Chapter 9 · Lesson 4 of 4 · about 12 minutes
Uncertainty, sensitivity and the mission total
Two curves can fit the measured points and predict different rates when the uncertain region overlaps a strong part of the spectrum. Report a range and say where it comes from. More decimal places do not resolve an untested mode, an unknown lot relationship or a poor choice of environment.
What you need to know
A rate is only as good as its weakest input, and the inputs differ in kind. The counting uncertainty of chapter 8 is a statistical term that shrinks with fluence. The sensitive-volume depth is a model term bounded by the X/5 and X/100 cases. The response curve below the lowest measured LET is an assumption bounded by the alternative curves that fit the data. The spectrum is a scenario: solar minimum or maximum, quiet or stormy, with or without an event week. Shielding and the application state (duty cycle, mode, bias) are design facts that change the inputs. The result to report is a best estimate with each term listed as a factor, the scenarios stated as cases, and a list of what was not quantified. A Monte Carlo over the quantified terms describes the model that was sampled. It says nothing about the modes the model does not contain.
The mission total is assembled by function rather than by part. A function fails when a part on its path fails in a way the architecture does not cover. The parts list gives the terms, the architecture gives the dependence between them, and the consequence is read at the function. A criticality analysis counts the populated devices, the replicated channels, the active states and the time at risk, and maps each part's response through the circuit to the function. It separates the events the system tolerates (a corrected word, an interrupt cleared by the scrubber) from those it does not (an uncorrectable word in a control table, an interrupt during a maneuver). Rates are summed only for compatible event definitions and only across independent hazards. Shared resources (clocks, supplies, the recovery controller itself) are counted once, as common causes.
The mission, the environment, the application and the lifetime together decide which uncertainty matters. A short mission with few susceptible parts can accept a factor of three on a rate whose μ is 10⁻³. A long mission with a thousand devices and a critical window every orbit cannot, and the same factor of three on U2's interrupt rate moves the recorder's availability from acceptable to marginal. A sensitivity analysis ranks the terms by their effect on the function-level trade, which tells the project where the next sample, beam shift or analysis hour will do the most good.
Worked example 9.4. The mission total for the recorder function
| Part and mode | Rate per device per day (SSO, X/5) | Devices | μ over 3 years | P(at least one) | Covered by |
|---|---|---|---|---|---|
| U1 shared regulator | 1.3 × 10⁻⁶ | 1 | 1.4 × 10⁻³ | 0.14 % | current limiter and independent recovery supply; verified in chapter 10 |
| U2 recorder memory | 140 | 1 | 1.54e+05 | 100.00 % | SEC-DED with interleaving and a scrub interval short compared with the anomaly pass |
| U2 recorder memory | 0.026 | 1 | 28.5 | 100.00 % | watchdog and the reset ladder; 87.5 s of downtime each |
| U3 precision analog | 1.7 × 10⁻⁵ | 2 | 0.0378 | 3.71 % | low-pass filter at the receiver; time qualification in software |
Three factors change the table. The X/100 geometry multiplies the U1 line by 3.7 and the U2 lines by about 1.08. Solar maximum lowers the galactic terms by a factor of two to three and also lowers the trapped-proton flux at low altitude, because the expanded atmosphere removes inner-belt protons and AP-8 MAX lies below AP-8 MIN there. An event week adds the burst of lesson 9.2 to every line. An extension to five years (chapter 11 lesson 3) multiplies every μ by 5/3. The term that decides the recorder's availability is the U2 interrupt line, which is proton-dominated and known to about a factor of two from the proton data. The term that decides the risk discussion is the U1 line, known only to a factor of 3.7 from geometry alone. The next beam shift therefore goes to U1 at angles, to constrain the depth, rather than to U2 at higher LET.
Exercise
Deliver the best estimate, the bounds with the conditions they depend on, the scenario comparison and the list of what the model leaves out together, because a best estimate on its own invites the reader to treat it as the answer.
Draw the map from each part's response through the circuit to the function it serves, then compare the modeled numbers against the criteria the mission has allocated to that function.
Going deeper
Separate variability from lack of knowledge
Counting statistics describe the sampling within a model. Variation across the device population, the gap between the model and the part, and mechanisms nobody observed are different kinds of uncertainty. A Monte Carlo propagation means something only when its input distributions and their dependence can be defended. A narrow numerical interval can depend on assumptions that dominate the real decision.
