Pick a device family, a dose rate class, your mission dose and your margin, and read the fraction of part types in that family that had a parameter out of its datasheet limit by that dose, how the family ranks against the others, how often it was enhanced at low dose rate, and for optocouplers how the current transfer ratio falls with displacement damage.
The curves are read at your specification dose, the mission dose times the design margin.
Every family with at least ten part types in the selected dose rate class, ranked by the fraction out of limit at your specification dose. Bars are the point estimate; the whisker is the 90 percent bootstrap band where one was computed.
Standard rate observations for the selected family split three ways. Cells with fewer than five part types are not shown. The era is the publication year of the test, which lags the design by several years.
| Stratum | Part types | Failed in test | Out by 10 krad | Out by 30 krad | Out by 100 krad | Dose where 10% are out |
|---|
Two different questions. How often did a part type, once someone tested it at low rate, turn out to be enhanced? And when it was, by how much? The first answer is biased upward by who gets tested; the second is not. For factors to apply to a high dose rate result by part class and coverage, the tests that answer the question directly, and a 3D model of the mechanism, see the ELDRS page.
| Family | Determinations | Flagged | Fraction (90% band) |
|---|
Published enhancement factors for bipolar linear circuits, the change in the sensitive parameter at low dose rate divided by the change at high dose rate at the same total dose, generally fall between about 1.5 and 20 at 10 mrad(Si)/s. The largest measured with both changes resolved is about 60: an LM139 comparator sold as a radiation hardened QML part, irradiated unbiased, whose input bias current changed by about 600 nA at 0.01 rad/s against about 10 nA at 50 rad/s at 20 krad [1], [2]. The same parts irradiated biased gave about 10 to 16. Larger values in print are lower bounds or ratios over a high dose rate change close to zero, and several parts were still becoming worse at the lowest dose rate tested, so the enhancement at orbital dose rates may be larger than any value measured at 10 mrad/s [3], [4].
| Part and parameter | Dose rates, total dose | Enhancement | Note |
|---|---|---|---|
| LT1185 negative regulator, output current, unbiased | 0.01 against 40 rad/s, 50 krad | more than 100 | lower bound: non-functional at low rate, no change at high rate [5] |
| LM139 comparator (RH QML), input bias current, unbiased | 0.01 against 50 rad/s, 20 to 30 krad | about 60 to 70 | high rate change about 10 nA, so about plus or minus 30 percent [1], [2] |
| AD590 temperature transducer, flatpack, output current | 0.01 against 100 rad/s, 20 krad | 52 | [2] |
| LM139 comparator (non-hardened), input bias current, biased | 0.02 against 50 rad/s, 100 krad | 44 | the same table lists 2385 for output off current, a ratio over a near-zero change [6] |
| RF25 lateral PNP transistor, excess base current | 0.001 against 294 rad/s, 20 krad | more than 30 at VEB 0.5 V, about 6 at 0.8 V | the factor depends on the measurement bias [7] |
| LM193 comparator, input bias current, unbiased | 0.002 against 25 rad/s, 15.5 krad | about 30 | still rising at 2 mrad/s [3] |
| OP42 JFET-input op amp, input offset voltage | 1 to 22 mrad/s against high rate | about 100 in failure dose | operates to 1 Mrad at high rate, fails near 10 krad at low rate [8] |
At the level of the oxide defects the enhancement is smaller, generally 2 to 8 in the density of interface traps over the base [9]; the circuit multiplies it because the excess base current of a PNP under the base oxide rises steeply with interface traps at the low bias points that set input bias current. The factor also depends on the total dose, the bias during irradiation, the hydrogen in the package cavity (which raises high dose rate damage more than low dose rate damage, so it compresses the ratio) and the parameter chosen, so a single value describes one test, not a part type [4], [10]. The distribution in the plot above is formed differently, as the ratio of doses to the first parameter out of limit, from the site's test records.
