space-rha.comShort courseChapter 8 · Data analysis and interpretation

Chapter 8 · Lesson 3 of 4 · about 18 minutes

Dose endpoints and population coverage

A part that passes the final exposure has a failure dose somewhere above it, and a part that passes one step and fails the next has a failure dose somewhere between them. Keep those distinctions rather than assigning every part an exact threshold, and do not let the average hide an early failure.

What you need to know

A cumulative test produces, for each part, a trajectory and a failure dose, the dose at which the parameter the circuit depends on left the application limit. A part that passes the final step has a failure dose somewhere above that step (right-censored). A part that passes one step and fails the next has a failure dose somewhere between them (interval-censored), and assigning it the midpoint invents resolution the test did not have. The circuit sets the application limit (chapter 7 lesson 3), and that limit is generally reached before the datasheet limit, so the same test can give different failure doses for different applications of the same part.

The test sample was drawn from the population of flight parts, and the claim the project needs is about the weak members of that population rather than its average. Three statistical statements answer three different questions. A confidence interval for the mean describes the center of the population. A prediction interval describes one more part. A tolerance bound describes a stated fraction of the population at a stated confidence, and that is the form a hardness claim takes. With n failure doses of sample mean x̄ and standard deviation s, the one-sided lower tolerance bound is x̄ − k s, where the factor k depends on n, on the covered fraction P and on the confidence γ, and comes from the noncentral t distribution. For five parts, k is 2.742 for 90 % of the population at 90 % confidence and 4.203 for 95/95. The figure shows how k falls as the sample grows. Organizations use the bound divided by the required dose (mission dose times the design margin) as a categorization criterion. Below one, the lot has not demonstrated the requirement at that coverage.

The bound assumes that the failure doses are normally (or lognormally) distributed, and five points cannot test that assumption. A mixture of lots, a bimodal process or one weak die can defeat it, and failing to reject normality with a tiny sample says little about the tail. The distribution-free alternative needs more parts (22 for 90/90, chapter 6 lesson 1), so the tolerance method is used instead, with its assumption stated and with provenance (one lot, one date code, one wafer where known) to support pooling the parts. The lognormal form is often closer to the physics for a parameter that degrades multiplicatively, and the two bounds are generally reported together.

Dose rate comes back in at this point. If the sample consists of high-dose-rate failure doses and the part is ELDRS-sensitive, the low-dose-rate characterization supplies a factor, the ratio of the LDR failure dose to the HDR failure dose on the parts that were run both ways. The bound is scaled by that factor, or the LDR data are used directly. For U3 the LDR leg (to 30 krad, lesson 6.2) crossed the 100 nA limit on most parts before 30 krad, at about 50 % of each part's HDR failure dose, and that factor is applied below.

Figure 8.3. One-sided normal tolerance factor k against sample size for three coverage and confidence combinations, from the noncentral t distribution. Five parts cost about one standard deviation more margin than twenty.
One-sided normal tolerance factor against sample size 5 10 15 20 25 30 0 1 2 3 4 5 6 7 8 parts tested k: lower bound = mean − k × standard deviation 99 % of the population, 90 % confidence 95 / 95 90 / 90
Try it: a tolerance bound for a TID failure-dose sample

Enter the failure doses, choose the covered fraction and the confidence, and compare the bound with the required dose. The lognormal bound is computed from the logarithms of the same doses.

Worked example 8.3. Five U3 parts against a 30 krad requirement

The five high-dose-rate failure doses of U3 (bias current beyond 100 nA, each interpolated on the part's trajectory between the bracketing 5 krad steps of lesson 7.3) are 42, 55, 61, 48, 70 krad(Si), with mean 55.2 and standard deviation 10.9. The mission dose is 15 krad behind the shielding, and the required dose with the design margin of 2 is 30 krad.

Coverage / confidencek (n = 5)Lower bound x̄ − k s, kradBound ÷ 30 krad
90/902.74225.20.84
95/954.2039.20.31
99/904.6664.20.14

Every part failed above 42 krad, yet the 90/90 bound is 25.2 krad, below the 30 krad requirement. The sample is too small and too scattered to demonstrate that 90 % of the lot exceeds 30 krad at 90 % confidence. The lognormal bound (calculator) is somewhat higher, but the conclusion stands. The ELDRS factor of 0.5 lowers the HDR bound further, to about 13 krad at 90/90, and that is the number the low-dose-rate flight exposure has to be judged against. The options are those of chapter 11: more parts from the flight lot, more shielding at U3 to lower the mission dose, a different part, or a documented acceptance of the residual risk with the numbers above in the record. The distribution-free statement from the same sample is weak. Five passes at 30 krad would give only 41 % confidence that fewer than 10 % of the lot fails below it.

A question to carry through the course

When, if ever, can the dose applied to a part during a proton SEE test be used as a proxy for a TID test?

Things to weigh here:

  • How many parts a proton SEE run exposes, typically two or three, against the sample a tolerance bound or a lot test needs.
  • Whether those parts came from the flight lot, and whether the parameter the circuit depends on was measured at all during the run.
  • For a bipolar part, that a proton run at high dose rate says nothing about the low-dose-rate response.
  • What the exposure can still support: a screen, or supporting evidence for the lot, rather than the lot-acceptance test itself.

Exercise

Keep each part's trajectory and its censoring, and tie each failure dose to the requirement of the circuit that uses the parameter.

