[ieee 6th annual reliability physics symposium (ieee) - los angeles, ca, usa (1967.11.6-1967.11.8)]...

9
FAILURE MECHANISMS AND DEVICE RELIABILITY J. S. Smith and J. Vaccaro Rome Air Development Center Griffiss Air Force Base, New York 13440 Summa This paper shows how expressions for failure mechanisms may be derived and related to device reliability. The necessaxy conditions and limi- tations of such relations are examined, and their applicability to practical device reliability problems is assessed. The role of the reliability physicist in furnishing needed information and understanding at the mechanism level is indicated. Introduction An important objective of reliability physics is an attempt to relate basic failure mechanisms to device reliability. Although reliability physics has gained general acceptance as a useful approach to the solution of device reliability problems, such relationships have not yet been too clearly drawn, even in principle. This has led to an uncritical acceptance and application of vaguely defined concepts on the part of the reliability engineer. A need, therefore, exists to clarify the practical uses and limitations of these failure mechanism relationships to device reliability. In the reliability physics approach, one attempts to determine the principal failure modes and mechanisms of a device by performing a com- parative analysis between failed devices, from either laboratory tests or actual use, and known good devices. From such analyses, approximate kinetic expressions for degradation resulting from a single physical or chemical mechanism have been derived. However, these expressions have been found more useful in confinning that a pos- tulated mechanism exists, agrees with observed device behavior, and is consistent with device physics, rather than in predicting device failure rates. A major problem in developing kinetic expressions has been that of estimating the in- fluence of process-induced (quality) defects on the degradation rate. When several mechanisms contribute to device degradation, as is usually the case, modeling rapidly becomes complex and impractical. For these reasons, investigations to date have been confined largely to the anal- ysis of single failure mechanisms, and to esti- mates of their relative contributions to the total device failure rate. A determination of device failure rate levels will, for the newer devices and technology at least, continue to come from laboratory testing, field results and related experience. It is the purpose of this paper to show how expressions for failure mechanisms may be derived and explicitly related to times to failure. The necessary conditions and practical limitations for such expressions are examined for both the single- and many-mechanism case. The influence of quality defects and screening tests are con- sidered, and the importance of establishing the actual failure mode distribution is emphasized. The most effective use of mechanism models, and the role of the reliability physicist in furnish- ing needed information and understanding is indicated. Single Failure Mechanism Case From the standpoint of analytical tracta- bility, the ideal situation prevails when a single time-dependent failure mechanism in a device interacts with a stress field. The rate of device degradation (not always reflected by changes in device parameter values) is then pre- sumed to be proportional to the rate governing the dominant failure mechanism and can be de- scribed in terns of an energy transfer. The most common mechanism underlying an irreversible change in solid state materials is that of diffusion, hence, degradation of solid state devices is often related to that process. Diffusion in solids is apt to proceed quite slowly, however, it is strongly influenced by the presence of crystal defects. The basic space time diffusion equatioi is given by ac = div D grad C at (1) where C = concentration of diffusing atoms D = diffusion coefficient, usually dependent on C The temperature dependence of D for a fixed concentration is given by D = A exp (- c/kT) (2) where A = related to the jumping frequency c = activation energy If the diffusing species is ionic, it can move under the influence of a field and the diffusion coefficient becomes D = A exp (- E/kT - Ee 8/2kT) (3) -1- I

Upload: j

Post on 27-Mar-2017

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

FAILURE MECHANISMS AND DEVICE RELIABILITY

J. S. Smith and J. VaccaroRome Air Development Center

Griffiss Air Force Base, New York 13440

Summa

This paper shows how expressions for failuremechanisms may be derived and related to devicereliability. The necessaxy conditions and limi-tations of such relations are examined, and theirapplicability to practical device reliabilityproblems is assessed. The role of the reliabilityphysicist in furnishing needed information andunderstanding at the mechanism level is indicated.

Introduction

An important objective of reliability physicsis an attempt to relate basic failure mechanismsto device reliability. Although reliabilityphysics has gained general acceptance as a usefulapproach to the solution of device reliabilityproblems, such relationships have not yet beentoo clearly drawn, even in principle. This hasled to an uncritical acceptance and applicationof vaguely defined concepts on the part of thereliability engineer. A need, therefore, existsto clarify the practical uses and limitations ofthese failure mechanism relationships to devicereliability.