Optocoupler current transfer ratio falls with displacement damage in the LED and the phototransistor, and it is the failure mode that has grounded more optocoupler applications than total dose. The corpus has hundreds of optocoupler test records but only a few dozen state a retention number against a fluence; those are fitted to a one parameter damage law and shown with their scatter, which is large and is the message.
How it works
Every record of a part type at one dose rate class is collapsed to one observation: the highest dose it passed and the lowest dose at which any parameter was out of limit. Parts that never failed stay open above their last pass.
The Turnbull estimate turns those censored intervals into the fraction of part types still in limit at each dose, with a bootstrap band that says how well the family curve is known.
Your mission dose times the design margin is the specification dose. The curve there gives the fraction of part types out of limit, and the verdict says whether heritage carries weight or the part needs a test.
Understanding the mechanisms
The priors above count how often part types failed; the model below shows why they fail, and why the dose rate, the temperature, the bias and the time after the exposure change the answer. Ionizing dose makes electron-hole pairs in the silicon dioxide of a part. In a MOS transistor the holes that escape recombination drift in the gate field and some are trapped near the silicon, where they pull the threshold voltage negative; the protons they free build interface traps, over seconds to hours, which in an n-channel transistor pull the threshold back the other way. After the exposure, electrons tunnel into the trapped holes and heat empties them while the interface traps stay, so the threshold can end above where it started (rebound). In the thick oxide over the base of a bipolar transistor the field is weak: at high dose rate charge builds up in the oxide, which turns protons back or, in the more recent models, holds electrons that recombine with the holes before they free protons, and at low dose rate it does not, so each rad makes more interface traps, which raise the base current and lower the gain. That is enhanced low dose rate sensitivity, and the hydrogen in the package, the passivation over the die and heating during the test all move it.
Where a paper gives a number (the pair generation constant, the charge yield against field, the tunneling and thermal emission fronts, the interface trap build-up time and its temperature dependence, the hydrogen anchors of a gated lateral PNP, the gain half-doses of lateral and substrate PNP transistors) the model uses it; every other number is illustrative, chosen so that the direction and rough size of each effect follows the papers listed below, not to predict any part. It cannot tell you what a given part will do: the priors above and a test do that. The views show one state four ways: the oxide as a block you can rotate, the transistor characteristic before the exposure and at the time shown, the history on a log time axis, and the same exposure repeated at every dose rate.
MOS gate oxides. Each rad makes 8.1e12 electron-hole pairs per cm3 of oxide [1]; the fraction of holes that escape recombination rises with the oxide field along the cobalt-60 charge yield curve of Shaneyfelt and co-workers [2], and at the weak fields of a bipolar oxide it falls to the Onsager value of about 6 percent at room temperature [3]. The long-lived trapped fraction of escaped holes, 0.10 in the model, lies inside the 0.018 to 0.16 measured by McLean and co-workers [5]; trapped charge and interface traps per escaped hole fall with field as in [2], [6], [12]. After the exposure, electrons tunneling from the silicon remove trapped holes behind a front about 0.2 nm deeper per decade of time [7], and thermal emission removes them behind a front kT ln(A T^2 t) through trap levels near 0.7 and 1.1 eV [8]; the anneal is slower at 0 V than under a positive gate [9], and under a negative gate part of the annealed charge returns [7], [10]. Interface traps form in two stages: holes free protons, which drift to the interface under a positive gate and react there, with a half time near 25 s at 2 MV/cm, a field dependence near E^-1.6, a dispersive time law and an activation energy near 0.8 eV [11], [13], [14]; they saturate at high dose [15] and anneal little in gate oxides below about 150 C [16]. The dose rate enters a MOS gate oxide mainly through time: the response to any history is the sum of each increment of dose annealing from its own time [4], so a slow exposure shows fewer trapped holes and more interface traps, and rebound after the exposure is the test concern; the accelerated aging step of TM1019 and ESCC 22900 exists for it [17]. The steep fall of the shift in oxides thinner than about 10 nm, and the exception of thick field oxides, which can show a true low dose rate enhancement, are from the review [6].
