Marcus P. Corrosion mechanisms in theory and practice

Подождите немного. Документ загружается.

latent manufacturing defects and long-term wearout [119]. These two causative

mechanisms are very different directly correlate to the two primary reasons for

performing accelerated aging noted in the first paragraph of this section. The

elapsed time to fail and cumulative failure percentile are often plotted using a

lognormal distribution (Fig. 7). Two statistical measures result: the mean time to

failure (MTTF, the time for 50% of the population to fail) and the population standard

deviation (calculated from the slope) [8,119]. Because the goal of accelerated

aging is to establish a relationship between the environmental parameters and

service life under operating conditions, the MTTF metric can be used to calculate

an important parameter known as the acceleration factor (AF). In terms of MTTF,

664 Frankel and Braithwaite

Table 1 Specifications for Three Standardized Atmospheric Corrosion Test Environments

The acceleration factors are experimentally measured as a function of the

environmental stresses using the techniques described in the previous subsection.

Figure 7 Example of a lognormal distribution of failure times and the definition of the

mean-time-to-failure (MTTF) metric.

Using triple-track test devices, the surface conductance, G, can be directly

measured. This parameter is a more fundamental indicator of the susceptibility to

electrolytic degradation mechanisms compared with MTTF because, as noted earlier,

electrolytic corrosion will occur only if the ohmic resistance between lines does

not consume the applied voltage difference and leakage current flows [111,120].

G is calculated by multiplying the leakage current by the area of insulator and

dividing by the applied voltage. The resistance change of the lines can also be used

as a measure of degradation (either electrolytic or open circuit). The important

assumption inherent in these types of tests is that the specific metric being measured

(e.g., surface conductance or wirebond resistance) is directly related (proportional

or inversely proportional) to failure rate and thus to MTTF [66,111,120].

The cumulative values for MTTF determined by any of these techniques for a

range of conditions permit the formulation of general relationships that describe the

dependence of failure rate on the stressing parameters. The effect of temperature, T,on

reaction or failure rate, R, can usually be described by an Arrhenius-type relationship:

Corrosion of Data Storage Devices 665

where E

is activation energy, k is Boltzmann’s constant, and A is an empirical constant.

A range of activation energies for failure of ICs has been reported, from 0.4 to 1.2 eV.

The acceleration factor associated with temperature becomes

Because of the concern that higher voltage levels will increase the internal

temperature (ohmic heating) and dry the device, most investigators use the normally

applied operating voltage during testing. However, for the corrosion mechanisms that

occur due to electrical bias, the corrosion rate is proportional to the current between

the conductors and therefore scales with applied voltage. As such, the MTTF

decreases as bias increases and an inverse relationship may exist [7,60,121,122]. For

the limited cases in which a higher voltage level is used, Shirley [123] has proposed

a general linear acceleration model where a and b are constants:

Unfortunately, the influences of RH and impurities are most poorly established

than the effects of temperature and bias. Many investigators have measured MTTF

or G for actual devices or test structures in high-temperature and -humidity

environments and developed empirical relationships by fitting the data to various

equations. Several commonly referenced variants are presented in Table 2. Each

expression in this table has a parameter A that represents a proportionality constant.

These equations can be used to determine an acceleration factor by considering the

ratio of field to test MTTF as was done in Eq. (3). Using this technique, the

proportionality constants cancel. Then by using Eq. (1), the service life under

operational conditions may be predicted after measuring lifetime under stress

conditions. The first four expressions in Table 2 are “Eyring” models because

the influences of T and RH are considered separately and then multiplied

together to get a combined equation. The last expression does not assume that the

effects of temperature and humidity are independent and has an activation energy

that is dependent on RH [124]. These models have been used in a number of

investigations to fit experimental data, and different values for the constants have

been determined for various systems and degradation modes [96,98,99,102–

105,119,121,125]. An example of how such models are applied is shown in

Figure 8, which is based on surface conductivity measurements that were made

with an Al-Cu triple-track structure in a HAST chamber [103]. The temperature

and humidity combinations were fit to various equations and those shown in

Figure 8 are for the last expression in Table 2, which contains an RH-dependent

activation energy. The correlation coefficient for this expression was found to be

higher than those for other models [103].