In the reliability physics approach, oneattempts to determine the principal failure modesand mechanisms of a device by performing a com-parative analysis between failed devices, fromeither laboratory tests or actual use, and knowngood devices. From such analyses, approximatekinetic expressions for degradation resultingfrom a single physical or chemical mechanism havebeen derived. However, these expressions havebeen found more useful in confinning that a pos-tulated mechanism exists, agrees with observeddevice behavior, and is consistent with devicephysics, rather than in predicting device failurerates. A major problem in developing kineticexpressions has been that of estimating the in-fluence of process-induced (quality) defects onthe degradation rate. When several mechanismscontribute to device degradation, as is usuallythe case, modeling rapidly becomes complex andimpractical. For these reasons, investigationsto date have been confined largely to the anal-ysis of single failure mechanisms, and to esti-mates of their relative contributions to thetotal device failure rate. A determination ofdevice failure rate levels will, for the newerdevices and technology at least, continue tocome from laboratory testing, field results andrelated experience.

It is the purpose of this paper to show howexpressions for failure mechanisms may be derivedand explicitly related to times to failure. Thenecessary conditions and practical limitations

for such expressions are examined for both thesingle- and many-mechanism case. The influenceof quality defects and screening tests are con-sidered, and the importance of establishing theactual failure mode distribution is emphasized.The most effective use of mechanism models, andthe role of the reliability physicist in furnish-ing needed information and understanding isindicated.

Single Failure Mechanism Case

From the standpoint of analytical tracta-bility, the ideal situation prevails when a

single time-dependent failure mechanism in a

device interacts with a stress field. The rateof device degradation (not always reflected bychanges in device parameter values) is then pre-sumed to be proportional to the rate governingthe dominant failure mechanism and can be de-scribed in terns of an energy transfer.

The most common mechanism underlying anirreversible change in solid state materials isthat of diffusion, hence, degradation of solidstate devices is often related to that process.Diffusion in solids is apt to proceed quiteslowly, however, it is strongly influenced bythe presence of crystal defects. The basic spacetime diffusion equatioi is given by

ac = div D grad Cat (1)

where C = concentration of diffusing atomsD = diffusion coefficient, usually

dependent on C

The temperature dependence of D for a fixedconcentration is given by

D = A exp (- c/kT) (2)

where A = related to the jumping frequency

c = activation energy

If the diffusing species is ionic, it can

move under the influence of a field and thediffusion coefficient becomes

D = A exp (- E/kT - Ee 8/2kT) (3)

-1-

I

Page 2: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

where E = electrical fielde = charge on the ion

8/2 = distance between equilibrium conditionand the activation energy barrier

A good example of the role of the diffusionprocess in device degradation is the case ofionic motion on the surface of a device underthe influence of an applied field.

Functionally, then, we can expect that thetime rate of concentration in a given region willfollow an Arrhenius-type relation of the form

c f(A exp (-C/kT)a3t

Let us consider a single mechanism of failurewhich is rate limited by the diffusion processand is expressed simply by

dq= A exp (- C/kT)

dt

where q is some measurable indicator of the rateof the diffusion process.

Integrating (5) we get

q ( T,t ) = qO ±J A exp (- E/kT) dt0

(6)

Next, it is necessary to relate q(t) to ameasurable device parameter, P. In a semicon-ductor device an electrical parameter P woulddepend on a number of fundamental device quan-tities such as doping profile, junction area,minority carrier lifetime., stress, temperature,etc. Let us express this relationship as

P (t) = fq ( sl....Sn.^Tst)J (7)

where Si are the various fundamental devicequantities and stress factors.

Assume the simple case where we can separatethe variables and express them as follows:

P (t) = q ( S1,....Sn) q ( t,T)

q1( Sl....Sn)% +f A exp(-%f)dt]()q 1(sl*....Sn) qo +

ql( Sl,..-..Sn) A exp(-E/kT)(t - to)

then q1 (Sl, ........n)qo = Po will be the valueof P at to. Then if we let t - to = tF ^ thetime to failure, and define P(t) = PF as thelevel of P at failure we have

t = . .. .

