Electromigration and reliability in submicron metallization and multilevel interconnection

176

Materials Chemisg, and Physics, 33 (1993) 176-188

Invited

Review

Electromigration and reliability multilevel interconnection

in submicron

meta lliza tion and

Thomas Kwok IBM Research Division, Thomas J. Watson Research Center, Yorktown Heights, NY 10598 (USA)

(Received

November

23, 1992)

Abstract There are increasing reliability concerns of electromigration-induced and thermal stress-induced failures in submicron interconnects and in multilevel interconnection with W studs. The electromigration characteristics of Al and Al-01 submicron lines, two-level Al-Cu lines with W studs, Al fine lines under pulsed current stressing at high frequencies, and Al and Al-Cu fine lines under temperature cycling have been systematically studied. Lifetime is affected by grain size, grain morphology and bend structure in submicron metal lines. The lifetime of W stud chains is less than a half of that of AI-Cu flat lines. The discontinuity of the supply of Cu at Al-Cul W interfaces accounts for most of the reduction in the electromigration resistance of W stud chains. Under pulsed current stressing at frequencies 50-200 MHz, our data indicate no drastic change in lifetime within this frequency range. However, lifetime increases with duty cycle as t,,ar-‘.‘, which is a remarkable improvement over an average current density model. Lifetime also depends explicitly on both current-on time and currentoff period. The extra thermal stress induced by temperature cycling shortens the lifetime of both Al and Al-01 fine lines by more than an order of magnitude. Our results also show that the addition of Cu in Al fine lines improves the resistance to thermal stress-induced failures, probably by the suppression of grain boundary sliding and migration.

Introduction A number of comprehensive review articles on electromigration in metallic thin films have been published [l-4]. In VLSI circuits, the increase in wiring density requires the use of a multilevel interconnection. This multilevel interconnection consists of metal lines with different linelengths, linewidths and film thicknesses, and contains geometrical structures such as bends, studs and vias. It is necessary to study the effects of metal line geometry and bend structure on electromigration resistance in order to define the maximum current density allowable in interconnects. Moreover, the linewidth of interconnects is shrinking into the submicron range to improve circuitry density and speed performance. The high aspect ratio of studs and vias also gives rise to current crowding and local heating. With the projected high current density in submicron interconnects and current

crowding in these geometrical structures, there are increasing reliability concerns of electromigration-induced failures in VLSI multilevel interconnection. A number of early studies on linelength [5,6] and linewidth [5,7-103 dependence of lifetime in fine metal lines have been reported. This paper summarizes our recent studies [ll-143 on the dependence of lifetime on linelength, linewidth, film thickness and number of bends in submicron metal lines. A two-level A1-43 interconnection has been fabricated utilizing full oxide planarization [15]. In this structure, CVD W studs formed by W etchback, were used to serve as vertical connections for interlevel vias due to the ability of CVD W to fill high aspect ratio studs. The electromigration characteristics of these W stud chains differ significantly from those of simple large Al-Cu lines [ 161. These characteristics, in conjunction with the inhomogeneity in local grain structure, can cause

Elsevier Sequoia

flux divergences that lead to void-open or extrusionshort circuit failures. The projected high current density in submicron interconnects in VLSI technology will stretch the current-carrying capability of Al-based metallurgies to the limit of that projected by the electromigration model under continuous direct current (d.c.) stressing. However, most of these very large scale integration (VLSI) circuits will operate under pulsed current conditions at a frequency as high as 500 MHz in the next few years. Recently, we have systematically studied electromigrationinduced failures under pulsed d,c. stressing at frequencies of Xl-200 MHz using a specially designed pulsed current test system with an environmental chamber [17]. Notching and voiding in Al [18] and Al-Si 1193 lines were first observed in dynamic random access memory (DRAM) products and these reliability problems were later identified as stress-induced failures in fine metal lines [19-211. Since then, various hypotheses or models have been proposed to account for this failure mechanism [21-231 and its kinetics [24-271. This failure is due to excessive stress induced in fine metal lines because of the large thermal mismatch among metallic and dielectric materials and the Si substrate. When coupled with electrical current stressing, electromigration-induced void-open failures accelerated by the presence of thin film stress are of growing importance. This paper also summarizes the results of our recent study on void-open failures in Al and Al-Cu fine lines under three different lifetime test conditions: temperature cycling with and without electrical current stressing, and electrical current stressing at constant temperatures [28].

Experimental To quantitatively evaluate these electromigration characteristics, several test sites were designed. They consist of single-level straight lines with different linelengths and linewidths, meandering lines with bends, and two-level fine lines with stud chains. As shown in Fig. l(a), each test site consists of three parallel metal lines. The central line is used as a test stripe connected to four terminals for electrical current passage and voltage measurements. The two outer lines are used as sensory stripes connected to terminals for extrusion-short monitors. In order to avoid the current crowding effect at the junction between the narrow test stripe and the large bond pad, an end contact segment of gradually decreasing width is used to

Fig. 1. Metal line test sites for lifetime measurements with voidopen and extrusion-short monitors: (a) straight line; (b) fine line with 32 bends.