Bipolar oxides. The space charge explanation of enhanced low dose rate sensitivity, slow holes held in the oxide at high rate that turn protons back from the interface, is from Fleetwood and co-workers [20] and Rashkeev and co-workers [21]; the bimolecular reactions that make the high rate growth sublinear are from Hjalmarson and co-workers [22]. The hydrogen anchors (interface traps at low and high rate in air, 1 percent and 100 percent H2, and the transition dose rate in each) are the gated lateral PNP data of Pease and co-workers [23], with about one decade of transition per decade of hydrogen [24]; heating moves the transition with the hole transport energy used by Hughart and co-workers, who also describe arriving protons passivating traps at high temperature [25]. No passivation, oxide and nitride, and silicon carbide passivation follow the split lots of Shaneyfelt and co-workers [26], [27], [28]. The separation of the enhancement into a true dose rate part and a time-dependent part, with interface traps that keep forming after a high rate exposure, follows [29], [30]; the elevated temperature optimum and its fall with dose follow [25], [31], and the shortfall of elevated temperature tests against the low rate damage follows [18], [19]. Interface traps repassivate when heated, with a barrier near 1.3 eV [32]. The emitter-base bias dependence is from [33]; the gain half-doses of lateral and substrate PNP transistors, the ratio of trapped holes between rates and the ideality of the excess base current from Schmidt and co-workers [34], [36]; the room temperature anneal of high rate trapped charge from [35]; the oxide charge at which it takes over the base surface from [37]; the dependence of the enhancement factor on the measurement bias from [38]; the flight data that tracked 10 mrad/s ground tests from [39]; and the TM1019 dose rate conditions and the overall picture from the review [40].
Two choices are the model's own and are marked as such in its source: it holds the low field yield at its room temperature value above 25 C, where [3] has no data and where the rise would put elevated temperature damage above the low rate damage, which the data do not show; and it adds the high temperature passivation term with values fitted to the elevated temperature data. It leaves out border traps, hydrogen effects in MOS gate oxides, thick MOS field oxides, the differences between cobalt-60, X-rays and protons, pre-irradiation elevated temperature stress, circuit parameters beyond the base current of one transistor, and the spread between parts and lots, which in practice is often larger than any effect the model shows.
What the curves mean. Each observation is one part type tested at one dose rate class: the highest dose it passed and the lowest at which any parameter was reported out of limit. Many never failed within the dose tested, so the data are interval censored, and the curves are the nonparametric maximum likelihood survival estimate (Turnbull) for that censoring: the fraction of part types in the family with every parameter still inside its limit at dose D. "Out of limit" is the author's judgment against the datasheet or SMD, which for a bipolar op amp is often an input bias current specification the application never needed. The curves are therefore conservative for a derated application and are exactly right for the question a lot acceptance test asks.
What they are for. Deciding which parts on a list need a test and which can be bought on heritage; putting a number on "that family never has a problem at 30 krad" before the lot acceptance test is planned; sizing an optocoupler CTR margin before the displacement damage test. They are not for qualifying a part. If 70 percent of tested part types in a family pass 50 krad, the remaining 30 percent do not; that family-level result does not establish the response of an untested part.
Before a part has been tested, the only honest statement about its total dose response is a statement about parts like it. This page makes that statement with numbers. From 6,639 total dose and ELDRS test records on 3,772 part and dose rate combinations, drawn from four decades of public REDW, RADECS and NASA test reports plus an internal set counted but not identified, it gives for each device family the fraction of part types that had a parameter out of its datasheet limit by a given dose, at the standard test rate and at low rate, with confidence bands; how often a family turned out to be enhanced at low dose rate and by how much; and, for optocouplers, how the current transfer ratio falls with displacement damage fluence. Release TDHP 2026-09.