F ql (Sly ...9S) A exp ( -C/iT)(9)

(4)

If the device lot is non-homogeneous, themean time to failure of N devices will be

N_ i PF Pi

tF =

NT 7t1 q(Si,...-Sn) Ai exp(-t6~kT)(10)

The subscript i denotes a variance in theactivation energy 6, the pre-exponential term A,and the initial parameter value P°. The dis-tribution of times to failure will then be de-termined by the distributions of these threeterms.

The mean time to failure over an extendedperiod may not be too meaningful, however, whenour concern is primarily with limited time in-tervals within such a distribution. Experienceindicates, for example, that the average timeto failure of semiconductor devices tends to in-crease with time, hence, changes in time to fail-ure during early life may be an important con-sideration in estimating times to failure at lateruse conditions. For convenience here we will con-sider times to failure to be constant, but differ-ent, over various short time intervals during theearly life of a device.

Let us now examine the expression for theinitial parameter P0 introduced in equation (9)above. This term is composed of a function whichdescribes the ideal behavior of a device param-eter. This ideal behavior may then be modifiedby the presence of a failure mechanism. If attime zero, Po is measured and found to be out ofspecification tolerance, then it is simply re-jected. Quite often, however, a potentially baddevice will be acceptable from a terminal measure-ment point of view. This may, in part, be due tothe fact that a vendor's specification on thetolerance of a device parameter is not tightenough from a reliability point of view.

The expression for P0 also serves as a guidein screening techniques. To cull the potentiallyunreliable device,a screen can be developed fromtwo points of view. The first and most common isto stress the device for a substantial period oftime, or at higher than use stresses, so as to

-2-

I

Page 3: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

r, ~ 1

enhance the failure mechanism contribution toideal device response and thus simplify detectionand screening. The expression also suggests,however, the alternate possibility that by suffi-ciently modifying ideal device response to par-ticular stresses, such as voltage and temperature,one can more easily detect the failure mechanismcontribution. Here, then, is an area whicb iscurrently occupying, and should continue tooccupy, the attention of the reliability phyjsicist.,i.e., the search for more sensitive indicators,q, and techniques for measuring them. Coupledwith this is the requirement for a detaileddefinition of initial conditions, q1(S1, . Sn)for ideal device behavior.

Summarizing the assumptions made above forthis simple case, we have

1. all failures will be due to a singlemechanism

2. q is linear in time and has an Arrheniustemperature dependence

3. A and E are not time dependent

4. equation (8) is separable in the terms(t,T) and (Sl;....Sn)

Consider one of the pitfalls here for reli-ability prediction. If the Arrhenius term,(A e - /kT) for a process was some undeterminednon-linear function of (A e a/kT) then the acti-vation energy would also be indeterminate. Inthis case we cannot assume that the slope of aln A vs l/T plot is the activation energy ofthe mechanism. The most questionable assumptionin any practical situation is, of course, thatouly a single mechanism is present in the device.

As an example of the single mechanism case,let us consider degradation of a semiconductorjunction device which fails due to drift of sur-face charge in a potential gradient. In thiscase, the device parameter, P, used as a failureindicator, is the reverse current leakage. Thisresults from an inversion layer growrth beneaththe oxide due to the surface charge accumulation.The rate of accumulation of charge density, Q,is therefore

aQ 1 2V

at Ro a x2

anad since Q = COV

(11)

where CO = oxide capacitance per unit areaRa = sheet resistance of the oxide surfaceV = electrostatic potential on the surface

Equation (f2) is of the same form as thediffusion equation, with V taking the place of aconcentration and l/CORB taking the place of thediffusion coefficient.

Consider the situation where the surfacepotential on this device is zero everywhere att = o. At this point a voltage large enough toinvert the surface is applied to a metal gateoverlying the junction. Then the solutionequation (2) for this initial condition is

ViX= 2 t/P C0 arg erfc-

V(13)

-0

where Vi is the voltage required to invert thesurface, Vo is the applied voltage, and X is thelength of the external inversion layer.