Fig. 2. An optical micrograph of the end contact segment structure linking the test stripe to the bond pad in the metal line test site.

link the test stripe to the bond pad as shown in the optical micrograph of Fig. 2. Fine line test site patterns were generated by optical lithography while submicron line test site patterns were generated by electron beam lithography. Samples were fabricated using the lift-off technique. The substrate used was p-type Si wafer coated with 100 nm thermally grown SiOz, 100 nm low pressure chemical vapor deposited S&N, and 200 nm low temperature deposited SiO,. Different thicknesses of Al or Al-4wt.%Cu films were deposited on top of these dielectric layers by evaporation from RF induction of Al or Al-Cu alloy sources, respectively. The base pressure was between 3-7 x lo--’ torr. The deposition rate was - 1.5 nm s-l at a substrate temperature of 35-40 “C. Patterned wafers without any passivation were annealed at 400 “C in forming gas ambient for periods of 7.5, 15, 30 and 60 minutes. Some patterned wafers were passivated with 6.5 pm sputtered quartz and then annealed at 400 “C in forming gas ambient for an hour. Transmission electron microscopy (TEM) was used to evaluate the microstructure of these samples.

178

A TEM specimen was prepared by mechanically grinding the sample from the back side, followed by wet chemical etching and argon ion milling. The accelerated lifetime test method is the only reliable and the most commonly used method for evaluating electromigration resistance of metal lines. The time-to-failure data usually follow a lognormal distribution, and the result is given in terms of a median time-to-failure (MTTF), or tSo, which is the time to reach the failure for 50% of a group of identical metal lines. In order to obtain the result of lifetime measurements in a reasonable time frame, the lifetime test is carried out under a set of accelerated test conditions. For electromigration-induced failures, the accelerated test condition would be a high electrical current density stressing level and an elevated temperature. The data are then extrapolated to the device operating condition by an ~rhenius-like empirical equation 129] tso =Aj --R exp(E,/kT)

(1)

whereA is a material constant,j the current density, n the current exponent, E, the activation energy for electromigration-induced failure, k the Boltzmann constant and T the absolute temperature. In our accelerated lifetime test with constant or pulsed current stressing, data on current and voltage measurements of each test stripe were taken every 15 to 30 minutes until the test stripe failed by either void-open or extrusion-short. For thermal stress-induced failures, the accelerated test condition could be a wide temperature cycle range with fast heating and cooling rates. A current of 0.01 MA cm-* was passed through each test stripe for less than a second for every 30 minutes to monitor the test stripe failure. This short-pulsed electrical current of low current density should not have imposed any electromigration-induced damage on the test stripe. Eight to twelve samples were used in each lifetime test conditions. Results and discussion A. Eflects of grain size on electromi~~tio~ lifetime The three-dimensional shape of a grain in a

polycrystalline metal film is rather complex and it is very difficult to define the grain size precisely. In the case where the grain size is constrained by the thickness and/or the linewidth of a metal strip, the grain diameters along and across the strip, and across the metal film are quite different. The average grain size reported in this paper is determined by using the linear intercept method

along and across the strip in a planar section of the metal strip. The microstructure in 2.0 pm wide Al-Cu lines after various heat treatments is shown in Fig. 3. The co~esponding average grain size and electromigration lifetime are plotted in Fig. 4. As revealed in Fig. 4, normal grain growth is observed when the grain size is much smaller than the metal line thickness. This implies that segregation of Cu in grain boundaries, Al&u second phase particles, preferred orientation of Al grains and grain-boundary grooves have little effect in reducing the grain growth-rate in this initial period of grain growth. However, grain growth slows down very rapidly when the grain size approaches the metal line thickness. Thus, an upper limit on the grain size after annealing is determined by the metal line thickness. The electromigration lifetime measurements for these annealed samples are carried out at an accelerated test condition with a current density of 5.0 MA cm-’ and temperature at 127 “C. Figure 4 also shows that the electromigration lifetime increases with increasing grain size. B. Dependence of electromigration lifetime on metal line geomehy and grain rno~~olo~

The lifetimes of Al-Cu submicron lines, 10-500 pm in length, have been measured under a current stressing of 2.5 MA cm-’ at 227 “C [13]. These Al-Cu lines are 0.75 pm wide and 0.54 pm thick. By best fitting the log-normal distribution to the time-to-failure data using the least square method, the estimated values of the median time-to-failure, tso, and the standard deviation of the log-normal failure distribution, o, are plotted in Fig. 5. The lifetime decreases with increasing linelength from 10-50 pm and then levels off between 50-500 pm. If we assume that defects of constant density and varying severities are distributed randomly in the metal line, and that lifetime is determined by the most severe defect in the metal line, then a decrease of lifetime with increasing linelength is expected because the probability of finding more severe defects is higher in a longer metal line. The insensitive of lifetime to linelength for the metal lines of length greater than 50 pm can be understood by assuming that electromigration-induced failure is caused by any defects beyond a certain severity, and that there is an abundance of such defects in these metal lines. The standard deviation of the log-normal failure distribution does not vary systematically with linelength. The lifetimes in Al and Al-Ku fine lines of linewidth ranging from 0.5 to 2.0 pm and film

179

0.4 pm

(e)

(d)

(c>

(b)

(a)

Fig. 3. Transmission electron micrographs of 2.0 pm wide Al-Cu lines annealed (b) 7.5 minutes; (c) 15.0 minutes; (d) 30.0 minutes; (e) 60.0 minutes.

- 25

0.B -

0 - lifetime T = 127’ C J=5XlO’A/cm’

0.2 -

0.0

at 400 “C for various time periods: (a) as deposited;

t

’

10

0

I

t

I

I

20

30

40

50

-

t *-

54

’

Y

metal line thickness 0

0 - 0.35 pm Al

60

A -

Annealing lime (1) at 400°C, in minutes

n

Fig. 4. Grain size and electromigration lifetime of 2.0 pm wide Al-Cu lines annealed at 400 “C for various time periods.

11 0.0

0.75pm

Al

- 0.90/un Al

‘.,

\

‘I ‘.

\.

