Chapter 8 · Lesson 4 of 4 · about 12 minutes
Uncertainty and the analysis handoff
A dataset can have tiny error bars and still say little about the flight use. Keep the counting uncertainty apart from the uncertainties in the dosimetry, the device population, the test conditions, the environment and the model that links them. The next thing to do is whatever reduces the uncertainty that changes the engineering conclusion.
What you need to know
A rate or a hardness number comes with several uncertainties of different kinds, and the list of them is more useful than the number. Counting uncertainty is statistical and shrinks with fluence. Dosimetry uncertainty is a calibration term common to every point taken with one setup, so it shifts the whole curve. Population uncertainty (lot, die revision, part-to-part variation) is an assumption about how the result transfers to the flight parts. Model uncertainty (which curve through the points, which sensitive-volume geometry, which spectrum) is a choice, and it is bounded by running the alternatives rather than by a distribution. Modes that were never observed cannot be in the budget at all and are listed as a gap. Combining these terms as if they were independent random errors produces a single percentage that hides the term that matters. The budget instead lists each term with its cause, its size as a factor on the result, whether it is correlated across points, and the action that would reduce it.
The analysis is done when it says what happens to the function, and a cross-section curve or a dose bound by itself does not say that. Designers need the error signatures (what an upset looks like at the interface, how long an interrupt lasts, which rung of the recovery ladder clears it), the parameter changes at end of life with the population statement that goes with them, the distribution of recovery times, and the conditions under which the evidence does not apply. For single events, the handoff contains the counts, live fluences, conditions, classification rules, response curves or bounds, and recovery times. For cumulative effects, it contains the individual trajectories, the failure-dose sample and its tolerance bound, the dose rate and temperature history of the test, and the application limits used. Both state what would reopen the assessment: a new lot, a changed bias, a mission extension.
Keep separate claims separate, even for one part. U1 has a destructive-event claim (none to the tested LET at the tested voltage and temperature, with its bound), a recoverable high-current claim (a cross-section curve with a recovery demonstration) and a cumulative claim (a dose bound on its regulation parameters). Each has its own budget and its own owner in the system analysis. Collapsing them into one hardness label loses what the system engineer needs in order to decide which of them to protect against, and how.
Worked example 8.4. The uncertainty budget of the U2 upset rate, and what the designers receive
| Term | Kind | Size, as a factor on the SSO heavy-ion rate | Correlated across points | Action that reduces it |
|---|---|---|---|---|
| Poisson counting at saturation | statistical | × 1.02 (95 % upper, 7,991 counts) | no | already small; more fluence would not help here |
| Counting on the rise (LET 3.49, 933 counts) | statistical | × 1.07 | no | a run at LET 2 would constrain the onset |
| Facility dosimetry | calibration | × 1.10 (quoted 10 %) | yes, whole curve | cross-calibration between facilities; none within one campaign |
| Sensitive-volume depth, X/5 to X/100 | model | × 1.08 for U2; × 3.7 for U1 | yes | angular data; more knowledge of the device structure |
| Alternative curve through the same points | model | × 1.32 | yes | the low-LET run above |
| Spectrum: solar minimum GCR only; no solar heavy ions | scenario | event cases added separately (lesson 9.2) | yes | state the cases; do not average them in |
| Lot and die revision of the flight parts | transfer | not quantified; same date code as tested | yes | procurement control; a confirmation run on the flight lot |
| Proton SEFI modes not exercised in the pattern used | gap | unquantified | a proton run in the flight mode |
For U2 the statistical terms are small. The budget is dominated by the choice of curve (a factor of 1.32), the geometry (a factor of 1.08) and above all by the proton term, which chapter 9 will show is a hundred times larger than the heavy-ion rate. For U1, whose onset falls where the spectrum drops steeply, the geometry term alone is a factor of 3.7, so both the best-estimate and the worst-case rates are delivered. The memory controller designer receives the per-bit curve with both geometries, the SEFI curve and the reset rung that clears each class, the multiplicity distribution (how many bits per event, and whether they fall in one 72-bit word), the proton cross sections, and a statement that the SEFI rate in the flight mode is a gap.
Exercise
Write an uncertainty budget in the form of the table in worked example 8.4: for each term, its cause, whether it is correlated across points, its effect on the result, and the action that would reduce it.
Hand the circuit and system owners the response data with their conditions, the bounds and the conditions they hold under, the application limits used, and the changes that would reopen the assessment.