The surface sheet resistance has the func-tional form

R a = POeE/kT

(14)

with failure occurring when X = d. In this caseX would be an indicator of the degradation rate.If the i th device of a lot contains some inver-sion Xi at time zero, then the time required forthe inversion to grow to the edge of the devicechip (assuming no guard ring) becomes

(d-Xi)2 R1 CO et = [F 4 Larg erfc (vi/vO0) 2

(15)

For convenience in illustrating this situa-tion with a numerical example, let us assume tFto be fairly constant over some interval of time.Then, if there were 10 devices in the lot, theaverage failure rate over that time intervalcould be expressed as

10 2 - Ci/kT

J Iiarg erfc (i/Vo)12 e (16)

i Cl (Xi d)2

av 1 a 2v0 a

where we have assumed a single Vo.and a for all(12) the devices, but a variance on R% , C01, V1 X

and Ei.

-3 -

Page 4: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

Using some reported data,3

C = 1 evRo = 1010.XCo = 4.25 x 15pf/microns2Vi = 5 voltsVo = 10 volts

and assuming a clean oxide and a vralue of d = 250microns, a failure rate of .03%/1000 hours at atemperature of 200 C is obtained. (It should benoted here that gormal oide sheet resistance isof the order 101 to 101t n rather than thevalue of 4 x 1020 o used in the referencedexample.)

The example thus illustrates how a failuremechanism can be related to device parametersand a failure rate estimate made. Let us examinenext, what can happen to the failure rate in thepresence of a quality defect. Suppose we havea region which acts as a high carrier generationrate site, such as a scratch. The time to fail-ure will now depend on the nature and severityof the scratch and its proximity to the junction.If it is very close to the junction, the time tofailure may be reduced by orders of magnitude.The critical need for a better understanding ofthe physics of quality defects for reliabilityimprovement and control, as well as to prediction,is apparent. Some attention is being devoted tosuch studies, but in view of the seriousness ofthe quality defect problem., particularly in inte-grated circuits, much more is obviously needed.

Our analysis of the single failure mechanismcase has led us, however, into the practicaldifficulty which arises in any attempt to predictfailure rates of modern semiconductor devices.In principle, if we know the distributions of.theva.rious physical parameters of our example, RtCO , and XV, etc., we can determine the distribu-tion of the times to failure, and hence, thedistribution of failure rates throughout the de-vice life. Semiconductor devices, however, aretypically very reliable except for the presenceof an extremely small percentage of defectivesin each lot. Hence, our real problem is not somuch how to predict the results of normal designor process variations, but how to base predic-tions on the very small number of defectiveswhich fail early and thus characterize only theearly tail of the failure distribution. Inaddition, there will always be the problem ofscreening out the device, which meets prescribedtolerances on specified physical parameters, yetfails early because of the presence of an un-detected and unpredictable defect.

It is clear then that a failure mechanismanalysis will be much more useful in confirmingthat a postulated mechanism exists, agrees withobserved device behavior and is consistent withwhat is known about the physics of the device.It also furnishes a-basis for determining thesensitivity of the kinetics of each failuremechanism to defects and variations in devicedesign and processing. And perhaps most

importantly, it offers a sound physical basis forthe design of reliability and quality assurancetests and screens, since coupled with careful,detailed failure analyses, it furnishes usefulinformation on the stress response and stresslimitations of a device.

The nay-1Mchanisms Case

In this section we consider the more generaland realistic case where several mechanisms pre-vail and the contributions from quality defectsare included. Modeling for this situation rapid-ly becomes complex and unwieldy, and its applica-bility is limited due to the large numbers of un-knowns and uncertainties in the infornation re-quired. We consider here a model which attemptsto account for the many factors and variableswhich must be considered. Our interest, however,is not so much to emphasize the model,, since itis simply a systematic cataloging of events andconditions, but rather to use it as a frameworkto point out limitations and gaps which the re-liability physicist nay help alleviate.