-I

l. ..-.._._._ ‘*._.-

---~------_-__;-_~~ ____._._._._-.-.-tI

o - 0.5 /u-n Al-h + - O.BS pm AI-Cu

I

I

I

0.5

1.0

1.5

I 2.0

Metol Line Linewidth in pm 251”

1.4

E

1 = 227’C J = 2.5 X 10' A/cm'

. 0'

0

1

100

Fig. 5. Dependence lines.

1

I

I

200 300 400 Metal line length in pm

oft,, and (eon linelength

,

500

in AI-Cu submicron

thickness from 0.35 to 0.9 pm have also been measured under a current stressing of 1.0 MA cm-’ at 182 “C [ll]. The median time-to-failure, cSO,is plotted in Fig. 6. As the linewidth is decreased, the lifetime initially decreases and then decreases beneath a critical width. The critical linewidth also varies significantly with the thickness of the metal

Fig. 6. Dependence of electromigration lifetime on linewidth and film thickness in AI-Cu submicron lines.

lines and lies between 0.625 and 2.0 pm for film thickness ranging from 0.35 to 0.9 pm. Figure 6 also indicates a significant improvement in the electromigration resistance of Al-& over Al metallization. We have also studied the microstructure in these fine metal lines [12, 301. The microstructure in Al-Cu fine lines of 0.5 pm film thickness is shown in Fig. 7. Only one grain was observed across both the stripe and the film in the 0.5 pm line. The grains were rectangular prisms in shape and their structure resembled bamboo as revealed in Fig. 7(a). Grain boundaries near the free surface tend to lie perpendicular to the surface. A large spread in the grain size distribution is observed in AlCu lines with linewidth down to 0.75 pm. An abrupt change in grain size is also observed in 2.0 pm wide lines. Their grain size distributions are also rather narrow. X-ray diffraction data also

180

(4

(b)

(~1

(d)

(4

Fig. 7. Transmission electron micrographs of AI-Cu fine lines with 0.5 pm film thickness after annealing at 400 “C for an hour: (a) linewidth 0.5 pm; (b) linewidth 0.625 pm; (c) linewidth 0.75 pm; (d) linewidth 1.0 pm; (e) linewidth 2.0 pm.

indicate a strong (111) preferred orientation of the Al grains. By tilting the specimen relative to the incident electron beam direction, only one to two grains across the film were observed in the submicron lines while a few grains across the film were observed in the 2.0 pm line. This implied that the columnar and bamboo grain structures could be enhanced in the annealed submicron lines. In general, for Al and Al-Cu thin films of thickness less than 1.0 pm, the grain size is constrained by the film thickness as the film thickness is comparable to the grain size in bulk materials. So a columnar grain structure was generally observed in these metallic thin films. In Al and AlCu submicron lines, the ultimate grain size and the final grain structure are determined by the linewidth in addition to the film thickness as both the linewidth and film thickness are smaller than the grain size. Thus, a bamboo grain structure was observed in those Al and Al-Cu submicron lines with a linewidth comparable to or smaller than the film thickness. These submicron metal lines were also found to have longer lifetimes than other metal lines. These results are summarized and illustrated in Fig. 8. Since the electromigration-induced open-circuit failure in the metal line is along its width, an increase of lifetime with increasing linewidth is expected because the probability of aligning failurecausing defects across the width is lower for a wider line. Moreover, when the linewidth is larger than the average grain size, the lifetime is also expected to increase with the linewidth since it becomes more difficult for a crack to propagate across a wide line. As the linewidth becomes comparable to or smaller than the grain size, the fine metal line assumes a more or less bamboo

’

0.0 0.0

.’

I

I

I

J

0.5

1.0

1.5

2.0

Linewidth

(0 ,a -

columnar

,o -

structure,

l

submicron

lines)

in pm

grain structure, Al subinicron

*,A

lines,

A,

bamboo

A-

grain

AI-CU

Fig. 8. Effects of linewidth and film thickness on the microstructure in Al-& submicron lines.

grain structure and the lifetime is again expected to increase. The effects of linewidth and film thickness on lifetime can thus be correlated with grain size and grain morphology in these metal lines. There are several structural factors contributing to the improvement of electromigration resistance in fine metal lines, especially those of submicron width and thickness. When the grain size is larger or comparable to the metal line width and thickness, the number of grain-boundary paths for diffusion along the metal line is small, thus the mass transport through the metal line crosssection is reduced. With large grain size and/or uniform grain structure in fine metal lines, the number of grain-boundary triple points is also small. The electromigration-induced atomic flux is minimized in the bamboo grain structure since most of the grain boundaries are perpendicular to the direction of current flow. C. Effect of bend structure on electromigration and reliability Electromigration lifetime measurements on AlCu submicron bend lines were carried out under a current stressing of 2.0 MA cm-* at 162 “C [14]. These Al-Cu lines are 0.75 pm wide, 0.54 pm thick and 960 ,um long. As shown in Fig. l(b), each Al-Cu submicron line contains 12 to 96 horizontal bends with a separation between adjacent bends ranging from 80 to 10 pm. Measured values of t5,, and u are plotted in Fig. 9. Lifetime decreases linearly with an increasing number of bends at a rate of 0.4% per bend. The standard

181

a 0

/ 20

/ 40

1 60

/ 80

0.6 v) 100

Number of bends

Fig. 9. Dependence of tSO and v on the number Al-Cu submicron lines.

