SEL Threshold Predictor

Preliminary analysis aid. Check the result against your own analysis and, where needed, a test on the intended part. Full disclaimer.

1. What This Tool Does

Beam time runs out, facilities offer one cocktail energy, and screening campaigns produce exactly one number per part: a cross section (or a null) at a single LET that is rarely the saturation value. This tool turns that number into a threshold estimate with confidence bounds whose coverage was checked by leaving each part out of the fit in turn (the tool page's calibration table): the LET at which the SEL cross section crosses 10⁻⁸ cm², conditioned on device class, era and population, without ever assuming your measurement was saturated. Four inputs are accepted: SEL observed at a test LET (events and fluence, or a cross section), a saturated cross section, and a null result.

2. Using It

• SEL observed at the test LET: enter events and fluence (Poisson uncertainty on σ is propagated with a Jeffreys prior), or enter σ directly. The prediction uses the iso-LET predictor at your test LET, then applies the hard constraint: latching at the test LET means the 10⁻⁸ crossing lies below it, so the distribution is truncated there. When most of the untruncated trend mass sits above the test LET the result is flagged as onset-dominated: your measurement caught the curve near its rise and the truncation is doing the heavy lifting. The cross-check line shows what treating the measurement as saturated would have given, and the corpus growth factor from your LET to saturation, which is why the tool does not do that.
• Saturated cross section: enter σ(80), the cross section at LET 80 or on the plateau. The tool uses the partially pooled class trend for the chosen era when the class has at least eight entries (bulk CMOS, ADC/DAC, MCU/processor, power, logic, FPGA, amplifier, flash, DRAM, BiCMOS, BCD), otherwise the global errors-in-variables trend, and flags the fallback. The global, OLS and CERN 2023 values are shown alongside.
• Null result: enter the fluence achieved and whether the test was hot. The tool updates the class prior P(SEL-susceptible) for the chosen population stratum, using the same censoring-credit construction that produced the priors: candidate onsets from the global onset CDF are weighted by the corpus probability of zero events at their cross section at the test LET, following Ladbury, Joplin and Lauenstein [2]. You get the posterior P(susceptible), the threshold distribution if the part is susceptible after all, and the demonstrated σ bound. A room-temperature null is credited at a hot-equivalent LET of 0.74 times the test LET.
• Elevated temperature: tick the box if the test was at or above 60 °C. Unticked, the tool reports the as-tested result and a hot-equivalent median scaled by 0.744 (the within-part factor from 20 room/hot pairs).
• Device class, era and stratum: the class selects the saturation trend and the susceptibility bucket; the era selects the class trend (modern, 2010 and later, is the default); the population stratum selects the prior for the null branch (modern is the default; the Space-RHA campaign stratum reflects customer-selected COTS parts and is more pessimistic for BiCMOS).
• SRAM: in the corpus, SRAM onset is uncorrelated with σ; the tool switches to a density-gated lognormal prior (below 1 Mbit legacy, GM 22; 1 Mbit and above, GM 1.6 with s_ln 0.9) and applies the same truncation logic.

3. Calibration

Leave-one-out validation across the corpus: each part's threshold predicted from a model trained without it. The nominal 90% intervals of the iso-LET predictors cover 94, 93, 91 and 92% of parts at test LET 20, 30, 50 and 80, with median error factors of 1.5 to 1.7. An external check on CERN parts with full Weibull curves covers 94%. The susceptibility priors pass posterior-predictive checks on held-out modern papers; held-out 2025-26 commercial campaigns exceed literature-trained predictions, which is why a dedicated stratum exists. Treat the intervals as honest; treat single medians as guidance, not clearance.

4. Default susceptibility priors (modern stratum)

Bulk CMOS 0.45, ADC/DAC 0.83, BiCMOS 0.39, SiGe 0.25, SRAM 0.83, DRAM 0.29, flash 0.54, SRAM-based FPGA 0.32, antifuse FPGA 0.05, SOI 0.06, SOS 0.07, BCD / smart power 0.42, unknown 0.35. Each is a posterior mean with a 90% credible interval shown in the tool; see the methodology for all strata.

5. Caveats

• The trend describes the historical population of SEL-susceptible parts; a part from a process family with unusual SEL topology can sit in the tails. Extreme legacy SRAMs and physically large multi-die devices populate the corpus tails for real reasons.
• Threshold is defined at σ = 10⁻⁸ cm² from the fitted curve, not the author-quoted first-observed-SEL LET, which typically runs higher. Extrapolated corpus thresholds were treated as intervals in the fit; for conservative use of your own data, floor an extrapolated threshold at the lowest tested LET.
• Test conditions matter: the corpus mixes temperatures and biases; worst-case hot, maximum-bias data should be compared against hot-equivalent predictions.
• A one-point prediction is a planning instrument: it tells you where to aim the next test, or what a mission rate might look like. It is not a substitute for a measured curve, and for system-level screening decisions use the SEL Test-LET Tool.

6. References