REDW Radiation Effects Distributions

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

1. What the page shows

How tested parts of each type are distributed in total dose hardness and latchup immunity, from about 7,900 records in the IEEE NSREC Radiation Effects Data Workshop compendia, 1992 to 2025 [1]. Four tabs: TID Hardness (survival curve of pass levels), SEL Immunity (distribution of latchup thresholds), Statistics Table (P10, median and P90 by device type), Parts List (the filtered records). Filters by grade, publication year, minimum TID and device type. The purpose is to know, before you have data on a part, what is typical, what is exceptional and what is unlikely for its class.

2. The data

Each REDW record is one device tested at one facility, as reported by its authors, from government laboratories, aerospace companies and universities across three decades. TID entries are pass or fail levels in krad(Si); SEL entries are threshold LETs or immunity to a maximum tested LET. Devices are sorted into 17 categories by primary function: BJT, op amp, comparator, voltage reference, linear regulator, power converter, ADC, DAC, FPGA, MOSFET, optocoupler, SRAM, flash, logic and interface, MCU and processor, sensor, other. Multifunction parts are placed by the function tested. Grade is COTS, or rad hard, space and QML, from the publication and datasheet; the boundary is ambiguous for some families and some records are surely misfiled.

This is not a random sample of electronics. It is what the radiation community chose to test: parts with a plausible space or military use, parts retested across lots because they looked promising, results interesting enough to publish, and a technology mix that shifts with the decades (more FPGAs and processors recently, more discrete bipolar early). Read every curve with that in mind.

3. TID survival curve

Fraction of the filtered devices that passed at or above each dose, dose on a log axis. The curve reads 0.60 at 50 krad if 60 percent of the devices were demonstrated functional at 50 krad or more. It starts at 1 and falls; a steep drop means a narrow spread of failure levels, a gradual one means wide variability.

Most TID records are right censored. "Passed 100 krad" means the test stopped at 100 krad with the part working; the failure level is somewhere above. Tests stop at facility limits (100 or 300 krad are common), at the program requirement plus margin, or when the beam time ran out. So the curve is a lower bound on the true survival:

A category showing 50 percent survival at 30 krad does not mean half its parts fail at 30 krad; it means at least half were shown to survive 30 krad. The bias is conservative in the right direction for screening and it grows toward the top of the dose range. Two other cautions on the pass levels: nearly all are high dose rate results (50 to 300 rad(Si)/s per MIL-STD-883 TM1019 [10]), which understate bipolar degradation where ELDRS applies [8], [9]; and some MOS parts recover with time, which makes the high rate test conservative in the other direction.

4. SEL threshold curve

Fraction of the filtered devices with no latchup at or above each LET. A value of 0.70 at 40 MeV cm2/mg means 70 percent showed no SEL at 40 or above.

The pile of points near 120 MeV cm2/mg is not physics. Most facilities top out at 80 to 120 with gold or uranium, and a part that does not latch at the maximum is recorded as immune to that LET or better. Those records are right censored too: the true threshold is at least the tested maximum and may not exist [6], [7]. Read a threshold near 120 as "clean to the facility limit."

Two things to carry when comparing with a mission. Behind typical shielding at GEO, the effective LET cutoff for most of the galactic cosmic ray population is about 30 to 40 MeV cm2/mg, so thresholds above that are usually treated as immune for that environment, while solar heavy ion events and thin shielding reach effective LETs near 100. And the REDW thresholds are mostly room temperature measurements. Latchup gets easier with temperature; a part clean at 25 C can latch at 85 or 125. Ask what temperature the record was taken at before you believe it for a hot box.

5. Percentiles and censoring

The Statistics Table reports P10, median and P90 of the reported values per category and test type, with the record counts N, N TID and N SEL (not every record has both). P10 is the conservative screen: 90 percent of tested parts met or exceeded it. The median is the typical part. P90 is the level only 10 percent exceeded.

Because the pass levels are censored, percentiles of the raw values understate the true distribution. The survival curves use a Kaplan-Meier style estimator adapted to this structure [2], [5]: failures step the curve down, censored records stay in the risk set without stepping it.

Sample size governs how far to trust a percentile. Op amps have hundreds of records and stable percentiles; sensors and processors have far fewer. Treat N below 20 with caution and N below 10 as anecdote. The current N is always displayed.

6. Filters: grade, year, device type

Grade. All parts mixes two populations with different design intent and gives a bimodal curve: COTS failing low, rad hard surviving high. COTS parts are built on processes optimized for cost, speed and power; their dose tolerance is incidental and varies by manufacturer, node and lot [2], [3]. Rad hard, space and QML parts are hardened by design (enclosed layout transistors, guard rings, triple wells, oxide processing [4]), screened for space, or built on MIL-PRF-38535 lines with radiation lot acceptance. Filter to the population you are actually choosing from.

Year. Mainstream CMOS went from about 800 nm to 5 nm over the span of the data, and the physics moved with it [2], [3], [5]. Thinner gate oxides trap less charge, so sub 100 nm parts often tolerate dose better; isolation oxides did not scale the same way and their leakage is now the usual failure; lower supply voltages shrink the margins; latchup shifted with well doping and epitaxial layers [6]. A part number tested in 1995 was a different die. For current commercial parts, filtering to about 2010 onward is more representative; for rad hard parts on stable lines, older data hold if the manufacturer confirms no process change. Narrowing the years costs sample size, and the N readout shows the price.

Device type. Chip filters on the TID and SEL tabs, or compare categories in the table. Categories with known ELDRS exposure (BJT, references, optocouplers) are marked in the category notes.

7. Using the distributions

The question the curves answer is: for this type of part, what fraction of tested parts would meet my requirement? If a mission needs 100 krad and 5 percent of COTS op amps have been shown to survive it, a COTS op amp is a high risk pick and the rad hard alternative should be priced in. If 80 percent of COTS logic passes the level, COTS logic is a reasonable plan. Comparing categories shows which subsystems will drive the radiation design. The percentiles give a defensible sentence for a review: "P10 TID for COTS voltage references in the REDW record is X krad; about 10 percent of tested parts failed below it."

Percentile	Meaning	Use
P10	90% of tested parts met or exceeded it	Conservative screen
Median	Half met or exceeded it	Typical expectation
P90	Only 10% exceeded it	Optimistic case

Compare the TID curve with the mission dose including its design margin (a factor of two is the conventional minimum), and the SEL curve with the effective LET environment for the orbit and shielding. The TID threshold slider sets that reference level and filters the parts list to records that passed it. A category where 90 percent of tested parts meet the requirement can still hand you a part that does not; lots, process changes and bias conditions decide the individual, and flight lot testing is still the rule for anything critical.

8. Limitations

• Right censoring in both TID and SEL, described above. The curves are lower bounds on hardness.
• Heterogeneous test conditions. Different facilities, sources, dose rates, bias configurations (worst case CMOS bias is usually all terminals grounded or V_DD with pins grounded, but programs vary), pass criteria (parametric against datasheet limits versus functional in the application), ion species and energies, temperatures, and latchup definitions (current step, destructive or not). Two records at "50 krad" may not mean the same thing.
• Lot to lot variation. One record is usually one lot and a few parts. Oxide thickness, doping and contamination move between lots; the same part number can pass 100 krad from one lot and fail at 20 from another [2], [4]. The distribution averages over that and cannot predict a lot.
• Selection bias as described in section 2: interest, survivorship, publication and technology mix.
• What it is for. Early screening, feasibility, technology trades, deciding what to test first. Not device characterization, not lot acceptance, not formal qualification, not worst case analysis for a critical function.

9. References

Back to the Distribution Data