space-rha.comShort courseChapter 8 · Data analysis and interpretation

Chapter 8 · Lesson 2 of 4 · about 19 minutes

Denominators and response fits

A device cross section, a bit cross section and an error rate are different quantities. State the unit and what was operating during the exposure. If only some of the bits or modes were exercised, dividing by every bit on the datasheet or by the whole elapsed time gives a result that looks better than the experiment supports.

What you need to know

A cross section is events divided by fluence. Everything else about it says what was counted and what was exposed. Per device, per bit and per exercised channel are different denominators. A per-bit value divides by the number of bits that were written, read and compared during the exposure, not by the capacity on the datasheet, and a per-device value for a part exercised in one mode is a value for that mode only. Report the unit with the number (cm² per device, cm² per bit), together with the operating mode, pattern, bias and temperature, and the live fluence it was divided by. Keep the conditions in separate bins until a model justifies pooling them, and when you aggregate over several devices, say how much each one contributed.

A response curve is an engineering model that connects the measured points to a spectrum. The Weibull form σ(L) = σ_sat [1 − exp(−((L − L0)/W)^s)] has four parameters: a saturation cross section, an onset L0 below which the response is taken as zero, a width W and a shape s. The form is empirical. Its use is that a rate calculation can integrate it over a spectrum (chapter 9). Its danger is that the same smooth form draws a confident curve through regions that were never measured. For a soft device the onset region drives the rate, because that is where the heavy-ion flux is largest, and it is generally the region with the fewest counts. A run with zero events does not give a zero cross section, and a zero cannot be plotted on a logarithmic axis. The run enters the fit as one that saw no events in its fluence, and it is drawn as an upper bound.

Fitting the counts rather than the cross sections uses all the information in the data. Each run contributes a Poisson likelihood for its count given σ(L) × fluence, and the fit maximizes the product of those likelihoods. A run with many counts therefore constrains the curve tightly, and a run with none still pulls it down. A least-squares fit to log cross sections weights every point equally and cannot use a zero-count run at all. The plotting convention used on this site, in the figure and in the calculator, follows the same statistics. The points are not connected, because the data say nothing about what happens between them. The fit is dashed because it is a model. Each point carries its exact Poisson interval, and a run without events is drawn as an open downward triangle at its bound.

Several curves can fit sparse points about equally well. The dataset in the figure is well sampled, and even so an alternative curve with a lower onset and a wider rise matches the points above LET 20 nearly as well as the best fit. The run at LET 0.9 that saw no events is what rejects it, and the two curves differ by a third in the rate in orbit (worked example 8.2). Where the lowest measured LET already shows a saturated response, the data do not determine the onset at all, and the rate below that LET is an assumption the analyst has made. The report should say so and take both curves into the rate calculation.

Figure 8.2. Heavy-ion SEU cross section of U2 (1 Gbit) per device, from the counts of worked example 8.2: filled points with exact 95 % Poisson intervals (smaller than the markers at these counts), an open downward triangle for the run with no events, and the Poisson maximum-likelihood Weibull fit, dashed.
Heavy-ion SEU cross section of U2 with a Weibull fit 0 10 20 30 40 50 60 10-7 10-6 10-5 10-4 0.001 0.01 0.1 1 LET in silicon, MeV·cm²/mg cross section, cm² per device Weibull fit no events: 95 % upper bound fit: σ_sat 0.396 cm², L0 1.02, W 11.6, s 1.34
Try it: fit a Weibull curve to counts

Edit the runs, add one, or remove the lowest-LET point and watch what the fit does to the onset. The fit maximizes the Poisson likelihood of the counts; the plot follows the house convention.

Worked example 8.2. The U2 SEU curve, per device and per bit

LET, MeV·cm²/mgUpsetsLive fluence, cm⁻²σ per device, cm²95 % interval
0.8901.0 × 10⁶no eventsbelow 3.0 × 10⁻⁶
3.499332.0 × 10⁴4.7 × 10⁻²4.4 × 10⁻² to 5.0 × 10⁻²
9.743,8982.0 × 10⁴1.9 × 10⁻¹1.9 × 10⁻¹ to 2.0 × 10⁻¹
21.176,9602.0 × 10⁴3.5 × 10⁻¹3.4 × 10⁻¹ to 3.6 × 10⁻¹
30.867,6512.0 × 10⁴3.8 × 10⁻¹3.7 × 10⁻¹ to 3.9 × 10⁻¹
48.157,8602.0 × 10⁴3.9 × 10⁻¹3.8 × 10⁻¹ to 4.0 × 10⁻¹
58.787,9912.0 × 10⁴4.0 × 10⁻¹3.9 × 10⁻¹ to 4.1 × 10⁻¹

The Poisson maximum-likelihood fit gives σ_sat = 0.396 cm² per device, L0 = 1.02, W = 11.6 and s = 1.34. The LET at a quarter of saturation is 5.6 MeV·cm²/mg. Divided by the 1,073,741,824 bits exercised, σ_sat is 3.7 × 10⁻¹⁰ cm² per bit. An alternative curve with L0 = 0.3 and W = 18.6 matches the saturated points nearly as well as the best fit. Its log-likelihood is lower by about 10,000, and almost all of that difference comes from the run at LET 0.89 that saw no events in 1.0 × 10⁶ ions per cm², where the alternative curve predicts about 9,696 upsets. Without that run the two curves would differ by only 467 units of log-likelihood, most of it from the point at LET 9.7. In a sun-synchronous orbit the alternative predicts 1.37 heavy-ion upsets per day against 1.04 for the best fit (X/5), a ratio of 1.32. With only the four points above LET 20 the two curves would be indistinguishable and the ratio would be a free choice. The clean run at the lowest LET is the most valuable run in the set.

