iτ = Te iω , so that T is

the magnitude, and ω the direction, of the remote load. This elasticity boundary value problem

has been solved analytically, giving the stress intensity factor (Rice, 1988)
− iε
K = (1 + 2iε ) π a (2a ) Teiω . (58)

9/13/2002 35
The energy release rate is obtained by inserting (58) into (50). A comparison of (51) and (58)

gives the mode angle:

ψ = ω + tan −1 (2ε ) + ε ln (l / 2 a). (59)

When ε = 0, the mode angle ψ equals the angle ω of the remote load. The two angles are

different when ε ≠ 0. Consider representative values, ε = −0.05, l = 1 nm, and 2a = 1 mm.

Under remote tension, ω = 0, one finds that ψ = 34° , indicating a significant shear component

near the crack tip, at the size scale of bond breaking zone.

As a second example, consider the debond of a thin film from a substrate. A film,

thickness h, is under a tensile inplane stress σ and initially bonded to the substrate. The film

may debond from the root of a channel crack, or from the edge of the film. When the debond

length exceeds several times the film thickness, the debonding process attains a steady-state, in

which the energy release rate is independent of the debond length. Under the plane strain

conditions, an elementary consideration gives the energy release rate

σ2h
G= . (60)
2E f

The debond crack is under the mixed mode conditions. The mode angle must be determined by

solving the elasticity boundary value problem. When the film and the substrate have similar

elastic constants, the stress intensity factors are given by (46), setting P = σh and M = 0. The

mode angle is ψ ≈ 52° . Next examine the effect of the modulus mismatch on the mode angle.

The complex stress intensity factor is given in the form

K = K h−iε e iω . (61)

The magnitude K relates to the energy release rate (60) through the general relation (50). The

angle ω depends on elastic mismatch between the film and the substrate, ranging between 40° to

9/13/2002 36
60° if the mismatch is not excessive, as tabulated in Suo and Hutchinson (1990). A comparison

of (61) and (51) gives the mode angle

ψ = ω + ε ln(l / h). (62)

For representative values, ε = −0.05, l = 1 nm, and h = 1 µm, the additional angle is

ε ln(l / h) = 20° .

3.3 Measuring Interfacial Fracture Energy

When a debond crack extends on a bimaterial interface, mixed mode conditions usually

prevail. The fracture energy is a function of the mode angle, Γ (ψ ) . The crack extends when the

energy release rate reaches the fracture energy: G = Γ(ψ ). Due to elastic mismatch in the two

materials, one should specify the length l in defining the mode angle ψ . The curve Γ (ψ ) has

been measured for a few interfaces (e.g., Cao and Evans, 1989; Wang and Suo, 1990; Liechti and

Chai, 1991). Figure 17 sketches the trend of the Γ (ψ ) curve. The fracture energy tends to

increase as mode angle approaches ± π / 2 . The energy cost for debond varies with the relative

proportion of opening and shearing modes, because, e.g., a large amount shear may promote

more inelastic deformation in the constituent materials, or promote near-tip sliding against

roughness if the interface is not perfectly flat. The curve Γ (ψ ) need not be symmetric with

respect to ψ = 0 , because the two materials on either side of the interface are dissimilar,

breaking the symmetry between ψ < 0 and ψ > 0 .

Many test configurations for measuring interfacial fracture energy exist. See reviews by

Hutchinson and Suo (1991), Evans and Hutchinson (1995), and Volinsky et al. (2002). For

example, the process of a tensile film debonding from a substrate is a means to determine the

interfacial fracture energy. To ensure that the crack is in the steady-state, one should introduce an

9/13/2002 37
initial debond crack of length several times the film thickness (Yu et al., 2001). The energy

release rate (60) depends on both the film stress and the film thickness. One may change the

film stress by varying the temperature, or by applying a mechanical load such as pulling the

substrate. One may change the film thickness during deposition. The combination of the stress

and the thickness that cause the debond crack to grow gives the fracture energy according to

(60). It is often more convenient to deposit a layer of different material to increase the energy

release rate. For example, Cr has been used as an overlayer for its high residual stress (Bagchi et

al., 1994). Thick epoxy overlayers have also been used (Shaffer et al., 1996). When an

overlayer is used, one should obtain a different energy release rate expression, and introduce an

initial debond crack of length several times of the combined thickness of the film and the

overlayer.

A class of test methods has been particularly versatile for thin film structures (Figure 18).

Suppose one needs to measure the interfacial fracture energy of a thin film of material 1, and a

substrate of material 2. One can sandwich the thin film between two substrates of material 1.

One may use a layer of adhesive, such as epoxy, to glue the bare substrate to the one covered

with the film. The substrates are much thicker then the film and the adhesive, so that the whole

specimen is easy to load. The interfacial fracture energy can be measured if the crack runs on

the desired interface. There are several advantages of methods of this kind. First, the methods

are applicable for thin films of any thickness, and can even measure the interfacial energy

between two thin films, when they are both sandwiched. Second, because the film is still bonded

to one substrate in the crack wake, the residual stress in the film is unrelieved, and therefore does

not contribute to the energy release rate. Third, the experimental data are relatively easy to

interpret, as explained below.

9/13/2002 38
 Let us return to Figure 18. Because the substrates are much thicker than the films, the

stress field in the substrate, far away from the films, is unaffected by the presence of the films,

and is the same as that around a crack in a homogeneous body, given in the form (39). The load

on the substrate can be represented by the two stress intensity factors KI and KII for a crack in a

homogenous body. Near the crack tip, the stress field is that of an interfacial crack (47),

characterized by the complex stress intensity factor K. The energy release rate is related to both

the far field and the near tip field, namely,

1  1 1  K2
( II )
1
2  E1 E2  cosh2 (πε )
G= K 2
+ K 2
= + . (63)
E1 I

The energy release rate can be calculated from the homogeneous specimen, neglecting the thin

films.

The mode angle, however, has to be determined by solving the boundary value problem

that includes the thin films. Let ψ ∞ be the mode angle of the remote load, namely,

tan ψ ∞ = KII / KI . The local mode angle ψ is defined by (51) using an arbitrary length l. The

local mode angle relates to the remote mode angle as

ψ = ψ ∞ + ω + ε ln (l / h), (64)

where h is the thickness of one of the films, and ω is an angle that depends on elastic properties

of the films and the substrates, as well as the ratios between various films. Suo and Hutchinson

(1989) solved the case of a single film between two substrates, and found that ω is less than 10°

provided the elastic mismatch is not too large. It is rather cumbersome to keep track of the local

mode angle, especially when several films are sandwiched, and when some of them deform

plastically. A common practice has been to specify the mode angle for the sandwich specimens

by the remote value ψ ∞ .

9/13/2002 39
 As an example, Figure 19 illustrates a four-point bend specimen. The method was

introduced by Charalambides et al. (1989) and Cao and Evans (1989) to determine interfacial

fracture energy, and developed by Dauskardt et al. (1998) for thin film structures relevant for

interconnects. The experimental procedure is as follows. Use two silicon wafers to sandwich

films of interest. Notch the top wafer to within a few microns of the sandwiched films with a

diamond wafering blade. Place the sample in a four-point bend fixture. Record load as a

function of displacement, and observe the crack propagation in an optical microscope. At a

certain load, a crack initiates from the notch root, approaches the interface, and then bifurcates

into two cracks to propagate on the interface. When the interface cracks are long compared to

the substrate thickness, the load-displacement curve exhibits a plateau—that is, the crack reaches

a steady state, in which the energy release rate is independent of the crack length.

At the steady state, both the energy release rate and the mode angle are obtained

analytically. As shown in Figure 20, the four-point bend configuration is the superposition of

two other configurations, one being pure mode I, and the other being pure mode II. The energy

release rates for the two configurations can be obtained from elementary considerations.

Consequently, the energy release rate for the four-point bend specimen is

21 M 2
G= , (65)
4 E H3

where M is the moment per substrate unit width, H the thickness of each substrate, and E the

plane strain modulus of the substrates. The mode angle for the four-point bend specimen is

ψ ∞ = tan−1 ( 3 / 2) ≈ 41° . (66)

Once the plateau load is measured experimentally, (65) is used to calculate the fracture energy Γ

for the mode angle 41°.

9/13/2002 40
 Also commonly used is the double-cantilever beam specimen (Oh et al., 1987; Kook and

Dauskardt, 2002). The energy release rate for the double-cantilever beam specimen of a

homogenous material is found in handbooks, given by

(Pa )2  H
2

H E  a
G = 12 3 1 + 0.677 , (67)

where a is the crack length, H the thickness of each beam, P the force per beam width, and E

the plane strain elastic modulus. The specimen is under the mode I condition, ψ ∞ = 0° . As

discussed before, a crack in a homogenous double-cantilever beam specimen tends to curve away

from the mid-plane (Fig. 14). By contrast, when films are sandwiched between the two beams,

creating a weak interface, the crack will propagate along the interface. In the experiment of Oh et

al. (1987), 750 nm copper films were evaporated onto two glass substrates (of size 75 by 25 by 2

mm). To confine crack propagation to one glass-copper interface, a layer of 10 to 20 nm

chromium was evaporated onto one substrate prior to depositing the copper film. The chromium

layer promoted adhesion. The copper films from the two substrates were then pressure diffusion

bonded. To load the sample, aluminum arms were attached by epoxy to the ends of the

specimen. The thin specimen led to high enough specimen compliance that the crack was easily

arrested, permitting multiple testing of one specimen. The experimentally measured load to cause

the crack to propagate rapidly gave the interfacial fracture energy according to (68). The crack

on the glass-copper interface displayed subcritical growth rates spanning from ~10 −8 to 10 −3 m/s,

in the range of the energy release rates from ~0.5 to 2.0 J/m 2 .

The measurements of interfacial fracture energy provide a tool for process control (Dr. Q.

Ma, of Intel Corporation, private communication). Many factors during the interconnect

fabrication process can affect adhesion. Because debonding usually occurs after a interconnect

structure is fabricated, when stresses arise from, e.g., thermal expansion mismatch. After a

9/13/2002 41
debond is identified, one has to modify the processing conditions, fabricate the interconnect

structure again, and examine if the new structure has the debond. The iteration is time

consuming. Alternatively, without fabricating the whole interconnect structure, one can measure

the interfacial fracture energy for the interface at question under various fabrication conditions,

and then select the condition that gives an adequate interfacial fracture energy.

The measurements of interfacial fracture energy also provide a tool to select materials to

promote adhesion. When copper was introduced into the interconnect structure in late 1990s,

new barrier layers are required to prevent copper diffusion into dielectrics, and to provide

adhesion of copper to the dielectrics. The sandwiched four-point bend specimens were used to

measure the interfacial fracture energy of the thin film structures with Ta and TaN as barrier

layers (Lane et al., 2000), and thereby to study the effects of interfacial chemistry, copper layer

thickness, and dielectric type on adhesion. More recently, numerous materials of low dielectric

constant have been developed to compete to be a replacement for silica in the interconnect

structure. To compare these materials, a commonly accepted method is desired to measure the

interfacial fracture energy. Rim et al., (2002) have evaluated adhesion of a carbon-doped oxide

in thin film structures.

Interfacial cracks are often susceptible to subcritical growth under static or cyclic load.

Environment strongly affects the growth rate. While the implications of subcritical debond

growth for long term reliability of the interconnect structure have yet been fully understood, the

fracture specimens discussed above have been used to study subcritical debond growth (Kook

and Dauskardt, 2002; Snodgrass et al., 2002).

3.4 Effects of Plasticity

9/13/2002 42
 The above discussions are confined in the Linear Elastic Fracture Mechanics (LEFM). A

main requirement of the LEFM is that the inelastic zone around the crack tip must be much

smaller than the feature size, in our case, the film thickness. The inelastic deformation within the

small zone is accounted for by the fracture energy. In computing the energy release rate, the

structure is taken to be elastic. When the inelastic zone is large compared to the film thickness,

however, we must consider the inelastic deformation more explicitly. The following examples

illustrate roles played by plasticity.

As a first example, consider two films, one brittle and the other ductile, sandwiched

between two silicon wafers (Fig. 21). The crack extends on the interface between one wafer and

the brittle film by a bond breaking process. So long as the brittle film thickness is large

compared to the atomic spacing, the bond breaking process is unaffected by the presence of the

ductile film. Consequently, the fracture energy Γ is well defined for the interface between the

wafer and the brittle film. As the crack extends, a material particle in the ductile layer plastically

deforms, and then elastically unloads. This hysteresis can be accounted for by the continuum

theory of plasticity. In the silicon wafers, far away from the films, the stress state is the same as

that for a crack in a homogeneous elastic material, with the energy release rate G representing

the amplitude of the external load. An elastic-plastic boundary value problem relates the

external load G to the bond breaking energy Γ and the plastic deformation properties (Suo et al.

1993; Lane et al. 2000). The hysteresis costs energy. The energy release rate to maintain steady

state crack growth, Gss , exceeds the interfacial fracture energy Γ . The ratio Gss / Γ depends on

the thicknesses of the two films, and on the yield strength of the ductile layer. When the ductile

layer is absent, Gss / Γ = 1, so that Γ can be determined experimentally by the external load. By

gradually increasing the ductile layer thickness, the effect of plastic deformation on Gss can be

9/13/2002 43
both measured and calculated. One may study deformation properties of the ductile layer this

way.

A second example involves a crack on a metal-ceramic interface (Figure 22). The metal

deforms plastically, which may be described by the continuum plasticity theory (Hutchinson and

Evans, 2000), or by dislocation dynamics (Needleman and Van der Giessen, 2001). Two classes

of failure modes are envisioned. The crack may extend by decohesion (Fig. 22a). The bond

breaking process may be characterized by a traction-displacement law. The crack tip is no

longer a structureless point, but spreads out into a cohesive zone. The deformation hysteresis in

the metal costs energy, as discussed above. Alternatively, the crack may extend by linking voids

(Fig. 22b). As the load increases and the crack tip blunts, the peak stress remains constant (a few

times the yield strength), but the strain keeps increasing (Shih et al., 1991). Voids open ahead of

the crack tip, and link together to extend the crack. The behavior is similar to ductile fracture of

bulk metals.

As a third example, consider a thin metal film sandwiched between two elastic substrates

(Figure 23). A debond crack preexists on one of the interfaces. The substrates are loaded to

open the crack. Let h be the film thickness, and rp be the plastic zone size upon crack

propagation. When h»rp , the substrates place no additional constraint on fracture, and the

behavior is the same as described in Fig. 22. When h < rp , however, the substrates constrain the

plastic flow in the film. As the load increases and the crack tip blunts, the level of the peak stress

in the metal increases, and the location of the peak stress is several times the film thickness

ahead of the crack tip (Varias et al., 1991). The high stress activates flaws ahead of the crack tip

into new debond areas, which link back to the main crack (Reimanis et al., 1991). When the film

thickness is very small, the dislocations emitted from the crack tip pile up at the substrate-film

9/13/2002 44
interfaces (Figure 24). As the load increases and the crack tip blunts, more dislocations pile up

and the stress at the crack tip can increase up to the theoretical strength (Hsia et al., 1994; Mao

and Evans, 1997). Even FCC metals can fracture by cleavage this way.

4 RATCHETING

4.1 Temperature Cycling as a Qualification Test

Temperature cycling has long been used as a test to qualify devices (Suhir, 2002). After

being cycled between two temperatures, say –55°C to 125°C for 1000 times, a device is tested

for its functions, and then sectioned and examined in microscopes for failure modes (e.g.,

distortion and cracking). If a failure mode is found, one makes a new device by modifying either

processing parameters, or geometries, or materials. The new device is temperature cycled again,

followed by the microscopy examination. The iterations do not guarantee quick convergence, as

the modifications made to avert one failure mode may cause another. These make-and-break

iterations are extremely time consuming, and are a bottleneck for innovation. Consequently, it is

urgent to understand various failure modes caused by temperature cycling.

Figure 25a and b are a photo of a flip-chip package and a schematic of its cross section. An

organic substrate is bonded to a silicon die on the face that contains transistors. The solder

bumps connect the interconnects on the die to the printed circuits on the substrate. Layers of

polymers (epoxy underfill and polyimide) lie between the substrate and the die. Figure 25c

illustrates part of the interconnect structure, with an aluminum film on silica. The feature size in

the interconnect structure is much smaller than the overall package. Before the die is bonded to

the substrate, a silicon nitride (SiN) film is deposited over the interconnect structure, covering

both the metal and the dielectric, serving as a barrier to environmental molecules.

9/13/2002 45
 Observed failure modes in such a package caused by temperature cycling include

• die-polymer or polymer-substrate debonding (Yan and Agarwal, 1998; Gurumurthy et al.,

1998),

• solder bump detachment (Lau, et al., 1998),

• silicon die cracking (Han, 2001; Michaelides and Sitaraman, 1999),

• SiN film cracking (Huang et al., 2000, 2002; Edwards et al., 1987; Alpern et al., 1994; Gee

et al., 1995), and

• metal film crawling (Huang et al., 2001; Alpern et al., 1994; Isagawa et al., 1980; Thomas,

1985).

The last two failure modes in the above list occur in the interconnect structure, and

provide the focus of the following discussion. Figure 26 shows the micrographs taken after

temperature cycles. Figure 26a is a plan view of the die surface near a corner of a test structure.

The structure has many aluminum pads. Cracks develop in the SiN film over some of the

aluminum pads. Figure 26b shows a patterned aluminum pad on silica, with a SiN film covering

both aluminum and silica. Cracks form in SiN over aluminum, but not in SiN over silica. Figure

26c shows a cross section of a crack in the SiN passivation film. Notice that the crack opens

substantially. Figure 26d is a cross section of two aluminum pads, which shift and break the SiN

film. The main experimental observations are summarized as follows.

