How large and how long could an analog transient be, with no test data on your part?
Pick the device family, the rail and how much of the tested population to cover. The advisor returns a bounding amplitude and pulse width fitted to 1,225 public heavy ion records, checks it against what the next stage tolerates, and sizes the mitigation to the bound.
Part and supply
How much of the population to cover
What the next stage tolerates
Pick the family and the rails. The envelope is the fitted quantile of the measured population at your chosen exceedance probability, so a 95 percent bound means roughly one part in twenty of that family would be expected to exceed it. The maximum actually observed is shown beside the fitted value, because where the two diverge the fit is extrapolating past the data and you should know it.
Bounding envelope
Fitted bound and what was observed
What this implies for the design
Cross section against capture threshold
An analog cross section is not a property of the part on its own. A test only counts an event if it crosses the discrimination level the tester chose, so the number that gets quoted is a joint property of the device and the oscilloscope setting. Raise the level and the cross section falls, the apparent threshold LET rises, and the rate estimate that depends on both falls faster still. This is the single most common way an analog number is misread. In the corpus behind this page, only 169 of 1,225 records state the discrimination criterion at all.
Inputs
Effective cross section at your threshold
Signature identifier
You have a captured waveform and you want to know what you are looking at, whether it is credible, and what to check next. Select what you observed. The output is a ranked set of hypotheses with the evidence for each and the measurement that would separate them, not a verdict.
What you observed
Hypotheses
Device families
The mechanism, the signature, the test conditions that change the answer, and the things the literature says are commonly got wrong. Fitted numbers are shown per family with the record count behind them, so you can see which are supported and which are borrowed from the pooled prior.
How it works
From public test records to a bound you can defend
Read the public record
1,225 heavy ion transient records over 641 parts, each read and adjudicated individually against a written specification, give an amplitude as a fraction of the rail and a width in seconds.
Fit each device family
A lognormal for amplitude and one for width, fitted by maximum likelihood with the censored records kept in. Sparse families are shrunk toward the pooled prior, and latched events are fitted as a separate population.
Bound at your coverage
The fitted quantile at the coverage you choose is the envelope. It is compared with the threshold and width your next stage tolerates, and a filter, if one is needed, is sized to it.
Why a distribution and not a rule of thumb the twenty microsecond rule
An analog transient is not a bit flip. It has an amplitude, a width and a shape, and whether it matters depends on what the next stage does with it. This page is about the situation most programs are actually in: no single event transient data on the part in hand, a schedule that will not accommodate a beam campaign, and a need to say something defensible about how large and how long the transients could be. It gives a bounding envelope by device family with a stated exceedance probability, the signatures that tell you which mechanism you are looking at, and mitigation sized against the bound rather than against a habit.
The bounding rule in circulation is roughly rail to rail, twenty microseconds, traceable to a RADECS 2001 hardness assurance paper and repeated ever since. It is not a worst case. Fitted to 1,225 publicly citable heavy ion records over 641 parts, twenty microseconds sits at the 77th percentile of measured operational amplifier transient widths. Close to one in four tested op-amps exceeds it.
The documented counterexamples are not obscure. The LM6144 produces transients to about 1.5 ms through a bias network that latches rather than glitches. The RH1014, a radiation hardened part, produced pulses roughly eight times longer than the commercial LM124 tested beside it. The corpus here contains an operational amplifier with a laser induced event longer than 600 ms.
So the useful output is not a single number but a distribution, and a choice of how much of it you want to cover. That is what the advisor above returns. It also separates two things that get conflated: a self recovering transient, whose width is set by charge collection and circuit bandwidth, and a latched or shutdown event, whose duration is set by a soft start or restart timer and which is a functional interrupt rather than a transient. Roughly 32 percent of linear regulator records that report a duration describe the second kind.