of bends

in

(4

in Fig. 10, there is no dramatic change in the grain structure from the straight line portion to the bend structure in these Al-Cu submicron lines [14]. Thus, the microstructure in the bend structure is believed to have little effect on the electromigration resistance of metal lines. Current density and temperature distributions in bend, stud and via structures have been calculated 131, 321. In these calculations, a finite element method [33] is used to obtain numerical solutions to differential equations governing the current and heat flows subject to appropriate boundary conditions. The cross-sectional view of a 1.75 pm wide and 1.75 pm long bend structure with 0.75 pm linewidth is shown in Fig. 11. As revealed in Fig. 12, there are two current density peaks located in the lower left and upper right bend corners of this bend structure. These peaks are of the same magnitude 2.2/J& where PO/ is the average current density in the metal line. They come mainly from the x-component of pool. The temperature distribution in this bend structure is illustrated by the contour in Fig. 13. The center region of the metal line has a small temperature increase of 2.0 K above the ambient temperature at the substrate end contact to chip carrier. Because of the larger cross-sectional area and thus lower average current density in this bend structure, the temperature increase in the bend structure is found to be 1.1 K, only about half of that in the center region of metal lines. Figure 13 also indicates that

J!+Ti

Fig. 11. A 1.75 pm wide and 1.75 pm long bend structure 0.75 pm tinewidth.

with

(b) Fig. 10. Transmission electron micrographs of 0.54 pm thick AlCu submicron lines with bend structures after annealing at 400 “C for an hour: (a) linewidth 0.5 pm; (b) linewidth 0.625 ,um.

deviation of the log-normal failure distribution increases with an increasing number of bends. The microstructure in the 0.5 pm and 0.625 pm wide Al-Cu lines of 0.5 pm film thickness are shown in Figs. 10(a) and IO(b), respectively. As revealed

Fig. 12. Current density distributions in Fig. 11.

in the bend structure

shown

103

high

I

.

E .E

2 102

\

e

.

s P

B b

T = 250” C \ Linewidth = 2.0 pm

.E 10’

Fig. 13. Temperature Fig. 11.

contour

in the bend structure

shown in

._ ;

loo

.

I

6

’

0

g,oo

2

3

,Il,l

‘10

Current density in MA/cm*

Fig. 15. Dependence chains.

of lifetime

D. Electromigration resistance in a two-level Al-& interconnection with W studs Electromigration lifetimes in two-level Al-Cu lines of 2.0 pm linewidth with W stud chains have been measured at 250 “C and with current stressing levels at 1, 2 and 4 MA cm-’ [16]. Each W stud chain contains 100 square studs of 1.8 pm width with 6.0 pm link length as shown in Fig. 14(a). Measured values of rso in W stud chains for these current densities are plotted on Fig. 15. Most electromigration-induced failures in W stud chains are void-opens occurring on second level Al-Cu lines at areas near Al-Cu/W interfaces. The lifetime increases with decreasing current density with a current exponent equal to 2.8. However, the average value of the current exponent for Al-Cu

density in W stud

T = 250’ C J = 1.0 M/cm=

Fig. 14. Electromigration test sites: (a) two-level AI-Cu fine lines with W stud chains; (b) Al-Cu meandering lines of rectangular bend shape.

the current density peaks in both the lower left and upper right bend corners do not cause any local heating problem. Indeed, they have very little effect in the overall temperature distribution in the bend structure due to their small magnitude. The decrease of electromigration lifetime with bend structure is probably due to the high current density peak and the associated flux divergence at the bend structure.

on current

10’ ’ 0.0

I 0.5

I

I

I

1 .o 1.5 2.0 Metal line linewidth in pm

I

I

2.5

3.0

Fig. 16. Effect of metal line linewidth on electromigration in W stud chains.

lifetime

flat lines is around 2.0 [34]. The higher value for the current exponent in W stud chains compared to Al-Cu flat lines is probably due to the peak current density increasing faster than the average current density. The electromigration lifetime test was also carried out on W stud chains of 1.0 pm linewidth under a current stressing of 1 MA cm-’ at 250 “C. Each W stud chain contains 200 square studs of 0.8 pm width with 3.0 pm link length. Measured values of tso in W stud chains for two different linewidths are plotted in Fig. 16. The lifetime increases by a factor of about 5 with increasing linewidth from 1.0 to 2.0 pm. Because the probability of aligning failure-causing defects across a wide line is lower than for a narrow line and it is also more difficult for a crack to propagate across a wide line, an increase of lifetime with increasing linewidth is expected. However, an increase in the range 35-85% was measured for Al-Cu straight lines [ll]. Current density peaks

183

at studs versus linewidth, based on our finite element calculations, are plotted in Fig. 17. The increase in current crowding at studs with decreasing linewidth accounts for the stronger linewidth dependence of lifetime in W stud chains than in Al-Cu flat lines. Under an accelerated current stressing of 4 MA lifetimes in singlecm-’ at 250 “C, electromigration level Al-Cu meandering lines of rectangular bend shapes have been measured. As shown in Fig. 14(b), the linewidth, link length and height of Al-Cu meandering lines are 1.0, 3.0 and 1.2 pm, respectively. Measured values of & in Al-Cu straight and meandering lines and W stud chains at 250 “C with 4 MA cm-’ are plotted in Fig. 18 for comparison. Because the linewidth is 1.0 pm for Al-Cu meandering lines and 2.0 pm for both Al-Cu straight lines and W stud chains, the projected value of tsO in Al-Cu meandering lines of 2.0 pm linewidth is estimated to be about three times longer [14]. Moreover, both Al-Cu straight and meandering lines are passivated by the first insulator layer while there is no second insulator layer to passivate second level metal lines of W

stud chains. In general, the lifetime of Al-Cu flat lines increases by an order of magnitude or more with a complete surface coverage [35]. Thus, the projected value of tsO in W stud chains with passivation is estimated to be about ten times its measured value. When an electric current is passed through a two-level Al interconnection with W studs at the device operating temperature, which is well below a half of the Al or W melting point, the resulting atomic flux J comes primarily from electromigration at grain boundaries [36]. For Al lines or W studs, J=

0.0

0.5

1 .o Stud

Fig.

