space-rha.comShort courseChapter 6 · Evidence strategy and facility selection

Chapter 6 · Lesson 1 of 4 · about 20 minutes

Decide what evidence the project needs

A beam run is worth its cost when the result changes what the project can defend. Start from the function, its radiation failure mode and the consequence, then decide whether the project needs a screen, an application bound, a response curve or a recovery demonstration.

What you need to know

A radiation test costs beam time, prepared samples, test hardware, people and calendar, and it is worth running only when the result changes what the project can defend. Before anyone asks which facility to book, the project has to decide which claim it needs to make. Four kinds of claim recur, and each calls for a different experiment.

Write the analysis before the run plan. For each consequential failure mode, record the quantity the follow-on analysis needs (an event count with its live fluence, a voltage limit, a parameter trajectory, a recovery time) and connect each planned run to one of them. When the beam is late, a run that feeds no analysis is the one to drop.

Three numbers, all fixed before the beam is on, decide how strong the result can be. The first is the fluence. With no events observed, the 95 % one-sided upper bound on a cross section is 3.00 divided by the live fluence (chapter 8 lesson 1), so a bound of 10⁻⁷ cm² needs 3.0 × 10⁷ ions per cm², and a bound ten times tighter needs ten times more fluence. The second is the number of parts. When every part in a sample passes a dose, the sample supports, with confidence C, only the statement that the fraction of the population failing below that dose is less than p, where n = ln(1 − C) / ln(1 − p). For 90 % confidence that fewer than 10 % fail, n = 22 parts; for 95 % confidence that fewer than 5 % fail, n = 59. Five parts support only weak distribution-free statements, so the tolerance methods of chapter 8 lesson 3 assume a distribution to get further. The third number is the set of conditions covered. More fluence or more parts cannot repair a missing condition afterward.

An existing report is a lead. Before you reuse it, compare what was tested with what will fly: the manufacturer, die revision and process (the mask set generally governs single-event response); the lot and date code (total dose and displacement response generally vary lot to lot, and for bipolar linear parts the dose-rate sensitivity varies too); the package and any backside thinning (which change what an ion reaches); and the bias, temperature, load, data pattern and operating mode of the test against the application. A report that omits the counts, the fluence, the live time or the recovery procedure can still support a screen, but it cannot support a rate. Record the basis and the limits of every transfer in the evidence ledger, so that a procurement change reopens the right comparison.

Try it: what a clean run can and cannot say

Enter the events counted and the live fluence of a run. The table gives the cross section with exact Poisson limits; the note below it gives the fluence a target bound needs before the beam is booked.

Worked example 6.1. The evidence U1 and U2 need

U1, the shared regulator, powers the recovery controller, so a high-current event on U1 is a question of what is lost. The project needs a bound on the rate of high-current events at the flight voltage and 125 °C, and a demonstration that the current limiter and the independent recovery supply restore the function. U2, the recorder memory under SEC-DED, is a question of how often. The project needs a cross-section curve against LET for upsets and for functional interrupts, and a proton cross section, because trapped protons dominate its rate in low Earth orbit (chapter 9).

Part and modeClaim typeQuantity the analysis needsMinimum experiment
U1 high-current eventresponse curve and boundevents and live fluence per LET at flight bias and 125 °C; no destructive outcomesix LETs to about 70 MeV·cm²/mg (xenon at 30° for the last), 10⁶ ions per cm² each, at two bias conditions; a clean run bounds σ below 3.0 × 10⁻⁶ cm²
U1 recoverydemonstrationcurrent-limiter trip time, rail decay, restart time, function after restartthe flight limiter and recovery supply on the bench, ten forced events with waveforms
U2 SEUresponse curveupsets per run with fluence, pattern and mode; multiplicity per eventseven LETs from the lowest cocktail ion up; 2 × 10⁴ ions per cm² at each LET above the onset
U2 SEFIresponse curveinterrupts per run, recovery step that cleared each10⁶ ions per cm² per LET, so that a clean run bounds σ below 3.0 × 10⁻⁶ cm²
U2 protonsresponse curvecross section at 200 MeV and one lower energy; TID accumulatedtwo energies, 10¹⁰ protons per cm² each; parts retired afterward
U3 TIDpopulation boundinput bias current against dose per part, HDR and LDR22 parts would support a 90/90 distribution-free claim; the project can afford five, so chapter 8 lesson 3 uses a tolerance bound

The U3 line shows the trade most projects face. The distribution-free sample size is unaffordable, so the claim will be built on five parts and a normal (or lognormal) assumption about the failure doses. That assumption is part of the evidence, and it is written into the plan now rather than discovered at the review.

The four teaching parts in this chapter

Four constructed parts run through chapters 6 to 12 so that each chapter's work can be followed on the same hardware: U1, a shared regulator that powers sensing, logic and the recovery controller; U2, a 1 Gbit recorder memory under SEC-DED error correction and periodic scrubbing; U3, a precision analog chain with an accuracy allocation; and U4, an optical interface that must keep its signal margin. Their response curves are stated in the build (constructed from a declared true curve with Poisson sampling), so that every number in these chapters can be traced; they are teaching cases, not commercial devices.