As will be discussed in the next subsection, to be truly effective, predictive

capabilities must be based on fundamental physical understanding. The models

presented in Table 2 are simply empirical correlations. Limited progress has been

made to date to improve this situation. Comizzoli [46] has provided a mechanistic

explanation and associated mathematical expression for the exponential dependence

of surface conductance (and MTTF) on relative humidity that is contained in some of

the models listed above. Pecht and Ko [83] developed a comprehensive model for

predicting absolute time to failure when microelectronic die metallization is corroded

by an electrolytic process. Their phenomenological underpinning involves ion

transfer based on Ohm’s and Faraday’s laws. The other key feature is their treatment

of the critical process involving moisture ingress that permits this model to be useful

for both PEM and CHP devices. A final reference that is relevant to this subject is

the work being performed by Graedel and co-workers [128]. This team is developing

and exercising a physically based atmospheric corrosion model referred to as

GILDES that conceptually can include all microelectronic corrosion processes (e.g.,

gas transport, adsorbed water layer, corrosion product layer, electrochemistry).

666 Frankel and Braithwaite

Table 2 Empirically Derived Models Describing the Influence of Temperature, T, and

Relative Humidity, RH, on Mean Time to Failure, MTTF

Concerns and Limitations

In general, the reliability engineer cannot ignore failures observed during testing.

The expense of accelerated testing and the often costly effort required to address the

problems highlighted by such tests are rationalized by the assumption that failures

have a finite probability of occurring during service if they are observed at all during

testing. In other words, the accepted belief is that improved test yields correlate with

longer service life. A device may actually be redesigned without any knowledge of

whether the failure mechanism occurs under operating conditions at a rate that can be

extrapolated from stress conditions or if it is just an artifact of the test. The prediction

of device lifetime in the field by extrapolation of data obtained under high-stress

conditions is extremely sensitive to the model chosen. The basic issue here goes back

to the use of empirical equations—one cannot reliably extrapolate outside the range

of testing conditions. Thus, the goal has to be to obtain data on the behavior in the

mildest conditions possible [97]. The authors of this chapter do not know of a sole

expert in this field who believes at this time that actual service life can be accurately

predicted and, as such, in practice, accelerated aging is presently useful only for

product qualification and identification of latent manufacturing defects and

Corrosion of Data Storage Devices 667

Figure 8 Surface conductivity as a function of temperature. RH is the relative humidity

by IEEE.)

design deficiencies. The remainder of this subsection provides support for these

statements.

Changing Failure Mechanisms as a Function of Environmental Stress This

classical issue is relevant to all applications of accelerated aging [97]. The

complexity inherent in microelectronic devices results in a multitude of coupled

processes that must be properly identified and accounted for in accelerated aging

models. Every process that affects a failure mechanism could have a different

sensitivity to the environmental parameters. Lall [129] has documented some of

the problems associated with the common use of temperature as a stress agent for

microelectronics. The wide variability in activation energy mentioned before is

indicative of different or changing mechanisms. Three relevant examples further

illustrate the issue of changing mechanisms. The first involves the formation of

conductive anodic filaments (CAFs) on PCBs described earlier. Work performed

subsequent to the discovery of CAF formation found that a threshold in the

temperature/humidity phase space may exist below which CAF growth does not take

place [130]. This threshold is apparently not encountered in typical use environments

and thus the phenomenon of CAF may be only an artifact of accelerated testing.