How the corpus became curves, what was left out, where the estimates are weak, and how the numbers compare with what the literature has said in words for thirty years.
The corpus is the site's radiation effects database: structured records extracted from the IEEE Radiation Effects Data Workshop proceedings (1992 on), the RADECS data workshop and conference proceedings, the NASA Goddard radiation test reports, and a set of internal test reports that may not be identified or quoted. Each record carries the highest dose at which the part met its specification, the lowest dose at which a parameter was reported outside its limit, the dose rate as printed, an ELDRS flag when the source made a determination, and the confidence of the extraction. Only records with the extraction confidence "extracted" (a number printed in the source) are used; records summarizing vendor claims or third party screens are excluded. That left 6,639 total dose and ELDRS records.
Internal records contribute to every fit and are counted in the census as internal. Nothing on this page or in the published asset identifies a part, a program, a document or a value from an internal record; the only per record values published are the anonymous optocoupler retention points, which carry a fluence and a fraction and nothing else.
Many part types appear in several records: several bias conditions, several parameters, several papers. Pooling records as if they were independent would weight a well studied part type twenty times over. So records are collapsed to one observation per part type per dose rate class, with L the maximum dose passed over all its records and U the minimum dose at which any parameter was reported out of limit. The failure dose of that part type is then known to lie in (L, U], or in (L, ∞) if it never failed. A record whose failure dose is below its pass dose (two parameters, two bias conditions) keeps the failure and sets L to half of it.
Two cleaning rules matter. A failure recorded at 0 krad is a pre irradiation out of spec measurement, of which the Goddard reports contain many; those are not radiation failures and are dropped, as are values below 0.5 krad, which are almost always the same thing or a unit error. Doses above 3000 krad are right censored at 3000: the corpus contains parts run to hundreds of megarads for accelerator use, and letting them into a lognormal fit stretches the distribution to describe a population no spacecraft carries.
The database records 916 distinct device type strings. A keyword mapper folds them into 22 families, applied in a fixed order so that "optocoupler" is caught before "transistor" and "MOSFET driver" lands in power management rather than power MOSFET. The mapping is in the build script and is imperfect at the edges: "controller" is ambiguous and "switch" more so. Families with fewer than five part types are not published. Grade comes from the part grade field (COTS, Mil and Hi-Rel, rad hard and QML), era from the publication year of the source (before 2005, 2005 to 2014, 2015 on), and process from the process technology field where the source gave one, which is a minority of records.
The dose rate is a free text field ("50 rad(Si)/s", "10 mrad(Si)/s", "0.05 to 0.11 krad(Si)/hour"). A parser converts every rate it finds to rad per second and takes the lowest quoted, since the lowest rate governs the ELDRS question. Classes: standard at 1 rad/s and above, intermediate between 0.01 and 1 rad/s, low at 10 mrad/s and below, and elevated temperature where the text says so. Records with no rate stated, about a third, are grouped with the standard class, because the great majority of unstated tests in this literature are Condition A tests; this dilutes the standard class slightly toward whatever the unstated tests were. Method 1019's boundary for enhanced response is 50 rad/s, not 1 rad/s; the 1 rad/s cut was chosen because the data between 1 and 50 rad/s are few and behave like Condition A. The intermediate class is published but is a mixed population and should be read as such.
With every observation an interval (L, U], the natural estimator is the nonparametric maximum likelihood estimate for interval censored data, due to Turnbull [1]: a probability mass on each cell of a dose grid, found by the EM iteration that assigns each observation's unit mass across the cells its interval covers in proportion to the current masses, then renormalizes. The grid is 1, 2, 3, 5, 7, 10, 15, 20, 30, 50, 70, 100, 150, 200, 300, 500, 1000 and 3000 krad. The survival curve is one minus the cumulative mass. It is flat where no failures were recorded and it takes no position on the shape between grid points; the pages interpolate in log dose for readout and stop the curves at 1000 krad, because the last cell holds the mass of every part type that was still passing when its test ended and says nothing about where those parts would have failed.