How the numbers are computed methodology
What was fitted, and to what
The corpus is 1,225 single event transient test records covering 641 distinct parts, drawn from publicly citable sources: IEEE Radiation Effects Data Workshop papers, NASA Goddard radiation test reports and RADECS proceedings. Records from sources that are not publicly citable were held out of the published fit entirely and used only as a sensitivity check; including them shifts the pooled amplitude location parameter by -0.010 and the pooled width location parameter by +0.087 in natural log units, which is to say it changes nothing material. Only heavy ion records are in the published model. Proton records (299) and pulsed laser records (183) are fitted separately, because proton transients are systematically shorter and smaller and because a laser result carries no LET equivalence.
How the measurements were pulled out
The measurements live in free text, not in structured columns. A first pass using pattern matching was audited against forty records adjudicated by hand and scored three correct out of forty: supply and bias voltages were being read as transient amplitudes in more than half the sample. That approach was abandoned. Every record was instead read and adjudicated individually against a written specification that separates a transient amplitude from a supply voltage, a discrimination threshold and a rated maximum, and separates a transient width from an irradiation time, a laser pulse width and a tester dead time. The result was scored against the independent hand adjudication on the overlapping records and agreed on 35 of 35 of them. The specification is published with the model so the judgment calls are inspectable rather than implicit.
The statistical model
Amplitude is expressed as a fraction of the rail, which is the quantity that transfers between parts with different supplies, and is capped at 2.0, being a rail to rail excursion on a split supply measured against one rail. Width is in seconds. Each is given a lognormal, fitted by maximum likelihood over interval censored observations, because a record saying the worst case lasted more than five microseconds is a left censored observation of the worst case and not a measurement of five microseconds. Discarding those records would bias the tail downward, which is precisely the tail the bound depends on. Families with few observations are shrunk toward the all family pooled prior with weight n over n plus 12, and any family whose bound is carried by the pooled prior rather than its own data is labeled as such in the advisor.
Self recovering transients and latched or shutdown events are fitted as separate populations. This is not a cosmetic split. Pooling them produces 99th percentile widths in the tens of milliseconds for operational amplifiers and tens of seconds for supervisors, which is an artifact of mixing a charge collection process with a soft start timer. Classification uses the recorded description of the event, and any event lasting a millisecond or more is treated as latched unless the record explicitly says the part recovered on its own.
Validation against published values check the numbers
The model was fitted without reference to the individual published results below, so the agreement is a check rather than a construction.
| Quantity | Published | This model | Source |
|---|---|---|---|
| Comparator median transient width | 0.6 to 1 us for low overdrive 139 family parts | 1.1 us fitted median | [johnston2000] |
| Op-amp worst case width, hardened part | 300 to 360 us measured on RH1014 | 300 us maximum observed, 204 us fitted 90th percentile | [larsson2004] |
| Op-amp millisecond outlier | 1.5 ms on LM6144 through a bias network latch | 17.2 ms fitted 99th percentile for self recovering transients | [boulghassoul2004] |
| Regulator shutdown duration | up to 100 ms on TPS73601 | 54 ms median latched duration, 1000 ms maximum observed | [poivey2006] |
| Comparator output reaches the rail | rail to rail change of state is the characteristic signature | 48 percent of records report the rail | [poivey2004] |
What this model does not do assumptions and limits
- It does not predict your part. It describes the population of parts that have been tested and published, which is a biased sample: parts get tested because somebody suspected a problem or because they were candidates for flight. A bound from this model is a statement about that population, and using it is a considered bet that your part belongs to it.
- It does not replace testing. Every guidance document in this field, including the NASA application notes this page draws on, says to take data, and they are right. The model exists for the case where that is not going to happen in time.
- It carries the capture thresholds of the underlying tests, which are mostly unstated. A width or amplitude only enters the corpus if somebody's trigger caught it, so the small event end of the distribution is under-represented by an unknown amount. This biases the fit toward larger events, which is the safe direction for a bound and the wrong direction for a rate.
- It says nothing about cross section against LET for your part, which is what a rate calculation needs. For that, either test, or use the family threshold LET distribution as a prior and accept the width of it.
References
Sources for the mechanisms, signatures and published values on this page.