17. Current

,

200

r.? AZ 2 .,5,,_~ . I

density

B

?? 1 ‘2 100

(W)

peak

I

.I

meandering

-

passivated

b

-

unpassivated

J = 4.0

studs

I

straight

+

2.0

stud

2.5

in pm

and

p&voted

T = 250’ -

1.5

width

versus

stud

width

0

chains

0.0

DAI

PAI

Nw

Dw

pw Z;“;

-%,

(3)

where the subscripts Al and W denote the parameters for Al lines and W studs, respectively. The measured and calculated values of Z:, and Z& do not usually differ by more than an order of magnitude. At 250 “C during an accelerated lifetime test,

Pw

rO.5,

NAI ~ 210-1 Nw

D*, = Dw

1024,

‘f

2

1021

(4)

W

C MA/cm2

*

/ 05

NA,

fi

lines

L .$ 5om

dw -----

d,,

lines

B

6 e $

-zz

Jw

dw -lo-‘, d AI

/

(2)

where N is the atomic density, D the diffusivity, p the resistivity and eZ* the effective charge. The uantity 6 is the effective boundary width (N 10 x ) for mass transport and d is the average grain size. For metallic thin films with less ideal or random grain structures, the grain-boundary parameters in eqn. (2) have to be replaced by an average value [3]. An effective geometrical factor taking into account the effect of grain orientation on mass transport should also be included in eqn. (2) [4]. The relative magnitude of two atomic fluxes JAI and Jw at Al/W interfaces can be estimated from the ratio of JAI and Jw, J Al

2.0

-&~hDjpe.Z’

I

I

1 .o

1.5

Metal

line linewidth

a 2.0

I 2.5

3.0

in pm

Fig. 18. Comparison of electromigration lifetimes and meandering lines and W stud chains.

in AI-Cu

straight

Thus, electromigration-induced atomic flux in W studs is vanishingly small as compared with that in Al lines. Since mass transport at Al/W interfaces is not balanced, voids and hillocks are expected to appear in Al lines at or near Al/W interfaces depending on the direction of current flow. This is illustrated in Fig. 19. It is known experimentally that the presence of Cu solute at Al grain boundaries can retard grainboundary diffusion and thus improve the elec-

184

Fig. 19. Unbalanced mass transport at the Al/W interface two-level AI-Cu interconnection with W studs. A(-)

e2nd level AI-Cu

in a

Bl + ) line

0

Fig. 20. Schematic drawing of an upper level AI-Cu line segment in a two-level Al-Cu interconnection with W studs.

tromigration resistance [37]. Analysis with the electron microprobe of Al-Cu lines after electromigration testing revealed that Cu and Al atoms migrate to the positive electrode at such relative rates that at the positive terminal the Cu concentration increases while it decreases at the negative terminal [ll]. It was also revealed that failure occurs in regions of the Al-Cu line which, presumably as a result of electromigration, are depleted of Cu [ 111. The A1,Cu second phase particles formed by Cu and Al serve as sources of Cu atoms to Al grain boundaries [38]. Electron microprobe measurements were carried out on a 300 pm long line segment of the second level Al-CU line which is connected to the first level Al-Cu lines by two W studs as shown in Fig. 20. The negative terminal or the cathode end is labelled as A( -) while the positive terminal or the anode end is labelled as B( + ). The normalized Al and Cu Ka X-ray intensities with respect to the Al-4wt.%Cu control contact pad along this Al-Cu line segment after electrical current stressing is shown in Fig. 21. Cu is found to be almost completely depleted within 50 ,xm at the cathode end A( - ). Figure 21 also indicates several Cu peaks due to the presence of A&u second phase particles along the 300 pm long Al-Cu line segment. As shown in Fig. 22, there is a discontinuity of the Cu supply at AlCu/W interfaces with W studs in a two-level AlCu interconnection. Thus, the observed void-open circuit failures occurred much earlier in these W stud chains than in Al-Cu meandering lines. E. Electromigration characteristics under pulsed direct current stressing Measured electromigration lifetimes for frequencies ranging from 50 to 200 MHz with a 50%

50

100 Line segment

150 A(-)

200 - B(t)

250

300

in pm

Fig. 21. Normalized Al and Cu Kcr X-ray intensity upper level AI-Cu line segment shown in Fig. 20.

along the

e-

Fig. 22. Discontinuity of Cu supply at the AI-Cu/W a two-level Al-Cu interconnection with W studs.

0'

0

I

50

I

I

100 150 Frequency in MHz

I

200

Fig. 23. Effect of frequency on electromigration fine lines under pulsed d.c. stressing.

interface

in

I 250 lifetime

in Al

duty cycle under a constant peak current density of 2.0 MA cm-’ are shown in Fig. 23. The lifetime was found to depend on frequency when the duty cycle, average and peak current densities were kept constant. But the data indicate no threshold frequency for a drastic change in electromigrationinduced failure rate [39]. However, our data suggest

185

that there can be more than one kind of relaxation process involving different types of defects during current-off periods. Electromigration lifetimes corresponding to different current-on times (t-on) with fixed current-off period (t-off) and different t-off with fixed t-on under a constant peak current density of 2.0 MA cmP2 are plotted in Fig. 24. Lifetime increases with t-off as tsoa (t-off)‘.’ and decreases with t-on as & a (t-on))‘.‘. Therefore, the dependence of lifetime on t-off is much stronger than on t-on. Figure 24 suggests that the formation of defects during the current-on time takes less than 5 ns while it takes much longer than 15 ns for a complete relaxation of defects during the current-off period. Electromigration lifetimes for different duty cycles at 50 MHz under a constant peak current density of 2.0 MA cm-* are plotted in Fig. 25.