• Cracking occurs in the SiN film at the die corners.

• The aluminum pads crawl toward the die center. The displacement reduces for pads

away from the die corners.

• Cracking occurs after temperature cycles, and becomes more extensive as the number of

cycles increases.

9/13/2002 46
 • Cracking occurs in the SiN film over the metal films, but not in the SiN film over silica.

• Cracking is more likely when the metal films are wide and the SiN film is thin.

Because the organic substrate has a larger thermal expansion coefficient than the silicon

die, upon cooling from the curing temperature, a shear stress develops on the silicon die. The

shear stress concentrates at the die corners, and decreases away from the corners. As illustrated

in Fig. 27, the entire range of cyclic temperature is below the epoxy curing temperature, so that

the shear stress at the die corners point toward the die center during cycling. It is not surprising

that failures occur at the die corners, and that aluminum pads crawl toward the die center.

Two questions, however, are inexplicable from these simple considerations. Why should

crawling and cracking occur at all? The shear stress on the die is transmitted through the

polymers, and is limited by the yield strength of the polymers, say 100 MPa. The yield strength

of the metal film exceeds 100 MPa (Nix, 1989). The fracture strength of the SiN thin film is on

the order 1 GPa (Ma et al., 1998). Why should cracks occur after many temperature cycles?

Unlike ductile materials, SiN does not have an intrinsic fatigue mechanism.

4.2 Ratcheting toward a Steady State

Huang et al. (2000) have recently discovered a mechanism that answers these questions.

As mentioned before, the silicon die and the packaging substrate have different thermal

expansion coefficients, inducing shear stresses, τ 0 , at the die corners upon curing (Fig. 28).

When the temperature cycles in a range below the curing temperature, the direction of the shear

stress τ 0 is biased toward the die center. For simplicity, τ 0 is taken to be constant, set by the

yield strength of the polymer, independent of position and temperature. The shear stress τ 0 is

partly sustained by a membrane stress σ in the passivation film, and partly transmitted to the

9/13/2002 47
metal film underneath as a shear stress in the metal, τ m . The metal film also has a large thermal

expansion mismatch with silicon and silica beneath, so that the metal film plastically yields in

every temperature cycle. Because the temperature change by itself causes the metal to yield,

even a very small shear stress in the metal, τ m , will cause the metal to deform in shear

plastically. The in-plane plastic strain in the metal film is constrained by the elastic substrate.

Consequently, the amount of plastic shear strain per cycle in the metal is small, as will become

evident later when we look at the model closely. The increment of the shear strain is in the same

direction as the shear stress on the passivation film τ 0 , pointing toward the center of the die.

Incrementally, the shear stress in the metal film τ m relaxes, and the membrane stress σ in the

overlaying silicon nitride film builds up. It is this evolving stress state that cracks the SiN film

after some cycles.

Figure 29 illustrates the evolution of the stress field as the temperature cycles. In the first

temperature cycle, the shear stress τ 0 from the organic substrate is transmitted into the metal

film underneath the SiN film, so that the stress in the SiN film is low, comparable to τ 0 (Fig.

29a). As the temperature cycles, the shear stress in the metal pad decreases, and the stress in SiN

accumulates. After many cycles, a steady state is reached, in which the metal no longer has any

shear stress (Fig. 29b). The shear stress τ 0 from the organic substrate is now all sustained by the

membrane stress in the SiN film. Figure 29c shows the free body diagram of the SiN film.

Denote σ A as the magnitude of the membrane stress in SiN at the two ends, tensile on the left

and compressive on the right. As discussed above, the film has the traction τ 0 on the top face

from the organic substrate, but no traction on the bottom face from the metal pad. Force balance

requires that

9/13/2002 48
 τ0 L
σA = , (68)
2t

where t is the thickness of the SiN film, and L the length of the aluminum pad. Although the

shear stress τ 0 is small, the length ratio L / t is typically very large. Consequently, the stress in

the SiN film can build to a high level to cause cracking. The magnitude of the stress in the SiN

film increases with the shear stress at the die corner and the width of the aluminum pad, but is

inversely proportional to the SiN film thickness. These are consistent with the experimental

observations. In the steady state, the membrane stress in SiN varies linearly with the position,

given by

σ (x ) = −2σ A x / L , (69)

where x is the distance from the center of the aluminum pad.

Finite element calculation of a plane strain model has confirmed the main features of the

model (Huang et al. 2000). As the temperature cycles, the shear stress in the metal reduces, and

the membrane stress in the SiN film increases. In the steady state, the shear stress in the metal

vanishes, the inplane stress in the metal cycles with the temperature, and the membrane stress in

the SiN film attains the distribution (69).

The deformation in the same direction, guided by a bias and facilitated by a cyclic load

(temperature change in this case), is known as ratcheting deformation (Bree, 1967; Jansson and

Leckie, 1992; Suresh, 1998). In addition to the example discussed here, several other examples

of ratcheting deformation in thin film structures have been discovered recently (He et al., 2000;

Begley and Evans, 2001; Karlsson and Evans, 2001).

To avert SiN cracking, an industrial practice is to change the shape and the size of the

metal pad. For example, slots are made in the metal pads, so that the SiN film is in contact with

the silica in the slots (Fig.30). To aid design, we would like to do calculations over thousands of

9/13/2002 49
temperature cycles, for three-dimensional structures, and for many geometrical designs.

However, because of the nonlinear material behavior, each temperature rise or drop has to be

divided into many small steps in the finite element calculation. A crude division would lead to

numerical ratcheting, i.e., the apparent ratcheting deformation due to accumulated numerical

errors for structures that should not suffer ratcheting at all. Consequently, direct finite element

calculation would require too much time to be practical.

In the steady state, the shear stress in the metal pad vanishes, and the SiN film is in a state

of plane stress elastic deformation. At the edges of the metal pad, the displacement of the SiN

film is constrained by the silica, and is set to be zero. The shear stress from the organic substrate

is distributed over the face of the SiN film, acting like a body force. Such a linear elastic plane

stress problem can be solved rapidly by the finite element method for any pad geometry. As an

illustration, Figure 31a shows the normalized maximum principal stress distribution at the steady

state by applying the uniform shear stress τ 0 in the direction of -45° from the x-axis. The

membrane stresses in the passivation film scale with τ 0 L / h p , where L is the width of the metal

film. The maximum principal tensile stress is very high near the left edge and the left upper

corner, where cracking always happens. Figure 31b shows a slotted structure, where the

membrane stress in the passivation film is greatly reduced.

The steady-state is a robust concept: it does not depend on the detailed mechanical

property of the metal, and the stress field is easy to calculate. The steady-state is applicable even

if the metal creeps, or the passivation film and the metal debond. Once the stress field in the

steady state is calculated, one can ascertain if the SiN film is safe from cracking. If SiN can

sustain the stress in the steady state, then the design is immortal: it can sustain any number of

temperature cycles without cracking.

9/13/2002 50
 This steady-state approach leaves two issues unresolved: the condition under which the

metal film undergoes ratcheting deformation, and the number of cycles needed to approach the

steady-state. We study these issues in the following sections. To gain insight, we examine

idealized structures and material behaviors.

4.3 Shakedown vs Cyclic Plastic Deformation

First consider a blanket metal film bonded to a substrate. The film has thermal expansion

coefficient α m , Young’s modulus Em , and Poisson’s ratio ν m . We model the film as a

nonhardening solid, assuming that the yield strength Y is independent of temperature, and is the

same under tension and compression. Subject to a temperature change, the film is under biaxial

stress σ . The substrate, being much thicker than the film, has negligible stress, so that only its

thermal expansion coefficient α s enters the consideration. The film and the substrate are well

bonded, matching the net strain at all time.

When the temperature changes by dT , the film and the substrate acquire thermal strains

α m dT and α s dT , respectively. This mismatch is compensated in different ways, depending on

whether the film is elastic or plastic. If the film is elastic, σ < Y , the mismatch in the thermal

strains is compensated by the change in the elastic strain in the film. The latter gives rise to the

change in the film stress:

E m (α m − α s )
dσ = − dT . (70)
1− ν m

If the film yields, however, the stress in the film is fixed at the yield strength, either in tension

σ = +Y or in compression σ = −Y , so that the elastic strain in the film no longer changes with

9/13/2002 51
the temperature. Consequently, the mismatch in the thermal strains is compensated by the plastic

strain in the film:

dε p = − (α m − α s )dT . (71)

Figure 32 plots the change in the film stress σ and the plastic strain ε p when the

temperature cycles between TL and TH . Let us start at the preset high temperature TH , with the

initial film stress σ 0 anywhere between ± Y .

(1) As the temperature drops, the film is elastic, and the film stress changes toward

tension according to Eq. (70). The plastic strain does not change.

(2) At some temperature TA , the film yields in tension, and the film stress is fixed at the

yield strength, σ = +Y . As the temperature drops, the plastic strain increases

according to (71).

(3) On reaching the preset low temperature TL , the temperature starts to rise. The film is

elastic, and the film stress changes toward compression. The plastic strain does not

change.

(4) At some temperature TB , the film yields in compression, and the film stress is fixed

at the yield strength, σ = −Y . The plastic strain decreases as the temperature rises.

(5) On reaching TH , the temperature starts to drop again. The film is elastic, and the

film stress changes toward tension. The plastic strain does not change. Afterwards,

the stress and strain state of the film cycles with the temperature.

Inspecting Fig. 32, we distinguish three behaviors. Imagine that TL is fixed, and TH is

allowed to change. If TL > TA , the film will be elastic during temperature cycling. If TH < TB ,

the film will undergo plastic deformation in the first cycle, but will be elastic in all subsequent

cycles. This behavior is referred to as shakedown. The film reaches the shakedown condition if

9/13/2002 52
 Em (α m − α s )(TH − TL )
<2. (72)
(1− ν m )Y
If the inequality sign in (72) is reversed, the film will undergo cyclic plastic deformation. During

the fabrication of the interconnect structure, the metal film inevitably yields as the temperature

drops from the processing temperature. Consequently, plastic yield during the first cycle should

not be of any concern, and the elastic and shakedown behaviors are practically indistinguishable.

However, as we will see below, the difference between shakedown and cyclic plasticity has a

significant consequence. The demarcation (72) between the two regimes, shakedown and cyclic

plasticity, is unaffected by the initial film stress.

4.4 Ratcheting Strain Rate

In the above, the plastic strain cycles because the film has to match the net inplane strain

to the substrate. By contrast, the shear strain in the film is not constrained by the substrate, so

that it may accumulate under certain conditions. Huang et al. (2001) have calculated the strain

rate due to ratcheting. Figure 33a illustrates a blanket metal film bonded on a semi-infinite

elastic substrate. A shear stress τ m acts on the surface of the metal film. For the time being, this

shear stress is taken to be spatially uniform, and remains constant as the temperature cycles. The

stress field in the structure is very simple. The semi-infinite substrate is subject to the same shear

stress τ m , but no other stress components. Let the coordinates in the plane of the film be x1 and

x2, the coordinate normal to the plane be x3, and the coordinate x1 coincide with the direction of

the shear stress τ m . As shown in Fig. 33b, the film is in a uniform stress state of a combination

of a biaxial stress and a shear stress:

σ 11 = σ 22 = σ m , σ 13 = τ m , σ 23 = σ 12 = 0 . (73)

9/13/2002 53
The biaxial stress σ m changes with the temperature. Obviously the uniform stress states in the

substrate and in the film satisfy the equilibrium conditions.

To model plastic deformation in the metal film, we adopt the J2 flow theory (Hill, 1950).

The deviatoric stress tensor, s ij = σ ij − σ kk δ ij / 3 , has the components

s11 = s 22 = σ m / 3, s33 = −2σ m / 3, s13 = τ m , s 23 = s12 = 0 . (74)

The Mises yield condition, 3sij sij / 2 = Y 2 , is specialized to

σ m 2 + 3τ m 2 = Y 2 . (75)

The yield condition is an ellipse on the ( σ m , τ m ) plane (Fig. 33c). The metal film is elastic when

the stress state is inside the ellipse, and yields when the stress state is on the ellipse. The stress

state outside the ellipse is unreachable. When the film yields, for a given τ m , the biaxial stress

can only be at one of the two levels:

σ m = ± Y 2 − 3τ m 2 . (76)

These two states, tensile and compressive, are indicated in Fig. 33c.

Figure 34 shows a coordinate plane spanned by the normalized shear stress, τ m / Y , and

E (α f − α s )(TH − TL )
the normalized temperature range, . The plane is divided into three
(1 − ν )Y
τm 1
regimes: plastic collapse, shakedown, and crawling. If ≥ , the film plastically deforms
Y 3

under the shear stress alone without the aid of the temperature change, and is in the plastic

τm 1
collapse regime. If < and the temperature range is small enough, the film may plastically
Y 3

deform during the first temperature cycle, but cycles elastically afterward. The film is in the

shakedown regime. This shakedown condition is represented in Fig. 34 by a shaded region below

9/13/2002 54
a quarter of an ellipse:

τ
= 2 1 − 3 m  .
E(α f − α s )(TH − TL ) 2

(77)
(1 − ν )Y Y

The J2 flow theory dictates that the plastic strain increment tensor be in the same direction

as the deviatoric stress tensor, namely, dε ij = sijp dλ , where dλ is a scalar. Consequently, from

Eq. (74) we obtain that

dε p dγ p
= , (78)
σ m / 3 2τ m

where γ = 2ε 13
p p
is the plastic shear strain in the film. Equation (78) is the key to the

understanding of metal film crawling. Its geometric interpretation is well known: the increment

of the plastic strain tensor points in the direction normal to the yield surface (Fig. 33c). During

plastic deformation, dε p has the same sign as the biaxial stress σ m , and dγ p has the same sign

as the shear stress τ m . When film yields in compression (σ m < 0) , the plastic in-plane strain

decreases, dε p < 0 . When film yields in tension (σ m > 0) , the plastic in-plane strain increases,

dε p > 0 . Because the film is bonded to the substrate, for a given temperature increment, dε p is

always finite. Consequently, in each cycle, γ p

increases by a finite amount in the direction of

τ m both when the film is in tension and in compression.

When the film yields, the plastic in-plane strain changes according to Eqn. (71). A

substitution into (78) gives the increment of the plastic shear strain:

6τ m
dγ p
=− (α m − α s )dT . (79)
σm

When the film plastically deforms in tension, σ m = + Y 2 − 3τ m and dT < 0 . When the film
2

9/13/2002 55
plastically deforms in compression, σ m = − Y 2 − 3τ m and dT > 0 . In either case, the plastic
2

shear strain increment is in the same direction of τ m , namely, dγ p

> 0 , giving rise to crawling.

In Figure 34, above the shakedown regime, if τ m = 0 , the temperature change causes

cyclic in-plane strain but no shear strain. However, this cyclic state is unstable. Any shear stress

τ m , however small, will cause the film to crawl. In the crawling regime, the in-plane strain cycles

plastically, but the shear strain increases by a fixed amount each cycle, given by

 E (α − α )(T − T ) 
 m m − 2 .
12(1 − ν m )τ m
 (1 − ν m ) Y 2 − 3τ m 
∆γ p
= s H L
(80)
 
Em 2

Figure 35 plots the ratcheting strain per cycle, ∆γ p , as a function of the shear stress, τ m .

We make an analogy between the strain per temperature cycle (i.e. the ratcheting rate) and the

strain per unit time (i.e. the strain rate). Figure 35 is thus analogous to the relation between strain

rate and stress in viscous flow. In general, we write the ratcheting rate law as ∂γ p / ∂N = f (τ m ) .

The shear stress should be bounded as τ m / Y < 1 / 3 ; otherwise the film has unlimited plastic

shear strain even without the temperature change. As τ m / Y → 1 / 3 , the ratcheting rate

becomes large for any given temperature ranges. Depending on the temperature range, we

distinguish several behaviors as follows.

E m (α m − α s )(TH − TL )
When > 2 , the temperature change by itself can cause the metal
(1 − ν m )Y
film to deform plastically in every cycle. When the shear stress is zero, the metal film undergoes

cyclic in-plane plastic deformation, and has no shear strain. When the shear stress is small, the

ratcheting rate is linearly proportional to the shear stress:

9/13/2002 56
 ∂γ p τ m
= . (81)
∂N η R

We call ηR the ratcheting-viscosity. A comparison between Eqs. (80) and (81) gives

 E (α m − α s )(TH − TL ) 
−1

12(1 − ν m ) 
− 2 .
Em
ηR=

(82)
(1 − ν m )Y
The linear ratcheting is analogous to the Newtonian viscous flow. The ratcheting-viscosity has

the dimension of elastic modulus, and increases when the temperature range decreases. We will

mainly use the linear ratcheting approximation (82) in the subsequent development.

E m (α m − α s )(TH − TL )
When < 2 , the metal film yields only when the shear stress exceeds
(1 − ν m )Y
a critical value, namely, when the quantity in the bracket in Eq. (80) is positive. As seen in Fig.

35, this behavior is analogs to the Bingham viscous flow.

E m (α m − α s )(TH − TL )
When = 2 , for small shear stress, equation (80) gives rise to a
(1 − ν m )Y

power law behavior, ∂γ p / ∂N ~ τ m .

3

4.5 The Transient Process to Approach the Steady State

Huang et al. (2002) have developed a one-dimensional model of ratcheting of the thin

film structures. As shown in Fig. 36, a metal film is in a thick elastic substrate, and the

passivation thin film is on the metal film and the substrate. A constant shear stress, τ 0 , due to the

packaging substrate, is applied on the passivation film. The metal film is of thickness H and

length L. The passivation film is of thickness h. Cycle the structure between TH and TL . The

shear stress in the metal film is τ m , and the normal stress in the passivation film is σ . Now both

9/13/2002 57
τ m and σ are allowed to vary with the number of cycles, N, and the position, x. We expect τ m

to vary slowly with N and x, so that we will use the ratcheting rate law obtained in the last

section.