The estimator has one property worth knowing when reading the curves. A part type that failed with no pass dose recorded is an interval (0, U], and the EM spreads its mass over every cell below U in proportion to what other parts did there. If other parts in the family failed at 3 krad, some of that mass lands at 3 krad. This is correct for the likelihood and it means the low dose end of a curve is shaped partly by parts whose report said only "failed by 30 krad".
Alongside the nonparametric curve each stratum carries an interval censored lognormal fit, by maximum likelihood over μ and ln σ with a Nelder-Mead search, and a partial pooling step that pulls small strata toward the all-family fit with weight n/(n + 8). The lognormal is a summary, not a claim about the population: the families are mixtures of soft and hard part types and the fitted log sigma runs to 2 and beyond, at which point the cap of 3 in the fit binds and the lognormal describes the tails badly. The pages show it as a dashed overlay on request and quote its median; every headline number comes from the nonparametric curve.
Bands are 5 to 95 percent percentile bootstrap intervals from 100 resamples of the part types in the stratum, recomputing the Turnbull estimate each time. They express how well the family curve is determined by the part types in hand. They do not express how far your part might sit from the family, which is the larger uncertainty and is the reason a family prior does not replace a test. Bands are computed for the family level curves in each rate class; the grade, era and process subsets carry point estimates only.
Two quantities. The susceptibility fraction is the number of part types flagged as showing enhanced low dose rate degradation over the number for which the source made a determination either way, with a Jeffreys Beta(1/2, 1/2) prior giving the mean and the 90 percent interval. The enhancement factor is computed for part types with a recorded failure dose at both the standard and the low rate as the ratio of the two; for part types that failed at low rate and never at the standard rate, the ratio of the highest standard rate pass dose to the low rate failure dose is a lower bound and is reported separately. The pages publish the sorted list of factors, without identities, because the distribution is more useful than its quantiles when the question is "how often is the factor above my margin".
Of 361 optocoupler displacement damage records, 19 state a CTR retention as a fraction of the pre irradiation value against a fluence (regular expression over the summary and the quoted source text, checked by hand). Proton fluences are converted to 1 MeV neutron equivalent with the silicon NIEL table used by the site's NIEL tool [2], Φeq = Φp NIEL(E)/1.95e-3. Silicon NIEL is a proxy: the damage sits in a GaAs or AlGaAs LED and in a silicon phototransistor, and the two do not scale alike with energy, but the sources give no better handle. The retention R is fitted to the displacement damage law
which follows from minority carrier lifetime degradation 1/τ = 1/τ0 + KτΦ when CTR is proportional to LED efficiency and hence to lifetime [3], [4]. Each point gives a K; the K are lognormal with median 1.9e-12 cm² and a log sigma of 2.2, which is the scatter of everything the sources did not control: forward current, temperature, phototransistor gain, and the LED technology. The by LED type fits (AlGaAs double heterojunction, other AlGaAs, GaAsP or amphoteric GaAs:Si) are shown when three or more points exist and are barely tighter. The prior is honest about being thin; it is on the page because the first question about an optocoupler in a proton environment is whether to bother testing, and the answer is yes, and this shows why.
The curves reproduce what the reviews say in words. Bipolar linear parts fail early against datasheet input bias and offset current limits: 30 percent of op amp part types were out of limit by 10 krad at the standard rate and 36 percent at low rate, with the low rate curve below the standard rate curve at every dose from 5 krad up, which is the ELDRS signature in aggregate [5], [6]. Voltage references and discrete bipolar transistors show the largest low rate penalty, consistent with the literature on lateral and substrate PNP degradation. COTS flash and other nonvolatile memory fail across a band from 10 to 100 krad, dominated by the charge pump and controller, matching the NAND flash test literature [7]. Rad hard and QML part types are out of limit by 30 krad in 9 percent of cases for op amps and 4 percent for SRAM, which is not zero: tightened SMD limits and post irradiation parametric limits produce recorded "failures" in parts that are working as designed. Among part types with both rates tested, the median enhancement factor is near 1 (most parts tested at both rates were tested because they were suspected, and half were not enhanced), the upper quartile exceeds 2.5 and the 90th percentile exceeds 5, which agrees with the range of factors reported for the classic ELDRS parts [5].
