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

Chapter 7 · Lesson 4 of 4 · about 13 minutes

Destructive modes and the run record

Destructive SEE testing spends devices to find where an application becomes unsafe. Set the test up to capture the onset rather than to find a dead part afterward, apply the electrical stress the application will see, and document the protection so that it does not hide what the experiment is trying to measure.

What you need to know

Destructive single-event tests spend parts to find the edge of the safe region. For a vertical power MOSFET the mechanisms are single-event burnout (a parasitic bipolar turn-on under the drain field) and single-event gate rupture (a field across the gate oxide driven by the ion plasma). For CMOS it is latchup with enough current to damage metal or junctions. The test raises the stress in steps, either the drain-source voltage at a fixed gate bias or the gate bias at a fixed drain voltage, with the ion and the angle held fixed. The aim is to see the onset rather than to find a dead part afterward, so the drain and gate leakage are measured after each exposure, and a gate stress test after irradiation shows oxide damage that has not yet become a short. The ion has to reach the epitaxial layer and the body under the gate, which for a power device means tens of micrometers of range below the die surface. The angle is therefore kept at normal incidence, and the range is checked as in chapter 6 lesson 3.

The result of such a test is a safe operating region: the highest stress at which no events occurred, for the tested ion, angle, gate condition and temperature, with the derating that the organization applies to it. A common practice states the safe operating area as a fraction (75 % is a widely used value) of the highest passing drain-source voltage at a reference LET near 37 MeV·cm²/mg, with the gate condition stated. The reference LET is a convention tied to the mission's heavy-ion spectrum. The fraction is a margin against part-to-part variation and the coarseness of the voltage steps. Rate models for burnout and rupture are generally far less certain than those for upsets, so the usual argument is avoidance by derating rather than a predicted rate.

A protection circuit is tested to confirm that it protects, and that is a separate question from what the part does. A current limiter that trips at 400 mA in 10 µs prevents one visible failure. It says nothing about latent damage from the current that flowed before the trip, or from the local capacitor that discharged into the fault (chapter 10 lesson 3), and it says nothing about burnout, which completes on the ion's own time scale. Parts that were protected are inspected afterward (leakage, gate stress, functional margin), because latent damage from latchup has been shown to appear as degraded leakage and later failure in parts that recovered on the bench.

The record is what survives the campaign. Every sample has an identity that follows it from procurement through preparation, exposure and post-test measurement to wherever it goes afterward. Every run has an identity that links the beam record, the facility dosimetry, the device configuration, the raw waveforms and the analysis version. The run table is machine-readable, with units in the header and missing fields marked as missing rather than as zero. Runs with no events, failed runs and deviations are in the table with the others. A reanalysis years later, when the mission changes or the mechanism is better understood, can only use what was kept. A summary spreadsheet is a view of the record; the raw waveforms, logs and configuration files are the record.

Worked example 7.4. The U1 input stage: a safe operating voltage from a step test

The input stage of U1 uses a 200 V vertical MOSFET. It is tested at LET 37 with the gate held at 0 V, at normal incidence and 25 °C. The drain-source voltage is raised in steps, one part per step, each exposed to 10⁶ ions per cm². The drain and gate leakage are measured after each exposure, and a gate stress test is run at the end.

Drain-source voltageOutcome
100 Vno event; leakage and gate stress unchanged
120 Vno event; leakage and gate stress unchanged
140 Vno event; leakage and gate stress unchanged
150 Vno event; leakage and gate stress unchanged
160 Vgate leakage rose during the exposure; gate rupture

The highest passing voltage V_SEE is 150 V at LET 37. With the 75 % convention the safe operating area for this gate condition is 112 V. The application switches an 80 V rail, and the off-state drain-source voltage, with its switching transients, stays below 100 V, inside the derated region. The test did not establish behavior at a higher gate bias, at elevated temperature, or for ions with a longer range at the same LET. Each of those would be a separate claim. The step size of 10 to 20 V places the onset somewhere between 150 and 160 V. The 75 % factor covers that coarseness and the part-to-part variation. It is a convention of the organization, and the report says so.

Exercise

Before exposure, write down for the mechanism you are testing the bias steps, what the protection circuit will measure, how many parts you can afford to lose and the post-test diagnostics each part receives.

Close the campaign with a manifest of the evidence and a path from the raw run data to every result in the report that someone else could follow.

Going deeper

Treat each destructive mechanism separately

Burnout, gate rupture and latchup depend differently on the geometry, the deposited charge and the field, and two ions with the same LET need not produce the same destructive response. Because a device can fail only once, a destructive test also mixes part-to-part variation with the randomness of the fluence to failure. Characterize each mechanism with its own test, and keep both the exposures that ended without failure and the failed samples in the analysis.

A sparse destructive data set does not support a precise orbital-rate curve, and a fixed voltage derating needs evidence behind it.

