What is publicly known about each part on your BOM, and what still needs a test?

What is in the index and what is kept out, how a part number is matched, how the process technology is decided and graded, where the priors come from, what makes a row decision sensitive, and where the screen stops.

The public parts index

The Space-RHA radiation database holds one row per part and one row per extracted test result, each result tied to the document it came from. The index published here is built from the public document classes only: NASA GSFC radiation test reports [1], IEEE NSREC Radiation Effects Data Workshop and RADECS data workshop papers [2], RADECS proceedings and IEEE TNS papers, and vendor catalogs. For each part the best public number per effect is kept: the highest total dose passed and the lowest failed; the highest LET with no latchup and the lowest LET at which latchup was seen; the lowest upset, transient and functional interrupt onsets and the largest saturated cross section; the lowest burnout or gate rupture LET and the highest clean one; the highest displacement damage fluence; the highest proton energy. Vendor claimed values and aggregator single-LET screens are excluded from those numbers and flagged. Up to four source documents are kept per part.

Test reports that are export controlled are treated differently: a part that has records in them carries a single flag, restricted records exist, and no number, title, year or count from those records is written to the site. The build checks its own output for restricted markers before it is allowed to publish. The process technology of a part is an attribute of the part, not a test result, and is published with its source class: stated in a report or datasheet, a manufacturer family, or an inference.

Matching a part number

The entered part number is normalized (upper case, no spaces, the carrier and packaging tail after a slash or hash dropped) and looked up exactly, then through the aliases recorded in the database. Failing that, a generic is tried: for TI the longest generic in TI's catalog that prefixes the orderable (TPS7A4701RGWR belongs to TPS7A47, not to a stripped TPS7A4701) [3]; for ADI the candidate rules that the reliability harvest validated (AD7176-2BRUZ to AD7176-2 then AD7176; LT1963AEST-3.3 to LT1963A then LT1963) [4]; and for everything else a progressively shorter base, no shorter than five characters. A match through a generic is labeled as such, because the tested grade or package may differ from yours, and the row lists related parts with data that share the prefix.

The process evidence ladder

The technology class matters because the priors hang on it: a bulk CMOS part and a bipolar part with the same function carry different latchup and total dose expectations. The class is decided by the first rung that answers:

1. Set by you. The device type and technology class selectors in a row override everything below.
2. Stated process. A process named in a test report or datasheet for the matched part (high confidence), or an inference already in the database (medium if rated high, otherwise low).
3. Manufacturer source. ADI's reliability data pages name a process family per generic; a family with a node or a named process (0.18um CMOS, 0.60um BCDMOS, SiGe, 2.0um GaAs HBT) is high confidence, a coarse bin (bipolar by emitter size, older CMOS) medium. TI's product information API gives the product family for every generic and, for amplifiers and logic, an architecture parameter (bipolar, CMOS, FET, or a logic family such as LVC or LS) that maps to a class at medium confidence; TI does not publish the process itself through the API [3].
4. Prefix lexicon. What parts with the same prefix and manufacturer turned out to be, learned from the whole database (next section). Medium confidence when the bucket agrees at 80 percent with five or more parts, otherwise low.
5. Device type default. The advisor's mapping from device type to the usual technology, low confidence.

The device type follows a similar order: the database record, TI's product family, the description, and a prefix vote when 60 percent of parts with the prefix share a type. Every rung that answered is listed in the row's evidence, whether or not it won.

The prefix lexicon

For every part in the database with a known process, the alphabetic prefix and the first digit of the numeric field (AD7, LTC2, TPS7, IRF5; logic families keep their letters, SN74LVC, CD4) are paired with the manufacturer and the technology class, and the counts are published per bucket. The lookup tries manufacturer plus prefix, then prefix alone, then the letters alone. Leave-one-out accuracy of the top class, measured on the database at build time: 80.6 percent with the manufacturer and prefix, 78.5 percent with the prefix alone, 73.5 percent with the letters alone. The same buckets vote on the device type. The lexicon is an aggregate: it carries no per part data and no source.

Priors

The latchup prior is the Extended Historical SEL Priors model on this site [5]: for each technology class, the posterior fraction of parts whose latchup onset falls below LET 75 MeV.cm2/mg at hot equivalent conditions, with its 90 percent interval and the number of parts behind it, in the population you choose (modern parts by default). Bipolar, GaN and GaAs parts have no parasitic pnpn path and are reported as not applicable. The total dose prior is the Total Dose Hardness Priors model [6]: the lognormal fit to the failure dose of the part's device family, by technology where the family has enough parts of that technology, giving the 10, 50 and 90 percent failure doses and, for your mission dose times margin, the fraction of parts expected to fail below it.

Judgments and decision sensitivity

Total dose: a public pass at or above dose times margin counts as covered (a recorded failure below it is still called out); a recorded failure below it is an issue; a pass short of it asks for an extension by test or lot acceptance; with no data the prior fraction failing by the mission dose decides between low prior risk and a test. Latchup: not applicable for the classes without a pnpn path; observed latchup asks for protection (sized with the SEL protection tool) or a different part; a clean result at or above the LET goal is covered; short of it, extend or accept with protection; with no data the prior decides between cheap insurance and a test. Upset and functional interrupt, transients, burnout and displacement damage are judged by device class in the same way, pointing to the rate tool, the analog SET guideline, the power derating advisor and the optocoupler CTR prior.

A row is decision sensitive when the technology class is not high confidence and the runner up class would change the picture: the latchup prior moves by 15 points or crosses between under 5 and over 20 percent, or the total dose p10 moves by a factor of two. That is the case to resolve first, by asking the manufacturer, reading the datasheet's process paragraph, or a decapsulation, before paying for a test that the class would have made unnecessary.