First look at the elastic passivation film. As shown in the inset in Fig. 36, the passivation

film is subjected to a membrane stress σ ( N , x) , a constant shear stress τ 0 on the top surface, and

a variable shear stress τ m ( N , x) on the bottom surface. The force balance of the differential

element requires that

∂σ τ m − τ 0
= . (83)
∂x h

Let u (N, x) be the displacement of the passivation film in the x-direction. Elasticity of the

passivation film requires that

∂u
σ = Ep , (84)
∂x

where E p is Young’s modulus of the passivation film.

Next look at the ratcheting metal film. The shear strain relates to the displacement as

u E m (α m − α s )(TH − TL )
γp = . We assume that the temperature range is large, > 2.
H (1 − ν m )Y
Consequently, we adopt the shear stress and plastic shear strain increment relation (13) as

∂u
τm = ηR . (85)
H∂N

The ratcheting-viscosity ηR is defined by Eq. (82).

Substituting Eq. (84) and (85) into Eq. (83) gives the governing equation

∂u ∂ 2u Hτ 0
= DR 2 + . (86)
∂N ∂x ηR

We introduce the ratcheting-diffusivity as

9/13/2002 58
 DR = E p hH / ηR . (87)

Equation (86) is a diffusion equation with a source term.

We assume no separation between the metal film and the substrate, so that the

displacement of the metal film at the two edges is the same as the displacement of the substrate.

As the substrate is semi-infinite and elastic, its displacement is negligible compared to the

accumulated displacement in the passivation film. Consequently, the boundary conditions for the

displacement of the passivation film are

u = 0 at x = ± L / 2 for all N. (88)

Because the temperature range is large enough to cause the metal to yield in every cycle,

so long as there exists a shear stress τ m in the metal, the metal will ratchet. After many

temperature cycles, the structure will reach the steady state, in which the shear stress in the metal

film vanishes, τ m = 0 , but the cycling temperature still causes the metal film to yield cyclically.

The steady state can be obtained by setting ∂u / ∂N = 0 and satisfying the boundary conditions

(88). In the steady state, τ 0 is fully sustained by the membrane stress in the passivation film,

giving rise to a linear membrane stress distribution (69).

Analogous to any diffusion problem, the characteristic number of cycles to reach the

steady state can be estimated by NC = L2 / DR , or

 L2   Em (α m − α f )(TH − TL ) 
NC = 
−1

 hH  12E p (1 − ν m ) 
− 2 .
Em
 
(89)
(1− ν m )Y

The number N C can be large mainly because the ratio L2 / (hH ) is large. If one replaces the

metal film with a plastically deformable polymer, the ratio of elastic modulus of the polymer and

9/13/2002 59
the passivation is small, which would significantly reduce N C . Note that N C is independent of

τ0.

In the above, the metal film is taken to be a non-hardening material in this model. This

corresponds to a safe engineering design, as the hardening metal film can reduce the ratcheting

effects (Huang et al., 2001). The model can also be easily extended to consider the non-linearity

of the ratcheting-viscosity by using the finite difference method. The stress concentration at the

edges needs to be better treated, as the maximum tensile stress in the passivation film is at the

corner. Fracture conditions also need to be considered more carefully. Several useful rules for

design are evident. First, one should aim to avoid metal film ratcheting altogether. Under cyclic

temperature, the metal film cycles elastically if temperature range is small and the yield strength

of the metal is large, according to (77). Second, if one cannot avoid ratcheting, one may aim to

design the metal film geometry such that the passivation film can sustain the steady state, in

which the stress distribution in the passivation film can be calculated by solving a linear elastic

plane stress problem. Third, if the characteristic umber N C is much larger than the number of

cycles required by the qualification test, one may design according to the transient stress state,

solving the stress field as a function of the number of cycles. The present model allows one to

evaluate these options in design.

4.6 Ratcheting-induced Crack Growth

Consider the structure illustrated in Fig. 10 again. In the present context, a brittle film

lies on a metal underlayer, which in turn lies on a rigid substrate. Cycle the structure between

temperatures TL and TH. Within the temperature and time of the experiment, the metal does not

creep, but can have time-independent plastic deformation. For simplicity, assume that the tensile

9/13/2002 60
stress in the uncracked blanket film, σ , is constant as the temperature changes. This happens in

practice when the elastic film and the substrate have a similar coefficient of thermal expansion,

and σ is due to the deposition process or an applied load. As temperature cycles, the metal layer

undergoes ratcheting plastic deformation. The stress field in the film relaxes in the crack wake,

and intensifies around the crack tip. By the ratcheting-creep analogy, we rewrite the scaling law

for the crack velocity (38) as the scaling law for crack growth per cycle (Liang et al., 2002a):

da Hhσ 2
= 0.6 , (90)
dN ηRΓ

where ηR is given by (82), and Γ is the fracture energy of the brittle film.

5 STRESS-INDUCED VOIDING

This section begins with a description of stress-induced voiding (SIV), followed by an

analysis of its main features: saturated void volume, resistance change, stress generation, void

initiation, void growth, and stress relaxation. We then give an analytical model of non-

hydrostatic stress state in an elastic-plastic interconnect line encapsulated in a dielectric. Recent

observation of SIV in blanket copper films is discussed. We finally consider the effects of low k

dielectrics and barriers. Okabayashi (1993) and Sullivan (1996) have provided two reviews on

SIV. This section will focus on physical descriptions of the phenomenon, leaving the design

implications to the next section, after we have studied electromigration.

5.1 Stress-Induced Voiding in Encapsulated Metal Line

Early aluminum lines had the width much larger than the thickness. They behaved like

blanket films. When narrow aluminum lines were introduced, voids were observed on wafers

held in ovens, or even on wafers left on shelves at room temperature (Curry, et al., 1984; Yue, et

9/13/2002 61
al., 1985; Turner and Wendel, 1985). The voids may be crack-like or rounded, grow with time,

and sever the interconnects (Fig. 37).

A basic understanding soon followed (Jones and Basehore, 1987; Hinode et al., 1990;

Greenebaum et al., 1991; Korhonen, et al. 1991; Sauter and Nix, 1992). Aluminum has a larger

thermal expansion coefficient than silica. Encapsulated in silica at a high temperature, on

cooling, a narrow aluminum line is in a tensile, triaxial stress state. The magnitude of the stress

in aluminum readily exceeds the yield strength of aluminum. Indeed, even a liquid with zero

yield strength, so encapsulated, would be under hydrostatic tension. Over a long time, a small

piece of a crystal is like a liquid in that they both creep under shear. An encapsulated metal line

under uniform hydrostatic tension is in a state of equilibrium: atoms have no motivation to

diffuse from one place to another. However, this equilibrium state is metastable. A flaw may act

like a void nucleus. If the stress near the flaw is low or vanishes, the nonuniform stress field

motivates aluminum atoms to diffuse away from the flaw, simultaneously growing the void and

relaxing the tensile stress (Figure 38). The metal line approaches a stable equilibrium state

consisting of voids and unstressed metal.

An accelerated testing method has also been developed, known as the high temperature

storage (HTS) test. The aluminum line is stress-free at a temperature T0 , which is close to the

silica deposition temperature. Samples are kept in an oven at some temperature T below T0 . The

thermal expansion mismatch volume scales with the temperature drop, T0 − T . It is this

mismatch volume that feeds to the void. The atomic diffusivity depends on the temperature as

D = D0 exp (−Q / kT ), where k is Boltzmann’s constant, Q the activation energy, and D0 the

frequency factor. As the temperature drops, the mismatch volume increases, but the diffusivity

decreases (Figure 39). Consequently, voids grow fastest at a particular temperature Tmax . During

9/13/2002 62
the storage test, the electrical resistance of a line is measured as a function of time. The line fails

if its resistance increases by, say 20%. The time to fail half of the lines, i.e., the median-time-to-

failure (MTF), is fit to a semi-empirical relation (McPherson and Dunn, 1987):

 Q
 kT  ,
C
MTF = N exp (91)
(T0 − T )
where C and N are fitting parameters, along with T0 and Q. This relation captures the above

physical considerations. The failure time approaches infinity when the test is carried either at the

stress-free temperature T0 , or at the absolute zero. To obtain the fitting parameters, the storage

tests are carried out at several temperatures, possibly around Tmax to shorten the testing time.

Equation (91) is then extrapolated to predict the failure time at, say, room temperature.

SIV is mitigated in aluminum interconnects in several ways. First, adding a few percent

of copper into aluminum reduces the diffusion rate by nearly two orders of magnitude (Ames et

al., 1970). Second, materials of negligible diffusion rates, such as W and TiN, are used as shunts

(Figure 38), so that a void in the aluminum line does not open the circuit, but just increases the

electrical resistance somewhat (Howard et al., 1978). Third, a short aluminum line cannot grow

a large void, because the volume of material to be relaxed is small (Filippi et al., 1995). These

measures were all originally developed to mitigate electromigration voiding.

The interest in SIV has resurged recently since copper lines entered the interconnect

structure. Voids have been observed in electron beam evaporated copper lines, and found to be

sensitive to the texture (Keller et al., 1997). On-chip copper lines are electroplated, and have

small initial grains. If the lines are not properly annealed before encapsulation, later annealing

will cause grain growth, densify copper, and induce voids (Hommel et al., 2001; Ogawa et al.,

2002). In aluminum interconnects, voids grow in the lines, but not in the tungsten vias. By

9/13/2002 63
contrast, in copper interconnects, vias themselves are made of copper, and are susceptible to

voiding (Fig. 40). Furthermore, compared to aluminum, copper has large yield strength and a

high degree of elastic anisotropy. Stress concentration arises at the junction of the copper grain

boundary and SiN top layer; voids have been observed even in blanket copper films (Shaw,

2001; Sekiguchi, 2001).

5.2 An Analysis of Stress-Induced Voiding

Despite efforts in for nearly two decades, there is no single model that accounts for all

aspects of SIV. Complicated factors (e.g., grain structures, flaws, interfaces, and atomic

transport mechanisms) play roles. Detailed information of such factors is incompletely known,

and will probably never be precisely controlled in fabrication. At this point, it seems desirable to

consider the physics of the phenomenon semi-quantitatively. In this spirit, the following sub-

models are given in outline.

Saturated Void Volume (SVV)

A encapsulated metal line is stress-free at a high temperature T0 , and is cooled to the

storage test temperature T. The total number of atoms in the line is constant at all time. Because

atomic transport does not change the volume of the metal, the thermal misfit strain is

accommodated by a combination of void space and elastic deformation. When the stress in the

line completely relaxes, the thermal expansion misfit is fully accommodated by the void space.

The saturated void volume (SVV) is

Vsv = 3∆α (T0 − T)V , (92)

9/13/2002 64
where V is the volume of the metal line, and ∆ α is the effective thermal expansion mismatch

strain. If a metal line were embedded in silicon, ∆ α = α m − α Si , where α m and α Si are the

thermal expansion coefficient of the metal and silicon. ( α m = 24 ×10 −6 K −1 for aluminum and

17 ×10 −6 K −1 for copper). In an interconnect structure, however, ∆ α depends on all materials in

the structure. For the silica-based interconnect structure, because the metals have much larger

thermal expansion coefficients than the surrounding materials (silica, silicon, silicon nitride, and

tantalum), the thermal expansion mismatch is insensitive to which exact value is used for these

other materials; we will use ∆ α = α m − 4 ×10 −6 K −1 in the following estimates. A representative

stress-free temperature is 400°C, and a representative storage test temperature is 200°C. Taking

these vales, one finds that Vsv / V = 1.2% for aluminum, and Vsv / V = 0.78% for copper. For a

line of size 1 by 1 by 10 µm, the saturated void, if it is a single spherical one, has radius 306 nm

for aluminum, and 265 nm for copper. The interconnect structure with low k dielectrics will be

discussed in Section 5.6.

Many voids may nucleate along the metal line, each growing to relax stress in its vicinity.

Once the stress completely relaxes, the sum of the volumes of all the voids remains constant as

the line is held at the constant temperature. Equation (92) gives this sum. The actual void

volume can be larger or smaller than the SVV. When a line is cooled to a low temperature for a

short time, the tensile stress partially relaxes, so that the total void volume is smaller than the

SVV. If the line is then brought to a high temperature, for some time the metal is under

compression, and the void volume is larger than the SVV. In case of a copper line, we assume

that the electroplated copper has been annealed to allow densification prior to encapsulation, so

that only thermal expansion mismatch contributes to void volume.

9/13/2002 65
Resistance Increase

A line fails when the voids significantly increase the resistance of the line. The resistance

change does not uniquely relate to the void volume. Small, isolated voids do not cause as much

resistance increase as a single large void that depletes a segment of the metal line. Even when

only one void is left near a via, the resistance change depends on geometry. Figure 41 compares

two identical voids, the only difference being the size of the tungsten studs relative to the void.

Evidently, the resistance increases in Fig. 41a is smaller than that in Fig. 41b.

When the void length exceeds several times the line width, the resistance increase can be

estimated from the void length. Let ρ be the resistivity of the interconnect, L its length, and A

its corss sectional area. The resistance of a void-free interconnect is R = ρ L / A , ignoring the

conductance due to the aluminide shunt layers. Let ρ1 and A1 be the resistivity and the cross-

sectional area of the aluminide layers. When a void of length L1 appears in the line, current

flows in the aluminide layers, and the resistance increases by ∆ R = ρ1 L1 / A1 , where the

resistance of aluminum of length L1 is neglected. Consequently, the relative resistance increase

is

∆R ρ1 L1 / A1
= . (93)
R ρL / A

Everything else being equal, the thicker the shunt layers, the smaller the resistance change.

Taking ρ1 / ρ = 10 , A1 / A = 0.1 , and L1 / L = 1.0% , one finds that ∆R / R = 100% .

Hydrostatic Tension in Encapsulated Metal Line

The stress field in the metal line is complicated in general, with tensor components and

nonuniformity along the line. To grow a large void, atoms must diffuse in the line over a length

much larger than the line width. On such a time scale, creep should have relaxed each point in

9/13/2002 66
the line to a hydrostatic state (Korhonen et al., 1991). That is, over a sufficient long time scale,

at an elevated temperature, a crystal behaves like a liquid. The hydrostatic stress in the metal

line prior to voiding can be calculated as follows. Figure 42 illustrates separately the metal line

and the dielectric. In the metal line, the stress state has components of an equal magnitude σ in

all directions. The stress causes the metal to expand with the volumetric elastic strain σ / Bm ,

where Bm is the bulk modulus of the metal. When the metal is modeled as an isotropic material,

the bulk modulus relates to Young’s modulus Em and Poisson’s ratio ν m as Bm = Em / 3(1− 2ν m ).

For a metal of cubic crystalline symmetry, the bulk modulus relates to its crystalline elastic

constants as Bm = (C11 + 2C12 ) / 3 . For aluminum, C11 = 108GPa, C12 = 62GPa, C44 = 28GPa ,

giving Bm = 77GPa . For Copper, C11 = 168GPa, C12 = 121GPa, C44 = 75GPa , giving

Bm = 136GPa .

Regard the encapsulation as an elastic solid with a hole subject to a normal traction of

magnitude σ . The dielectric is in a state of plane strain deformation. The traction causes the

cross-sectional area of the hole to reduce by ∫ un ds , where un is the normal component of the

elastic displacement at the hole surface. Let A be the cross-sectional area of the hole. Linearity

of the elastic field requires that

∫
1 σ
un ds = − . (94)
A µ eff

This equation defines an effective modulus µeff . This effective modulus depends on the shape of

the hole, as well as the elastic constants of the dielectric, the barrier, and the silicon substrate.

The effective modulus can be readily calculated using the finite element method (Section 5.6). If

the cross section is a long rectangle, rather than a square, the system is more compliant, leading

to a small µeff . In the idealized case, when the hole has a circular cross section, and the

9/13/2002 67
dielectric is infinite, the stress field in the dielectric is known analytically, and we find that the

effective modulus is µeff = E d / 2(1+ ν d ), coincident with the shear modulus of the dielectric.

Let θ be the volumetric strain caused, e.g., by the mismatch in thermal expansion

coefficients, or densification, or atomic diffusion. This volumetric strain is accommodated by

the difference in the volume change of the interconnect line and that of the dielectric hole.

Combining the two contributions, one relates the hydrostatic stress in the metal, σ , to the

volumetric mismatch strain between the metal and the encapsulation, θ , namely,

σ = Bθ , (95)

with the effective modulus B given by

1 1 1
= + . (96)
B Bm µeff

If the cross section is a long rectangle, the effective modus is small, leading to a low hydrostatic

stress. In the limiting case when the interconnect is a blanket film, the hydrostatic stress

vanishes in this liquid-like model. For an interconnect of a squre cross section, taking

µeff = 30GPa for silica, the effective modulus is B = 22GPa for aluminum/silica structure, and

B = 24GPa for copper/silica structure. The mismatch strain is mainly accommodated by the

elastic deformation of the dielectric, so that the effective moduli of the two structures are similar,

despite the nearly factor 2 difference in the bulk moduli of aluminum and copper. The effective

modulus reduces if the cross section is a long rectangle. Assume that the volumetric strain is

entirely due to the thermal expansion mismatch between the metal and the dielectric,

θ = 3∆α (T0 − T ). For the temperature drop T0 − T = 200K , the hydrostatic stress is

σ = 260MPa in aluminum line, and σ = 187MPa in copper line. Some authors have treated the

dielectric as a rigid material, and taken the bulk modulus of the metal, Bm , as the effective

9/13/2002 68
modulus. This would significantly overestimate the effective modulus ( Bm / B = 3.5 for

aluminum and Bm / B = 5.7 for copper).