Say which uncertainties were propagated, which were bounded by scenarios and which were left open.
A larger parts list is not a single multiplier
One particle upsetting several bits, changes of operating mode and a redundant recovery path all create dependencies between the terms. The duty cycle can change the susceptibility, what is observed, or the consequence, and a powered-off part can still accumulate damage or sit at a vulnerable bias. A functional-state model or a fault tree has to represent those distinctions before the exposure of the parts can be turned into a probability of mission success.
Keep the model boundary, the conditional failure probabilities and the definitions of the critical windows with the system result.
The common mistake
A simulated distribution cannot validate the distribution that was assumed to produce it, and it cannot supply evidence about failure modes that were left out.
A count of parts says nothing about the architecture, the criticality of each part, the common causes or what the recovery covers.
Takeaway
- A sensitivity analysis sends the scarce samples, beam time and analysis hours to the uncertainties that drive the function-level trade.
- Taking the mission, environment, application and lifetime (MEAL) together prevents a low-LET or short-test shortcut justified by the orbit label alone.
Check yourself
Why is the U1 geometry factor more important to the project than the U2 geometry factor, when U2 has a thousand times the events?
U2's events are covered by the code and the scrubber, and its availability is set by the proton term. U1's events are uncovered destructive-class events whose probability enters the risk decision directly. Sensitivity is judged at the function, not at the part.
What does a Monte Carlo over the counting and dosimetry terms tell the project?
The spread of the rate under the sampled model, which here is a few tens of percent. It says nothing about the geometry or curve choices, which are scenarios, or about the proton SEFI gap, and those dominate.
Which of the four lines in the table would change if the recorder were duplicated?
The U2 and U3 lines double in device count. The U1 line does not, because U1 is shared, and it becomes the common cause that limits what duplication gains.
A colleague says that a factor of three on a rate is always acceptable, because rates are order-of-magnitude numbers anyway. How does the course decide whether a factor of three matters?
The mission, the environment, the application and the lifetime together decide which uncertainty matters. A short mission with few susceptible parts can accept a factor of three on a rate whose μ is 10⁻³; a long mission with a thousand devices and a critical window every orbit cannot, and the same factor of three on U2's interrupt rate moves the recorder's availability from acceptable to marginal. The judgment is made at the function, through a sensitivity analysis that ranks the terms by their effect on the function-level trade and sends the next sample, beam shift or analysis hour where it does the most good.
A reviewer asks for the table of worked example 9.4 to be recomputed at solar maximum as the conservative case. Why is solar maximum not simply the worst case here?
Solar maximum lowers the galactic terms by a factor of two to three, and at low altitude it also lowers the trapped-proton flux, because the expanded atmosphere removes inner-belt protons and AP-8 MAX lies below AP-8 MIN there. The spectrum is a scenario rather than a single conservative number, so the result is reported as cases: the quiet background at solar minimum for the galactic and trapped terms, and the event week of lesson 9.2 added to every line as a separate case. Averaging the cases into one spectrum would hide the case that sizes the scrubber and the recovery.
The U2 upset line shows a μ of about 1.5 × 10⁵ over three years and a probability of 100 %, yet the recorder function is not at risk from it. What separates the events the system tolerates from those it does not, and what does that do to the mission total?
The mission total is assembled by function rather than by part: a function fails when a part on its path fails in a way the architecture does not cover. A criticality analysis separates the events the system tolerates (a corrected word, an interrupt cleared by the scrubber) from those it does not (an uncorrectable word in a control table, an interrupt during a maneuver), and only the uncovered outcomes, mapped from each part's response through the circuit, enter the sum, and only across compatible event definitions and independent hazards. The upsets on the U2 line are the input to the chapter 10 question of whether the code and the scrub interval keep uncorrectable words rare, not a term in the function's μ.
Further reading
- 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.
- D. L. Hansen et al., "Correlation of prediction to on-orbit SEU performance for a commercial 0.25-µm CMOS SRAM," IEEE Trans. Nucl. Sci., vol. 54, no. 6, pp. 2525-2533, 2007. doi:10.1109/TNS.2007.908787.
- 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.
- K. A. LaBel et al., "Single event effect criticality analysis," NASA Goddard Space Flight Center, 1996.
- NASA, "Avionics Radiation Hardness Assurance (RHA) Guidelines," NASA/TM-20210018053, 2021. NTRS 20210018053.