PartFunctionThis chapter's step
U1 · shared powerA shared regulator powers sensing, logic and the recovery controller.Decide which high-current and functional modes the evidence has to tell apart. Check ion penetration, the bias and temperature conditions, and what the test can observe before choosing a source.
U2 · recorder memoryA memory stores science data under ECC and periodic scrubbing.Define the endpoints: bit, physical event, logical word and SEFI. Choose the patterns, modes and readout that expose both multiple-bit events and interrupts.
U3 · precision analogA precision analog chain has to hold its accuracy and timing allocation in the application.Set the circuit error endpoint first. Justify the TID source, dose rate, bias, temperature, measurement timing and any transfer between lots, and ask the SET question separately where it applies.
U4 · optical interfaceAn optical interface has to deliver enough signal margin for the whole mission.Choose the material, the spectrum and the electrical or optical endpoint. A proton or neutron fluence generally does not produce the same displacement damage in every emitter and detector.

Exercise

For each consequential failure mode, write down the claim the test has to support before you choose a source.

For every archived report you reuse, record in the evidence ledger what the transfer is based on and where it stops applying.

Going deeper

Write down what the test assumes

A test is an inference about a population from a sample, so the plan should say which population, which operating conditions and which mechanisms the experiment can reach. Bounding conditions help to reveal a susceptibility, and application tests show what it would do to the circuit. Neither covers a mode that was never exercised, a signature the instrumentation could not see, or a technology other than the one tested.

Keep the question of whether the model is right separate from the question of what its parameters are, and write down which result would show the proposed assurance argument to be wrong.

A transfer comes with conditions

More event counts reduce the counting error, but they say nothing about whether a different population responds the same way. Pooling data across lots, or updating a prior with new data, requires a justified relationship between the populations and a prior that can be defended. Keep unlike lots and incompatible configurations identifiable in the data, so that an attractive combined estimate cannot hide a vulnerable subgroup.

A report that omits counts, fluence or recovery details may still support a screen, but not a quantitative prediction.

The common mistake

No campaign of finite size exercises every mechanism or every state of a complex device.

A customary sample size does not by itself show that a result transfers, and it does not remove process and lot variation.

Takeaway

  • The claim the project needs decides the source, the sample, the instrumentation and when to stop.
  • A result reused from the wrong population can outweigh every later statistical refinement.

Check yourself

A part number matches a report in the archive. What three things must match before the SEE result transfers, and what additionally for a TID result?

For SEE: the die revision and process (mask set), the package or die preparation that sets what the ions reached, and the bias, temperature and operating mode. For TID: the lot (and date code) as well, because dose response generally varies lot to lot, and for bipolar parts the dose rate of the test.

A run plan calls for 10⁶ ions per cm² at each LET. What is the tightest cross-section bound a clean run can state at 95 % confidence?

3.00 divided by 10⁶, about 3.0 × 10⁻⁶ cm². A tighter bound needs proportionally more fluence, and a clean result cannot be tightened afterward.

Why is a lot of five parts unable to support a 90/90 claim without a distribution assumption?

With all five passing, the distribution-free confidence that fewer than 10 % of the population fails below the dose is only 1 − 0.9⁵ = 41.0 %, and a 90/90 claim needs 22 parts. The tolerance methods reach the claim with five parts by assuming a normal or lognormal failure distribution, and that assumption becomes part of the argument.

To make up for testing U2 at one bias and one temperature only, a colleague proposes to raise the SEFI fluence from 10⁶ to 10⁷ ions per cm² per LET. What does the extra fluence buy, and what does it not?

Ten times the fluence tightens the bound a clean run can state by a factor of ten, since the 95 % bound is 3.00 divided by the live fluence. It tightens that bound only for the bias, temperature and mode that were run. The set of conditions covered is the third of the lesson's three numbers, and more fluence or more parts cannot repair a missing condition afterward; the untested bias and temperature generally need runs of their own, or a defensible argument that the response does not depend on them.

Why does the lesson ask for the analysis to be written before the run plan, when the facility only needs the run plan?

Because a run earns its place by the quantity it hands to a follow-on analysis: an event count with its live fluence, a voltage limit, a parameter trajectory or a recovery time. Writing that mapping first tells the team which observables and denominators each run has to record, and which run to drop when the beam is late. A run planned without a named analysis tends to produce data that no later step can use, and the gap (a missing live time, an unrecorded recovery procedure) is found at the review instead of at the plan.

The U1 recovery demonstration in worked example 6.1 uses no beam at all. Why is it in the evidence plan, and what would an ion run add to it?

A recovery demonstration answers what happens next: whether the flight limiter and the independent recovery supply restore the function, and how long that takes. Ten forced events on the bench, with waveforms, give the trip time, the rail decay and the restart time that the downtime analysis of chapter 9 multiplies by the event rate (lesson 7.3). The ion run supplies the rate and shows whether an ion-induced event has the same signature as the forced one; it generally cannot replace the bench demonstration, because a handful of beam events does not exercise the recovery path repeatedly under controlled conditions.

Further reading

  • NASA, "Avionics Radiation Hardness Assurance (RHA) Guidelines," NASA/TM-20210018053, 2021. NTRS 20210018053.
  • K. A. LaBel et al., "Emerging radiation hardness assurance (RHA) issues: a NASA approach for space flight programs," IEEE Trans. Nucl. Sci., vol. 45, no. 6, pp. 2727-2736, 1998. doi:10.1109/23.736521.
  • 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. Koga, "Single-event effect ground test issues," IEEE Trans. Nucl. Sci., vol. 43, no. 2, pp. 661-670, 1996.
  • JEDEC JESD57A, "Test procedures for the measurement of single-event effects in semiconductor devices from heavy-ion irradiation," 2017.
  • ESCC Basic Specification No. 25100, "Single event effects test method and guidelines."