We begin by categorizing the types offailures which may occur in a given lot ofdevices using the following symbols:

FM = devices with long term, time andtemperature dependent mechanismshaving the same functional form

DFm = as above, but with defects presentwhich accelerate the kinetics of theFM

Dv= devices with defects which causefailure dependent only upon thenumber of times a pulse of stressis applied

DR = devices which fail at random timeswith unknown stress dependence

No claim is made here that this classifica-tion is complete or optimum, but simply that itis adequate for the purposes of this discussion.If we ascribe contributions to the time to fail-ure from each of the above types of failure, thenthe average time to failure of the failed devicesmay be expressed as:

n1 tFM + n2 tDFM + n3 t D + n4 tD (17)n, + n2 + n3+ n4

where ni are the number of failures from eachcategory, in a user's lot of devices, and ti arethe corresponding average times to failure ofthose categories. We have assumed here that thecontributions from each categor,y are independent.

The information required to calculateaverage time to failure under use conditionsusing equation (17) is itemized as follows:

-4-

m mr.--i

I

Page 5: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

1. the percentage of failures in the lotUfor each of the categories listed in the tineperiod of concern, i.e., the imodal distributionof failures

2. the tinme to failure contribution associ-ated with each of these rmodes

3. if the failures are observed under lab-oratory conditions, then the relevance of thesefailures to use conditions must be established.

These requirements contain several importantimplications. The modal distribution of failures1may radically change from laboratory to use con-ditions. If the design of screen tests, burn-inand life tests are not realistic, such failuresmay be negligible in actual operation. Carefulanalvsis of device failures from screens, bumn-inor life tests should be conducted and comparedwith simnilar device failures from the field toobserve any changes in failure mode distributionand to determine the effectiveness of screens andtest design. Without extensive life testing and/or field failure information, the failure contri-butions of each failure mode can only be estimatedfrom an accurate identification of failure mech-anisms for each mode and their stress dependenciesSuch analyses constitute another major area ofpotential contributions from the reliabilityphysicist.

Consider for example how an erroneousassumption concerning temperature dependence offailure mechanisms and the distribution of failuremodes can affect failure prediction. The distri-bution of failure modes of an integrated circuitcan change from predominantly temperature-depen-ent in lab tests to predominantly temperature-independent in the field. Figure 1 illustratesthe modal distribution of integrated circuitfailures actually experienced on one militarysystem. Note the shift in distribution from theFM and DFM types after burn-in to the Dy, and DRtypes in the field. It should be noted herethat the burn-in failure data was inadequate topermit a clear distinction between FM and DFMtypes, but together they constituted an unmis-takable majority of the total failures. Screensand burn-in were apparently not very effectivein screening out Dy and D from the users lot.High temperature test data based on this burn-indata might yield points on an Arrhenius plot inFigure 2.

The curvature of the line connecting pointson the plot is often interpreted as being due tothe activation energies of two different failuremechanisms. It is not often recognized that thecurvature of the line may also be due to failurecontributions which are entirely temperature in-dependent and, hence, have zero activation ener-

gies. The ambiguity in this case must be resolvedthrough careful failure analysis and comparisonof the stress dependencies of the failure mechan-isms in lab tests with those in the field.Figure 3 lists some of the more common mechanismsand defects and the stress dependencies typically

observed in integrated circuits.

The model introduced here, then,takes intoaccount the possibility of changes in failuremode distribution, and attempts to utilize avail-able data in estimating time to failure underuse conditions.

To determine the ni of equation (17)., thepopulation distribution in the users lot, we

will need some measure of the efficienty of thescreens used for each mode of failure. For sim-plicity, let us assume only bad devices are re-jected by our screens. We will also assune thatfailures at any screen are independent of whathappened to the device at previous screens.

At each,screen,sj, a number,Fs. I of devicesfail. Let E. be the percent that f il due tothe i th defect or failure mechanism. The totalfailures due to the i th reason are the sum ofthose failing for the i th reason at each screen,and is given by

ni = E' Fs + E Fs + ....... E sj1l 2 ~2

If we consider also the number, A.,fail from any add-on, maximum rated liefor the i th reason, the efficiency,Cs_each screen can be expressed as J

E+ FeSi_s~ ~ ~ ~

G Fs, + Ai

lI=j

(18)

thattest

, of

(19)

where k = number of screens,the s. screens are considered in sequence,and "I = dummy variable

The number of devices left in the user'slot, having an i th failure mechanism or defect,Xi, is given by

< Ni + AiXi = (1 -El.) L

3 N(20)

where L = total devices in user's lotcNi + Ai=- most probable % ux devices in i th

N category

When failures in the i th category from allscreens have been considered, Xl, we get

k

Xti=j=l

C Ni+ Ai>(1 _E(i L (21)

-5-

I -

-r-I...