The second example is from an investigation of conduction in printed circuit

boards by Takahashi [130]. He studied the various conduction paths in boards using

an AC impedance technique during humidified exposures without bias and identified

both ohmic and diffusion-controlled processes. He then made the point that

because multiple conduction paths and mechanisms exist at a single temperature

and humidity condition, it is possible for conduction processes determined in

stress tests to have no bearing on failure mechanisms under field conditions. The

final example concerns the observation that the glass-transition temperature of the

encapsulating plastic in PEM devices cannot be exceeded in testing because

a large difference in water permeability results. Nevertheless, sometimes this

requirement is satisfied. A few studies have shown that data obtained from THB

and HAST/pressure cooker testing can be fit to a single mathematical relationship,

indicating that the failure mechanisms may be the same [68,99,104].

Evolving and Improving Technology The continuing rapid evolution of

microelectronic technology leads to several related difficulties. The first is that the

dominant failure mechanisms change and thus long-term field failure information

is not necessarily relevant to state-of-the art devices. Historically, during accelerated

aging, two types of moisture-related phenomena have been observed: distributed Al

track corrosion and Au wire/Al bondpad interfacial degradation. The root cause of

track corrosion was probably moisture penetration through defects in the protective

passivation layer and the presence of contamination. Modern best commercial

practice has effectively eliminated passivation defects and thus track corrosion as a

significant failure mechanism. Now, the exposed wirebonds are the prime

susceptibility. Lall [129] documents another important factor: field failure

information is becoming more limited because the reliability of state-of-the art

devices has now improved to the point that it no longer limits useful system lifetime.

The final factor is that posttest analyses of aged devices have often been limited to

simply confirming a failure, not determining root cause. Thus, the ability to correlate

the results with true service life or identify actual failure mechanisms has not been

developed to any significant extent.

668 Frankel and Braithwaite

Effect of Humidity and Device Type A specific disconcerting aspect of the

existing T/H models that is easy to identify is their wild divergence at low RH

values. Figure 9 shows humidity acceleration factors relative to 85% RH calculated

for each of the expressions in Table 2. In these calculations, temperature was

normalized out by setting its value to 25°C. There is relatively good corres-

pondence of the models in the high-RH range where the experimental data exist.

However, at low RH, the models differ by many orders of magnitude. One possible

explanation is that none of the mathematical relationships proposed have a

phenomenological basis. For example, water adsorption on the metallization

may be the key process and, if so, the rate response should be sigmoidal (which

none of these equations simulate). Osenbach [17] describes another physical

basis that involves the observation that surface leakage current is minimal below

about 50% RH and has an exponential dependence above this level. Few exper-

imenters examine the RH range less than 50% because of the long times needed

to obtain failure. Unfortunately, this RH range is where most operating conditions

exist, especially considering the local heating associated with power dissipation.

Furthermore, there is no reason to expect different systems to exhibit similar

temperature or, for this topic, different humidity relationships. For instance, silicone-

encapsulated devices fail at much lower rates in a given environment than

devices encapsulated in epoxy [60,96,112]. This occurs despite the faster transport

of moisture in silicones and could result from low impurity content and

improved adhesion to the die surface [75,112]. Silicone-encapsulated parts may

therefore display a different humidity acceleration factor than unencapsulated or

epoxy-encapsulated parts [66].

Corrosion of Data Storage Devices 669

Figure 9 Humidity acceleration factor as a function of RH relative to 85% for the various

models shown in Table 2.

Effect of Contamination As discussed previously, contamination at

metallization surfaces is a critical corrosion factor. For example, the level of

contamination affects the ionic strength in the adsorbed water layer that, in turn,

influences surface conductance. Contamination also affects the critical relative

humidity, which, for some salts, can be quite low. Finally, the type of contamination

often determines the corrosion mechanism (e.g., whether electrolytic anodic or

cathodic corrosion of Al predominates). In many accelerated aging approaches, this

dominant factor is uncontrolled and often uncharacterized. In fact, T/H testing may

really be a measure of the cleanliness of the part prior to testing. This observation

is an example of why Osenbach’s second assumption noted earlier (testing of

representative devices) is probably not, in general, valid. To address this problem,

many investigators have either purposely contaminated the part prior to exposure or

added controlled contamination to the stressing environment [42,65,74,106,107,122,

131–133]. These approaches are useful for making comparative assessments of

various structures. However, it is a difficult task to generate a model that predicts

behavior under conditions of lower contamination levels. Pitting potentials of various

metals have been found to decrease linearly with a logarithmic increase in chloride

concentration [134]. Predicting part lifetime for a given surface contamination

concentration from such information, however, remains quite challenging.