Going deeper
What a sampled model cannot show
Sampling the parameters of a model explores only that model. It cannot find a mechanism that is missing from it, a wrong geometry or a population it does not describe. Confidence statements from different sources should not be multiplied or combined unless they have a joint probabilistic meaning. Alternative physical models, cases with and without correlation, and separate quantiles of the environment show the uncertainty that a narrow fit interval cannot.
Rank the possible follow-up measurements by how much each could change the decision, and leave the unresolved assumptions visible rather than folding them into one blanket factor.
Dependence and the limits of each claim
Circuit behavior can depend on parameters that shift together, on the state and load of the device, and on the damage accumulated so far. Taking the worst case of every parameter independently can describe a part state that cannot occur, and assuming independence can also miss a common cause. Hand the system analysis response cases that could physically occur together, note the dependencies that are not resolved, and state the exact conditions under which each measured or statistical claim holds.
The analysis that receives the data should be able to tell measured performance, interpolation by a model, extrapolation and engineering judgment apart, rather than finding all four flattened into one hardness number.
The common mistake
A Monte Carlo histogram is only as good as the input distributions, the correlations and the physical model behind it.
A report summary or a generic pass label does not replace an argument made for the specific application.
Takeaway
- Spend the effort on the uncertainty that changes the conclusion, which is often not the one easiest to measure.
- The analysis is useful when it connects the measured response of the part to what happens in the application.
Check yourself
Why is a dosimetry uncertainty of 10 % not added in quadrature with the Poisson intervals of each point?
Because it is common to every point taken with one setup. It moves the whole curve, and the rate with it. Treating it as independent noise on each point would make the average look more stable than the calibration allows.
A reviewer asks for one number, the uncertainty of the U2 rate. What should the reply be?
A best estimate with a stated worst case (geometry × 1.08, curve × 1.32), the proton term that dominates the total, and a list of what is not quantified. A single percentage would describe only the counting term, which is the smallest.
What distinguishes the handoff from a label that says the part was tested?
It contains the response data with their conditions and limits, the recovery behavior, the population statement for cumulative effects and the changes that would reopen the assessment, so that the circuit and system analyses can use what applies and see what does not.
To tighten the U2 heavy-ion rate, a colleague proposes another long run at high LET, where the response is saturated, to bring the counting uncertainty down. What does the budget of worked example 8.4 say about that proposal?
The counting term at saturation is already × 1.02 with 7,991 counts, and more fluence there would not help. The terms that dominate the heavy-ion rate are the choice of curve (× 1.32) and the geometry (× 1.08), both model terms that fluence at saturation does not touch, and above all the proton term, which chapter 9 shows is about a hundred times the heavy-ion rate. The action column of the budget points elsewhere: a run at LET 2 to constrain the onset, angular data to constrain the depth, and a proton run in the flight mode to close the SEFI gap.
The reviewers of U1 ask for a single hardness statement for the part. Why does the course deliver three separate claims, and what goes with each?
U1 has a destructive-event claim (none to the tested LET at the tested voltage and temperature, with its bound), a recoverable high-current claim (a cross-section curve with a recovery demonstration) and a cumulative claim (a dose bound on its regulation parameters). Each rests on different evidence, carries its own uncertainty budget and has its own owner in the system analysis, and each is protected against differently. Collapsing them into one hardness label loses what the system engineer needs in order to decide which of them to protect against, and how.
The proton SEFI mode was not exercised in the pattern used. A reviewer suggests assigning it a nominal factor, "to be conservative", and folding it into the budget. Why does the course list it as a gap instead?
A mode that was never observed cannot be in the budget at all, because there is no measurement to attach a factor to, and a number assigned without data would hide the gap behind a term that looks quantified. The budget lists the gap with its cause and the action that closes it, a proton run in the flight mode, and the handoff states that the SEFI rate in the flight mode is a gap, so that the analysis that receives the data can tell measured performance, interpolation, extrapolation and engineering judgment apart. A Monte Carlo over the quantified terms would not reveal the missing mode either, since sampling the parameters of a model explores only that model.
Further reading
- 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.
- E. L. Petersen, J. C. Pickel, J. H. Adams and E. C. Smith, "Rate prediction for single event effects: a critique," IEEE Trans. Nucl. Sci., vol. 39, no. 6, pp. 1577-1599, 1992.
- 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.