Void Initiation

Next consider a void-like flaw in a metal subject to a hydrostatic stress. The void can

change its size by relocating atoms from the void surface into the metal far away from the void.

In doing so, the void increases its surface area, and the remote stress does work. The former

increases the free energy, but the latter decreases it. Will the void shrink or enlarge? For the

time being, we assume that the void volume changes slightly, so that the number of atoms

transported into the line is so small that the stress σ does not change. Take the solid without the

void but under the same applied stress as the ground state, at which the system has zero free

energy. The current state has a void of radius a. Let γ be the surface energy per unit area. In

creating the void, a surface of area 4π a2 is exposed, raising the free energy of the system by

γ 4π a2 . At the same time, atoms occupying the volume 4πa3 / 3 relocate, allowing the remote

stress to do work σ 4πa / 3 . Thus, the free energy of the system, relative to that of solid with no
3

void, is

4 3
F(a) = 4π a γ − π a σ .
2
(97)
3

Figure 43 sketches the free energy as a function of the void radius. When a = 0, F = 0, which is

the ground state. When the void is small, the surface energy dominates, and the void shrinks to

reduce the free energy. When the void is large, the stress dominates, and the void enlarges to

reduce the free energy. The two behaviors are separated at a critical void radius a * , at which the

free energy maximizes. The critical void radius is given by a* = 2γ / σ . Taking representative

9/13/2002 69
values, γ = 1J/m 2 and σ = 200MPa , we find that the critical radius is a* = 10nm . If the pre-

existing flaw size a is known, the flaw can grow if the stress exceeds the value

σ void = 2γ / a . (98)

The use of this condition will be critiqued in Section 6.3.

Void Growth and Stress Relaxation

Because the volumetric strain due to thermal expansion mismatch is on the order of 10-2,

for a void to grow to a size to sever the line, a long length of the metal has to be relaxed. The

long range diffusion samples averages of various quantities, and is a much more robust process

than void initiation. Korhonen et al. (1993) formulated a model to study atomic diffusion in an

encapsulated metal line. To study mass transport over a length exceeding several times the line

width, one can assume that the stress state is hydrostatic, varying along the line and with time.

Let the x axis coincide with the line. The stress in the line is a function of position and time,

σ (x,t ). The nonuniformity of the stress motivates atoms to diffuse to the region of large tension.

When an atom is removed from a stress-free solid, and then added to another solid under

hydrostatic tension σ , the free energy reduces by Ωσ , where Ω is the volume per atom in the

metal. Consequently, the driving force to relocate atoms in the metal line is

∂σ
f =Ω . (99)
∂x

Let J be the atomic flux, i.e., the number of atoms crossing per unit area per unit time. Write

Einstein’s relation as

Df
J= . (100)
ΩkT

9/13/2002 70
The diffusivity D is taken to be the average along the aluminum line. The flux divergence gives

rise to the volumetric strain per unit time, ∂θ / ∂t = Ω∂J / ∂x . Recall that the mismatch

volumetric strain relates to the stress by σ = Bθ . Consequently, the stress rate is proportional to

the flux divergence:

∂σ ∂J
= BΩ . (101)
∂t ∂x

A combination of the above equations gives

∂σ DBΩ ∂ 2 σ
= . (102)
∂t kT ∂x 2

This partial differential equation, together with suitable initial and boundary conditions, governs

the evolving pressure distribution, σ (x,t ). Solutions of practical interest can be found in

(Korhonen et al., 1993).

With reference to Figure 38, assume that the flaw is already large enough so that the

stress near the flaw is zero. Equation (102) is identical to the usual diffusion equation, with the

diffusivity-like quantity DBΩ / kT . The solution to this initial-boundary value problem is well

known. The time for a void to relax a segment of interconnect of length l scales as

l 2 kT
tl ~ , (103)
DBΩ

The void fails a line by increasing the resistance of the line to a designated value. Assume that

void of length L1 corresponds to the failure. To grow a void of length L1 , the segment of the

metal line need be relaxed is about l = L1 / 3(α m − α d )(T0 − T ). Inserting into (103), we obtain

the lifetime

Q
exp 
1
t life ∝
B[(α m − α d )(T0 − T )]
2
kT . (104)

9/13/2002 71
This expression looks similar to (91). As noted before, the effective modulus B depends on the

cross section shape, and the dielectric type. Other things being the same, a structure with a

smaller B will lead to a longer lifetime.

5.3 Nonhydrostatic Stress State in Encapsulated Metal Line

The presence of the yield strength causes a nonhydrostatic stress state in the metal line,

and also increases the magnitude of all stress components. According to (98), a higher stress can

activate a smaller flaw to grow into a void. The stresses in encapsulated elastic plastic metal

lines have been studied by using finite element methods (Jones and Basehore, 1987; Greenebaum

et al., 1991, Gouldstone et al., 1998).

To gain insight into this phenomenon, we give an analytical solution for an idealized

model, in which a long metal line of cylindrical cross section, radius a, is encapsulated in an

infinite dielectric (Figure 44). The dielectric is elastic, with Young’s modulus Ed , Poisson’s ratio

v d , and thermal expansion coefficient α d . The corresponding properties for the metal are Em ,

v m , and α m . Now consider the metal and the dielectric separately as a cylinder and an infinite

solid with a cylindrical hole. The metal is under a uniform, triaxial stress field, with a

longitudinal component σ L and two transverse components of equal magnitude σ T . The

transverse stress σ T also acts on the hole surface in the dielectric. The dielectric is taken to be

infinite, and the longitudinal stress in the dielectric is zero. Inside the dielectric, the stress field

decays away from the hole. According to the Lame solution, the radial stress σ r and the hoop

stress σ θ in an infinite elastic solid subject to a stress on a hole surface are given by

(Timoshenko and Goodier, 1970)

σ r (r ) = σ T (a / r) , σ θ (r) = −σ T (a / r ) ,
2 2
(105)

9/13/2002 72
where r is the radial distance. Consequently, the stress field in the composite is determined once

σ L and σ T are determined.

Next look at the strains. When the temperature changes by ∆T , the stresses change by

∆ σ T and ∆ σ L . The change in the longitudinal strain in the dielectric is caused by thermal

expansion:

∆ ε L = α d ∆T .
d
(106)

On the hole surface, the radial and the hoop stresses in the dielectric are σ r = σ T and σ θ = −σ T .

The change in the hoop strain is the sum of that caused by thermal expansion and that caused by

the stresses:

1 + νd
∆ εθ = α d ∆T − ∆σ T .
d
(107)
Ed

When we look at the strains in the metal, the analysis will depend on whether the metal

has yielded. First assume that the metal is elastic. The changes in the strains in the metal are

1 2ν
∆ ε L = α m ∆T + ∆ σ L − m ∆σ T ,
m
(108)
Em Em

νm (1 − ν m )
∆ εθm = α m ∆T − ∆σL + ∆σ T , (109)
Em Em

Because the metal and the dielectric are well bonded, so that the strains across the interface are

compatible: ∆ ε dL = ∆ ε mL and ∆ εθd = ∆ εθm . Solving two coupled linear algebraic equations, we

obtain that

∆σT = −
(1 + ν m )(α m − αd )∆T , (110)
(1+ ν m )(1 − 2ν m ) + (1 + ν d )
Em Ed

∆ σ L = −Em (α m − α d )∆T + 2ν m ∆σ T . (111)

When the metal and the dielectric have similar elastic constants, the stresses become

9/13/2002 73
 E
∆σT = − (α − α d )∆T , (112)
2(1 − ν ) m

E
∆σ L = − (α − α d )∆T . (113)
1− ν m

So far we have assumed that the metal is elastic. For representative values, Em = 1011 Pa ,

α m − α d = 20 × 10 −6 K −1 , and ∆T = 500K , the stress is about 1 GPa, which exceeds typical yield

strength of metals. We next allow the metal to yield plastically. Recall that the metal is under a

triaxial stress state (σ L , σ T ,σ T ) . It is an empirical fact that superimposing a hydrostatic stress

does not affect plastic deformation of a metal. Imagine that we superimpose a hydrostatic state

(−σ T ,−σ T ,−σ T ) to the metal, so that the metal is under a uniaxial stress state (σ L − σ T ,0,0 ).
We assume that metal is nonhardening, and the yield strength Y is independent of the

temperature. One can readily modify the model to allow the yield strength to vary the

temperature and strain, or even include the effect of creep. The metal and the dielectric are

perfectly bonded. Consequently, the metal yield when the difference in the longitudinal and the

transverse stress reaches the yield strength:

σ L − σT = Y . (114)

After yielding, when the temperature changes, the two stresses can still change, provided

they change by the same amount, ∆ σ L = ∆σ T . Let the plastic strain in the longitudinal direction

be ε p . Because plastic deformation conserves volume, the plastic strains in either of the two

transverse directions must be − ε p / 2 . When the temperature changes by ∆T , after yielding, the

strains in the metal change by

1− 2ν m
∆ ε L = α m ∆T + ∆ε + ∆σ T ,
m p
(115)
Em

9/13/2002 74
 1 p 1 − 2ν m
∆ εθ = α m ∆T − ∆ ε + ∆σT .
m
(116)
2 Em

The metal and the dielectric are bonded, so that ∆ ε dL = ∆ ε mL and ∆ εθd = ∆ εθm . Solving for the

transverse stress, we obtain that

3(α m − α d )∆T
∆σT = − . (117)
3(1 − 2ν m ) 2(1+ ν d )
+
Em Ed

Note that this expression can be written as ∆ σ T = −3B(α m − α d )∆T , with the effective elastic

modulus given by (96).

Figure 45 sketches the stress build up as the temperature drops from the stress-free

temperature T0 . When the temperature is above a certain level, Te , the metal is elastic, and the

stresses increases according to (110) and (111). The temperature Te is determined by inserting

the stresses (110) and (111) into the yield condition (114), giving

Em (α m − α d )(T0 − Te ) Ed (1+ ν m )(1 − 2ν m )

=1+ . (118)
Y Em (1 + ν d )

At Te , the transverse stress is

E d (1+ ν m )
σ T (Te ) = Y, (119)
E m (1 + ν d )

and the longitudinal stress is at a level Y above the transverse stress. As the temperature drops

below Te , the stresses continue to build up, now with a different slope given by (117).

Figure 46 sketches the longitudinal stress as a function of the temperature, assuming the

metal to be elastic, or liquid-like, or elastic-plastic. At a given temperature below T0 , the elastic

metal acquires a high stress, and the liquid-like metal acquires a low stress. The stress in the

elastic-plastic metal falls between the two limits. Because we assume that the yield strength Y is

9/13/2002 75
a constant, the two lines for the plastic and liquid-like metals are parallel. The longitudinal stress

in the elastic-plastic metal line at temperature T is

 Ed (1 + ν m )
σ L (T) = 3B(α m − α d )(T0 − T ) + 1+
 Em (1+ ν d )
Y. (120)

The first term is the hydrostatic tension in a liquid-like metal. The second term is due to the

presence of the yield strength. A representative value of the yield strength at 200°C is 150MPa,

which contributes significantly to the stress in the metal line.

5.4 Voiding in Blanket Copper Film

Shaw et al. (2001) and Sekiguchi et al. (2001) have observed voids in blanket copper

films. In the experiment of Shaw et al., sketched in Figure 47, the copper film was plated on

SiO2 dielectric, and capped with a silicon nitride layer. An annealing at 400°C closed all the

voids, and left the copper film under compression. The film was then rapidly cooled and held at

200°C for 2 hours; voids grew at the copper/nitride interface.

Recall that SIV has been widely reported for encapsulated narrow aluminum lines, but

not for blanket aluminum films. As illustrated in Fig. 48a, in a blanket metal film, creep can also

relieve the biaxial stress by dislocation motion and atomic diffusion. Consequently, voiding

(Fig. 48b) must be regarded as a stress-relieving process in competition with creep. For a void to

initiate from a small flaw, the stress must be high enough to compete with the surface energy.

The wafer curvature measurements have indicated that the stress in copper is typically higher

than in aluminum, due mainly to the larger yield strength in copper films (Baker et al. 2001,

Shen et al., 1998). Furthermore, the elastic stiffness for copper is more anisotropic than

aluminum. For a crystal of cubic symmetry, a single number characterizes the degree of elastic

anisotropy:

9/13/2002 76
 C12 + 2C44
ξ= . (121)
C11

For an isotropic elastic solid, ξ = 1. Using the elastic constants listed in Section 5.2, one finds

that ξ = 1.1 for aluminum, and ξ = 1.6 for copper. Finite element calculation has shown that the

stress is high at the intersection of copper grain boundary and SiN cap layer (X.H. Liu, of IBM,

private communication).

For the void to grow in size, atoms must diffuse from the void surface and plate on

copper grain boundaries. The observation of voiding implies that creep is a less effective process

to relieve the stress. This may be understood by comparing Figure 48a and b. Assume that

dislocations motions are constrained in the small volume. To relieve the stress in the copper

film, both creep and voiding require atoms to diffuse on the copper grain boundaries. In

addition, creep requires that atoms diffuse on the copper/silica or copper/nitride interface, but

voiding requires that atoms diffuse on the void surface. If the interface diffusion is slower than

the surface diffusion, and if the grain boundary diffusion is not too much faster than the interface

diffusion, then voiding will be a more effective means to relieve stress than creep. After a long

enough time, allowing atoms to diffuse on the interfaces, voids should disappear, driven by the

surface energy, much like in powder sintering. Models of grain boundary diffusion in elastic

solids (Thouless et al., 1996; Gao et al., 1999) can be adapted to model this phenomenon.

5.5 Slit-like Voids

As illustrated in Figure 37, voids in an interconnect can have different shapes. A slit-like

void significantly increases electric resistance with a small amount of mass transport, and

therefore may be responsible for earlier failure. Models of void shapes are briefly considered

here. Figure 49a illustrates a void of radius a in a crystal under remote stress σ . For simplicity,

9/13/2002 77
we will only consider the uniaxial stress state. Assume that atomic transport inside the solid is

much slower than on the void surface, so that the void can change shape as atoms relocate on its

surface, and the solid is elastic. The surface energy favors a rounded void, but the elastic energy

favors an elongated void. The stress at point A is higher than that at B, motivating atoms to

diffuse on the surface. Express the relative importance of the elastic and the surface energy

variations with a dimensionless number (Sun et al., 1994):

σ 2a
Λ= . (122)
Eγ

Under no stress, the void has a rounded shape maintained by the surface energy. When the

crystal is under stress, two behaviors are expected. If the stress is small ( Λ small), the void

reaches an equilibrium shape close to an ellipse. If the stress is large ( Λ larger), the void can

never reach equilibrium; noses emerge and sharpen into crack tips. The critical value occurs at

Λ c ≈ 1 . Taking σ = 1GPa , a = 100nm , E = 70GPa , and γ = 1J/m 2 , we find from (122) that

Λ = 1.4. The mechanism operates only under very large stress.

To place this mechanism into perspective, recall other mass transport mechanisms by

which a void evolves. At elevated temperatures, surface tension and stress may relocate mass by

diffusion on a surface and a grain boundary, and by diffusion and creep in a crystal lattice. First

consider diffusion on surfaces and boundaries (Figure 49b). A void on a grain boundary normal

to an applied tensile stress grows as atoms diffuse from the pore surface into the grain boundary

(Chuang et al. 1979). When atoms diffuse faster on the grain boundary than on the void surface,

the void cannot maintain the shape of spherical caps, and will become slit-like. Inserting one

atom to the grain boundary, the tensile stress does work Ωσ . The elastic energy variation per

atom relocation scales with Ωσ 2 / Y . When atoms are mobile on the grain boundary, and the

grains can drift relative to each other to accommodate the space, the elastic energy variation is

9/13/2002 78
much smaller than the work done by the stress associated with the insertion of atoms,

Ωσ 2 / Y «Ωσ . The effect of elasticity of the grains is negligible: the models assume rigid

grains. However, when the grain boundary is absent, or atoms are immobile on the grain

boundary, the tensile stress cannot do work Ωσ , and the elastic energy variation may become

important.

Next consider a single crystal under a uniaxial tension. A void in the crystal may change

shape for several reasons. Elastic distortion causes small changes; large changes require mass

transport. As illustrated in Fig. 49c, if the crystal creeps rapidly, the void elongates in the

longitudinal direction (e.g., Budiansky et al., 1982). By contrast, if atoms diffuse rapidly on the

void surface and the crystal is elastic, the void elongates in the transverse direction.

5.6 The Effects of Low k Dielectrics and Barriers

A Low k dielectric often has a small elastic modulus and a large thermal expansion

coefficient. For example, SiLK has Young’s modulus 2.45 GPa and thermal expansion

coefficient 66 × 10 −6 K -1 (Martin et al., 2000). Will the metal line be under compression on

cooling from the fabrication temperature? How small are the effective modulus B and the

magnitude of the stress? Does the thin barrier layer between the metal and the dielectric play

any role? Because different dielectrics will be used in coming years, it is important to have a

convenient method to answer the above questions quantitatively (Ma et al., 1995; Hau-Riege and

Thompson, 2000).

Figure 50 illustrates the cross section of several long metal lines embedded in a dielectric.

In calculation, one can set the thermal expansion coefficient of every material by its own value

minus α si . Because the Si substrate is bulky, the structure is under a state of plane strain

9/13/2002 79
deformation. The adjacent lines provide stiffness to each other. One can take a unit cell

containing a single metal line, and prescribe zero displacement on the cell walls. Similar to Fig.