CORROSION OF MAGNETIC DATA-STORAGE COMPONENTS

The critical metallic components of advanced magnetic and magneto-optic (MO)

storage devices—thin-film metal disks, inductive or magnetoresistive heads, and

MO layers—are all susceptible to corrosion and each has been a subject of

considerable study. Several review articles covering corrosion of magnetic-storage

media may be found in the literature [135,136].

Thin-Film Magnetic Disks

As described in the technology overview section, the carbon overcoat layer on

thin-film disks typically does not fully cover the underlying layers as a result of

intentional roughening of the disk. The lack of coverage has two implications for the

corrosion behavior. First, the Co-based magnetic layer and perhaps even the NiP

substrate are exposed at small regions and can corrode. Furthermore, the overcoat

layer, which is often sputter-deposited carbon, can be somewhat conductive and quite

noble in comparison with the exposed areas of magnetic alloy. The unfavorable

anode-to-cathode area ratio can therefore result in aggressive galvanic corrosion.

Figure 10 shows potentiodynamic polarization curves measured in DI water

[135,137]. The plated Co-8%P material was considered for use as the magnetic

alloy in thin-film disks when they were first developed. Like pure Co, it is not very

corrosion resistant and does not readily passivate. The corrosion potential of two

different sputter-deposited carbon thin films is seen to be about 600 mV higher than

that of plated CoP. Although the nature of C thin films can change drastically as a

function of deposition conditions, the two C films sustain reasonably large cathodic

670 Frankel and Braithwaite

currents. In fact, galvanically induced corrosion of the small active areas of the

magnetic alloy by the C overcoat is the primary mechanism of corrosion of thin-film

disks. When an uncoated CoP film is exposed to an aggressive gaseous environment

at 25°C containing 70% RH and 10 ppb Cl

gas, a 30-Å-thick uniform corrosion

product forms [135,137]. However, a C-coated CoP sample forms a high density

of localized corrosion product particles that are several μm in diameter. This

dimension is 100 times larger than the head-disk separation in advanced disk files.

Corrosion can thus be minimized by decreasing the galvanic mismatch between

the magnetic layer and the overcoat. As shown in Figure 10, a sputter-deposited

Co-17%Cr film has a corrosion potential much closer to C and is spontaneously

passive in the DI water droplet. In deaerated Na

, the corrosion current was

found to decrease and the corrosion potential increased as the Cr content in CoCr

alloys increased from 0 to 20% [136,138]. Nonconducting overcoats can also

reduce corrosion but do not eliminate it if they are not totally covering, a requirement

that is quite difficult given the thickness limitations of the overcoat and the roughness

of most disks.

The trend in magnetic recording is, of course, to continually higher densities.

As with microelectronic devices, higher density is achieved by shrinking the

dimensions, including the separation of the magnetic medium and the sensor in the

head, which has implications relative to corrosion. In order to bring the head and

disk closer, the head must fly closer to the disk, and the carbon overcoat thickness

must decrease. The decrease in fly height means that less corrosion product

accumulated on the disk surface will cause detrimental interactions with the head.

