space-rha.comShort courseChapter 7 · Test execution and observation

Chapter 7 · Lesson 1 of 4 · about 14 minutes

Freeze the plan and prove the observation chain

Decide what the test should reveal, which signals will show it and when a run should stop, and write the functional and parametric limits down before anyone has seen the data. A team that has to invent its criteria after a surprising result weakens both its explanation and its design response.

What you need to know

The run plan is frozen before the beam comes on, so that the pass and fail criteria are written before anyone has seen the data. For each run it states the device and its identity, the sample allocation, the bias, temperature, load, clock and configuration, the operating mode exercised, the ion or the proton energy, the target fluence or dose, the flux, the observables, the limits that define an event or a parametric failure, the stopping rule, and the recovery procedure. The matrix generally holds two kinds of condition. A bounding run finds the mechanism under the harshest planned condition. An application run, at the flight bias and load, tells the project whether that mechanism matters in the circuit. The two have different purposes, and both are planned.

The stopping rule decides how the result can be analyzed. A run that delivers a fixed fluence and counts the events is analyzed with the fixed-exposure Poisson interval of chapter 8. A run that stops at the first event, as a latchup search does, is a different sampling scheme, and its interval has to be computed for that scheme. If you continue a run because the result looks good, or stop early because it does not, you have changed the sampling, and an interval computed as if the exposure had been fixed no longer applies. Deviations do happen during a campaign. They are logged with their reason and carried into the analysis.

The experiment sees the device only through the tester, the cables and the software. An event the measurement chain cannot see is absent from the data, and a clean run proves nothing about it. For that reason the chain is demonstrated end to end before irradiation. A known fault is injected at the observation point (a pulse at the comparator input, a corrupted word written to the memory, a forced current step at the regulator), and the record is checked for the complete signature, including the timestamp, the waveform length and the recovery sequence. Trigger thresholds, sampling rates, record lengths and polling intervals together decide which events get recorded and which do not. A long transient can be captured at its leading edge and lost in its tail, and a burst of upsets can overflow a buffer so that the count saturates while the beam continues. The bench demonstration shows that the chain works for the signature that was injected. It says nothing about the detection probability for signatures that were not, so the plan lists which ones were.

Time in which the device is blind while the beam stays on lowers the measured sensitivity. If the device under test resets and is blind for 15 s while the beam continues at 1.0 × 10⁴ ions per cm² per second, 1.5 × 10⁵ ions per cm² arrive unobserved during each reset. Over a 10-minute run with 4 resets the live fraction is 90.0 %, so a cross section computed with the full fluence comes out 10 % too low. A SEFI cross section computed that way is biased in the direction that favors the device. Beam-on, device-valid and acquisition-ready are three separate time series, and the fluence in the denominator is integrated over the intervals in which all three hold (chapter 7 lesson 2). Tester health is monitored separately from device health, so that a tester fault is neither counted as a device event nor allowed to hide one.

Worked example 7.1. The frozen run matrix for U1

RunIon (LET at the wells)BiasTemperatureFluenceStop ruleObservables and limits
1 to 6Xe at 30° (about 68 at the wells), then Xe, Ag, Kr, Cu and Ar at 0°, in that sequencemaximum input, full load125 °C10⁶ per cm² eachfixed fluence; a run stops at the third high-current event, and the fluence to that point is recordedinput current at 1 MS/s, trip above 400 mA, output voltage, controller heartbeat
7 to 12the same six ionsnominal input, flight load125 °C10⁶ per cm² eachas above; run only if runs 1 to 6 show eventsas run 1
13Xe at 0°maximum input25 °C10⁶ per cm²fixed fluencetemperature dependence of the threshold
R1 to R10none (bench)flight25 °C and 125 °Cforced high-current eventten repetitionslimiter trip time, rail decay to below 0.5 V, restart time, output within specification after restart

Before run 1 the chain is checked. A 500 mA step is forced through the input with a programmable load, and the record is examined for the full current waveform, the trip, the rail decay and the restart, all on one time base. The event definition is written down in advance: an input current above 400 mA for more than 100 µs that does not clear without the limiter. A current step that clears by itself is logged as a transient and kept separate. Each run's record includes the beam-on, device-valid and acquisition-ready intervals, so that the live fluence can be reconstructed (lesson 7.2).

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.Rehearse the current capture, the isolation, the rail collapse and the restart on the bench. Keep a high-current SEFI open as a competing explanation for any latchup-like symptom.
U2 · recorder memoryA memory stores science data under ECC and periodic scrubbing.Time-tag the addresses, syndromes, patterns and each recovery step. Measure the intervals in which a readout or reboot leaves the monitor blind.
U3 · precision analogA precision analog chain has to hold its accuracy and timing allocation.Measure each part's own baseline and its trajectory after every dose step, with the measurement timing and the anneal history controlled. Exercise the receiving function during transient tests.
U4 · optical interfaceAn optical interface has to keep enough signal margin through the mission.Control the bias, drive current, measurement timing and temperature. Record the source energy or spectrum and the dosimetry, and keep the TID and displacement-damage contributions separate.