47, remove the metal line from the structure, and leave behind a composite structure of silicon,

the dielectric, and the barrier layer. Apply a uniform normal traction σ on the hole surface, and

a temperature change ∆T . The structure is linear elastic, so that the change in the area of the

cross section is linear in the temperature change and the applied stress. Write

∫
1 σ
un ds = 2α eff ∆T − . (123)
A µeff

The two coefficients α eff and µeff can be computed by solving the elasticity boundary value

problem. The effective modulus B is still given by (96). The effective thermal expansion

mismatch is

3∆α = 3(α m − α Si ) − 2α eff . (124)

When the metal has no voids and is under a hydrostatic stress, the stress is given by

σ = 3B∆ α (T0 − T ). When the metal grows voids, and the stress in the metal is fully relieved, the

SVV is given by Vsv = 3∆α (T0 − T)V . When the dielectric and the barrier have thermal

expansion coefficient similar to silicon, α eff = 0 , and (124) recovers (92). When the dielectric

has a large thermal expansion coefficient, and the barrier is negligibly thin, α eff will dominate

(124). On cooling, the metal is under compression.

The boundary value problem such as that in Fig. 50 is easily solved by using the finite

element method. To appreciate the trend, consider an idealized structure similar to that in Fig.

44. The metal line has a circular cross section of radius a, surrounded by a barrier of thickness

hb , which in turn is surrounded by an infinite dielectric. This problem can be solved

9/13/2002 80
analytically. The effective modulus µeff relates to the shear modulus of the dielectric and the

barrier:

µeff = µ d + µb hb / a , (125)

The effective thermal expansion coefficient is

µ da
µ b hb
(1 + ν d )(αd − α Si ) + (1 + ν b )(α b − α Si )
α eff = µd a . (126)
+1
µ b hb

When (µ d / µb )(a / hb ) » 1, the dielectric prevails: µeff = µ d and α eff = (1 + ν d )(α d − α Si ). When

(µ d / µb )(a / hb ) « 1, the barrier prevails: µeff = µb hb / a and α eff = (1 + ν b )(α b − α Si ). A

representative set of magnitudes are µd / µb = 0.01, a / hb = 10 , (µ d / µb )(a / hb ) = 0.1. When the

barrier prevails, the saturation void volume is comparable to that in copper/silica structure.

Taking µb = 60GPa and a / hb = 10 , we obtain that µeff = 6.0GPa . For copper interconnect, the

barrier provides the effective modulus B = 5.7GPa . Even with a dielectric of negligible

stiffness, the barrier itself can provide significant constraint. This fact has significant

implications for electromigration resistance, as described in the next section.

6 ELECTROMIGRATION

In service, an interconnect line carries an intense electric current. The current motivates

metal atoms to diffuse in the direction of electron flow. The process, known as electromigration,

has been the most menacing and persistent threat to interconnect reliability. Before this threat

was recognized, electromigration had been one of the curious cross-field transport phenomena,

studied for its own sake (e.g., Huntington and Grone, 1961). For the last three decades, the

evolution of the electromigration research has been intertwined with that of the interconnect

9/13/2002 81
structure; see successive reviews coming out of the IBM T.J. Watson Research Center (d’Heurle

and Rosenberg, 1973; d’Heurle and Ho, 1978; Ho and Kwok, 1989; Hu et al., 1995; Rosenberg

et al., 2000).

In the recent decade, as the mechanical effect on electromigration is better formulated

(Korhonen et al., 1993), mechanics has become an integral part of the electromigration research.

Stress both originates from, and affects, electromigration. There is no fundamental difference

between the stress induced by thermal expansion mismatch and that induced by electromigration.

Stress may open voids in the metal, and break the encapsulation (Chiras and Clarke, 2000). Low

k dielectrics also have low stiffness, which significantly influences electromigration (Lee, et al,

2001; Hau-Riege and Thompson, 2000). As miniaturization continues, the diffusion time will be

shorter. At least some of the interconnects will have to be designed to be immortal, on the basis

of a stable state, rather than a transient process (Suo, 1998; Andleigh et al., 1999).

This section focuses on issues specific to the reliability of encapsulated interconnect

lines. At this writing, the design of the interconnect structure is in a state of flux. Materials,

architectures, and processes will change. Models and experiments based on first principles are

particularly valuable in guiding future development. This section draws on the previous ones.

Sections 2 and 3 come in when we consider cracking and debonding caused by electromigration.

Section 5 comes in when we consider voiding caused by electromigration. Section 5.6 is

particularly relevant when we consider the effects of low k dielectrics. Even though Section 4,

which is on ratcheting, dose not come in directly, the conscientious reader will soon feel the

pains shared by the two widely used accelerated tests: thermal cycling and electromigration.

Similarity will also be found in the models of the two phenomena.

9/13/2002 82
6.1 Electromigration in Encapsulated Metal Line

Figure 51 illustrates an aluminum line encapsulated in a silica dielectric. The tungsten

vias link aluminum lines between different levels. The titanium aluminide layers shunt the

electric current where voids deplete a segment of aluminum. Figure 52 sketches the stress

distribution along the metal line, assuming no voids. As the test temperature, the thermal

expansion mismatch induces a triaxial tension in the aluminum line, as discussed in Section 5.

At time t = 0, subject the line to a direct electric current. Electrons enter the aluminum line from

one tungsten via (cathode), and leave from another (anode). The electron flow causes aluminum

atoms to diffuse toward the anode. The atomic transport builds up tension near the cathode, and

compression near the anode. The stress gradient promotes atoms to diffuse back to the cathode.

After some time, the interconnect reaches a steady state: the backflow balances the

electromigration, and net atomic transport stops. The interconnect will last provided aluminum

does not grow voids at cathode, or extrude at the anode.

Next consider the case that the interconnect grows voids. Figure 53 summarizes the main

observations. Thermal expansion misfit causes a tensile stress in aluminum. Each void grows

and relaxes the stress in its vicinity (Figure 53a). After the line is subject to an electric current

(Figure 53b), the voids exhibit extraordinarily complex dynamics: they disappear, re-form, drift,

change shape, coalesce, and break up (Marieb et al., 1995). A particular behavior has captured

much attention. A void sometimes takes the shape of a slit, lying inside a single grain, severing

the metal line (Sanchez et al.,1992; Rose, 1992). Arzt et al. (1994) reported that a void is round

initially; it moves, grows, and then changes shape to become a slit. The void dynamics have

been simulated (e.g., Wang et al., 1996; Kraft et al., 1997; Zhang et al. 1999).

9/13/2002 83
 Despite the commotion in the transient, the end state is simple (Filippi, et al., 1995;

2002). After a sufficiently long time, the metal line reaches a steady state (Figure 53c). Only a

single void remains near the cathode; voids in the middle of the line have now been filled or

swept into the cathode void. The total number of aluminum atoms is constant in the line. As the

void at the cathode grows, atoms diffuse into the rest of the line, inducing a linear distribution of

pressure (Fig. 54). The pressure gradient balances the electron wind. The metal line behaves

like a column of a liquid in the gravitational field. The electron wind playing the role of the

gravity.

Imagine several identical lines, each carrying a different current. Figure 55 illustrates the

resistance change as a function of time. The resistance increases as the void enlarges. Two

long-term trends are expected: the resistance either saturates or runs away. Evidently, the larger

the current, the higher the saturation resistance. For a line carrying a large current, a large

pressure gradient is needed to stop mass diffusion. The high pressure in the metal near the

anode, in its turn, may cause cracking or debonding in the encapsulation materials (Chiras and

Clarke., 2000; Lee et al., 2002). Afterwards, the metal extrudes and the cathode void grows

further, so that the resistance increases indefinitely. Figure 56 and 57 show two examples of

extrusion. The interconnect is immortal if a circuit tolerates a variable resistance in an

interconnect up to its saturation level.

6.2 Basic Electromigration Characteristics

Electron Wind

9/13/2002 84
 When a direct electric current flows in a metal, the moving electrons impart momentum

to metal atoms. The electron wind force per atom, fe , is proportional to the electric current

density j. Write (Huntington and Grone, 1961)

fe = Z eρ j ,
*
(127)

where e is the electronic charge, and ρ the resistivity of the metal. Recall that ρj is the electric

field in the metal line, and eρ j has the dimension of force. The dimensionless number Z* is

known as the effective charge number. The sign convention makes Z* < 0 , because the electron

wind force is in the direction of the electron flow, opposite to that of the electric current density

j.

The electron wind force is small. When an atom moves by an atomic distance b , the

work done by the force, fe b , is much below the average thermal energy kT. Taking

representative values b = 10 −10 m , Z* = −2 , ρ = 3 × 10 −8 Ωm , j = 1010 A/m 2 , we find that

−8
fe b = 5 × 10 eV . The average thermal energy is kT = 0.04 eV at 500K. Recall that the energy

barrier for atomic motion in metals is on the order of 1 eV. Consequently, the electron wind

causes metal atoms to diffuse. The atomic flux is linearly proportional to the electron wind

force.

Atomic Diffusion

Einstein’s equation relates the atomic flux J (i.e., the number of atoms crossing unit area

per unit time) to the electron wind force:

DZ * eρ j
J= , (128)
ΩkT

9/13/2002 85
where D = D0 exp (−Q / kT ) is the diffusion coefficient, Q is the activation energy, and Ω is the

volume per atom in the metal. The equation is analogous to Equation (100) for diffusion

motivated by the stress gradient.

In a metal line, atoms may diffuse along several paths: the bulk crystal, the grain

boundaries, the interface between the metal and the encapsulation, and the interface between the

metal and the native oxide. Figure 58a sketches the cross-section of a wide metal line, width w

and thickness h. The metal line has columnar grains of diameter d. The effective Z* D , which

enters (128) to give the total atomic flux, results from all three kinds of paths (Hu et al., 1999):

1 1 1
Z* D = ZB* DB + ZG* DGδ G  1 −  + 2ZI* DI δ I  +  ,
d
(129)
d w w h

where the subscripts B, G, and I signify the bulk, the grain boundary, and the interface; δG and

δ I are the width of the grain boundary and the interface, respectively. The melting temperatures

of aluminum and copper are 933K and 1358 K, respectively. The temperatures at which

electromigration is of concern are below 500 K. For fine metal lines, bulk diffusion is

negligible. When a metal line is narrow, the grains form a bamboo structure, with all the grain

boundaries perpendicular to the flux direction (Fig. 58b). For such a bamboo line, the grain

boundaries do not contribute to the mass transport; the only remaining diffusion path is the

interfaces. As miniaturization continues, a large fraction of atoms will be on the interfaces, so

that the effective diffusivity will increase (Equation 129).

Median-Time-to-Failure (MTF)

The most widely used method to characterize electromigration is to measure the median

time-to-failure (MTF). Subject a batch of lines to a known current density and temperature. The

lines grow voids. When a void sever a line, the current stops and the line fails. Due to the

9/13/2002 86
statistical nature of void formation, failure times scatter even for a batch of carefully processed

lines. The MTF is taken as the time at which 50% of the lines have failed. In the interconnect

structure with shunt layers, voids never stop the current. A line is taken to have failed when the

resistance increases by, say, 20%.

MTF measurements provide a convenient way to evaluate effects of various parameters

on electromigration. For example, it has been shown that Al lines containing a few percent of

Cu last about 80 times longer than pure Al lines (Ames, et al., 1970), that intermetallic shunts

greatly increase electromigration resistance (Howard et al., 1978), that narrow Al lines of

bamboo grain structures last longer than polycrystalline wide Al films (Vaidya et al., 1980), and

that Cu lines last longer than Al lines (Rosenberg et al., 2000).

To fail lines within a time suitable for laboratory measurements, one has to accelerate the

test by using temperatures and electric current densities higher than those used in service. The

MFT can be measured for several temperatures and current densities. The results are fit to an

empirical relation (Black, 1969):

Q
MFT = Aj− n exp  , (130)
kT

where A, n and Q are fitting parameters. The exponent n typically ranges between 1 and 2. The

activation energy Q determined this way is consistent with those determined by other methods.

The use of MFT to project service lifetime raises at least two objections. First, the values

of the MFT are measured under accelerated conditions. Any conclusion drawn by extrapolation

to the service condition is suspect. Second, the interconnect structure has little redundancy. The

whole structure fails by the earliest failure of a small number of lines. Valid measurements of

the low-percentage-failure tail of the time-to-failure curve involve large numbers of samples

(Gall et al., 2001).

9/13/2002 87
Line Edge Drift

Another useful experiment involves line edge drift. In their original experiment, Blech

and Kinsbron (1975) deposited a gold film on a molybdenum film, which in turn was deposited

on an oxide-covered silicon wafer (Figure 59). The molybdenum served as the electrodes.

Because molybdenum has much higher resistance than gold, in the molybdenum/gold bilayer, the

electric current mainly flew in gold. Gold atoms diffused in the direction of the electron flow.

The edge of the gold film at the cathode drifted, and atoms piled up at anode edge. The edge of

the gold film drifted at a constant velocity at a given current and temperature.

The drift velocity v relates to the atomic flux as v = ΩJ . Using (128), one expresses the

drift velocity as

DZ * eρ j
v= . (131)
kT

Rewrite the above equation as

 vT   D Z* eρ  Q
ln  = ln  0
 j   k  kT
− . (132)

It is convenient to plot the experimentally measured ln (vT / j ) as a function of 1 / T . The data

points often fall on a straight line. The slope of the line determines the activation energy Q, and

the intercept determines the product D0 Z * . An additional independent experiment is needed to

separate D0 and Z* . We will discuss such experiments later.

Compared to the MTF measurements, the drift velocity measurements require the

observation of the edge displacement. In return, the drift velocity measurements provide the

values of D0 Z * . Furthermore, the drift velocity does not have the large statistical scatter

9/13/2002 88
associated with void formation. In principle, a single sample can be used to measure the drift

velocities at various temperatures. Such measurements are sufficient to determine Q and D0 Z * .

The drift experiment has been used to determine the dominant diffusion paths for

aluminum (Hu et al., 1993) and copper (Hu et al., 1999; Liniger et al., 2002). It has been used to

study alloying effects in copper lines (Lee, et al. 1995). The drift experiment led Blech (1976)

himself to discover a quantitative mechanical effect on electromigration. This landmark

discovery has since become a guiding light in the interconnect structure design and

electromigration research. The remainder of the section describes this discovery and its

consequences.

6.3 Balancing Electron Wind with Stress Gradient

On measuring the drift velocity in an aluminum line, Blech (1976) found that the line

edge drifted only when the current density was above a threshold value. He tested lines between

30 and 150 µm long. The threshold current density was inversely proportional to the line length.

The product of the line length L and the threshold current density jc was a constant,

jc L = 1.26 × 10 A/m . The jc L product increased when the aluminum lines were encapsulated in
5

silicon nitride. Blech attributed the observations to the pressure buildup near the anode. The

pressure buildup in aluminum was evident as it sometimes cracked the overlying SiN. The

pressure also caused a long aluminum extrusion through a hole etched in the SiN near the anode.

Blech used aluminum lines 25 µm wide and 0.46 µm thick, with columnar grains. Such

wide films could develop biaxial stresses, but not triaxial stresses. Atoms left the cathode region,

where tension arose. Atoms arrived in the anode region, where compression built up. At a point

in the line, atoms diffuse along the grain boundaries to equalize the two inplane stress

9/13/2002 89
components. The inplane stresses could not be readily relaxed because diffusion on the

aluminum-oxide interfaces is negligible. The stress gradient drove atoms to diffuse from the

anode (in compression) to cathode (in tension). The direction was opposite to the electron wind.

When the stress gradient balanced the electron wind, net atomic diffusion stopped. Incidentally,

Blech and Kinsbron (1975) did not report any threshold current density for gold lines.

Presumably diffusion on gold surface was so fast that the gold film could not build up significant

stresses during the electromigration experiment.

Denote the equal biaxial stress by σ , which is a function of the position x along the line

and time t. An atom in the metal line is subject to both the electron wind and the stress gradient.

The net force is the sum of (99) and (127), namely,

∂σ
f = Z eρ j + Ω
*
. (133)
∂x

The evolution of the stress field σ (x,t ) will be discussed later. We now focus on the steady

state. After a sufficiently long time, the net atomic diffusion stops, and the net driving force

vanishes, f = 0 . From (133), the line reaches the steady state when the stress in the line has a

constant slope, so that the stress needed to balance the electron wind force is

Ω∆ σ = Z * eρ jL, (134)

where ∆ σ is the stress difference between the two ends of the line. For a very small current, the

steady stress ∆ σ is small. Wide aluminum lines have finite yield strength, which has been

determined by the wafer curvature measurements (Nix, 1989). The lines drift when the stresses at

both ends exceed the yield strength. Take Ω = 1.66 × 10 −29 m 3 , Z* = −2 , ρ = 3 × 10 −8 Ωm ,

∆ σ = 10 Pa , and we find that jc L = 1.73 × 10 A/m .

8 5

The mechanical constraint effect on electromigration lifetime had been reported before

Blech’s work. For example, Ainslie et al. (1972) developed equation (134) to estimate the stress

9/13/2002 90
needed to stop electromigration. These authors used a value Z* = - 60 in their estimate, and

found an extremely large stress, which could not be sustained by the encapsulation. They

abandoned this interpretation, and instead tried to explain the mechanical effect in terms of the

effect of stress on diffusivity. The Blech experiment eliminated any diffusivity related

interpretation, because diffusion stopped when the current was below the threshold. Indeed, once

the threshold current is measured, together with the yield strength measured by the wafer

curvature method, Equation (134) determines the value of Z* , without being convoluted with the

diffusivity.