State which population the bound describes, the fraction covered, the confidence, and the evidence for the distribution you assumed.

Going deeper

Resolution and independence the data do not have

Placing every between-step failure at the midpoint understates the uncertainty. Repeated measurements of one part at successive doses are correlated observations of that part, not new independent samples of the lot. A model of the degradation curve and a model of the distribution of failure doses answer different questions, and both depend on assumptions about monotonic behavior, measurement error, sampling and extrapolation beyond the measured dose or time.

Keep the parameters that shift together as a set, so that the circuit analysis can evaluate a part state that could actually occur rather than a collection of unrelated worst cases.

Mixtures and transfer

A mixture of lots, a broad or bimodal distribution, or a thick tail can defeat the small-sample models that are convenient to use. Failing to reject normality with five parts is weak evidence about the far tail. Use the provenance of the parts, the variation seen historically and what is known of the mechanism to decide whether pooling is justified, and check how much the bound moves under a plausible out-of-family population instead of accepting the favorable model by default.

A fitted lower tail depends on the assumed relation between the sample and the flight population and on the model used. More data from the same sample do not test that relation.

The common mistake

A part that survived the last step gives a lower bound on its failure dose; its failure dose itself is not known.

No statistical formula can show that the assumed distribution is right or that the pooled parts came from one population.

Takeaway

  • A cumulative-effects claim is about the response of the lot against the application limit, and the average failure dose of the sample is only part of that.
  • The claim the project needs is about the flight population, and in particular its weakest members.

Check yourself

Why does the 90/90 bound fall below 30 krad when the lowest failure in the sample is 42 krad?

Because the bound describes the lot rather than the sample. With five scattered points the tenth percentile of the population is uncertain, and the factor k = 2.74 charges that uncertainty against the mean. More parts (k = 1.77 at twenty) or less scatter would raise the bound.

Which failure dose does a part that passed 45 krad and failed at 60 krad contribute?

An interval-censored value between 45 and 60 krad. A likelihood-based fit uses the whole interval; a quick analysis states which end it used and why. The midpoint is a convention, not a measurement.

What does the normal tolerance bound assume that the sample cannot verify?

That the lot's failure doses follow one normal (or lognormal) distribution: a single population, no mixture, no bimodality. The supporting evidence is provenance (one lot, one date code) and the manufacturer's knowledge of the process.

A colleague computes the 95 % confidence interval for the mean failure dose of the five U3 parts, finds it well above 30 krad, and concludes that the lot meets the requirement. Which question did that interval answer, and which one did the project ask?

A confidence interval for the mean describes the center of the population, a prediction interval describes one more part, and a tolerance bound describes a stated fraction of the population at a stated confidence, which is the form a hardness claim takes. The project asked about the weak members of the flight lot, so the relevant number is x̄ − k s with k = 2.742 for 90/90 at n = 5, which gives 25.2 krad against the 30 krad requirement even though the sample mean is 55.2 krad. A tight interval on the mean generally says little about the weak tail of the lot.

Why does the course use a tolerance bound that assumes a normal or lognormal distribution instead of a distribution-free statement, when five parts cannot verify the assumption?

The distribution-free alternative needs more parts than the sample has, 22 for 90/90 (chapter 6 lesson 1), and from five parts it is weak: five passes at 30 krad would give only 41 % confidence that fewer than 10 % of the lot fails below it. The tolerance method buys a stronger statement with the assumption of one normal (or lognormal) population, so the assumption is stated, provenance (one lot, one date code, one wafer where known) supports pooling the parts, and the normal and lognormal bounds are generally reported together. The bound is also checked under a plausible out-of-family population rather than accepting the favorable model by default.

The 90/90 bound of 25.2 krad was computed from high-dose-rate failure doses. Why is it not the number the flight exposure of U3 is judged against, and where does the correction come from?

U3 is ELDRS-sensitive, and the low-dose-rate leg of lesson 6.2 crossed the 100 nA limit on most parts before 30 krad, at about 50 % of each part's HDR failure dose. That factor of 0.5 lowers the HDR bound to about 13 krad at 90/90, and that is the number the low-dose-rate flight exposure has to be judged against; alternatively the LDR data are used directly. The factor comes from parts that were run both ways and applies to this part and this parameter, which is also why a proton SEE run at high dose rate says nothing about the low-dose-rate response of a bipolar part.

Further reading

  • MIL-HDBK-814, "Ionizing dose and neutron hardness assurance guidelines for microcircuits and semiconductor devices," 1994.
  • D. B. Owen, "A survey of properties and applications of the noncentral t-distribution," Technometrics, vol. 10, no. 3, pp. 445-478, 1968.
  • R. Ladbury and M. J. Campola, "Statistical modeling for radiation hardness assurance: toward bigger data," IEEE Trans. Nucl. Sci., vol. 62, no. 5, pp. 2141-2154, 2015.
  • R. L. Pease, R. D. Schrimpf and D. M. Fleetwood, "ELDRS in bipolar linear circuits: a review," IEEE Trans. Nucl. Sci., vol. 56, no. 4, pp. 1894-1908, 2009.
  • M. A. Xapsos et al., "Inclusion of radiation environment variability in total dose hardness assurance methodology," IEEE Trans. Nucl. Sci., vol. 64, no. 1, pp. 325-331, 2017. doi:10.1109/TNS.2016.2607021.
  • MIL-STD-883, Test Method 1019, "Ionizing radiation (total dose) test procedure."