‘I

fixed t-off,

vary t-on

n - fixed t-on,

0 -

vary t-off

const. T -

J, -

2 MA/cm2

25O’C

* On-time

Fig. 24. Dependence time and current-off stressing.

or off-time

in nsec

of electromigration period in Al fine

lifetime on current-on lines under pulsed d.c.

Lifetime increases with duty cycle as t,, ar -'.' which is a remarkable improvement over an average current density model (t50 ar -‘.O) suitable for pulsed d.c. stressing in the kHz region [34, 401. Different relaxation processes of defects which happened in different time frames can account for the improvement of lifetimes under pulsed current conditions with a different frequency range [41]. The enhancement of lifetimes under pulsed current conditions also depends on current density [18]. Our lifetime data for both pulsed d.c. stressing at 100 MHz with a 50% duty cycle and constant d.c. stressing for different peak current densities are shown in Fig. 26. Under continuous d.c. stressing, the current density exponent (n) increases -3) as the current density slightly (from -2 to 1 to -2 MA cm-*. These results increases from indicate that local heating due to Joule heating can increase the mass transport and thus decrease the electromigration lifetime in Al lines. For pulsed d.c. stressing below -3 MA cm-*, the current density exponent is found to be only - 1, which implies that Joule heating can be dissipated effectively during current-off periods. For pulsed d.c. stressing above -3 MA cm-2, the current density exponent jumps to -2 indicating that the current-off period is not long enough to dissipate all the Joule heating generated. F. Thermal stress-induced void-open failures by temperature cycling Lifetimes in Al and Al-Cu fine lines due to thermal stress-induced failures under two different temperature cycle conditions have been measured [28]. Temperature cycle condition A was carried out in a Delta 9023 environmental chamber with

3

1

‘\

‘Y \

“1 ‘\

pulsed dc

f - 50 MHz T - 250°C

dc - const.

J, -

I

‘,

.-2 E .c

J

2

?

L

100 MHz

0 -

r -

50%

l

T -

250°C

-

pulsed

dc

dc

T’

2 MA/cm’

I 2

f -

3

4

51189

102

;0o

Duty cycle

Fig. 25. Dependence of electromigration lifetime in Al fine lines under pulsed d.c. stressing.

2

J

1

1

J peak in MA/cm2

on duty

cycle

Fig. 26. Effect of Speak on electromigration lines under pulsed d.c. stressing.

lifetime

in Al fine

186

liquid Nz cooling at the fastest rate. The temperature was cycled between - 60 “C and 215 “C at 2 cycles per hour as shown in Fig. 27. Temperature cycle condition B was also carried out in the same environmental chamber but with air cooling. The temperature was cycled between 40 “C and 280 OC at -0.33 cycles per hour. The heating up and cooling down times are 20 and 160 minutes, respectively. Thermal stress-induced failures in Al fine lines during temperature cycling under condition A were detected. All the failures were observed to be due to the void-open in Al fine lines, The measured value of tso was found to be -416 hours. In the case of 1000 hours of temperature cycling under condition B, the thermal stress-induced void-open failure was detected in only one of the eight Al fine lines and in none of the eight Al-Cu fine lines. Thus, temperature cycling at a lower temperature range with a faster rate is found to further shorten the lifetime in fine metal lines. Lifetimes in Al fine lines due to thermal stressinduced and electromigration-induced failures under temperature cycle conditions A and B with a current stressing of 1.0 MA cm-* have also been measured. Electromigration lifetime measurements in Al fine lines were also carried out with the same electrical current stressing at 140, 215 and 250 “C. Their measured values of median time-to-failure for these temperature cycle conditions and constant temperatures are plotted in Fig. 28. The lifetime increases with decreasing temperature with an activation energy equal to - 0.5 eV from eqn. (1). The projected electromigration lifetime corresponding to the lifetime test under the temperature cycle condition can be estimated from eqn. (1) with the weighted average temperature and expressed as tso =

s

Aj -* exp(E,/kT)f,(T)

dT

(3

250

2501 200

1. '.. Y

150 .c g ,o

'..*

.._, '_

100

0

200

2

4

1

,..' ,s"

6

150

,,)/

.e..'

8

10

100

12

14

16

Time in minutes

Fig. 27. Heating and cooiing curves in an environmental with fiquid N2 cooiing.

chamber

8 :

5

g

J -

l

1.0 MA/cm*

t I

01

___A_-------_

I

lOOI

102

I

2

l/T

3

in lop5

4

5

K-’

Fig. 28. Void-open failures accelerated by temperature cycling. (---) indicates the temperature cycling range and its effective value of fse

and

s

h(T) dT=l

where ft(Q is the fraction of total temperature cycling time at temperature Tand can be calculated from the heating and cooling curves as shown in Fig. 27. The integration in eqns. (5) and (6) is performed along the entire temperature cycle range. By comparing electromigration lifetimes for temperature cycle conditions and constant temperatures, our results show that the extra thermal stress induced by temperature cycling shortens the lifetime by almost an order of magnitude in condition B and by approximately two orders of magnitude in condition A. It is speculated that the drastic decrease in lifetime for temperature cycle conditions is due to the thermal stress-induced cracking. In another set of experiments, temperature cycling was first carried out on sixteen Al fine lines under condition A for 72 hours. Eight Al fine lines were randomly selected among those fourteen Al fine lines without thermal stress-induced voidopen failure and also without visual damage (no voids observed optically). The electromigration lifetime test was then carried out on these eight Al fine lines together with another eight Al fine lines which had not undergone any temperature cycling. Their measured values of tSo for a current stressing of 1.0 MA cmW2 at 140 “C are plotted in Fig. 29. Our data indicate a reduction of - 15% in the electromigration resistance of Al fine lines after pre-testing temperature cycling. It is speculated that this is caused by the subtle structure damage in these Al fine lines during temperature cycling.