Adecrease in carbon thickness is also potentially deleterious to corrosion resistance

Corrosion of Data Storage Devices 671

Figure 10 Potentiodynamic polarization curves of thin-film disk materials in a droplet

of DI water. Parameter definitions: DCC DC sputtered C; RFC RF sputtered C; CoCr, Co +

l7%Cr; CoP, CO + 8%P; Co, pure Co. (From Ref. 135.)

because it is harder to cover the magnetic layer with a very thin carbon overcoat. On

the other hand, in order to fly closer, the disk roughness must also decrease. The trend

toward smoother disks is beneficial to corrosion resistance because the magnetic

layer can then be more easily covered by the carbon overcoat. Figure 11 shows the

time trends for carbon overcoat thickness and disk roughness average (RA), as well

as the ratio of the two. The average roughness of magnetic disks has decreased faster

than the carbon overcoat thickness. Therefore, the ratio of COC thickness to disk

roughness increased significantly during the 1990s, indicating that the COC should

now provide a better covering despite being much thinner. In essence, the disk

structure now tends to be more like Figure 3a than 3b. As a result, disk corrosion

should be less of a problem than in the past, and this trend will apparently continue

for the next few years.

Figure 12 presents polarization behavior in a droplet of water for more recent

disk materials and structures than those shown in Figure 10 [139]. The disk had low

roughness and a thin carbon overcoat. First note the two curves from full disk

structures. One is from a lubricated disk, the other from an unlubed disk. The lube

decreases the current only slightly, perhaps as a result of water displacement. The

anodic current flowing at a given potential from the unlubed disk is about 100 times

less than that of the blanket magnetic layer on glass. It is clear that the magnetic layer

is rather well covered by carbon. In the past, it was assumed that the anodic

current measured from a disk originated largely from spots not covered by COC.

However, for the disks studied here, the anodic signal is not coming only from the

small amount of uncovered magnetic layer because a blanket layer of C (unlubed) on

glass exhibits an electrochemical signal that is a large fraction (about half as large) of

the signal measured on the unlubed disk. Carbon is not really an extremely noble

material. More accurately, the oxidation of C is kinetically hindered. Carbon can, in

fact, oxidize to form various species, such as carbonyl, carboxyl, phenol, quinone, or

672 Frankel and Braithwaite

Figure 11 Trend of carbon overcoat (COC) thickness and disk texture or average roughness

(RA) with time. Also shown, with right axis, is the ratio COC/RA. (Data from HMT

Technology, Inc.)

lactone [140]. The polarization curve for the carbon film is similar to that of a

spontaneously passive metal. It supports a reasonable amount of cathodic current,

which is the cause of galvanic corrosion. At anodic potentials, there is some steady-

state oxidation current, which is probably associated with a low rate of C oxidation.

From the polarization curves, it could be suggested that the difference in current

between the unlubed disk and the C blanket is the corrosion current of the magnetic

layer. The finding that carbon oxidation is now a significant fraction of the total

current flowing to a thin-film disk is reasonable given the trend of the increasing ratio

of carbon thickness to disk roughness. However, despite the good coverage, very

small areas of magnetic layer are still exposed in the disk structure. Therefore,

corrosion will still occur if the disks are exposed to aggressive environments (such as

chloride) or if problems in disk texture control occur.

An electrochemical study of the protectiveness of carbon overcoats as thin as

5 nm has been reported [141]. Polarization resistance of disk structures in

0.5 M sulfuric acid was measured. Disks coated with hydrogen-containing ion

beam deposited carbon were found to have much lower corrosion rates than those

coated with standard sputtered C of equivalent thickness. The composition of the

sputtered carbons was varied by changing the sputtering gas. The corrosion rate

depended on the type of carbon: ion beam C-H < sputtered C-H, < sputtered

C-N < sputtered C. The different corrosion rates may be a result of differences in

both coverage and catalytic activity of the carbon layers.

Novotny and colleagues [136,142] have described methods for estimating

lifetimes of disks in drives using accelerated T/H tests. In the first approach, the

extent of corrosion of a disk following T/H exposure was determined by a measurement

of the surface Co concentration on the disk surface using Auger electron spectroscopy

Corrosion of Data Storage Devices 673

Figure 12 Polarization curves in a water droplet for a lubed and unlubed disk as well as

a blanket magnetic layer and carbon layer on glass. (From Ref. 141.)