space-rha.comShort courseChapter 5 · BOM scrub and parts review

Chapter 5 · Lesson 3 of 6 · about 12 minutes

Finding evidence and deciding whether it transfers

Search by manufacturer and exact part first, then by documented aliases and related technology if you have to. Public databases, vendor reports and workshop records can save a test campaign, provided you keep the original report so another engineer can see what was measured.

A search hit is a lead. Open the report.

Start the search with the manufacturer and the exact part number. Widen it to documented aliases and related technology only when the exact search comes up empty, and say so when you do. Public databases, vendor reports and workshop records can save you a test campaign. Keep the original report and its revision with the finding, so another engineer can see what was actually measured.

An empty search result means only that you found nothing. It is neither evidence of immunity nor, by itself, a reason to test every part.

Capture the evidence, not only the headline.

From each report, pull out the device identity, the number of samples and where they came from, the source and its dosimetry, the particle energy or the LET and range, the bias, temperature and operating modes, the fluence or dose reached, the endpoints that were measured, what recovery was observed and what the test could not see. Note which statements are measured results, which are vendor guarantees, which are analogies to other parts and which are expectations from the population of similar parts. A database summary can leave out the one condition that decides whether the data apply.

If a report is missing something you need, ask the author for it before you repeat the irradiation.

Tested does not mean tested for this use.

Put the tested device and the flight application side by side and compare them on identity, population, radiation source, exposure, operating conditions, the function that was observed and the recovery behavior. For each line, ask whether the test answers the question you need answered. A mismatch is a reason to investigate further, and an unknown is something to go and resolve.

The comparison exists to surface questions. Counting the rows that match and treating the count as a qualification score produces a number that means nothing.

Make each mismatch actionable.

For every difference between the test and the flight use, write down why it matters and what would close it: an analysis, information from the supplier, a targeted test or a design change. A good match on one variable does not make up for a failure mode the test could not see. A test run at a different condition can sometimes bound the flight condition, but only with a physical argument that is specific to the effect.

Record the claim the data support and the claim that is still open.

Exercise on the demonstration board

Select U1 on the demonstration board. Its no-SEL result was taken at one operating condition. Now suppose the flight mode in question is the hot case at full load. Decide which rows of the side-by-side comparison that changes and what verification it opens.

Going deeper

Evaluate how the record was constructed.

Look for censored failures, a stopping rule that ended the run early, excluded events, dead time in the monitoring, coarse measurement resolution, error classes lumped together and configuration changes partway through the test. What is known about similar parts can guide the plan, but it is not a measurement of this part. A vendor guarantee covers a stated scope under stated conditions, and the test evidence has to be tied to the specific claim you want to make.

When two reports disagree, keep both and explain the difference. Choosing the more favorable headline hides the disagreement without resolving it.

Check what the test could see, and whether its response transfers.

A test with zero events constrains only the events that the instrumentation and the workload could have detected, under the conditions that were run. Before you transfer the result, establish that the beam reached the sensitive volume with enough energy and range, that the angles covered are the ones that matter, that the test bias and temperature map onto the flight ones, how the response depends on those conditions, and whether the samples represent the parts you will fly. Put a number on the uncertainty only after you have settled the model and the class of event it describes.

More fluence at the same setting tightens the count. It does nothing for a mode the test could not see, a geometry it did not cover or a population it does not represent.

The common mistake

The typical behavior of similar parts, or the highest level a part has ever passed, cannot stand in for the source report. Identical test conditions are not required for a physical bound, and they are not enough for one either; the bound needs the argument.

Takeaway

  • Use the databases to find the evidence, then read the report to learn what it covers.
  • Compare the test and the flight use on the conditions that control the mechanism.

Check yourself

A search of the public databases for the exact part number returns nothing. One colleague takes that as a sign that the part is probably fine; another says every such part has to go to the beam. What does the course say?

An empty search result means only that you found nothing; it is neither evidence of immunity nor, by itself, a reason to test every part. Widen the search to documented aliases and related technology, and say so when you do; what happens next depends on what could stop an important function and on which piece of information would change the design decision (lesson 5.5). A missing record is a finding, not a susceptibility and not an immunity (lesson 5.4).

A database summary gives a part as free of latchup to a high LET. Why does the course insist on opening the original report, and what would you pull from it?

A database summary can leave out the one condition that decides whether the data apply. From the report, pull the device identity, the number of samples and where they came from, the source and its dosimetry, the particle energy or the LET and range, the bias, temperature and operating modes, the fluence or dose reached, the endpoints measured, the recovery observed and what the test could not see, and note which statements are measured results, which are vendor guarantees, which are analogies to other parts and which are expectations from the population of similar parts. Keep the report and its revision with the finding, and if something you need is missing, ask the author before repeating the irradiation.

A reviewer builds the side-by-side comparison of the test and the flight use, counts the rows that match and reports the fraction as a qualification score. What is wrong with that?

The comparison exists to surface questions; counting the rows that match and treating the count as a score produces a number that means nothing. A good match on one variable does not make up for a failure mode the test could not see, and for every mismatch the finding should say why it matters and what would close it: an analysis, information from the supplier, a targeted test or a design change. Record the claim the data support and the claim that is still open.

The archived test of a part reached a modest fluence, so the team proposes to repeat it at a much higher fluence to strengthen the evidence. When does that help, and when does it not?

More fluence at the same setting tightens the count; it does nothing for a mode the test could not see, a geometry it did not cover or a population it does not represent. A test with zero events constrains only the events the instrumentation and the workload could have detected under the conditions that were run, so before transferring the result establish that the beam reached the sensitive volume with enough energy and range, that the angles covered are the ones that matter, that the test bias and temperature map onto the flight ones, and that the samples represent the parts you will fly. Put a number on the uncertainty only after the model and the class of event it describes are settled.

U1's no-SEL result on the demonstration board was taken at reduced supply and load at 25 °C (lesson 5.1), and the flight mode in question is the hot case at full load. Can the test bound the flight condition?

Not without an argument, and the argument runs the wrong way here. A test at a different condition can sometimes bound the flight condition, but only with a physical argument specific to the effect, and for latchup the course treats higher temperature and higher voltage as the more severe condition (lesson 5.4), so a cooler, lower-bias run does not bound the hot, full-load case. The rows for exposure and operating conditions change, and the verification it opens is a screen at the flight corner, maximum voltage and 125 °C with ions that reach the wells (lesson 5.2), or a manufacturer record of the exact die under those conditions, as in the archive report of lesson 5.4.

Further reading