Downloaded from https://www.everyspec.com MIL-HDBK-816 NOTICE 1 of time has elapsed between the characterization measurements and the data analysis or when the person or organization undertaking the data analysis is different from the persons who took the data in the first place. Some of the most common reasons for suspecting that data may be bad are discussed below. 5.2.4.1.1 Outlying devices. It is not uncommon, in a batch of measurements, to find that one or more devices out of the sample show failure fluences or post-irradiation parameter values that are significantly outside the values that would be expected from the mean value and variance for the rest of the sample.
(Outlying devices may occasionally be found also in the pre-irradiation measurements. The assumption here is that the examination of the pre-irradiation data has eliminated any such devices from the characterization sample.) Although the identification of an “outlier” can sometimes be difficult, a plot of the cumulative probability distribution for the radiation fluences at which the devices fail (or the post-irradiation parameter values) will usually show that the device in question is not part of the population distribution that is characteristic of the rest of the sample. A more detailed discussion of cumulative probability plots may be found in textbooks on statistics.
MIL-HDBK-814 also contains an appendix on statistical techniques. A discussion of how to identify outliers may be found in ASTM E178. 5.2.4.1.2 Outlying lots. In this situation, data is available for a number of lots and an examination of the data shows that one or more of the lots are showing an unusual mean value or standard deviation. Again, a cumulative probability plot can be made of the data to evaluate whether the suspect lot or lots belong to a different population of lots. 5.2.4.1.3 Sufficiency of data. The first step in preparation for data analysis is to see if there is sufficient data for deriving end-point limits.
The recommendation in these guidelines is that the characterization sample should consist of three different lots taken at least one month apart, with five wafers per lot and five parts per wafer taken one from each quadrant and one from the center. The minimum sample size that is recommended is five different lots with five samples per lot. The data should also contain a range of values and radiation stresses. Thus, for example, if step-stress measurements were performed and all the parts failed in a single “bin” (where a bin is defined as the interval between one fluence and the next higher fluence), there is not enough data for an analysis.
If it happens that there is no way to acquire sufficient data, special analysis techniques may be required or some worst case assumptions may be needed to supplement the data. Sometimes it may happen that the number of lots used for obtaining the characterization data is not given in the data. For such cases, a method exists for estimating, from the data itself, what the effective sample size of the device population is and this effective sample size can be introduced into the end-point limit calculations (A. I. Namenson, 1979, see 2.3). 5.2.4.1.4 Abrupt failures.
The data should be examined to see if abrupt failures are likely at the radiation levels that will be required for the lot acceptance tests. If abrupt failures are expected to be the dominant failure mode at these radiation levels, then the mean value and the standard deviation of the fluences to abrupt failure in the characterization data should be used to estimate the radiation level at which the part can be used safely. The fluences to abrupt failure are then used to check for outlying devices, outlying lots, and the sufficiency of the data. 5.2.4.1.5 Unexpected functional dependencies.
If a parameter value does not vary smoothly with either bias conditions or radiation dose, the data may be suspect. 5.2.4.1.6 Systematic or nonrandom effects. The sampling of lots and parts must be a truly random representation of future lots if valid lot acceptance end-points are to be determined. It pays, therefore, to check whether the data is random in nature. Some typical causes of nonrandom effects are discussed in the sections which follow. 5.2.4.1.6.1 Correlations with respect to time. Lot-to-lot variations should be checked to see if they correlate with time.
Examples of such correlations are an average shift with respect to date of manufacture or simply a smooth variation with respect to the date of manufacture. Usually, a look at the data is sufficient to determine if such an effect is occurring. However, more rigorous techniques, such as time-series correlations, for example, exists which can check whether the data varies smoothly with respect to time or has seasonal variations. 5.2.4.1.6.2 Unexpected magnitudes of lot-to-lot variations. If lot-to-lot variations of either the mean value or the standard deviation are much larger or much smaller than expected, it is cause for concern.
If a number of lots are very similar, then it may be that the sampling over lots is not truly random. There may be, SUPERSEDES PAGE 30 OF MIL-HDBK-816 30