Page 6: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

where k= nurmLber of screensXt = most probable nunuber of devices of

i th category left in user's lotafter all screens

Then the ni from equation (19) are given b,

ml

i= t

i-1

m2

n2 = ) t

i=l

m3

3 7i=l

m"4n4_ =

i-1

iXt

Xit

where ml .....,imn4 represent the number in ea

category, i, which have the same functional foof tinme dependence.

The average timne to failure associated withe t5> , for example, could be expressed as

m3

11 t I m3

m3 n3

1=1Hence from equations (19), (21) and (23)

we have, for example.,

rn2

n2 i=,

i=

field and observed field failures for the actuallot will furnish corrective data. If the screensand bum-in are efficient and the devices highlyreliable, the first estimates of times to failurefrom the model will not be very far off.

Again, an important key to this method willbe a careful, detailed analysis of device fan1l-ures by the reliability physicist. Efficiencyof screens and cataloging of failures into propercategories are critically dependent upon theaccuracy of the failure analyses. It should be

(22) noted that equation (19) for the efficiency ofscreens is predicted on analyzing the failuresfrom each subsequent screen to determine thenurnber of i th defectives that escaped previousscreens. This approach warrants considerablymore attention than it has received in the past,since valuable information on the design ofeffective screens is lost without it. Field

Lch failures, as they occur, and relevant previous)rm experience, must also be carefully analyzed to

determine whether the failure mode distributionis significantly different from that of labora-

Lth tory tests. This information can then be usedto determine stress susceptibility of the variousmodes, and serve asabasis for improving screensand test design, as well as estimating failurerates.

(233)

kl (24)

EiF5.1 - (Ni+Ai)

j-l *F Fs!, A

and

In summary then, this method attempts toestimate average time to failure as a function of

1. the number of failures occurring duringscreening and bumn-in, and add-on life tests.

2. a careful, detailed analysis of thesefailures and assignment to categories accordingto functional form of time and stress dependence.

3. an estimate of the contribution to devicefailure of each failure category from an analysisof stress dependence and kinetics of the mech-anism.

4. determination of screening efficiencyof each screen.

m2_ 1 - -9

tDFM =-

n2ii=l

Xt tDt FMj~

as the nunber,n 2' of DFM failures left in lotafter screen and bumn-in tests, and tD is theircontribution to the average time to faVlure.

An important assumption implied in the abovemodel is that all potential failures will havebeen eliminated by the end of the bum-in period.Clearly then, the estimate of the failure timederived from this model will be in error to the

extent arising from neglect of the contributionof failures from the field. Hence, it can onlyrepresent a first step in an iterative correctionprocess. Experience with similar devices in the

5. monitoring and analysis of field fail-ures (and relevant previous experience) to detectchanges in failure mode distribution from thatof laboratory tests, and correction of early timeto failure estimated.

Even if we ignore interactions between mech-anisms, however, this model, in addition to sheercomplexity,, is still subject to the practicaldifficulties mentioned previously for the singlemechanism case. Indeed, when a defect or mech-anism has been investigated to the point of es-tablishing a mean time to failure, a more effi-cient screen or some corrective action is likelyto result. The model does, however, furnish a

framework for systematicaLly assessing progress,problems and gaps in required information. Whileit is too impractical for predicting absolutevalues of times to failure or failure rates, itcan be useful in determining the relative contri-

-6-

m

i

Page 7: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

butions of specific failure mechanisras to devicefailure.

Conclusions

Expressions for failure mechanisms may be

derived and explicitly related to time to failure.