187

stress-induced ures.

5 e

J -

and electromigration-induced

faii-

1.0 MA/cm2

._c Conclusions

.-i! 6 0 S

5_

0 - with pre-testing l

-

without

-;oo

temp.

any temp.

cycling

cycling

I

I

I

100

0

Temperature

300

200 in “C

Fig. 29. Effect of thermal stress-induced damage gration resistance. (- - -) indicates the pre-testing cycling range.

1

/

J -

on electromitemperature

I

1.0 MA/cm’ .

.

- Al lines * - AI-CU

l

- - 100 L 0

- AI-k

I 50

lines at constant

temp

Al lines, temp. cycle condition lines, temp. cycle condition

I

I

100 150 Temperature

I

I

200 in ‘C

250

Fig. 30. Effect of Cu impurity on thermal (- - -) indicates the temperature cycling value of tSO.

300

stress-induced failures. range and its effective

Lifetime measurements in Al and Al-Cu fine lines due to thermal stress-induced and electromigration-induced failures were also carried out under the temperature cycle condition B with a current stressing of 1.0 MA cm-‘. Their measured values of t,, are plotted in Fig. 30. Our data indicate that the addition of 4 wt.% Cu in Al fine lines improves the resistance to both thermal stressinduced and electromigration-induced failures by -50 times under temperature cycling. However, the Cu addition only improves the electromigration resistance alone by -25 times at constant temperatures. This is probably due to the segregation of Cu atoms into Al grain boundaries, which not only reduces grain-boundary diffusivity but also suppresses grain-boundary sliding and migration [42]; both factors improve resistances to thermal

In Al-Cu submicron lines, the electromigration lifetime decreases with increasing linelength and then levels off beyond a critical length. As the linewidth of Al or Al-Cu fine lines decreases, the lifetime initially decreases and goes through a minimum, and then increases beneath a critical width. The lifetime also increases with decreasing film thickness. The critical length and width appear to decrease with decreasing linewidth and film thickness, respectively. Our overall results indicate that as the linewidth and the film thickness decrease to the submicron range, the grain morphology becomes an important parameter to control the lifetime. We can also attribute the improvement of electromigration resistance in the submicron metal lines to the columnar and bamboo grain structures. The high current density peak and the associated flux divergence in the bend structure can account for the observed decrease in electromigration resistance in metal lines with bend structures. Measured electromigration lifetimes indicate that, among other factors, the discontinuity of Cu supply at Al-Cu/W interfaces contributes most significantly to the reduction in the electromigration resistance of W stud chains. Our results also indicate that the lifetime is shorter in W stud chains and Al-Cu meandering lines than in Al-Cu straight lines. This is probably due to the current density peaks at studs, which not only impose a large electromigration driving force but also induce a Joule heating problem; both factors increase the mass transport. In general, the effect of the metal line geometry on electromigration lifetime in interconnects can be a complex function of the metal line geometrical dimensions relative to the defect size and density, grain size and grain morphology, metallurgy, patterning technique and metal deposition conditions. As multilevel interconnection consists of metal lines of different geometry and with different numbers of bends and studs, when one defines the maximum current density allowable in the interconnects, one should consider the weakest metal line geometry with the largest number of bends and studs. Results of our finite element calculations can provide a general approach to optimize the aspect ratio and geometries of stud and via structures for minimal current crowding and local heating. This approach can also be applied to other

188

structures and materials with different material properties which may be employed to form multilevel interconnection. Under pulsed d.c. stressing, lifetime increases with t-off as &,,a (t-off)*.* and decreases with t-on as tsoa(t-on)-‘.‘. The data indicate no threshold frequency for a drastic change in lifetime. However, our data suggest that there can be more than one kind of relaxation process involving different types of defects during current-off periods. Lifetime also increases with duty cycle as tSoa r -2.7, which is a remarkable improvement over an average current density model suitable for the kHz region. Different relaxation processes of defects happening in different time frames can account for the improvement of lifetime at a different frequency range. All the thermal stress-induced failures in Al and Al-Cu fine lines under temperature cycle conditions are due to the void-open. Temperature cycling at a lower temperature range with a faster rate is found to further shorten the lifetime in fine metal lines. By comparing electromigration lifetimes for temperature cycle conditions and constant temperatures, the extra thermal stress induced by temperature cycling shortens the lifetime of fine metal lines by one to two orders of magnitude. Our data indicate a reduction of - 15% in the electromigration resistance of Al fine lines after pre-testing temperature cycling even though these Al fine Lines are without visual damage. Our results also show that the addition of 4 wt.% Cu in AI fine lines improves the resistance to both thermal stress-induced and electromigration-induced failures by -50 times. References and R. Rosenberg, Physics of Thin Films, Vol. 7, Academic Press, New York, 1973, p. 257. P. S. Ho, F. M. d’Heurle and A. Gangulee, Electra and Thermo-Transport in Metals and Alloys, TMS-AIME, New York, 1977, p. 108. F. M. d’Heurle and P. S. Ho, 7lrin Films: Interdi~sion and Reactions, Wiley-Interscience, New York, 1978, p. 243. T. Kwok and P. S. Ho, Diffusion Phenomena in Thin Films, Noyes, Park Ridge, NJ, 1988, p. 369. B. N. Agarwala, M. J. Attardo and A. P. Ingraham, I. Appl. Phys., 41 (10) (1970) 3945. A. 3. Learn and W. H. Shephard, Proc. 9th IEEE Int. Reliabili~ Physics Symp., 1971, p. 129. G. A. Scoggan, B. N. Agarwala, P. P. Peressini and A. Brouillard, Proc. 13th IEEE Int. Reliability Physics Symp., 1975, p. 151. S. Vaidya, T. T. Sheng and A. K. Sinha, Appl. Phys. Lett., 36 (1980) 464. E. Kinsbron, Appt. Phys. Let&, 36 (1980) 968. K. Eden, W. Roth and H. Beneking, Microelectron. En&, I (1983) 263.