Exercise

Record the denominator and the unit as fields of their own in the data, and show each step from the raw count to the normalized cross section.

With the fit, show the uncertainty that comes from the counts, the LET range the data actually cover, and what an alternative curve that fits nearly as well would do to the rate.

Going deeper

When the exposure varies during the run

The expected count is the integral over time of the flux, the sensitivity and the detection efficiency, taken only over the states in which the device could register an event. One count divided by the total fluence estimates a single cross section only when those quantities were steady through the run. If the device switched modes, if the beam covered only part of the die, or if the device was blind to events in some states, those periods need their own likelihood terms, or the result has to be left as a bound.

Check the normalization convention before applying both a facility effective-fluence correction and a geometric correction of your own, so that the same correction is not applied twice.

Extrapolation and geometry

A Weibull fit is an empirical description and says nothing by itself about the sensitive-volume mechanism. The four parameters are correlated, and when the onset or the saturation region is poorly sampled, a fit that looks precise can have several parameter sets that describe the data equally well. The upper confidence limits on the individual points do not combine into a confidence band on the curve or into a validated bound on the mission rate. Data taken at angle, or with different ion species, may call for a different model.

Take every physically plausible response curve into the rate calculation rather than hiding the disagreement between them inside one fitted curve.

The common mistake

The bit count on the datasheet and the wall-clock duration of the run are not the exposed population and the live time; those have to be measured.

A tight fit says nothing about whether the curve is physically right, or about the LET range where nothing was measured.

Takeaway

  • A correct count divided by the wrong exposure gives a wrong cross section.
  • A fitted curve takes its assumptions into the rate calculation along with the data.

Check yourself

Why is a zero-event run drawn as a triangle rather than omitted?

Because it constrains the curve: the fit must not predict more than about three events in that run's fluence. Omitting the run lets the fit rise below the lowest measured LET, and a zero cannot be plotted on a log axis, so the triangle at the bound is the right symbol.

A per-bit cross section was computed with the full 1 Gbit, but the test pattern exercised half the array. What is wrong with the number?

The denominator. The exposed and compared population was half the bits, so the per-bit value is a factor of two too low. Per-bit values divide by the bits exercised, and the pattern and mode are stated with them.

What determines whether the onset L0 is a measured quantity or an assumption?

Whether there are runs with few or no events below the rising part of the curve. With a zero-event run below the onset and a low-count run on the rise, the data bound L0. With the lowest point already saturated, L0 is free and the rate depends on the analyst's choice.

Why does the course fit the Weibull curve to the counts and live fluences rather than to the cross sections, and what does a least-squares fit to log cross sections lose?

Each run contributes a Poisson likelihood for its count given σ(L) × fluence, and the fit maximizes the product of those likelihoods, so a run with 7,991 counts constrains the curve far more tightly than one with 933, and a run with no events still pulls the curve down. A least-squares fit to log cross sections weights every point equally and cannot use a zero-count run at all, since a zero cannot be plotted on a logarithmic axis. In worked example 8.2 the clean run at LET 0.89 is what separates the best fit from the alternative curve, by about 10,000 units of log-likelihood; without that run the two curves differ by only 467 units, most of it from the point at LET 9.7, and a least-squares fit to the log cross sections would not see the run at all.

A colleague argues that with seven runs and Poisson intervals smaller than the markers, the U2 curve is fully determined and the rate is known to a few percent. What does the alternative curve of worked example 8.2 show?

The intervals describe each point, not the curve between and below the points. An alternative curve with L0 = 0.3 and W = 18.6 matches the saturated points nearly as well as the best fit (L0 = 1.02, W = 11.6) and predicts 1.37 heavy-ion upsets per day in a sun-synchronous orbit against 1.04, a ratio of 1.32. Only the run at LET 0.89 that saw no events in 1.0 × 10⁶ ions per cm² rejects it, where the alternative predicts about 9,696 upsets; with only the four points above LET 20 the two curves would be indistinguishable. The upper limits on the individual points do not combine into a confidence band on the curve or a validated bound on the rate, so both curves are taken into the rate calculation.

During a run the device was blind to events for part of the exposure (the resets that followed each SEFI), and the beam covered only part of the die. Why is the total count divided by the total fluence not a cross section?

The expected count is the integral over time of the flux, the sensitivity and the detection efficiency, taken only over the states in which the device could register an event, and one count divided by the total fluence estimates a single cross section only when those quantities were steady through the run. The blind periods and the uncovered part of the die need their own likelihood terms, or the result has to be left as a bound. The normalization convention is checked before applying both a facility effective-fluence correction and a geometric correction of your own, so that the same correction is not applied twice.

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.
  • 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.
  • 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.
  • JEDEC JESD57A, "Test procedures for the measurement of single-event effects in semiconductor devices from heavy-ion irradiation," 2017.
  • 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.