This was first shown for the case of a single,time-dependent failure mechanism and illustratedwith an example of semiconductor degradation.The dominating influence of quality defects on

the kinetics of degradation was indicated. A

general expression for estimating times to failurewas then considered for the case of several inde-pendent mechanisms, including contribution fromdefects. The source of infornation for this model

was actual failures observed from screen., burn-inand add-on life tests, failure causes beingascribed by careful, detailed failure analyses.From the consideration of these models, it was

concluded that:

1. at the failure mechanism level they areineffective for practical predictions of absolutefailure rate, but can be useful in determiningthe relative contributions of specific Lailuremechanisms to device failure. This is particular-ly true for the newer devices and technologywhich are immature from a reliability point ofview.

2. they are most useful in confirming thata postulated failure mechanism exists, agreeswith observed device behavior and is consistentwith what is knownm about the physics of a device.

3. they furnish a basis for determiningthe sensitivity of the kinetics of each failuremechanism to defects and variations in devicedesign and processing, thus offering a sounderphysical basis for the design of reliability andquality assurance tests and screens.

4. the use of Arrhenius models for ex-trapolating failure rates to use conditions isjustified only when it has been determined thatthe dominant failure mechanisms under those con-ditions are primarily temperature dependent.

Within the context of the models, severalareas requiring the attention of the reliabilityphysicist are indicated. These are:

1. the need for more sensitive indicatorsof device degradation at the mechanism level,and techniques for measuring them.

2. a more detailed definition of initialconditions for ideal device behavior.

3. a need for a better understanding ofthe physics of quality defects.

4. a more careful detailed analysis offailed devices, particularly in establishingfailure mode distributions at various periods in

the life of a device, and in evaluating screeningeffectiveness.

References

1. Sinnotti M. J., "Solid State for Engineers,John W7iley & Sons, New York 1958.

2. RADC-TR-66-36, "Failure Mechanisms in SiliconPlanar Epitaxial Transistors," Sello, Blech,Grove, 1ay 1966.

3. Ibid p. 11 - 14.

-7-

Page 8: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

MODAL DISTRIBUTION

BURN- I N FA I LURES

FM DFM Dy

TYPE FAILURES

DR

F I ELD FA I LURES

45

1 8

1'1 , 1 [1I.FM DFM Dy

TYPE FAILURES

DR

FIGURE 1 - INTEGRATED CIRCUIT FAILURES IN A MILITARY SYSTEMf

EFFECT

lnIL

OF TEMPERATURE INDEPENDENT FAILURES

ARRHENIUS PLOT

temperatureindependentfailures

FIGURE 2

-8-

63

1 4 131 0

imir

I,

m

F,

.._

Page 9: [IEEE 6th Annual Reliability Physics Symposium (IEEE) - Los Angeles, CA, USA (1967.11.6-1967.11.8)] 6th Annual Reliability Physics Symposium (IEEE) - Failure Mechanisms and Device

Failure Mode

intermetallic formation

cracks in chip

thin metal at oxidesteps

bond degradation

peeling aluminum

metal shorts to chip

interconnect-corrosion

inversion

accumulation

bond short to substrate

scratches-metallization

contamination

lead short - can or chip

lead opens

cracks - chip

Initiating Cause

bonding

bonding

process

contamination

poor adherence

pinholes-oxide

contamination

contamination

contamination

process

process

package leak

lead wires

lead wires

process

Figure 3 -

FailureIndicatoro

opens

opens

opens

opens

opens

shorts

opens

shorts

FT

shorts

shorts

shorts

opens

opens

Typical Integrated Circuit Failures

T:Suspected Failure Mechanism

(1O

chemical reaction & diffusion y(

strain relaxation u:

current density - melting u

electrolytic action y

mechanical or thermal shock n

diffusion i

chemical reaction (diffusion) j

-diffusion limited

diffusion limited

mechanical deformation

current density - melting

diffusion

mechanical deformation

mechanical fracture

mechanical fracture

ime Accelerating Stressependence Thermal Non-Thermal)nt1J;-)

es yes-Arrhenius

mknown no mechanical

nknown indirectly current den:

res yes-Arrhenius voltage

tone no thermo-mech

yes yes voltage

yes yes

yes yes-Arrhenius

yes yes-Arrhenius

none no freq. - mec

none no current dei

yes yes-Arrhenius

none no

none no shock- fr

none no mechanical

sity

1.

ch.

nsity

eq.