1 F. M. d’Heurle

2

8 9 10

11 T. Kwok,Proc. IstInt. ULSISci. Tech. Symp., ECS, Pennington, PA, 1987, p. 593. 12 T. Kwok, Proc. 4th IEEE Int. VLSI Multilevel Int~connection Co& 1987, p. 456. 13 T. Kwok, Froc. 26th IEEE In!. Reliability Physics Symp., 1988, p. 185. 34 T. Kwok, J. Finnegan and D. Johnson, Proc. 5th IEEE Int. VLSI Multilevel Interconnection Co@., 1988, p. 436. 1.5 D. Moy, M. Schadt, C. Hu, F. Kaufman, A. Ray, N. Mazzeo, E. Baran and D. Pearson, Proc. 6th IEEE Int. VLSI multilevel Interconnection Con&, 1989, p. 26. 16 T. Kwok, C. Tan, D. Moy, J. J. Estabil, H. R. Rathore and S. Basavaiah, Proc. 7th IEEE Int. VLSI Multilevel Interconnection Co& 1990, p. 106. 17 T. Kwok, R. Kaufman, B. Davari, Materials Reliability Issues in M~c~electroni&s, MRS, Pittsburgh, PA, 1991, p. 85. 18 S. J. O’Donnell, J. W. Bartling and G. Hill, Proc. 22nd IEEE Int. Reliability Physics Symp., 1984, p. 9. 19 J. Curry, G. Fitzgibbon, Y. Guan, R. Muollo, G. Nelson and A. Thomas, Proc. 22nd IEEE Int. Reliability Physics Symp., 1984, p. 6. 20 J. T. Yue, W. P. Funsten and R. V. Taylor, Proc. 23rd IEEE ht. Reliabi~~ Physics Symp., 198.5, p, 126. 21 T. Turner and K. Wended, FYOC.23rd IEEE Znt. Reliability Physics Symp., 1985, p. 142. 22 K. Hinode, N. Owada, T. Nishida and K. Mukai, J. vuc. Sci. Technol. B, 5 (1987) 518. 23 C.-Y. Li, R. D. Black and W. R. LaFontaine, Appl Phys. Lett., 53 (1988) 31. 24 J. W. McPherson and C. F. Dunn, J. Vuc. Sci. Technol. E, 5 (1987) 1321. 25 F. G. Yost, D. E. Amos and A. D. Romig, Proc. 27th IEEE Int. Reliability Physics Symp., 1989, p. 193. 26 T. D. Sullivan, Appl. Phys. Lett., 55 (1989) 2399. 27 H. Kaneko, M. Hasunuma, A. Sawabe, T. Kawanoue, Y. Kohanawa, S. Komatsu and M. Miyauchi, Proc. 28th IEEE Int. Rel~abili~ Physics Symp., 1990, p. 194. 28 T. Kwok, K. K. Chan, H. Chan and J. Sinko, J. I/&. Sci Technol. A, 9 (1991) 2523. 29 J. R. Black, IEEE Trans. Electron Devices, 16 (1969) 338. 30 T. Kwok, C. Y. Ting and J. U. Han, Proc. 2nd IEEE Int VLSI Multilevel Interconnection Conf., 1985, p. 83. 31 T. Kwok, T. Nguyen, P. Ho and S. Yip, Proc. 25th IEEE Znt. Reliabili~ Physics Symp., 1987, p. 130. 32 T. Kwok, T. Nguyen, S. Yip and P. Ho, Proc. 5th IEEE Int. VLSI Multilevel Interconnection Co& 1987, p. 252. 33 0. C. Zienkiewicz, The Finite Element Method, McGraw-Hill, London, 1977. 34 P. B. Ghate, Proc. 20th IEEE Int. Reliability Physics Symp., 1982, p. 292. 35 J. R. Lloyd and P. M. Smith, J. Vat. Sci. Technol. A, I (1983) 455. 36 H. B. Huntington, Diffusion in Solids - Recent Development, Academic Press, New York, 1975, p. 303. 37 I. Ames, F. M. d’Heurle and A. Gangulee, IBM J. Res. Dev., 4 (1970) 461. 38 T. Kwok, P. S. Ho and H.-C. W. Huang, J. vuc. Sci. Technol. A, 2 (1984) 241. 39 T. Noguchi, K. Hatanaka and K. Maeguchi, Proc. 6th IEEE Int. VLSI Multilevel Interconnection Co& 1989, p. 183. 40 J. M. Towner and E. P. van der Ven, F’roc. 21st IEEE Int. Reliability Physics Symp., 1983, p. 36. 41 K. Hiraoka and K. Yasuda, fioc. 7th IEEE Int. VLSI Multiievel Inte~onnection Co&, 1990, p. 120. 42 G. H. Bishop, R. J. Harrison, T. Kwok and S. Yip, J. Appl. Phys., 53 (1982) 5596.