August_EDFA_Digital

edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 22 NO. 3 12 In the following sections, the case of merging different products impacted by the same failure is studied. CASE STUDY: PREDICTING EXPECTED FAILURE QUANTITY FROM FIELD RETURNS Recently, a rare case happened of a same failure faced on a dozen automotive semiconductor products. During a short period of a time, a manufacturing machine was improperly working and submitted parts to an abnormal phosphorus rate. This phosphorus generated an inter dielectric delamination phenomena, more or less impor- tant according to the design of the impacted products (more precisely according to edge seal design andnumber of metal layers). Story following can be easily imagined: humidity started to penetrate into dies, leading to field failures. [3] Previous studies performed on this topic showed that the model that provided the best fitting on field failures of the dozen of impacted products, was mostly a defec- tive subpopulation (DS) Frechet model. [4,5] Of course, this samemodel reinforces the previous findings about a same failure. The concern is that sensitivity-to-failure-level is different from a product impacted to another one, so that DS Frechet model parameters are different for all the products: a Wilcoxon group homogeneity test launched on the elementary models statistically does prove this difference between the products in term of field model, actually linked to this different sensitivity. Thus, it is not so obvious to tell that the failure seen by the products is really the same one, without having conducted a rigorous structured problem solving approach as presented previ- ously with 8D, Is-Is Not, 5 x Whys tools. A Frechetmodel is widely similar to the typical Weibull model in the extent that it is fitting with a specific case of generalized extreme value distribution, as Weibull distri- bution. Frechet density f ( t ), distribution F ( t ) and quantile equations are as follows: where: t is time, µ location, σ scale parameter, q quantile probability. What ismore interesting is tohighlight the linkbetween aDSmodel and the physical phenomenon. TheDSdensity f’ ( t ) and DS distribution F’ ( t ) equations are as follows: A DS model is fitting with only a population portion failing: this portion is expressed in the p parameter seen in the formulas. But a DS model also corresponds to a failure rate that increases thendecreases fromamaximum failure rate event. This feature may be fully explained by the failure itself and how it happens: indeed, failure rate increases as humidity rate increases in the dies through delamination. But, humidity rate also decreases with engine heating, in field, and failure rate may decrease, too. So, a competitive situation is set between humidity increase in timeand its decreasewithmileages cumulated. Typically, there is no competitive situation between time and mileage, so that field model density and distri- bution can be expressed indifferently in time or mileage unit in the extent that a conversion coefficient is used from a standard driver’s behavior driving about 1200 km per month. In this delamination case, amodeling per portion would be preferable, but is made impossible by the fact that themaximumfailure rate event is variable, depending on car usage or on product application, and should have also to be defined within a confidence interval to take its variability into account: the lesser the car utilization or heat generated by product application, the higher the humidity rate, the faster failures happen, and the slower failure rate is seen decreasing; at the inverse, if car ismore used or if heat given off by component in application is high, failure rate will be seen quickly decreasing. A DS model allows this increasing then decreasing failure rate, and confidence interval onmodel parameters can express variability on maximum failure rate event. In this delamination case, DS equations are expressed in time and time 0 is set in failingmanufacturing stepwhen delamination is created. From delamination generation and before product life start in field, there is a period duringwhich, even if humidity starts to penetrate, failures cannot happen on the parts not yetmounted in cars: thus, in field, the first life period sees an abnormally higher failure rate generated by this humidity excess: this early life period of about 600 days, fitting with about a 2 year life duration, is cautiously excluded from field modeling. While a DSmodel perfectly explains the failure, predic- tions stay difficult: because of ‘time versus mileage’ com- petition referred previously, maximum failure rate event may be badly estimated, and the fitting maximum failure quantity may be wrongly over or under estimated (see