Parts get tested and written up when someone needs them or when they are interesting, and interesting means soft or surprising. The corpus over represents parts that failed relative to the population of parts that exist. The failure fractions are therefore upper bounds on what a random part from the catalog would show and fair estimates of what a part somebody wanted to fly would show. The ELDRS susceptibility fraction has the stronger form of this bias, since low rate tests are expensive and are run on suspects; the enhancement factors do not, since they condition on the test having been run. "Out of limit" is the author's judgment against a datasheet, and authors differ on whether a 10 percent excursion in a non critical parameter counts. Publication year is a poor proxy for design year. And the family mapper is a keyword mapper.
Three uses are intended. Screening a parts list: a family at your specification dose with a failure fraction under 10 percent and a parameter you can derate is a candidate for heritage; anything above 30 percent is a test. Setting expectations for a lot acceptance test: the family fraction out of limit at the specification dose says how surprised to be by a failing unit, and the lot acceptance report generator then evaluates the sample on its own. And the optocoupler prior sizes the CTR margin the circuit must tolerate before the displacement damage test is planned. The parts advisor's category statistics and this page describe the same corpus from two directions and should agree; where they do not, this page's censoring treatment is the more careful one.
This release describes the first parameter out of limit and nothing about how far parameters move, which is the question a derated application actually asks. The extraction that would answer it, parameter versus dose curves from the tabular test reports, is specified and not yet run. The optocoupler prior is 19 points. The intermediate rate class is a mixture. The families are coarse. The corpus is not the population. Document these limitations, the values used, and the reasoning for their selection.
[1] Turnbull, B. W., The Empirical Distribution Function with Arbitrarily Grouped, Censored and Truncated Data, Journal of the Royal Statistical Society B 38(3), 290, 1976.
[2] Jun, I., Xapsos, M. A., Messenger, S. R., Burke, E. A., Walters, R. J., Summers, G. P. and Jordan, T., Proton Nonionizing Energy Loss (NIEL) for Device Applications, IEEE Transactions on Nuclear Science 50(6), 1924, 2003.
[3] Johnston, A. H. and Miyahira, T. F., Characterization of Proton Damage in Light-Emitting Diodes, IEEE Transactions on Nuclear Science 47(6), 2500, 2000.
[4] Rax, B. G., Lee, C. I., Johnston, A. H. and Barnes, C. E., Total Dose and Proton Damage in Optocouplers, IEEE Transactions on Nuclear Science 43(6), 3167, 1996.
[5] Pease, R. L., Schrimpf, R. D. and Fleetwood, D. M., ELDRS in Bipolar Linear Circuits: A Review, IEEE Transactions on Nuclear Science 56(4), 1894, 2009.
[6] Johnston, A. H., Swift, G. M. and Rax, B. G., Total Dose Effects in Conventional Bipolar Transistors and Linear Integrated Circuits, IEEE Transactions on Nuclear Science 41(6), 2427, 1994.
[7] Nguyen, D. N., Guertin, S. M., Swift, G. M. and Johnston, A. H., Radiation Effects on Advanced Flash Memories, IEEE Transactions on Nuclear Science 46(6), 1744, 1999.
[8] MIL-STD-883, Test Method 1019, Ionizing Radiation (Total Dose) Test Procedure, current revision; the dose rate conditions and the ELDRS provisions referred to on this page.
[9] IEEE Radiation Effects Data Workshop records 1992 to 2024, RADECS proceedings 1991 to 2023, and NASA GSFC radiation test reports, as indexed in the site's database; the public sources of the corpus.