Keep the evidence apart from its interpretation

The raw records are kept as recorded and are not edited. The parsing and classification code is versioned, and every derived result can be regenerated from the raw data. An interval that is excluded, or an observation that has an alternative reading, is recorded with its reason rather than deleted. When a correction changes the counts or the live fluence, it is carried through the uncertainty, the fits and the system conclusions, and the earlier version is kept.

That provenance is what makes a reanalysis possible when the understanding of the mechanism, the flight configuration or the mission environment changes.

The common mistake

A fast current limiter stops the visible failure. It does not rule out damage from the current that flowed before it tripped, or burnout, which completes on the ion's own time scale.

A summary spreadsheet does not replace the raw signatures, the configuration files and the calibration records.

Takeaway

  • A destructive test ends with either a safe operating region the data support or a vulnerability the application cannot accept.
  • A traceable record lets a later engineer question and reuse the evidence without repeating the campaign.

Check yourself

Why is the gate stress test performed after exposures that showed no change in leakage?

Because gate rupture can leave oxide damage that does not yet conduct. The stress test (a gate voltage above the operating value, held for a defined time, with the leakage measured) exposes weakened oxide before the part is credited as a pass.

A current limiter kept U1 alive through ten latchups on the bench. Why are those parts not flown?

Latent damage. Parts that recover from latchup have shown degraded leakage and later failures, so a part that has latched is inspected and generally retired. The flight parts come from a population that has never latched.

What goes in the run table for a run with zero events?

Everything a run with events has: the identities, the configuration, the beam record, the live fluence, the acquisition intervals and a statement of what was observable. Its live fluence sets the bound the run supports.

The U1 input-stage MOSFET passed at 150 V and failed at 160 V at LET 37 with the gate at 0 V, and the application stays below 100 V. A colleague argues that the 75 % derating to 112 V is excessive because the onset is known to within 10 V. What does the argument miss?

The 75 % factor covers more than the step size. One part was exposed per step, so each voltage is a single sample, and a destructive test mixes part-to-part variation with the randomness of the fluence to failure; the fraction is the organization's margin against both, and the report says so. The result also holds only for the tested ion, angle, gate condition and temperature: a higher gate bias, an elevated temperature or an ion with a longer range at the same LET are each a separate claim, since two ions with the same LET need not produce the same destructive response. Rate models for burnout and rupture are generally far less certain than those for upsets, which is why the argument is avoidance by derating rather than a predicted rate.

The U1 limiter trips at 400 mA within 10 µs, and a colleague concludes that the input-stage MOSFET is protected against single-event burnout as well. Why does the limiter not settle that question?

Burnout completes on the ion's own time scale, generally before a limiter that trips in 10 µs can act, so the limiter prevents one visible failure (a sustained high-current state) and says nothing about a mechanism that has finished by the time it responds. It also says nothing about the current that flowed before the trip or about the local capacitor that discharged into the fault (chapter 10 lesson 3). Confirming that the protection protects is a separate test from finding what the part does, which is why the MOSFET is stepped to its own safe operating voltage and the protected parts are inspected afterward for leakage, gate stress and functional margin.

Two years after the U1 campaign the mission changes and a reviewer asks for the raw current waveforms. The summary spreadsheet is complete, but the waveforms, logs and configuration files were discarded to save space. What has been lost?

The record. A summary spreadsheet is a view of the record; the raw waveforms, logs and configuration files are the record, and a reanalysis can use only what was kept. Without them the events cannot be reclassified under a revised definition, the live fluence cannot be rebuilt from the beam-on, device-valid and acquisition-ready intervals (lesson 7.1), and a result that depended on an excluded interval or an alternative reading of a signature cannot be checked. Derived results can be regenerated from the raw data when the parsing and classification code is versioned; nothing regenerates the raw data, which is why it is kept as recorded and not edited.

Further reading

  • J. L. Titus and C. F. Wheatley, "Experimental studies of single-event gate rupture and burnout in vertical power MOSFETs," IEEE Trans. Nucl. Sci., vol. 43, no. 2, pp. 533-545, 1996.
  • L. Z. Scheick, "Testing guideline for single event gate rupture (SEGR) of power MOSFETs," JPL Publication 08-10, 2008.
  • F. W. Sexton, "Destructive single-event effects in semiconductor devices and ICs," IEEE Trans. Nucl. Sci., vol. 50, no. 3, pp. 603-621, 2003.
  • H. N. Becker, T. F. Miyahira and A. H. Johnston, "Latent damage in CMOS devices from single-event latchup," IEEE Trans. Nucl. Sci., vol. 49, no. 6, pp. 3009-3015, 2002.
  • JEDEC JESD57A, "Test procedures for the measurement of single-event effects in semiconductor devices from heavy-ion irradiation," 2017.