6.4 No-Voiding Condition for Encapsulated Line

We now return to Figure 52 for a quantitative discussion. On cooling from the stress-free

temperature T0 to a temperature T, the metal line and the encapsulation have a mismatch

volumetric strain 3∆α (T0 − T ). When the structure is held at T for a sufficiently long time, the

line will evolve into one of the two equilibrium states: either a uniform hydrostatic stress state

σ = 3∆α (T0 − T)B with no voids, or a stress-free state with a saturated void volume,

SVV = 3∆α (T0 − T )V . Here V is the volume of the metal line. Section 5.6 prescribes ways to

calculate the effective thermal expansion coefficient mismatch ∆ α and elastic modulus B . The

evolution is facilitated by atomic transport. At a finite time, the metal line is in a transient state.

The stress state is non-hydrostatic and nonuniform. Voids and unrelaxed metal coexist.

Of the two equilibrium states, the hydrostatic state is metastable, but the state with SVV

is stable. Consider the hydrostatic state first. In Figure 52, this hydrostatic stress field is labeled

as t = 0 . Section 5.2 prescribes a critical stress for void initiation, σ void = 2γ / a , where a is the

radius of the pre-existing flaw size, and γ the surface energy. The condition can be modified for

9/13/2002 91
flaws on interfaces, or junctions between the grain boundaries and the interfaces. It is assumed

that the annealing will close all voids provided the uniform hydrostatic stress is below the critical

stress, σ < σ void . For a copper/low k structure, taking ∆ α = 13 × 10 −6 K −1 , T0 − T = 200K ,

B = 6 × 10 Pa , γ = 1J/m and a = 5nm , we find that the hydrostatic stress is σ = 0.47 × 10 Pa ,

9 2 8

and the critical stress is σ void = 4 × 10 8 Pa . This comparison seems to suggest that the compliant

dielectrics reduce B, making the stress-induced voiding less likely.

The above void initiation condition is overly simplistic, and is risky. It assumes an

artificial scenario: the metal reaches the hydrostatic state first, and then the flaws start to

question themselves whether to enlarge or shrink. In reality, during annealing, the elastic

anisotropy between metal grains, as well as the elastic inhomogeneity between the metal and the

encapsulation, can momentarily cause very high stresses at various junctions. If atomic transport

is slow in relaxing the metal into the hydrostatic state, the high stresses may cause the flaws to

grow. The situation is worse than voiding in blanket films discussed in Section 5.4. A stress-

free, void-free blanket film is the stable equilibrium state. By contrast, in an encapsulated metal

line, the uniform hydrostatic state is only metastable; the voids, once initiated, will grow until the

entire line is stress-free. Experiments of Hau-Riege (2001) showed that voids formed in copper

lines even at very small electric current.

With caution, we pursue this risky model further, asking whether a superimposed electric

current will initiate voiding. After the electric current is applied, the metal line evolves to a

steady-state, in which the stress is linearly distributed. The number of atoms is conserved in the

metal line. If no voids form, atoms lost in the cathode region relocate to the anode region. By

symmetry, the middle of the line nether loses or gains atoms, and remains to have the stress

σ (L / 2,t ) = 3∆α (T0 − T )B at all time. The steady stress state as t → ∞ in the line is

9/13/2002 92
 Z *e ρj  L
Ω  2
σ (x,∞ ) = 3∆α (T0 − T)B − x − . (136)

In Figure 52, this steady state stress field is labeled as t = ∞ .

We assume that voids do not initiate if the steady stress at the cathode is below the

critical stress, σ (0,∞ ) < σ void , namely,

Z* eρ
3∆α (T0 − T )B + jL < σ void . (137)
2Ω

Figure 60 sketches this condition on the plane spanned by the temperature T and the product jL.

On this plane, Equation (137) is the straight line. Below the straight line, the current density is

low and the temperature is high, voids will not form. Above the straight line, the void will form.

6.5 Saturated Void Volume (SVV)

A conservative model assumes that the voids grow in the metal line (Fig. 53). Figure 54

sketches the stress evolution in such a metal line. Before the line carries the electric current,

voids may already grow in the line due to thermal stresses. This stress distribution is labeled as

t = 0 in Figure 54. After the line carries electric field for some time, all the voids along the line

disappear, and a single void is left in the cathode edge. The void relieves the stress at the

cathode, σ (0,t ) = 0; the capillary pressure is negligible. The stress at the anode becomes

compressive. When the stress gradient balances the electron wind, the stress is linearly

distributed along the line:

σ (x,∞ ) = − Z* eρ jx / Ω, (138)

This stress distribution is labeled as t = ∞ in Figure 54. The steady stress field is unaffected by

the temperature, to the extent that the temperature dependencies of the materials constants in

(138) are negligible. In contrast to the void-free steady state (136), the steady state with a

9/13/2002 93
cathode void is stable. No further change will occur in the line until the current density or the

temperature changes.

The void at the cathode end reaches the saturated volume. The thermal expansion

mismatch contributes 3∆α (T0 − T )V to the SVV. The steady stress field σ contributes

A ∫0 (σ / B)dx to the SVV, where A is the cross-sectional area of the metal line. Adding the two
L

contributions, the SVV is (Korhonen et al., 1993; Suo 1998)

Vsv Z *e ρ
= 3∆α (T0 − T) + jL. (139)
V 2ΩB

The SVV increases when the temperature decreases and the current density increases. The void

volume affects the resistance increase, as discussed in Section 5.2. Failure can be set to the

condition that Vsv / V reaches a critical value.

On the plane spanned by T and jL, Equation (139) looks the same as that shown on Fig.

60, but with the intercept replaced by T0 − (Vsv / V ) / (3∆α ). The slope of the line is proportional

to B. Everything else being equal, jc L increases with B. In Section 5.6, we estimated that when

the encapsulation changes form Ta/silica to Ta/SiLK, the effective modulus drops by about a

factor of 4. In a recent experiment, Lee et al. (2002) found that jc L = 3.7and1.2 × 105 A/m for

the two kinds structures.

When the temperature drops, the thermal stress increases, and the critical current density

decreases. The slope of the straight line in Figure 60 is 6ΩB∆α / Z* eρ . An estimate gives the

value of this slope to be 2 × 103 Am −1 /K . That is, decreasing the temperature by 100 K, one

expects a decrease of the jc L product by 2 × 105 Am −1 . This decrease may be of practical

concern, because the jc L is measured at an elevated temperature in an accelerated test, and is

9/13/2002 94
often reported without qualification. I am unaware of any experimental support of this trend.

Incidentally, in Blech’s (1976) drift experiment, jc L increased as the temperature dropped from

500 to 200 K. He used wide aluminum lines. The high values of jc L at low temperatures should

be due to the corresponding high yield strength.

In writing (139), we have ignored the volume change due to chemical reaction between

aluminum and titanium. Any volume expansion associated with such a reaction would reduce the

saturation void volume. Similarly, encapsulation materials lead to a small or even negative ∆ α

will reduce the SVV.

The copper interconnect structures in use today have very thin barrier layers. They are

ineffective as shunts: the resistance abruptly increases when a void at the cathode grows across

the via (Hau-Riege, 2001; Wang and Filippi, 2001). A more appropriate failure criterion should

be the void volume Vsv itself, rather than the volumetric strain Vsv / V , reaches a critical value.

Neglecting the thermal strain in (139), one finds that Vsv ∝ jL2 / B . Hau-Riege (2001) has

suggested that jL2 / B , instead of jL, be used as a figure of merit. The void-at-via problem can be

mitigated by methods to generate a void off the via, at a location uncritical to the interconnect

(Hu et al., 2001). Of course, it remains an option to provide a low-resistance shunt to the

interconnect structure.

6.6 Experimental Determination of Z , ∆ α (T − T0 ), B , and D

*

No reliable method exists to compute Z* from first principles; the parameter is measured

experimentally. Progress has been made in recent years to compute diffusivity D form first

principles; however, interconnects have multiple diffusion paths, and computations involving

interfaces and grain boundaries are unreliable. Although one can compute ∆ α and B as

9/13/2002 95
described in Section 5.6, it is desirable to measure such important parameters experimentally,

especially when input material parameters needed for computation are unreliable. As discussed

above, the drift experiment with unpassivated lines gives the product Z* D . Additional

independent experiments are needed to separate them, and to determine B and ∆ α .

Void in encapsulated metal lines can be observed in situ using high voltage scanning

electron microscopy (Marieb et al. 1995). When the void at the cathode is large enough, its

growth can be observed as the displacement of line edge at the cathode, similar to the original

Blech-Kinsbron (1975) drift velocity experiment. Figure 61 plots the SVV as a function of the

jL product, as predicted by Equation (139). If the saturation void volume is measured

experimentally at several current densities, the slope of the line in Figure 61 determines the ratio

Z* / B , and the intercept gives 3∆α (T0 − T ).

Figure 62 sketches the void volume as a function of time. After the line attains the steady

state, one can stop the electric current. The stress gradient in the line will drive atoms to diffuse

back, and the void will shrink. The same line can be tested at various current densities and

temperatures. Similar to the stress relaxation described in Section 5.2, the process of

approaching steady state is governed by a diffusion equation. Both the electron wind and the

pressure gradient drive the atomic flux:

D  * ∂σ 
ΩkT  ∂x 
J= Z e ρj + Ω . (140)

The diffusivity is taken to be an average in the aluminum line. Inserting (140) into (101) leads to

the diffusion equation (102) for the stress field σ (x,t ). The solution to this classical initial-

boundary value problem takes the form of an infinite series, the slowest decaying term being

proportional to exp (−t / τ ), with

9/13/2002 96
 L2 kT
τ= . (141)
π 2 DBΩ

Note that the time scale is independent of the current density. A large current density transports

mass at a high rate, but also needs to transport more mass to reach the steady state. An

experimental measurement of the relaxation time determines the product DB . In combination

with the Z* / B measured by the saturation volume, and Z* D by the edge drift of a line with

encapsulation of negligible stiffness, one can separate the three parameters Z* , D and B.

The number Z* can also be determined, on the basis of the steady stress distribution

(138), by measuring the stress in the metal line using synchrotron-based x-ray microdiffraction

(Wang et al., 1998), or by measuring the fringe stress field in silicon using Raman spectroscopy

(Ma et al., 1995).

6.7 No-Extrusion Condition for Encapsulated Line

In the steady-state with SVV, the compressive stress in the metal line at the anode is

σ A = −Z *e ρjL / Ω . (142)

Using L = 100 µm, j = 1.0 × 1010 A / m 2 , Z* = 2 , e = 1.6 × 10 −19 C , ρ = 3 ×10 −8 Ωm , and

Ω = 1.66 × 10 −29 m 3 , one finds that the steady-state compressive stress at the anode is σ A = -600

MPa. The large stress may cause cracking or debonding of the encapsulation. For example,

Filippi et al. (1995) tested 100 µm aluminum lines under various current densities, and reported

that aluminum did not extrude under 0. 56 × 1010 A / m 2 , but did under 1.0 × 1010 A / m 2 .

Section 2.2 has outlined the application of fracture mechanics to structures of small

feature size. The fabrication process controls the geometry all the way to the feature size, so that

9/13/2002 97
crack-like flaws must be smaller than the feature size. Use the linewidth, w, as a representative

size scale, the no-extrusion condition is stated as (Suo, 1998)

σ A2 w
β < Γ, (143)
E

Here Γ is the fracture energy of an encapsulation material or an interface, E Young’s modulus of

a material, and β a dimensionless parameter depending on ratios of various elastic moduli and

lengths that describe the anode. The values of β can be calculated by solving elasticity boundary

value problems containing cracks. Taking representative values β = 0.25 , Γ = 4J/m 2 ,

E = 70 GPa , and w = 0.5 µm, one finds from (2) that the encapsulation can sustain anode

pressure up to σ A = 1.5GPa .

In the above, thermal stresses in the passivation have been ignored. Upon cooling from

the processing temperature, a large compressive hoop stress arises in the oxide, resulting from

the thermal expansion misfits between the aluminum and the oxide, and between the silicon

substrate and the oxide. When an electric current is supplied, the volume of aluminum near the

anode increases, which first compensates its thermal contraction, and then goes beyond.

Consequently, in the stable state, the stress in the oxide is due to the pressure in the aluminum,

and the thermal misfit between the oxide and the silicon substrate. The latter is negligible.

Two other circumstances should be mentioned in this connection. Some dielectrics may

develop large intrinsic stresses at the deposition temperature. They are unrelated to the volume

change of aluminum, and persist under temperature change and current. A second circumstance

involves a line tested under a current up to the stable state, and then brought to a higher

temperature. Instantaneously the thermal expansion misfit between aluminum and oxide adds

9/13/2002 98
more tensile stress in the insulator. In such cases, the additional stresses must be included in the

no-extrusion condition.

Compliant low k dielectrics present a new challenge. As a first approximation, we

model the Ta barrier as a cylindrical pressure vessel of radius a and thickness hb , and neglect the

dielectric. The hoop stress in the barrier σ θ relates to the internal pressure σ A as σ θ = σ A a/ hb

Assume that crack-like flaw size is on the order of the barrier thickness hb . The no-extrusion

condition may be written as

σ A2 a 2
β < Γ. (144)
Ehb

The above estimates are intended to illustrate a basic modeling approach. Because the

structural details at line ends are complicated, accurate calculation is difficult. A more reliable

method is to carry out experiments with known current density, and use the electromigration

generated stress to test the condition for extrusion (Chiras and Clarke, 2000).

6.8 The Immortal Interconnect

Provided a circuit tolerates a variable resistance in an interconnect up to its saturation

level, the interconnect is immortal (Suo, 1998; Andleigh et al. 1999). The interconnect evolves

into a stable state to adapt to newly established temperature and current. The immortality gives a

simple perspective on reliability. One can focus on the stable state itself, rather than the transient

process to reach it. The perspective plays down the roles of the time scale, the rate processes,

and the microstructure of the metal. No longer need the microstructure be optimized for slow

mass transport or low void nucleation rate. No longer need the median-time-to-failure (MTF) be

measured, for the lifetime is infinite. No longer need a magic alloy be sought to slow down

9/13/2002 99
diffusion rate, for the diffusivity is irrelevant. In short, the reliability is warranted by energetics,

rather than kinetics. Having thermodynamics on our side, the design is much more robust.

The use of the stable state is limited in two ways: the metal may extrude at the anode, or

the saturation resistance may be prohibitively large. For a short line, the saturation void volume

is small compared to the via size, giving rise to a small resistance change. Consequently, the

stable state finds ready applications in short lines. The saturation void volume increases when

the effective modulus B is small (Equation 139). The interconnect structure with low k

dielectrics and thin barriers will no doubt require attention. The copper interconnect structures in

use now have very thin barrier layers. They are ineffective as shunts: the resistance abruptly

increases when a void at the cathode grows across the via. An immortal interconnect would

require low-resistance shunts.

As miniaturization continues, larger fraction of atoms will be on the metal/encapsulation

interface. The effective diffusivity will be larger (Equation 129), and the diffusion time shorter.

The day may arrive when some of the interconnects must be designed to be immortal.

Investigation into ways to limit the resistance increase and to prevent metal extrusion will

prepare us for such a day.

ACKNOWLEDGEMENTS

The work of my group in this area has been supported over the years by the National

Science Foundation, Office of Navel Research, Defense Advanced Research Projects Agency,

New Jersey Science and Technology Commission, Institute for Materials Research and

Engineering, and Intel Corporation. Dr. Z.Y. Huang has kindly helped me prepare the figures.

9/13/2002 100
REFERENCES

N.G. Ainslie, F.M. d’Heurle, O.C. Wells. Coating, mechanical constraints, and pressure effects

on electromigration. Appl. Phys. Lett. 1972, 20, 173-174.

P. Alpern, V. Wicher, R. Tilgner. A simple test chip to assess chip and package design in the

case of plastic assembling. IEEE Trans. Compon. Pack. and Manuf. Techno. A, 1994, 17,

583-589. Correction in 1995, 18, 862-863.

J.M. Ambrico and M.R. Begley, The role of initial flaw size and plasticity in adjacent layers in

channeling crack growth in thin films. Submitted for publication. 2002.

J.M. Ambrico, E.E. Jones and M.R. Begley, Cracking in thin multi-layers with finite width and

periodic architectures. Int. J. Solids Structures, 2002, 39, 1443-1462.

I. Ames, F. d’Heurle, R. Horstmann, Reduction of electromigration in aluminum films by copper

doping. IBM J. Res. Develop. 1970, 14, 461-463.

V.K. Andleigh, V.T. Srikar Y.J. Park, C.V. Thompson. Mechanism maps for electromigration-

induced failure of metal and alloy interconnects. J. Appl. Phys. 1999, 86, 6737-6745.

E. Arzt, O. Kraft, W.D. Nix, J.E. Sanchez. Electromigration failure by shape change of voids in

bamboo lines. J. Appl. Phys. 1994, 76, 1563-1571.

A. Bagchi, G.E. Lucus, Z. Suo, A.G. Evans. A new procedure for measuring the decohesion

energy for thin ductile films on substrate. J. Mater. Res. 1994, 9, 1734-1741.

S.P. Baker, A. Kretschmann A, E. Arzt. Thermomechanical behavior of different texture

components in Cu thin films. Acta Mater. 2001, 49, 2145-2160.

M.R. Begley and A.G. Evans, Progressive cracking of a multilayer system upon thermal cycling.

J. Appl. Mech., 2001, 68, 513-520.

9/13/2002 101
J.L. Beuth, Cracking of thin bonded films in residual tension, Int. J. Solids Structures, 1992, 29,

1657-1675.