Exercise

Write the run matrix for a part of your own, with the observables, the limits, the stopping rule and the contingency actions, and have it approved before beam-on.

Before irradiation, inject a representative event and show that the chain records the whole pulse and that the tester reports its own health separately from the device.

Going deeper

Decide the stopping rules in advance

A screening run, a fixed-exposure rate measurement and a run stopped on an event are three different sampling schemes. The record says which one was used, and stopped or failed runs stay in it. Choosing to continue because a result looks attractive invalidates an interval calculated as though the exposure had been fixed in advance.

The plan names the likelihood or interval method that goes with each stopping rule, and every deviation is logged with its reason.

Measure what the chain can and cannot see

Trigger thresholds, sampling rates, buffering, polling and the rearm time of the waveform capture together set which events are recorded. Calibrate that over the range of amplitudes, widths and timings you expect, including overload and burst conditions. An injected pulse validates only the path and the range it exercised; internal failure modes it did not reach remain untested.

Keep the analog traces and the raw comparison records, so that the detection and classification assumptions can be examined again later.

The common mistake

A long list of test conditions says nothing about the internal states that none of them exercised.

A bench checkout that passes shows the chain works for the fault that was injected. It is not a measured detection probability for every event signature.

Takeaway

  • The frozen plan is what lets a physical observation be turned into an inference that can be defended afterward.
  • The experiment measures the device only as the tester, the cables and the analysis software let it through.

Check yourself

A latchup search stops at the first event after 3 × 10⁵ ions per cm². Why is the fixed-exposure Poisson interval not the right one?

Because the exposure was not fixed in advance. Stopping at the first event is a different sampling scheme with a different likelihood, and the fluence to the first event is itself the random variable. The interval is computed for that scheme, or the run is reported as a bound with the scheme stated.

The tester logs show a 2 s gap in acquisition during a run with no events. What does the gap do to the result?

It removes 2 s of fluence from the denominator. The clean-run bound is computed with the live fluence rather than the delivered fluence, and the gap is flagged as an interval in which an event could not have been seen.

What does a successful injected-fault test establish?

That the chain captures the injected signature end to end, with the timestamp and the record length checked. It establishes nothing about signatures that were not injected, so the plan lists which ones were.

Run 3 of the U1 matrix reaches its 10⁶ ions per cm² with no events, the beam is stable, and the team lets it run on to tighten the bound. What has changed?

The sampling scheme. The plan fixed the exposure in advance so that the result could be analyzed with the fixed-exposure Poisson interval; continuing because the result looks good (or stopping early because it does not) makes the exposure depend on the data, and an interval computed as though it had been fixed no longer applies. The deviation is logged with its reason and carried into the analysis, and the report states the scheme that was used. The same discipline is why the matrix writes its stop rule (a halt at the third high-current event, with the fluence to that point recorded) before the beam is on.

In a 10-minute U2 run at 1.0 × 10⁴ ions per cm² per second the device resets 4 times and is blind for 15 s each time. A colleague notes that the resulting 10 % error is no larger than the dosimetry uncertainty and leaves the cross section uncorrected. What is wrong with that?

The two are different kinds of error. The live fraction is 90.0 %, and a cross section computed with the full fluence comes out 10 % too low in the direction that favors the device; that is a known bias, and a known bias is removed from the denominator rather than absorbed into an uncertainty. The dosimetry uncertainty, generally quoted near 10 % (lesson 7.2), is a separate term that stays in the budget after the correction. The 4 resets are also SEFI events in their own right, and they are counted rather than corrected away (lesson 7.2).

The U1 event definition counts an input current above 400 mA for more than 100 µs that does not clear without the limiter, and logs a step that clears by itself as a transient. Why draw the line before the run rather than sort the records afterward?

Because the pass and fail criteria have to exist before anyone has seen the data; a team that invents them after a surprising result weakens both its explanation and its design response. The two outcomes also carry different mission consequences, so mixing them would inflate the high-current count or hide the transients, and the classification of chapter 7 lesson 3 depends on current, functional state and the recovery step being recorded together. The 500 mA step forced before run 1 shows that the chain records the whole signature on one time base, so that the definition can be applied to what was captured.

Further reading

  • JEDEC JESD57A, "Test procedures for the measurement of single-event effects in semiconductor devices from heavy-ion irradiation," 2017.
  • ASTM F1192, "Standard guide for the measurement of single event phenomena (SEP) induced by heavy ion irradiation of semiconductor devices."
  • R. Koga, "Single-event effect ground test issues," IEEE Trans. Nucl. Sci., vol. 43, no. 2, pp. 661-670, 1996.
  • R. Ladbury, "Statistical properties of SEE rate calculation in the limits of large and small event counts," IEEE Trans. Nucl. Sci., vol. 54, no. 6, pp. 2113-2119, 2007.
  • NASA, "Avionics Radiation Hardness Assurance (RHA) Guidelines," NASA/TM-20210018053, 2021. NTRS 20210018053.