J.L. Beuth and N.W. Klingbeil, Cracking of thin films bonded to elastic-plastic substrates. J.

Mech. Phys. Solids, 1996, 44, 1411-1428.

J.R. Black, Electromigartion—a brief survey and some recent results. IEEE Trans. Electron

Devices, 1969, ED-16, 338-347.

I.A. Blech, Electromigration in thin aluminum films on titanium nitride. J. Appl. Phys. 1976, 47,

1203-1208.

I.A. Blech, E. Kinsbron. Electromigration in thin gold films on molybdenum surfaces. Thin

Solid Films. 1975, 25, 327-334.

I.A. Blech, E.S. Meieran. Direct transmission electron microscope observation of

electrotransport in aluminum thin films. Appl Phys Lett. 1967, 11, 263-265.

J. Bree. Elastic-plastic behaviour of thin tubes subjected to internal pressure and intermittent

high-heat fluxes with application to fast-nuclear-reactor fuel elements. J. Strain Analysis,

1967, 2, 226-238.

B. Budiansky, J.W. Hutchinson, S. Slutsky. Void growth and collapse in viscous solids.

Mechanics of Solids, The Rodney Hill 60th Anniversary Volume, p. 13-45 (edited by H.G.

Hopkins and J. Sewell), Pergamon Press, Oxford (1982).

H. C. Cao, A.G. Evans. An experimental study of the fracture resistance of bimaterial interface.

Mechanics of Materials. 1989, 7, 295-305.

P.G. Charalambides, J. Lund, A.G. Evans, R.M. McMeeking. A test specimen for determining

the fracture resistance of bimaterial interfaces. J. Appl. Mech. 1989, 56, 77-82.

9/13/2002 102
S. Chiras, D.R. Clarke. Dielectric cracking produced by electromigration in microelectronic

interconnects. J. Appl. Phys. 2000, 88, 6302-6312.

T.-J. Chuang, K.I. Kagawa, J.R. Rice, L.B. Sills. Non-equilibrium models for diffusive cavitation

of grain interfaces. Acta Metall. 1979, 27, 265-284.

R.F. Cook and E.G. Liniger. Stress-corrosion cracking of low-dielectric-constant spin-on-glass

thin films. J. Electochem. Soc. 1999, 146, 4439-4448.

J. Curry, G. Fitzgibbon, Y. Guan, R. Muollo, G. Nelson, A. Thomas. New failure mechanisms

in sputtered aluminum-silicon films. 22rd Ann Proc. IEEE Int Reliability Phys. Symp.

1984, 6-8.

R.H. Dauskardt, M. Lane, Q. Ma, N. Krishna. Adhesion and debonding of multi-layer thin film

structures. Engineering Fracture Mechanics, 1998, 61, 141-162.

F.M. d’Heurle and P.S. Ho. Electromigration in thin films. pp. 243-303. In Thin Films:

Interdiffusion and Reactions. J. Poate, K.N. Tu, J. Mayer, eds. John Wiley, New York

(1978).

F.M. d’Heurle and R. Rosenberg. Electromigration in thin films. pp. 257-310. In Physics of

Thin Films, Vol. 7, G. Hass, M. Francombe, and R. Hoffman, eds. Academic Press, New

York (1973).

D.R. Edwards, K.G. Heinen, S.K. Groothuis, J.E. Martinez. Shear-stress evaluation of plastic

packages. IEEE Trans. Compon. Hybrids. Manuf. Technol. 1987, 12, 618-627.

F. Erdogan and G.C. Sih, On the crack extension in plates under plane loading and transverse

shear. Journal of Basic Engineering, Transactions of the ASME, Vol. 85, 519-527 (1963).

A.G. Evans, M.D. Drory, M.S. Hu, The cracking and decohesion of thin films. J. Mater. Res.

1988, 3, 1043-1049.

9/13/2002 103
A.G. Evans J.W. Hutchinson. The thermomechanical integrity of thin films and multilayers.

Acta Metall. Mater. 1995, 43, 2507-2530.

R.G. Filippi, G.A. Biery, and R.A. Wachnik, The electromigration short-length effect in Ti-

AlCu-Ti metallization with tungsten studs. J. Appl. Phys. 1995, 78, 3756-3768 (1995).

R.G. Filippi, R.A. Wachnik, C.-P. Eng, D. Chidambarrao, P.-C. Wang, J.F. White, M.A.

Korhonen, T.M. Shaw, R. Rosenberg, T.D. Sullivan. The effect of current density, stripe

length, stripe width, and temperature on resistance saturation during electromigration testing.

J. Appl. Phys. 2002, 91, 5787-5795.

M. Gall, C. Capasso, D. Jawaranl, R. Hemandez, H. Kawasaki, P.S. Ho. Statistical analysis of

early failures in electromigration. J. Appl. Phys. 2001, 90, 732-740.

H. Gao, L. Zhang, W.D. Nix, C.V. Thompson, E. Arzt. Crack-like grain-boundary diffusion

wedges in thin metal films. Acta Mater. 1999, 47, 2865-2878.

S.A. Gee, M.R. Johnson, K.L. Chen. A test chip design for detecting thin-film cracking in

integrated-circuits. IEEE Trans. Compon. Pack. and Manuf. Technol. B, 1995, 18, 478-484.

A. Gouldstone, Y.L. Shen, S. Suresh, C.V. Thompson. Evolution of stresses in passivated and

unpassivated metal interconnects. J. Mater. Res. 1998, 13, 1956-1966.

B. Greenebaum, A.I. Sauter, P.A. Flinn, W.D. Nix. Stress in metal lines under passivation -

comparison of experiment with finite-element calculations. Appl. Phys. Lett. 1991, 58, 1845-

1847.

C.K. Gurumurthy, J. Jiao, L.G. Norris, C.-Y. Hui, E.J. Kramer. A thermo-mechanical approach

for fatigue testing of polymer bimaterial interfaces. ASME Trans. J. Electron. Pack., 1998,

120, 372-378.

9/13/2002 104
J.B. Han, Flip-chip BGA design to avert die cracking. J. Electronic Packaging, 2001, 123, 58-

63.

S.P. Hau-Riege. Immortality of Cu damascene interconnects. pp. 21-32, in Sixth International

Workshop on Stress-Induced Phenomena in Metallization. S.P. Baker, M.A. Korhonen, E.

Arzt, P.S. Ho, eds. Am. Inst, Phys. (2001).

S.P. Hau-Riege and C.V. Thompson. The effects of the mechanical properties of the

confinement material on electromigration in metallic interconnects. 2000, 15, 1797-1802.

M.Y. He, A.G. Evans, J.W. Hutchinson. The ratcheting of compressed thermally grown thin

films on ductile substrates. Acta Mater. 2000, 48, 2593-2601.

R. Hill. The mathematical theory of plasticity, Clarendon Press, Oxford, 1950.

K. Hinode, I. Asano, T. Ishiba, Y. Homma. A study on stress-induced migration in aluminum

metallization based on direct stress measurements. J. Vac. Sci. Technol. B8, 1990, 495-498.

P.S. Ho and T. Kwok, Electromigration in metals. Rep. Progr. Phys. 1989, 52, 301-347.

M. Hommel, A.H. Fischer, A. v. Glasow, A.E. Zitzelsberger. Stress-induced voiding in

aluminum and copper interconnects. pp. 157-168, in Sixth International Workshop on

Stress-Induced Phenomena in Metallization. S.P. Baker, M.A. Korhonen, E. Arzt, P.S. Ho,

eds. Am. Inst, Phys. (2001).

J.K. Howard, J.F. White, P.S. Ho. Intermetallic compounds of Al and transitions metals: effect

of electromigration in 1—2- µ m-wide lines. J. Appl. Phys. 1978, 49, 4083-4093.

K.J. Hsia, Z. Suo, and W. Yang, Cleavage due to dislocation confinement in layered materials. J.

Mech. Phys. Solids., 1994, 42, 877-896.

C.-K. Hu, L. Gignac, S.G. Malhotra, R. Rosenberg. Mechanisms for very long electromigartion

lifetime in dual-damascene Cu interconnections. Appl. Phys. Lett. 2001, 78, 904-906.

9/13/2002 105
C.-K. Hu, K.P. Rodbell, T.D. Sullivan, K.Y. Lee, D.P. Bouldin. Electromigration and stress-

induced voiding in fine Al and Al-alloy thin-film lines. IBM J. Res. Develp. 1995, 39, 465-

497.

C.-K. Hu, R. Rosenberg, K.Y. Lee. Electromigration path in Cu thin-film lines. Appl. Phys.

Lett. 1999, 74, 2945-2947.

C.-K. Hu, M.B. Small, P.S. Ho. Electromigration in Al(Cu) two-level structures: Effect of Cu

and kinetics of damage formation. J. Appl. Phys. 1993, 74, 969-978.

M.S. Hu and A.G., Evans, The cracking and decohesion of thin films on ductile substrates. Acta

Metall., 1989, 37, 917-925.

M. Huang, Z. Suo, Q. Ma, and H. Fujimoto, Thin film cracking and ratcheting caused by

temperature cycling, J. Mater. Res., 2000, 15, 1239-1242.

M. Huang, Z. Suo and Q. Ma, Metal film crawling in interconnect structures caused by cyclic

temperatures. Acta Mater., 2001, 49, 3039-3049.

M. Huang, Z. Suo and Q. Ma, Ratcheting induced cracks in thin film structures. J. Mech. Phys.

Solids., 2002, 50, 1079-1098.

R. Huang, J.H. Prévost, Z.Y. Huang, Z. Suo. Channel-cracking of thin films with the extended

finite element method. Manuscript submitted for publication. 2003.

H.B. Huntington and A.R. Grone. Current-induced marker motion in gold wires. J. Phys. Chem.

Solids. 1961, 20, 76-87.

J.W. Hutchinson and A.G. Evans. Mechanics of materials: top-down approaches to fracture.

Acta Mater. 2000, 48, 125-135.

J.W. Hutchinson and Z. Suo, Mixed-mode cracking in layered materials. Advances in Applied

Mechanics, 1991, 29, 63-191.

9/13/2002 106
M. Isagawa, Y. Iwasaki, T. Sutoh. Deformation of Al metallization in plastic encapsulated

semiconductor devices caused by thermal shock. Proc. Int. Reliability Physics Symp., 1980,

171-177.

S. Jansson, F.A. Leckie. Mechanical-behavior of a continuous fiber-reinforced aluminum matrix

composite subjected to transverse and thermal loading. J. Mech. Phys. Solids, 1992, 40, 593-

612.

R.E. Jones, M.L. Basehore. Stress analysis of encapsulated fine-line aluminum interconnect.

Appl. Phys. Lett. 1987, 50, 725-727.

A.M. Karlsson, A.G. Evans. A numerical model for the cyclic instability of thermally grown

oxides in thermal barrier systems, Acta Mater. 2001, 49, 1793-1804.

R.R. Keller, J.A. Nucci, D.P. Field. Local textures and grain boundaries in voided copper

interconnects. J. Electronic Mater. 1997, 26, 996-1001.

S.-Y. Kook and R.H. Dauskardt. Moisture-assisted subcritical debonding of a polymer/metal

interface. J. Appl. Phys. 2002, 91, 1293-1303.

M.A. Korhonen, R.D. Black, C.-Y. Li. Stress relaxation of passivated aluminum line

metalizations on silicon substrate. J. Appl. Phys. 1991, 69, 1748-1755.

M.A. Korhonen, P. Borgesen, K.N. Tu, C.-Y. Li. Microsturcture based statistical model of

electromigration damage in confined line metallizations in the presence of thermally

induced stresses. J. Appl. Phys. 1993, 73, 3790-5004.

O. Kraft and E. Arzt. Electromigartion mechanisms in conductor lines: void shape changes and

slit-like failure. Acta Mater. 1997, 45, 1599-1611.

M. Lane, R.H. Dauskardt, N. Krishna, I. Hashim. Adhesion and reliability of copper

interconnects with Ta and TaN barrier layers. J. Mater. Res. 2000, 15, 203-211.

9/13/2002 107
M. Lane, R.H. Dauskardt, A. Vainchtein, H. Gao. Plasticity contributions to interface adhesion

in thin film interconnect structures. J. Mater. Res. 2000, 15, 2758-2769.

J. Lau, C.P. Wong, J.L. Prince, W. Nakayama. Electronic Packaging, Design, Materials,

Process, and Reliability, McGraw Hill, New York, 1998.

B. Lawn, Fracture of Brittle Solids. Cambridge University Press (1993).

K.-D. Lee, X. Lu, E.T. Ogawa, H. Matsuhashi, P.S. Ho. Electromigration study of Cu/low k

dual-damascene interconnects. 40th Ann Proc. IEEE Int Reliability Phys. Symp. 2002, 322-

326.

K.L. Lee, C.-K. Hu, K.N. Tu. In situ scanning electron microscope comparison studies on

electromigartion of Cu and Cu(Sn) alloys for advanced chip interconnects. J. Appl. Phys.

1995, 78, 4428-4437.

F.Z. Li, C.F. Shih, A. Needleman. A comparison of methods for calculating energy release rates.

Engineering Fracture Mechanics. 1985, 21, 405-421.

J. Liang, R. Huang, J.H. Prévost, and Z. Suo, Concomitant thin film cracking and underlayer

creep. Submitted to Experimental Mechanics. 2002a

J. Liang, R. Huang, J.H. Prévost, and Z. Suo, Evolving crack patterns in thin films with the

extended finite element method. Submitted to Int. J. Solids Structures. 2002b

K.M. Liechti, Y.S. Chai, Biaxial loading experiments for determining interfacial fracture-

toughness. J. Appl. Mech. 1991, 58, 680-687.

E. Liniger, L. Gignac, C.-K. Hu, S. Kaldor. In situ study of void growth kinetics in electroplated

Cu lines. J. Appl. Phys. 2002, 92, 1803-1810.

X.H. Liu, Z. Suo, Q. Ma, H. Fujimoto, Cracking and debonding in integrated circuit structures.

Engng Fracture Mech., 2000, 66, 387-402.

9/13/2002 108
T.C. Lu, J. Yang, Z. Suo, A.G. Evans, R. Hecht and R. Mehrabian. Matrix cracking in

intermetallic composites caused by thermal expansion mismatch. Acta Metall. Mater. 1991,

39, 1883-1890.

Q. Ma, S. Chiras, D.R. Clarke, and Z. Suo, High resolution determination of the stress in

individual interconnect lines and the variation due to electromigration. J. Appl. Phys. 1995,

78, 1614-1622.

Q. Ma, J. Xie, S. Chao, S. El-Mansy, R. McFadden, H. Fujimoto. Channel cracking technique

for toughness measurement of brittle dielectric thin films on silicon substrates. Mater. Res.

Soc. Symp. Proc. 516, 331-336 (1998).

S.X. Mao, A.G. Evans. The influence of blunting on crack growth at oxide/metal interfaces.

Acta Mater. 45, 1997 4263-4270.

T. Marieb, P. Flinn, J.C. Bravman, D. Gardner, M. Madden. Observations of electromigration

induced void nucleation and growth in polycrystalline and near-bamboo passivated Al lines.

J. Appl. Phys. 1995, 78, 1026-1032.

S.J. Martin , J.P. Godschalx, M.E. Mills, E.O. Shaffer, P.H., Townsend. Development of a low-

dielectric-constant polymer for the fabrication of integrated circuit interconnect. Advanced

Materials, 2000, 12, 1769-1778.

P.P.L. Matos, R.M. McMeeking, P.G. Charalambides, M.D. Drory. A method for calculating

stress intensities in bimaterial fracture. Int. J. Fracture. 1989, 40, 235-254.

J.W. McPherson and C.F. Dunn. A model for stress-induced notching and voiding in very large-

scale integrated Al-Si (1%) metalization. J. Vac. Sci. Tech. 1987, B5, 1321-1325.

S. Michaelides, S.K. Sitaraman. Die cracking and reliable die design for flip-chip assemblies.

IEEE Trans. Adv. Pack. 1999, 22(4), 602-613.

9/13/2002 109
N. Moes, J. Dolbow, T. Belytschko. A finite element method for crack growth without

remeshing. Int. J. Numerical Methods in Engineering. 1999, 46, 131-150.

M. Morgen, E.T. Ryan, J.-H. Zhao, C. Hu, T. Cho, P.S. Ho. Low dielectric constant materials

for advanced interconnects. JOM, 1999 September, 37-40.

A. Needleman , E. Van der Giessen. Discrete dislocation and continuum descriptions of plastic

flow. Mater. Sci. Eng. A309, 2001, 1-13.

W.D. Nix. Mechanical-properties of thin-films. Metall. Trans., 1989, 20A, 2217-2245.

E. T. Ogawa, J. W. McPherson, J. A. Rosal, K. J. Dickerson, T. - C. Chiu, L. Y. Tsung, M. K.

Jain, T. D. Bonifield, J. C. Ondrusek, and W. R. McKee. Stress-induced

voiding under vias connected to wide Cu metal leads. Proceedings of 2002 IEEE

International Reliability Physics Symposium.

T.S. Oh, R.M. Cannon, R.O. Ritchie, Subcritical crack-growth along ceramic-metal interfaces.

J. Am. Ceram. Soc. 1987, 70, C352-C355.

H. Okabayashi. Stress-induced void formation in metallization for integrated circuits. Mater.

Sci. Eng. 1993, R11, 191-241.

T.R. Reid, The Chip. Simon & Shuster, New York (1984).

I.E. Reimanis, B.J. Dalgleish, A.G. Evans. The fracture-resistance of a model metal ceramic

interface. Acta Metall. Mater. 1991, 39, 3133-3141.

J.R. Rice. Elastic fracture mechanics concepts for interfacial cracks. J. Appl. Mech. 1988, 55,

98-103.

J.E. Rim, F. Shi, D-I. Lee, F. Ohta and R.H. Dauskardt. Adhesion and debonding of low-k

carbon-doped oxide films in multi-layer thin film structures. J. Mater. Res. Submitted.

9/13/2002 110
J.H. Rose. Fatal electromigration voids in narrow aluminum-copper interconnect. Appl. Phys.

Lett. 1992, 61, 2170-2172.

R. Rosenberg, D.C. Edelstein, C.-K. Hu, K.P. Rodbell. Copper metallization for high

performance silicon technology. Annu. Rev. Mater. 2000, 30, 229-262.

J.E. Sanchez, L.T. McKnelly, J.W. Morris. Slit morphology of electromigration induced open

circuit failures in fine line conductors. J. Appl. Phys. 1992, 72, 3201-3203.

A.I. Sauter, W.D. Nix. A study of stress-driven diffusive growth of voids in encapsulated

interconnect lines. J. Mater. Res. 1992, 7, 1133-1143.

A. Sekiguchi, J. Koike, S. Kamiya, M. Saka, K. Maruyama. Void formation by thermal stress

concentration at twin interfaces in Cu thin films. Appl. Phys. Lett. 2001, 79, 1264-1266.

E.O. Shaffer, F.J. Mcgarry, L Hoang, Designing reliable polymer coatings. Polymer Eng. Sci.

1996, 36, 2375-2381.

T.M. Shaw, L. Gignac, X.-H. Liu, R.R. Rosenberg, E. Levine, P. Mclaughlin, P.-C. Wang, S.

Greco, G. Biery. Stress voiding in wide copper lines. pp. 177-183, in Sixth International

Workshop on Stress-Induced Phenomena in Metallization. S.P. Baker, M.A. Korhonen, E.

Arzt, P.S. Ho, eds. Am. Inst, Phys. (2001).

Y.L. Shen YL, S. Suresh, M.Y. He, A. Bagchi, O. Kienzle, M. Ruhle M, A.G. Evans. Stress

evolution in passivated thin films of Cu on silica substrates. J. Mater. Res. 1998, 13, 1928-

1937.

C.F. Shih, R.J. Asaro, N.P. O’Dowd. Elastic-plastic analysis of cracks on bimaterial interfaces .3.

large-scale yielding. J. Appl Mech. 1991, 58, 450-463.

9/13/2002 111
J.M. Snodgrass, D. Pantelidis, M.L. Jenkins, J.C. Bravman, R.H. Dauskardt. Subcritical

debonding of polymer/silica interfaces under monotonic and cyclic fatigue loading. J. Mater.

Res. Submitted.

E. Suhir, Accelerated life testing (ALT) in microelectronics and photonics: its role, attributes,

challenges, pitfalls, and interaction with qualification tests. J. Electronic Packaging. 2002,

124, 281-291.

T.D. Sullivan. Stress-induced voiding in microelectron metallization: void growth models and

refinements. Annu. Rev. Mater. Sci. 1996, 26, 333-364.

S. Suresh, Fatigue of Materials, 2nd edition. Cambridge University, Cambridge, 1998.

B. Sun, Z. Suo and A.G. Evans, Emergence of crack by mass transport in elastic crystals stressed

at high temperatures. J. Mech. Phys. Solids., 1994, 42, 1653-1677.

Z. Suo, Stable state of interconnect under temperature change and electric current. Acta

Materialia , 1998, 46, 3725-3732.

Z. Suo and J. He, unpublished research (2002).

Z. Suo and J. W. Hutchinson, Sandwich specimens for measuring interface crack toughness.

Materials Science and Engineering, A107, 135-143 (1989).

Z. Suo and J. W. Hutchinson, Interface crack between two elastic layers. International Journal

Fracture, 1990, 43, 1-18.

Z. Suo, C.F. Shih, and A.G. Varias, A theory for cleavage cracking in the presence of plastic

flow. Acta Metall. Mater. 1993, 41, 1551-1557.

H. Tada, P.C. Paris and G.R. Irwin, The Stress Analysis of Cracks Handbook, Del Research, St.

Louis, MO. (1985).

9/13/2002 112
R.E. Thomas. Stress-induced deformation of aluminum metallization in plastic molded

semiconductor-devices. IEEE Trans. Compon. Hybrids. Manuf. Technol. 1985, 8, 427-434.

M.D. Thouless. Cracking and delamination of coatings. J. Vac. Sci. Technol. 1991, A9 2510-

2515.

M.D. Thouless, A.G. Evans, M.F. Ashby, J.W. Hutchinson. The edge cracking and spalling of

brittle plates. Acta Met. 1987, 35, 1333-1341.

M.D. Thouless, H.H. Yu, Z.J. Zhao, W. Yang. Damage nucleation during electromigration along

an isolated interface in an elastic medium. J. Mech. Phys. Solids. 1996, 44, 371-387.

S.P. Timoshenko and J.N. Goodier. Theory of Elasticity. McGraw-Hill Book Company (1970).

T. Turner and K. Wendel, The influence of stress on aluminum conductor life. 23rd Ann Proc.

IEEE Int Reliability Phys. Symp. 1985, 142-147.

S. Vaidya, T.T. Sheng, and A.K. Sinha. Linewidth dependence of electromigration in

evaporated Al-0.5%Cu. Appl. Phys. Lett. 1980, 36, 464-466.

A.G. Varias, Z. Suo and C.F. Shih, Ductile failure of a constrained metal foil. Journal of the

Mechanics and Physics of Solids, 1991, 39, 963-986.

R.P. Vinci, S.P. Baker, S.P., editors, Mechanical Properties in Small Dimensions, MRS Bulletin,

2002, 27, 12-53.

A.A. Volinsky, N.R. Moody, W.W. Gerberich. Interfacial toughness measurements for thin

films on substrates. Acta Mater. 2002, 50, 441-466.

J.S. Wang and Z. Suo, Experimental determination of interfacial toughness curves using Brazil-

nut-sandwiches, Acta Met. 1990, 38, 1279-1290.

P.-C. Wang, G.S. Cargill, I.C. Noyan, C.-K. Hu. Electromigration-induced stress in aluminum

conductor lines measured by x-ray microdiffraction. Appl. Phys. Lett. 1998, 72, 1296-1298.

9/13/2002 113
P.-C. Wang, R.G. Filippi. Electromigration threshold in copper interconnects. Appl. Phys. Lett.

2001, 78, 3598-3600.

W. Wang, Z. Suo, and T.-H. Hao, A simulation of electromigration-induced transgranular slits. J.

Appl. Phys. 1996, 79, 2394-2403.

X. Yan, R.K. Agarwal. Two test specimens for determining the interfacial fracture toughness in

flip-chip assemblies. ASME Trans., J. Electron. Pack. 1998, 120, 150-155.

T. Ye, Z. Suo and A.G. Evans, Thin film cracking and the roles of substrate and interface, Int. J.

Solids Structures. 1992, 29, 2639-2648.

H.H. Yu, M.Y. He, J.W. Hutchinson. Edge effects in thin film delamination. Acta Mater. 2001,

49, 93-107.

J. Yue, W. Funsten, R. Tayler, Stress induced voids in aluminum interconnects during IC

processing. 23rd Ann Proc. IEEE Int Reliability Phys. Symp. 1985, 126-134.

Z.C. Xia, J.W. Hutchinson. Crack patterns in thin films. J. Mech. Phys. Solids 48, 2000, 1107-

1131.

Y.W. Zhang, A.F. Bower, L. Xia, C.F. Shih. Three dimensional finite element analysis of the

evolution of voids and thin films by strain and electromigration induced surface diffusion, J.

Mdch. Phys. Solids. 1999, 47, 173-199.

9/13/2002 114
 FIGURE CAPTIONS

Figure 1 A transmission electron microscope image of a cross-section of an interconnect

structure (Courtesy of John Mardinly, of Intel Corporation).

Figure 2 A plan view of a silicon nitride film of about 1 µm thick grown on a silicon

substrate. The tensile stress in the film causes channel cracks. The contrast

indicates that the film also partially debonds from the substrate. (Courtesy of Q.

Ma, of Intel Corporation).

Figure 3 A crack in a body subject to a distribution of load that opens the crack. Also

shown in the coordinate in a plane normal to the crack front.

Figure 4 The singular stress field is valid in the K-annulus. Inside the inner circle, the

atomistic bond-breaking process invalidates the linear elasticity theory. Outside

the outer circle, the boundary conditions invalidate the singular stress field.

Figure 5 A cylinder is embedded in a matrix. The cylinder is under compression. The

tensile hoop stress in the matrix causes cracking. The stress intensity factor is

plotted as a function of the crack length.

Figure 6 (a) When the crack is stationary in an elastic body, the load is linear in the

displacement. The area under the curve represents the elastic energy stored in the

body. (b) For the two copies of the elastic body, each containing a stationary

crack, the body with a larger crack area ( A2 ) has a larger compliance.

Figure 7 (a) A crack in a freestanding sheet. (b) a channel crack in a thin film bonded to a

substrate.

9/13/2002 115
Figure 8 A micrograph of cracked film surface. Initial cracks are introduced by a scratch.

When a bending load is applied, the cracks propagate in the direction normal to

the bending stress.

Figure 9 When the film is above a wide cavity, the crack driving force greatly increases.

Figure 10 A structure consists of an elastic overlayer, an elastic-plastic underlayer, and a

rigid substrate.

Figure 11 The driving force for channel crack as a function of the membrane force. Several

underlayer behaviors are compared: elastic, perfectly plastic, elastic-plastic,

elastic-perfectly-plastic, and power law.

Figure 12 Debond from an edge of the film, driven by residual tensile stress in the film.

When the debond front approaches the other edge of the film, the debond driving

force decreases, so that the debond arrests before it reaches the film edge.

(Courtesy of Q. Ma, of Intel Corporation).

Figure 13 A crack in a body subject to both the opening and shearing loads.

Figure 14 (a) The crack in a double-cantilever beam tends to curve away from the mid-

plane. (b) A crack slightly away from the mid-plane has a mode II component in

the direction that drives the crack further away from the mid-plane.

Figure 15 (a) Schematic of a thin film spalling from an adherent substrate, driven by the

tensile stress in the film. The crack runs along a path parallel to the interface. (b)

The effective axial force and bending moment caused by the residual stress.

Figure 16 Mud-crack pattern simulated by using the extended finite element method

(XFEM).

Figure 17 A schematic of Γ (ψ ) .

9/13/2002 116
Figure 18 A thin layer of material 2 is sandwiched between two substrates of material 1,

using a glue layer. A crack is on one of the interface. Near the crack tip, the

stress field for the interfacial crack prevails. Far away from the crack, the stress

field for a crack in homogeneous material prevails.

Figure 19 Four-point bend configuration. Thin films of interest are sandwiched between

two silicon substrates. On bending, crack runs on one of the interfaces.

Figure 20 The four-bend configuration is the superposition of two other configurations, one

being pure mode I, and the other being pure mode II.

Figure 21 A brittle film and a ductile film are sandwiched between two silicon wafers. The

crack runs on the interface between one silicon wafer and the brittle film. As the

crack tip passes by, a material particle in the ductile layer undergoes a history of

deformation: elastic (A), yielding (B), plastic straining (C), and elastic unloading

(D).

Figure 22 A crack on a metal-ceramic interface. The plastic deformation may be described

by continuum plasticity theory or by discrete dislocation dynamics. (a) Crack

grows by a bond breaking process. (b) Crack tip blunts, and voids open ahead of

the crack tip.

Figure 23 A thin metal film is sandwiched between two ceramic plates. The plastic

deformation in the metal is constrained, leading to high triaxial stresses several

film thicknesses ahead of the crack tip.

Figure 24 A very thin metal film is sandwiched between two ceramic plates. Dislocations

emitted from the crack tip pile up at the interface. As the applied load increases,

the stress at the blunted crack tip can be very high, leading to cleavage.

9/13/2002 117
Figure 25 (a) A photo of a flip-chip package. (b) A schematic of cross section of the

package. (c) Magnified view of the interconnect structure.

Figure 26 (a) The plan view of the surface of a die, near a die corner (Huang et al., 2000).

(b) Patterned aluminum pads. Cracks form in SiN over aluminum, but not in SiN

over silica (Huang et al., 2002). (c) A cross section shows a crack in SiN

(Courtesy of Dr. J.B. Han). (d) A cross section shows shifted aluminum pads

(Huang et al., 2001).

Figure 27 Temperature as a function of time. The cyclic temperature range is below the

curing temperature of epoxy.

Figure 28 At a corner of the die, the organic substrate applies a shear stress on the surface of

the die. The shear stress points to the die center. The thermal expansion misfit

between aluminum and silicon causes the aluminum pad to yield repeated during

temperature cycles. Directed by the biased shear stress, the aluminum pad shear

by a small amount each temperature cycle. After many cycles, the accumulated

shear significantly distort the SiN film, leading to cracking.

Figure 29 The evolving stress state in the SiN film and the aluminum pad. (a) Before

temperature cycling, the shear stress due to the organic substrate is transmitted to

the aluminum pad, and the structure does not distort much. (b) After many

temperature cycles, the shear stress in aluminum vanishes, and the structure

distorts greatly. (c) The shear stress from the organic substrate is balanced by the

membrane stress in SiN.

Figure 30 (a) An aluminum pad on silica surface. (b) and (c) The aluminum pads are

slotted. In each case, a SiN film covers both aluminum and silica.

9/13/2002 118
Figure 31 The normalized mximum priciple stress distribution, (σ max h p ) / (τ 0 L ) . The length

L is taken to be the width of the metal film. (a) a rectanglar aluminm pad; (b) a

slotted aluminum pad (Huang et al. 2002).

Figure 32 Plastic strain and stress change with the temperature during a thermal cycle of

blanket aluminum film on a silicon substrate.

Figure 33 (a) A blanket aluminum film on a silicon substrate. (b) The stress state in the

aluminum film consists of biaxial inplane stresses and a shear stress. (c) the yield

condition on the plane of (σ m , τ m ).

Figure 34 Various regimes on the plane spanned by the normalized shear stress in metal and

the normalized temperature range.

Figure 35 Shear stress as a function of the ratcheting strain rate.

Figure 36 Illustration of one-dimensional structure of passivation film on the metal film and

substrate. The insets show the different element and the stress state of the

passivation film, and the stress and flow in the metal film.

Figure 37 Voids grow in aluminum lines encapsulated in silica after some storage time.

Figure 38 An encapsulated aluminum line is under triaxial tension. The stress vanishes at a

flaw. The stress gradient motivates aluminum atoms to diffuse in the line,

simultaneously grow the void and relax the stress.

Figure 39 As the temperature drops from the stress-free temperature T0 , the misfit strain

increases, but the atomic diffusivity decreases. For a given time duration, the

void size maximizes at an intermediate temperature Tmax .

Figure 40 In the copper interconnect structure, vias are also made of copper, and are

susceptible to stress-induced voiding.

9/13/2002 119
Figure 41 (a) Void length smaller than tungsten stud diameter. (b) Void length larger than

tungsten stud diameter.

Figure 42 Separately consider the metal line as a cylinder and the dielectric as a solid with a

hole.

Figure 43 The free energy as a function of the void radius.

Figure 44 The metal is under a triaxial stress state, with the transverse stress different from

the longitudinal stress. The dielectric is subject to a radial stress at the surface.

Figure 45 The stresses build up as the temperature drops.

Figure 46 Compare stress in three cases: elastic metal, liquid-like metal, and elastic-plastic

metal.

Figure 47 When a blanket copper film under tension is held at an elevated temperature,

voids grow on the copper/nitride interface to relax the stress in the copper film.

Also sketched is the stress in the copper film as the temperature cycles.

Figure 48 Two processes to relieve stress in copper film: (a) creep, and (b) voiding.

Figure 49 A void changes its shape (a) in an elastic solid, (b) on a grain boundary, (c) in a

viscous solid.

Figure 50 An interconnect structure on a silicon substrate, showing the metal lines, the

barriers, and the dielectric.

Figure 51 A schematic of an interconnect, embedded in a silicon dioxide matrix, on a silicon

substrate.

Figure 52 Schematic of the stress field σ (x,t ) in an encapsulated metal line, assuming no

voids.

9/13/2002 120
Figure 53 (a) Upon cooling from the processing temperature, voids form to partially relax
thermal stress. (b) Subject to a current, atoms diffuse toward the anode, and a
void grows at the cathode end. (c) The line reaches the stable state, with a single
void left at the cathode end, a linear pressure distribution in the rest of the line,
and no further mass diffusion.

Figure 54 Schematic of the stress field σ (x,t ) in an encapsulated metal line, assuming that

voids grow and move in the line.

Figure 55 Resistance increases as a function of time. Under a small current, the resistance

saturates. Under a large current, the resistance runs away.

Figure 56 The aluminum line in the middle carries an electrical current in the direction

perpendicular to the paper. The pressure in the aluminum line causes the silicon

dioxide to crack, and aluminum to extrude (Courtesy of W.K. Meyer).

Figure 57 Extrusion from an encapsulated aluminum line (Courtesy of S. Chiras and D.R.

Clarke).

Figure 58 (a) The cross section of a wide metal line. (b) A narrow line with a bamboo grain

structure.

Figure 59 Schematic of the edge drift experiment.

Figure 60 The product jc L as a function of the temperature.

Figure 61 The normalized saturation void volume as a function of the jL product.

Figure 62 The nornamized void volume as a function of time, first approaches the saturation

under a current, and then relaxes to the thermal misfit volume after the current is

stopped.

9/13/2002 121